Upload folder using huggingface_hub

.gitignore

@@ -45,6 +45,7 @@ models-mnt
 /server
 /simple
 /batched
+/batched-bench
 /export-lora
 /finetune
 /speculative
@@ -106,3 +107,6 @@ tests/test-tokenizer-1-bpe
 rocblas.dll
 hipblas.dll
 koboldcpp_hipblas.so
+
+# Jetbrains idea folder
+.idea/

CMakeLists.txt

@@ -356,6 +356,8 @@ add_library(ggml OBJECT
             ggml.h
             ggml-alloc.c
             ggml-alloc.h
+            ggml-backend.c
+            ggml-backend.h
             k_quants.h
             k_quants.c
             ${GGML_SOURCES_CUDA})

Dockerfile

@@ -2,10 +2,17 @@ FROM python
 WORKDIR /app
 COPY . .
 RUN apt update \
- && apt install build-essential wget libopenblas-dev make -y \
- && make LLAMA_OPENBLAS=1 \
+ && apt install build-essential wget libopenblas-dev make cmake -y \
+ && mkdir build \
  && wget https://huggingface.co/TheBloke/Pygmalion-2-7B-GGUF/resolve/main/pygmalion-2-7b.Q6_K.gguf \
- && apt remove build-essential wget make -y \
- && rm -fr *.bat convert-* ci docs examples otherarchs tests
+ https://github.com/mozilla/sccache/releases/download/v0.5.4/sccache-dist-v0.5.4-x86_64-unknown-linux-musl.tar.gz \
+ && tar -vxzf sccache-dist-v0.5.4-x86_64-unknown-linux-musl.tar.gz \
+ && mv sccache-dist-v0.5.4-x86_64-unknown-linux-musl/sccache /usr/bin/sccache\
+ && cd build \
+ && cmake .. -DCMAKE_C_COMPILER_LAUNCHER=sccache -DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
+ && cmake --build . \
+ && cd .. \
+ && apt remove build-essential wget make cmake -y \
+ && rm -fr *.bat convert-* ci docs examples otherarchs tests sccache-dist-v0.5.4-x86_64-unknown-linux-musl*
 
  ENTRYPOINT ["python", "koboldcpp.py", "pygmalion-2-7b.Q6_K.gguf", "--port", "7860", "--smartcontext"]

Makefile

@@ -372,6 +372,8 @@ endif # LLAMA_NO_K_QUANTS
 #there's no intrinsics or special gpu ops used here, so we can have a universal object
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 	$(CC)  $(CFLAGS) -c $< -o $@
+ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+	$(CC)  $(CFLAGS) -c $< -o $@
 
 #version 2 libs
 ggml_v2.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
@@ -402,7 +404,7 @@ ggml_v2-opencl-legacy.o: otherarch/ggml_v2-opencl-legacy.c otherarch/ggml_v2-ope
 	$(CC) $(CFLAGS) -c $< -o $@
 
 # intermediate objects
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h otherarch/llama-util.h
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h otherarch/llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 common.o: common/common.cpp common/common.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -427,7 +429,7 @@ gpttype_adapter_cublas.o: $(GPTTYPE_ADAPTER)
 clean:
 	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state gguf gguf.exe main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp_default.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp_hipblas.dll koboldcpp_default.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so koboldcpp_hipblas.so
 
-main: examples/main/main.cpp build-info.h ggml.o $(KQ1) ggml-alloc.o llama.o common.o console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp build-info.h ggml.o $(KQ1) ggml-alloc.o ggml-backend.o llama.o common.o console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
@@ -438,11 +440,11 @@ gguf: examples/gguf/gguf.cpp build-info.h ggml.o llama.o $(OBJS)
 
 
 #generated libraries
-koboldcpp_default: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_default: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o ggml-backend.o grammar-parser.o $(OBJS)
 	$(DEFAULT_BUILD)
 
 ifdef OPENBLAS_BUILD
-koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o ggml-backend.o grammar-parser.o $(OBJS)
 	$(OPENBLAS_BUILD)
 else
 koboldcpp_openblas:
@@ -450,7 +452,7 @@ koboldcpp_openblas:
 endif
 
 ifdef FAILSAFE_BUILD
-koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ3) ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ3) ggml-alloc.o ggml-backend.o grammar-parser.o $(OBJS)
 	$(FAILSAFE_BUILD)
 else
 koboldcpp_failsafe:
@@ -458,7 +460,7 @@ koboldcpp_failsafe:
 endif
 
 ifdef NOAVX2_BUILD
-koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ2) ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ2) ggml-alloc.o ggml-backend.o grammar-parser.o $(OBJS)
 	$(NOAVX2_BUILD)
 else
 koboldcpp_noavx2:
@@ -466,7 +468,7 @@ koboldcpp_noavx2:
 endif
 
 ifdef CLBLAST_BUILD
-koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o $(KQ1) ggml-alloc.o ggml-backend.o grammar-parser.o $(OBJS)
 	$(CLBLAST_BUILD)
 else
 koboldcpp_clblast:
@@ -474,7 +476,7 @@ koboldcpp_clblast:
 endif
 
 ifdef CUBLAS_BUILD
-koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o grammar-parser.o $(CUBLAS_OBJS) $(OBJS)
+koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o ggml-backend.o grammar-parser.o $(CUBLAS_OBJS) $(OBJS)
 	$(CUBLAS_BUILD)
 else
 koboldcpp_cublas:
@@ -482,7 +484,7 @@ koboldcpp_cublas:
 endif
 
 ifdef HIPBLAS_BUILD
-koboldcpp_hipblas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o grammar-parser.o $(HIP_OBJS) $(OBJS)
+koboldcpp_hipblas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o ggml-backend.o grammar-parser.o $(HIP_OBJS) $(OBJS)
 	$(HIPBLAS_BUILD)
 else
 koboldcpp_hipblas:
@@ -490,15 +492,15 @@ koboldcpp_hipblas:
 endif
 
 # tools
-quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o $(KQ1) ggml-alloc.o
+quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o $(KQ1) ggml-alloc.o ggml-backend.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gptj: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gptj: ggml.o llama.o $(KQ1) ggml-alloc.o ggml-backend.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gpt2: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gpt2: ggml.o llama.o $(KQ1) ggml-alloc.o ggml-backend.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_neox: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_neox: ggml.o llama.o $(KQ1) ggml-alloc.o ggml-backend.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_mpt: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_mpt: ggml.o llama.o $(KQ1) ggml-alloc.o ggml-backend.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
 

Package.swift

@@ -1,10 +1,10 @@
-// swift-tools-version:5.3
+// swift-tools-version:5.5
 
 import PackageDescription
 
 #if arch(arm) || arch(arm64)
 let platforms: [SupportedPlatform]? = [
-    .macOS(.v11),
+    .macOS(.v12),
     .iOS(.v14),
     .watchOS(.v4),
     .tvOS(.v14)
@@ -41,12 +41,13 @@ let package = Package(
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
+                "ggml-backend.c",
                 "k_quants.c",
             ] + additionalSources,
             resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32"]),
+                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
                 .define("GGML_USE_K_QUANTS"),
                 .define("GGML_USE_ACCELERATE")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+

colab.ipynb

@@ -33,7 +33,7 @@
         "!nohup ./cloudflared-linux-amd64 tunnel --url http://localhost:5001 &\r\n",
         "!sleep 10\r\n",
         "!cat nohup.out\r\n",
-        "!python koboldcpp.py model.ggml --usecublas 0 mmq --gpulayers $Layers --hordeconfig concedo\r\n"
+        "!python koboldcpp.py model.ggml --usecublas 0 mmq --gpulayers $Layers\r\n"
       ]
     }
   ],

common/CMakeLists.txt

@@ -5,6 +5,8 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
+    sampling.h
+    sampling.cpp
     console.h
     console.cpp
     grammar-parser.h

common/common.cpp

@@ -107,6 +107,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     std::string arg;
     gpt_params default_params;
     const std::string arg_prefix = "--";
+    llama_sampling_params & sparams = params.sampling_params;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -184,7 +185,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.top_k = std::stoi(argv[i]);
+            sparams.top_k = std::stoi(argv[i]);
         } else if (arg == "-c" || arg == "--ctx-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -216,73 +217,73 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.top_p = std::stof(argv[i]);
+            sparams.top_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.temp = std::stof(argv[i]);
+            sparams.temp = std::stof(argv[i]);
         } else if (arg == "--tfs") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.tfs_z = std::stof(argv[i]);
+            sparams.tfs_z = std::stof(argv[i]);
         } else if (arg == "--typical") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.typical_p = std::stof(argv[i]);
+            sparams.typical_p = std::stof(argv[i]);
         } else if (arg == "--repeat-last-n") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.repeat_last_n = std::stoi(argv[i]);
+            sparams.repeat_last_n = std::stoi(argv[i]);
         } else if (arg == "--repeat-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.repeat_penalty = std::stof(argv[i]);
+            sparams.repeat_penalty = std::stof(argv[i]);
         } else if (arg == "--frequency-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.frequency_penalty = std::stof(argv[i]);
+            sparams.frequency_penalty = std::stof(argv[i]);
         } else if (arg == "--presence-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.presence_penalty = std::stof(argv[i]);
+            sparams.presence_penalty = std::stof(argv[i]);
         } else if (arg == "--mirostat") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.mirostat = std::stoi(argv[i]);
+            sparams.mirostat = std::stoi(argv[i]);
         } else if (arg == "--mirostat-lr") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.mirostat_eta = std::stof(argv[i]);
+            sparams.mirostat_eta = std::stof(argv[i]);
         } else if (arg == "--mirostat-ent") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.mirostat_tau = std::stof(argv[i]);
+            sparams.mirostat_tau = std::stof(argv[i]);
         } else if (arg == "--cfg-negative-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.cfg_negative_prompt = argv[i];
+            sparams.cfg_negative_prompt = argv[i];
         } else if (arg == "--cfg-negative-prompt-file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -294,16 +295,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
-            if (!params.cfg_negative_prompt.empty() && params.cfg_negative_prompt.back() == '\n') {
-                params.cfg_negative_prompt.pop_back();
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(sparams.cfg_negative_prompt));
+            if (!sparams.cfg_negative_prompt.empty() && sparams.cfg_negative_prompt.back() == '\n') {
+                sparams.cfg_negative_prompt.pop_back();
             }
         } else if (arg == "--cfg-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.cfg_scale = std::stof(argv[i]);
+            sparams.cfg_scale = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -512,7 +513,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "--ignore-eos") {
             params.ignore_eos = true;
         } else if (arg == "--no-penalize-nl") {
-            params.penalize_nl = false;
+            sparams.penalize_nl = false;
         } else if (arg == "-l" || arg == "--logit-bias") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -524,7 +525,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             std::string value_str;
             try {
                 if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
-                    params.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
+                    sparams.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
                 } else {
                     throw std::exception();
                 }
@@ -627,6 +628,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
+    const llama_sampling_params & sparams = params.sampling_params;
+
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
@@ -659,19 +662,19 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.frequency_penalty);
     printf("  --mirostat N          use Mirostat sampling.\n");
     printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)sparams.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)sparams.mirostat_tau);
     printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
     printf("                        modifies the likelihood of token appearing in the completion,\n");
     printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
@@ -682,7 +685,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        negative prompt to use for guidance. (default: empty)\n");
     printf("  --cfg-negative-prompt-file FNAME\n");
     printf("                        negative prompt file to use for guidance. (default: empty)\n");
-    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
     printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
     printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
     printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
@@ -690,7 +693,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --no-penalize-nl      do not penalize newline token\n");
     printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
     printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
     printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
@@ -840,7 +843,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+        params.sampling_params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
     }
 
     {
@@ -932,127 +935,6 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
     return result;
 }
 
-//
-// Sampling utils
-//
-
-llama_token llama_sample_token(
-                  struct llama_context * ctx,
-                  struct llama_context * ctx_guidance,
-                  struct llama_grammar * grammar,
-               const struct gpt_params & params,
-        const std::vector<llama_token> & last_tokens,
-         std::vector<llama_token_data> & candidates,
-                                   int   idx) {
-    const int n_ctx   = llama_n_ctx(ctx);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-
-    const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
-    const float   top_p           = params.top_p;
-    const float   tfs_z           = params.tfs_z;
-    const float   typical_p       = params.typical_p;
-    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-    const float   repeat_penalty  = params.repeat_penalty;
-    const float   alpha_presence  = params.presence_penalty;
-    const float   alpha_frequency = params.frequency_penalty;
-    const int     mirostat        = params.mirostat;
-    const float   mirostat_tau    = params.mirostat_tau;
-    const float   mirostat_eta    = params.mirostat_eta;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    llama_token id = 0;
-
-    float * logits = llama_get_logits_ith(ctx, idx);
-
-    // Apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    candidates.clear();
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
-
-    if (ctx_guidance) {
-        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
-    }
-
-    // apply penalties
-    if (!last_tokens.empty()) {
-        const float nl_logit = logits[llama_token_nl(ctx)];
-        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
-
-        llama_sample_repetition_penalty(ctx, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
-                last_n_repeat, repeat_penalty);
-        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
-                last_n_repeat, alpha_frequency, alpha_presence);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    if (grammar != NULL) {
-        llama_sample_grammar(ctx, &cur_p, grammar);
-    }
-
-    if (temp <= 0) {
-        // Greedy sampling
-        id = llama_sample_token_greedy(ctx, &cur_p);
-    } else {
-        if (mirostat == 1) {
-            static float mirostat_mu = 2.0f * mirostat_tau;
-            const int mirostat_m = 100;
-            llama_sample_temp(ctx, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-        } else if (mirostat == 2) {
-            static float mirostat_mu = 2.0f * mirostat_tau;
-            llama_sample_temp(ctx, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-        } else {
-            // Temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
-            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
-            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
-            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
-            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
-            llama_sample_temp(ctx, &cur_p, temp);
-
-            {
-                const int n_top = 10;
-                LOG("top %d candidates:\n", n_top);
-
-                for (int i = 0; i < n_top; i++) {
-                    const llama_token id = cur_p.data[i].id;
-                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
-                }
-            }
-
-            id = llama_sample_token(ctx, &cur_p);
-
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
-        }
-    }
-    // printf("`%d`", candidates_p.size);
-
-    if (grammar != NULL) {
-        llama_grammar_accept_token(ctx, grammar, id);
-    }
-
-    return id;
-}
-
 //
 // YAML utils
 //
@@ -1204,6 +1086,8 @@ std::string get_sortable_timestamp() {
 
 void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    const llama_sampling_params & sparams = params.sampling_params;
+
     fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
     fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
     fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
@@ -1250,21 +1134,21 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
     fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
-    dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str());
-    fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale);
+    dump_string_yaml_multiline(stream, "cfg_negative_prompt", sparams.cfg_negative_prompt.c_str());
+    fprintf(stream, "cfg_scale: %f # default: 1.0\n", sparams.cfg_scale);
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
-    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.frequency_penalty);
     dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
     fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
-    const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx));
+    const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 
     dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str());
@@ -1277,7 +1161,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
 
     fprintf(stream, "logit_bias:\n");
-    for (std::pair<llama_token, float> lb : params.logit_bias) {
+    for (std::pair<llama_token, float> lb : sparams.logit_bias) {
         if (ignore_eos && lb.first == logit_bias_eos->first) {
             continue;
         }
@@ -1301,30 +1185,30 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
-    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
-    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau);
-    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
     fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
     fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
     fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
     fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
     fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
-    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
     fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
-    fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false");
+    fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
     fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
-    fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.presence_penalty);
     dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
     fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
     fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
     fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
     dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
     fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
-    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.repeat_penalty);
 
     fprintf(stream, "reverse_prompt:\n");
     for (std::string ap : params.antiprompt) {
@@ -1342,15 +1226,15 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
-    fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
+    fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
-    fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z);
+    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
     fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
-    fprintf(stream, "top_k: %d # default: 40\n", params.top_k);
-    fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p);
-    fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p);
+    fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }

common/common.h

@@ -4,6 +4,8 @@
 
 #include "llama.h"
 
+#include "sampling.h"
+
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
@@ -49,7 +51,6 @@ struct gpt_params {
     int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
-    int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
     int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
     float   rope_freq_base                  = 0.0f; // RoPE base frequency
     float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
@@ -67,13 +68,8 @@ struct gpt_params {
     int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
-
-    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-
-    // Classifier-Free Guidance
-    // https://arxiv.org/abs/2306.17806
-    std::string cfg_negative_prompt;       // string to help guidance
-    float       cfg_scale         = 1.f;   // How strong is guidance
+    // // sampling parameters
+    struct llama_sampling_params sampling_params;
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
     std::string model_draft       = "";                              // draft model for speculative decoding
@@ -115,7 +111,6 @@ struct gpt_params {
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
     bool instruct          = false; // instruction mode (used for Alpaca models)
-    bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
@@ -180,36 +175,6 @@ std::string llama_detokenize_bpe(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens);
 
-//
-// Sampling utils
-//
-
-// this is a common sampling function used across the examples for convenience
-// it can serve as a starting point for implementing your own sampling function
-//
-// required:
-//  - ctx:    context to use for sampling
-//  - params: sampling parameters
-//
-// optional:
-//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
-//  - grammar:       grammar to use for sampling, ignore if NULL
-//  - last_tokens:   needed for repetition penalty, ignore if empty
-//  - idx:           sample from llama_get_logits_ith(ctx, idx)
-//
-// returns:
-//  - token:      sampled token
-//  - candidates: vector of candidate tokens
-//
-llama_token llama_sample_token(
-                  struct llama_context * ctx,
-                  struct llama_context * ctx_guidance,
-                  struct llama_grammar * grammar,
-               const struct gpt_params & params,
-        const std::vector<llama_token> & last_tokens,
-         std::vector<llama_token_data> & candidates,
-                                   int   idx = 0);
-
 //
 // YAML utils
 //

common/sampling.cpp

@@ -0,0 +1,166 @@
+#include "sampling.h"
+
+llama_sampling_context::~llama_sampling_context() {
+    for (auto & it : sequence_contexts) {
+        if (it.second.grammar != NULL) {
+            llama_grammar_free(it.second.grammar);
+            it.second.grammar = NULL;
+        }
+    }
+}
+
+llama_sampling_context llama_sampling_context_init(
+        const struct gpt_params & params,
+                  llama_grammar * grammar) {
+  llama_sampling_context result;
+
+  result.params = params.sampling_params;
+  result.grammar = grammar;
+  return result;
+}
+
+// Note: Creates the context if it doesn't exist, so this always return something.
+llama_sampler_sequence_context & llama_sampling_get_sequence_context(
+              llama_sampling_context & ctx_sampling,
+        const llama_seq_id             seq) {
+    const auto it = ctx_sampling.sequence_contexts.find(seq);
+    if (it != ctx_sampling.sequence_contexts.end()) {
+        return it->second;
+    }
+    llama_sampler_sequence_context new_ctx = {
+        2.0f * ctx_sampling.params.mirostat_tau,
+        ctx_sampling.grammar != NULL ? llama_grammar_copy(ctx_sampling.grammar) : NULL,
+    };
+    return ctx_sampling.sequence_contexts.insert({seq, new_ctx}).first->second;
+}
+
+bool llama_sampling_context_reset(
+              llama_sampling_context & ctx_sampling,
+        const llama_seq_id             seq) {
+    const auto it = ctx_sampling.sequence_contexts.find(seq);
+    if (it == ctx_sampling.sequence_contexts.end()) return false;
+    if (it->second.grammar != NULL) {
+        llama_grammar_free(it->second.grammar);
+        it->second.grammar = NULL;
+    }
+    ctx_sampling.sequence_contexts.erase(it);
+    return true;
+}
+
+llama_token llama_sampling_sample(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_sampling_context & ctx_sampling,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+        const                      int   idx,
+                          llama_seq_id   seq) {
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+    const llama_sampling_params & params = ctx_sampling.params;
+    const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const float   repeat_penalty  = params.repeat_penalty;
+    const float   alpha_presence  = params.presence_penalty;
+    const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+
+    float * logits = llama_get_logits_ith(ctx, idx);
+
+    // Apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    candidates.clear();
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+    if (ctx_guidance) {
+        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    }
+
+    // apply penalties
+    if (!last_tokens.empty()) {
+        const float nl_logit = logits[llama_token_nl(ctx)];
+        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+
+        llama_sample_repetition_penalty(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    llama_sampler_sequence_context & ctx_seq = llama_sampling_get_sequence_context(ctx_sampling, seq);
+
+    if (ctx_seq.grammar != NULL) {
+        llama_sample_grammar(ctx, &cur_p, ctx_seq.grammar);
+    }
+
+    if (temp <= 0) {
+        // Greedy sampling
+        id = llama_sample_token_greedy(ctx, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_seq.mirostat_mu);
+        } else if (mirostat == 2) {
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &ctx_seq.mirostat_mu);
+        } else {
+            // Temperature sampling
+            size_t min_keep = std::max(1, params.n_probs);
+            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
+            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
+            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
+            llama_sample_temp(ctx, &cur_p, temp);
+
+            {
+                const int n_top = 10;
+                LOG("top %d candidates:\n", n_top);
+
+                for (int i = 0; i < n_top; i++) {
+                    const llama_token id = cur_p.data[i].id;
+                    (void)id; // To avoid a warning that id is unused when logging is disabled.
+                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                }
+            }
+
+            id = llama_sample_token(ctx, &cur_p);
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+        }
+    }
+
+    if (ctx_seq.grammar != NULL) {
+        llama_grammar_accept_token(ctx, ctx_seq.grammar, id);
+    }
+
+    return id;
+}

common/sampling.h

@@ -0,0 +1,108 @@
+#pragma once
+
+#include "llama.h"
+
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+// sampling parameters
+typedef struct llama_sampling_params {
+    int32_t top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typical_p         = 1.00f; // 1.0 = disabled
+    float   temp              = 0.80f; // 1.0 = disabled
+    float   repeat_penalty    = 1.10f; // 1.0 = disabled
+    int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   frequency_penalty = 0.00f; // 0.0 = disabled
+    float   presence_penalty  = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    float   mirostat_eta      = 0.10f; // learning rate
+
+    bool    penalize_nl       = true;  // consider newlines as a repeatable token
+
+    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+
+    // Classifier-Free Guidance
+    // https://arxiv.org/abs/2306.17806
+    std::string cfg_negative_prompt;   // string to help guidance
+    float       cfg_scale     = 1.f;   // How strong is guidance
+
+    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
+
+} llama_sampling_params;
+
+// per-sequence sampler context
+typedef struct llama_sampler_sequence_context {
+    float mirostat_mu; // mirostat sampler state
+    llama_grammar * grammar;
+} llama_sampler_sequence_context;
+
+// general sampler context
+typedef struct llama_sampling_context {
+    ~llama_sampling_context();
+
+    // parameters that will be used for sampling and when creating
+    // new llama_sampler_sequence_context instances
+    llama_sampling_params params;
+
+    // map of sequence ids to sampler contexts
+    std::unordered_map<llama_seq_id, llama_sampler_sequence_context> sequence_contexts;
+
+    // when non-NULL, new instances of llama_sampler_sequence_context
+    // will get a copy of the grammar here
+    // note: only the pointer is stored here, it is not a copy of
+    //       the grammar and shouldn't be freed
+    llama_grammar * grammar;
+} llama_sampling_context;
+
+#include "common.h"
+
+// Create a new sampling context instance.
+llama_sampling_context llama_sampling_context_init(
+        const struct gpt_params & params,
+                  llama_grammar * grammar = NULL);
+
+// Fetches the sampler context for the specified sequence id (defaults to 0).
+// If the context for that sequence id doesn't already exist, it will be created with
+// default values based on the parameters in the ctx_sampling argument.
+llama_sampler_sequence_context & llama_sampling_get_sequence_context(
+              llama_sampling_context & ctx_sampling,
+        const llama_seq_id             seq = 0);
+
+// Reset the sampler context for the supplied sequence id (defaults to 0).
+// This is necessary to reuse a sequence id or free memory used by sequences
+// that are no longer required.
+bool llama_sampling_context_reset(
+              llama_sampling_context & ctx_sampling,
+        const llama_seq_id             seq = 0);
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+// Note: When using multiple sequences, it is the caller's responsibility to call
+//       llama_sampling_context_reset when a sequence ends
+//
+// required:
+//  - ctx:          context to use for sampling
+//  - ctx_sampling: sampling-specific context
+//
+// optional:
+//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
+//  - seq:           sequence id to associate sampler state with
+//
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
+//
+llama_token llama_sampling_sample(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_sampling_context & ctx_sampling,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+        const                      int   idx = 0,
+                          llama_seq_id   seq = 0);

convert-bloom-hf-to-gguf.py

@@ -0,0 +1,238 @@
+#!/usr/bin/env python3
+# HF bloom --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+# Supported Models:
+#   https://huggingface.co/bigscience/bloom-1b7
+#   https://huggingface.co/bigscience/bloom-3b
+#   https://huggingface.co/bigscience/bloom-7b1
+#   https://huggingface.co/Langboat/bloom-1b4-zh
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Bloom model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
+    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "BloomForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.BLOOM
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Bloom")
+n_embed = hparams.get("hidden_size", hparams.get("n_embed"))
+n_head = hparams.get("n_head", hparams.get("num_attention_heads"))
+gguf_writer.add_context_length(hparams.get("seq_length", n_embed))
+gguf_writer.add_embedding_length(n_embed)
+gguf_writer.add_feed_forward_length(4 * n_embed)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(n_head)
+gguf_writer.add_head_count_kv(n_head)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0)  # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+# params for qkv transform
+n_head_kv = hparams.get("n_head_kv", n_head)
+head_dim = n_embed // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    has_lm_head = True
+    if "lm_head.weight" not in model_part.keys() and "output.weight" not in model_part.keys():
+        has_lm_head = False
+
+    for original_name in model_part.keys():
+        data = model_part[original_name]
+        name = re.sub(r'transformer\.', '', original_name)
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data = np.concatenate(
+                (qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                 qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                 qkv_weights[:, 2, :, :].reshape((-1, n_embed))),
+                axis=0
+            )
+            print("re-format attention.linear_qkv.weight")
+        elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+            qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
+            data = np.concatenate(
+                (qkv_bias[:, 0, :].reshape((n_embed,)),
+                 qkv_bias[:, 1, :].reshape((n_embed,)),
+                 qkv_bias[:, 2, :].reshape((n_embed,))),
+                axis=0
+            )
+            print("re-format attention.linear_qkv.bias")
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+        if not has_lm_head and name == "word_embeddings.weight":
+            gguf_writer.add_tensor("output.weight", data)
+            print(name, "=>", "output.weight" + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))  # noqa
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-mpt-hf-to-gguf.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python3
+# HF mpt--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "MPTForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.MPT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layers"]
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_context_length(hparams["max_seq_len"])
+gguf_writer.add_embedding_length(hparams["d_model"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
+gguf_writer.add_head_count(hparams["n_heads"])
+gguf_writer.add_layer_norm_eps(1e-05)
+if hparams["attn_config"]["clip_qkv"] is not None:
+    gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
+gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but
+# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to
+# accomodate some "reserved" tokens; this is causing problems down the line in
+# llama.cpp, so we pad the vocab with dummy tokens:
+
+vocab_size = hparams["vocab_size"]
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Cannot map tensor '" + name + "'")
+            continue # for the sake of compatibility with some old published models, don't quit
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+        # note: MPT output is tied to (same as) wte in original model;
+        # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+        if new_name == "token_embd.weight":
+            gguf_writer.add_tensor("output.weight", data)
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-refact-hf-to-gguf.py

@@ -0,0 +1,263 @@
+#!/usr/bin/env python3
+# HF refact--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if "NO_LOCAL_GGUF" not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
+import gguf
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert a Refact model to a GGML compatible file"
+    )
+    parser.add_argument(
+        "--vocab-only",
+        action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile",
+        type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model",
+        type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype",
+        type=int,
+        choices=[0, 1],
+        default=1,
+        nargs="?",
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f"Error: {args.model} is not a directory", file=sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf"
+
+print("gguf: loading model " + dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTRefactForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH = gguf.MODEL_ARCH.REFACT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+# Get refact feed forward dimension
+hidden_dim = hparams["n_embd"]
+inner_dim = 4 * hidden_dim
+hidden_dim = int(2 * inner_dim / 3)
+multiple_of = 256
+ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Refact")
+# refact uses Alibi. So this is from config.json which might be used by training.
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+
+gguf_writer.add_feed_forward_length(ff_dim)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+# params for qkv transform
+n_head = hparams["n_head"]
+n_head_kv = 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for i in range(block_count):
+        if f"transformer.h.{i}.attn.kv.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.attn.kv.weight"]
+            model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[
+                : n_head_kv * head_dim
+            ]
+            model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[
+                n_head_kv * head_dim :
+            ]
+            del model_part[f"transformer.h.{i}.attn.kv.weight"]
+        if f"transformer.h.{i}.attn.q.weight" in model_part:
+            model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[
+                f"transformer.h.{i}.attn.q.weight"
+            ]
+            del model_part[f"transformer.h.{i}.attn.q.weight"]
+        if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+            model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim]
+            model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:]
+            del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if (
+            ftype == 1
+            and data_dtype == np.float32
+            and name.endswith(".weight")
+            and n_dims == 2
+        ):
+            data = data.astype(np.float16)
+
+        print(
+            new_name
+            + ", n_dims = "
+            + str(n_dims)
+            + ", "
+            + str(old_dtype)
+            + " --> "
+            + str(data.dtype)
+        )
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

examples/CMakeLists.txt

@@ -25,6 +25,7 @@ else()
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
     add_subdirectory(batched)
+    add_subdirectory(batched-bench)
     add_subdirectory(speculative)
     add_subdirectory(parallel)
     add_subdirectory(embd-input)

examples/batched-bench/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET batched-bench)
+add_executable(${TARGET} batched-bench.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/batched-bench/README.md

@@ -0,0 +1,51 @@
+# llama.cpp/example/batched-bench
+
+Benchmark the batched decoding performance of `llama.cpp`
+
+## Usage
+
+There are 2 modes of operation:
+
+- `prompt not shared` - each batch has a separate prompt of size `PP` (i.e. `N_KV = B*(PP + TG)`)
+- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
+
+```bash
+./batched-bench MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
+
+# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
+./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 0 99
+
+# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 1 99
+
+# custom set of batches
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32
+```
+
+## Sample results
+
+- `PP` - prompt tokens per batch
+- `TG` - generated tokens per batch
+- `B` - number of batches
+- `N_KV` - required KV cache size
+- `T_PP` - prompt processing time (i.e. time to first token)
+- `S_PP` - prompt processing speed (`(B*PP)/T_PP` or `PP/T_PP`)
+- `T_TG` - time to generate all batches
+- `S_TG` - text generation speed (`(B*TG)/T_TG`)
+- `T` - total time
+- `S` - total speed (i.e. all tokens / total time)
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   128 |    128 |    1 |    256 |    0.108 |  1186.64 |    3.079 |    41.57 |    3.187 |    80.32 |
+|   128 |    128 |    2 |    512 |    0.198 |  1295.19 |    5.029 |    50.90 |    5.227 |    97.95 |
+|   128 |    128 |    4 |   1024 |    0.373 |  1373.96 |    6.878 |    74.44 |    7.251 |   141.23 |
+|   128 |    128 |    8 |   2048 |    0.751 |  1363.27 |    7.344 |   139.43 |    8.095 |   252.99 |
+|   128 |    128 |   16 |   4096 |    1.570 |  1304.68 |    8.455 |   242.23 |   10.024 |   408.60 |
+|   128 |    128 |   32 |   8192 |    3.408 |  1201.73 |    8.801 |   465.40 |   12.209 |   670.96 |
+|   128 |    256 |    1 |    384 |    0.107 |  1196.70 |    6.329 |    40.45 |    6.436 |    59.67 |
+|   128 |    256 |    2 |    768 |    0.194 |  1317.45 |   10.239 |    50.00 |   10.433 |    73.61 |
+|   128 |    256 |    4 |   1536 |    0.366 |  1399.03 |   13.960 |    73.35 |   14.326 |   107.22 |
+|   128 |    256 |    8 |   3072 |    0.751 |  1363.92 |   15.110 |   135.54 |   15.861 |   193.69 |
+|   128 |    256 |   16 |   6144 |    1.569 |  1304.93 |   18.073 |   226.64 |   19.642 |   312.80 |
+|   128 |    256 |   32 |  12288 |    3.409 |  1201.35 |   19.223 |   426.15 |   22.633 |   542.93 |

examples/batched-bench/batched-bench.cpp

@@ -0,0 +1,251 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+// mutates the input string
+static std::vector<int> parse_list(char * p) {
+    std::vector<int> ret;
+
+    char * q = p;
+
+    while (*p) {
+        if (*p == ',') {
+            *p = '\0';
+            ret.push_back(std::atoi(q));
+            q = p + 1;
+        }
+
+        ++p;
+    }
+
+    ret.push_back(std::atoi(q));
+
+    return ret;
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
+        printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
+        printf("  example: %s ggml-model-f16.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        return 1 ;
+    }
+
+    int n_kv_max     = 2048;
+    int is_pp_shared = 0;
+    int n_gpu_layers = 0;
+    int mmq          = 0;
+
+    std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
+    std::vector<int> n_tg = { 128, 256, };
+    std::vector<int> n_pl = { 1, 2, 4, 8, 16, 32, };
+    //std::vector<int> n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, };
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        n_kv_max = std::atoi(argv[2]);
+    }
+
+    if (argc >= 4) {
+        is_pp_shared = std::atoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        n_gpu_layers = std::atoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        mmq = std::atoi(argv[5]);
+    }
+
+    if (argc >= 7) {
+        n_pp = parse_list(argv[6]);
+    }
+
+    if (argc >= 8) {
+        n_tg = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_pl = parse_list(argv[8]);
+    }
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    model_params.n_gpu_layers = n_gpu_layers;
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed      = 1234;
+    ctx_params.n_ctx     = n_kv_max;
+    ctx_params.n_batch   = 512;
+    ctx_params.mul_mat_q = mmq;
+
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    llama_batch batch = llama_batch_init(n_kv_max, 0);
+
+    // decode in batches of ctx_params.n_batch tokens
+    auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token  + i,
+                nullptr,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
+                0, 0, 0, // unused
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+            if (ret != 0) {
+                LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
+                return false;
+            }
+        }
+
+        return true;
+    };
+
+    // warm up
+    {
+        batch.n_tokens = 16;
+
+        for (int i = 0; i < batch.n_tokens; ++i) {
+            batch.token[i]  = 0;
+            batch.pos[i]    = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+    }
+
+    LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
+    LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
+
+    for (        int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
+        for (    int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
+            for (int i_pl = 0; i_pl < (int) n_pl.size(); ++i_pl) {
+                const int pp = n_pp[i_pp];
+                const int tg = n_tg[i_tg];
+                const int pl = n_pl[i_pl];
+
+                const int n_ctx_req = is_pp_shared ? pp + pl*tg : pl*(pp + tg);
+
+                if (n_ctx_req > n_kv_max) {
+                    continue;
+                }
+
+                batch.n_tokens = is_pp_shared ? pp : pl*pp;
+
+                for (int i = 0; i < batch.n_tokens; ++i) {
+                    batch.token[i]  = 0;
+                    batch.pos[i]    = i;
+                    batch.seq_id[i] = 0;
+                    batch.logits[i] = false;
+                }
+                batch.logits[batch.n_tokens - 1] = true;
+
+                const auto t_pp_start = ggml_time_us();
+
+                llama_kv_cache_tokens_rm(ctx, -1, -1);
+
+                if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
+                    LOG_TEE("%s: llama_decode() failed\n", __func__);
+                    return 1;
+                }
+
+                if (is_pp_shared) {
+                    for (int32_t i = 1; i < pl; ++i) {
+                        llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
+                    }
+                }
+
+                const auto t_pp_end = ggml_time_us();
+
+                const auto t_tg_start = ggml_time_us();
+
+                for (int i = 0; i < tg; ++i) {
+                    batch.n_tokens = pl;
+
+                    for (int j = 0; j < pl; ++j) {
+                        batch.token[j]  = 0;
+                        batch.pos[j]    = pp + i;
+                        batch.seq_id[j] = j;
+                        batch.logits[j] = true;
+                    }
+
+                    if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
+                        LOG_TEE("%s: llama_decode() failed\n", __func__);
+                        return 1;
+                    }
+                }
+
+                const auto t_tg_end = ggml_time_us();
+
+                const int32_t n_kv = n_ctx_req;
+
+                const float t_pp = (t_pp_end - t_pp_start) / 1000000.0f;
+                const float t_tg = (t_tg_end - t_tg_start) / 1000000.0f;
+                const float t    = t_pp + t_tg;
+
+                const float speed_pp = is_pp_shared ? pp / t_pp : pl*pp / t_pp;
+                const float speed_tg = pl*tg / t_tg;
+                const float speed    = n_kv / t;
+
+                LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
+            }
+        }
+    }
+
+    llama_print_timings(ctx);
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/batched.swift/.gitignore

@@ -0,0 +1,9 @@
+.DS_Store
+/.build
+/Packages
+xcuserdata/
+DerivedData/
+.swiftpm/configuration/registries.json
+.swiftpm/xcode/package.xcworkspace/contents.xcworkspacedata
+.netrc
+batched_swift

examples/batched.swift/Makefile

@@ -0,0 +1,6 @@
+.PHONY: build
+
+build:
+	xcodebuild -scheme batched_swift -destination "generic/platform=macOS" -derivedDataPath build
+	rm -f ./batched_swift
+	ln -s ./build/Build/Products/Debug/batched_swift ./batched_swift

examples/batched.swift/Package.swift

@@ -0,0 +1,22 @@
+// swift-tools-version: 5.5
+// The swift-tools-version declares the minimum version of Swift required to build this package.
+
+import PackageDescription
+
+let package = Package(
+    name: "batched_swift",
+    platforms: [.macOS(.v12)],
+    dependencies: [
+        .package(name: "llama", path: "../../"),
+    ],
+    targets: [
+        // Targets are the basic building blocks of a package, defining a module or a test suite.
+        // Targets can depend on other targets in this package and products from dependencies.
+        .executableTarget(
+            name: "batched_swift",
+            dependencies: ["llama"],
+            path: "Sources",
+            linkerSettings: [.linkedFramework("Foundation"), .linkedFramework("AppKit")]
+        ),
+    ]
+)

examples/batched.swift/README.md

@@ -0,0 +1,4 @@
+This is a swift clone of `examples/batched`.
+
+$ `make`
+$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -0,0 +1,255 @@
+import Foundation
+import llama
+
+let arguments = CommandLine.arguments
+
+// Check that we have at least one argument (the model path)
+guard arguments.count > 1 else {
+    print("Usage: swift MODEL_PATH [PROMPT] [PARALLEL]")
+    exit(1)
+}
+
+let modelPath: String = arguments[1]
+let prompt: String = arguments.count > 2 ? arguments[2] : "Hello my name is"
+let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(arguments[3])! : 1
+
+// total length of the sequences including the prompt
+let n_len: Int = 32
+
+// init LLM
+llama_backend_init(false)
+defer {
+    llama_backend_free()
+}
+
+let model_params = llama_model_default_params()
+guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
+    print("Failed to load model")
+    exit(1)
+}
+
+defer {
+    llama_free_model(model)
+}
+
+var tokens = tokenize(text: prompt, add_bos: true)
+
+let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel)
+
+var context_params = llama_context_default_params()
+context_params.seed = 1234
+context_params.n_ctx = n_kv_req
+context_params.n_batch = UInt32(max(n_len, n_parallel))
+context_params.n_threads = 8
+context_params.n_threads_batch = 8
+
+let context = llama_new_context_with_model(model, context_params)
+guard context != nil else {
+    print("Failed to initialize context")
+    exit(1)
+}
+
+defer {
+    llama_free(context)
+}
+
+let n_ctx = llama_n_ctx(context)
+
+print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
+
+if n_kv_req > n_ctx {
+    print("error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", n_kv_req)
+    exit(1)
+}
+
+var buffer: [CChar] = []
+for id: llama_token in tokens {
+    print(token_to_piece(token: id, buffer: &buffer) ?? "", terminator: "")
+}
+
+print("\n")
+
+var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0)
+defer {
+    llama_batch_free(batch)
+}
+
+// evaluate the initial prompt
+batch.n_tokens = Int32(tokens.count)
+
+for (i, token) in tokens.enumerated() {
+    batch.token[i] = token
+    batch.pos[i] = Int32(i)
+    batch.seq_id[i] = 0
+    batch.logits[i] = 0
+}
+
+// llama_decode will output logits only for the last token of the prompt
+batch.logits[Int(batch.n_tokens) - 1] = 1
+
+if llama_decode(context, batch) != 0 {
+    print("llama_decode() failed")
+    exit(1)
+}
+
+for i in 1 ..< n_parallel {
+    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+}
+
+if n_parallel > 1 {
+    print("generating \(n_parallel) sequences ...\n")
+}
+
+var streams: [String] = .init(repeating: "", count: n_parallel)
+var streamBuffers: [[CChar]] = .init(repeating: [], count: n_parallel)
+var i_batch = [Int32](repeating: batch.n_tokens - 1, count: n_parallel)
+
+var n_cur = batch.n_tokens
+var n_decode = 0
+
+let t_main_start = ggml_time_us()
+
+while n_cur <= n_len {
+    // prepare the next batch
+    batch.n_tokens = 0
+
+    // sample the next token for each parallel sequence / stream
+    for i in 0 ..< n_parallel {
+        if i_batch[i] < 0 {
+            // the stream has already finished
+            continue
+        }
+
+        var n_vocab = llama_n_vocab(model)
+        var logits = llama_get_logits_ith(context, i_batch[i])
+
+        var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab))
+
+        for token_id in 0 ..< n_vocab {
+            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
+        }
+
+        var candidates_p: llama_token_data_array = .init(
+            data: &candidates,
+            size: candidates.count,
+            sorted: false
+        )
+
+        let top_k: Int32 = 40
+        let top_p: Float = 0.9
+        let temp: Float = 0.4
+
+        llama_sample_top_k(context, &candidates_p, top_k, 1)
+        llama_sample_top_p(context, &candidates_p, top_p, 1)
+        llama_sample_temp(context, &candidates_p, temp)
+
+        let new_token_id = llama_sample_token(context, &candidates_p)
+
+        // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+        // is it an end of stream? -> mark the stream as finished
+        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+            i_batch[i] = -1
+            // print("")
+            if n_parallel > 1 {
+                print("stream \(i) finished at n_cur = \(n_cur)")
+            }
+
+            continue
+        }
+
+        let nextStringPiece = token_to_piece(token: new_token_id, buffer: &streamBuffers[i]) ?? ""
+
+        // if there is only one stream, we print immediately to stdout
+        if n_parallel == 1 {
+            print(nextStringPiece, terminator: "")
+        }
+        streams[i] += nextStringPiece
+
+        // push this new token for next evaluation
+        batch.token[Int(batch.n_tokens)] = new_token_id
+        batch.pos[Int(batch.n_tokens)] = n_cur
+        batch.seq_id[Int(batch.n_tokens)] = Int32(i)
+        batch.logits[Int(batch.n_tokens)] = 1
+
+        i_batch[i] = batch.n_tokens
+
+        batch.n_tokens += 1
+
+        n_decode += 1
+    }
+
+    // all streams are finished
+    if batch.n_tokens == 0 {
+        break
+    }
+
+    n_cur += 1
+
+    // evaluate the current batch with the transformer model
+    if llama_decode(context, batch) != 0 {
+        print("llama_decode() failed")
+        exit(1)
+    }
+}
+
+if n_parallel > 1 {
+    print("\n")
+    for (i, stream) in streams.enumerated() {
+        print("sequence \(i):\n\n\(prompt)\(stream)\n")
+    }
+}
+
+let t_main_end = ggml_time_us()
+
+print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n")
+
+llama_print_timings(context)
+
+private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
+    let n_tokens = text.count + (add_bos ? 1 : 0)
+    let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
+    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos)
+    var swiftTokens: [llama_token] = []
+    for i in 0 ..< tokenCount {
+        swiftTokens.append(tokens[Int(i)])
+    }
+    tokens.deallocate()
+    return swiftTokens
+}
+
+private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
+    var result = [CChar](repeating: 0, count: 8)
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
+    if nTokens < 0 {
+        if result.count >= -Int(nTokens) {
+            result.removeLast(-Int(nTokens))
+        } else {
+            result.removeAll()
+        }
+        let check = llama_token_to_piece(
+            model,
+            token,
+            &result,
+            Int32(result.count)
+        )
+        assert(check == nTokens)
+    } else {
+        result.removeLast(result.count - Int(nTokens))
+    }
+    if buffer.isEmpty, let utfString = String(cString: result + [0], encoding: .utf8) {
+        return utfString
+    } else {
+        buffer.append(contentsOf: result)
+        let data = Data(buffer.map { UInt8(bitPattern: $0) })
+        if buffer.count >= 4 { // 4 bytes is the max length of a utf8 character so if we're here we need to reset the buffer
+            buffer = []
+        }
+        guard let bufferString = String(data: data, encoding: .utf8) else {
+            return nil
+        }
+        buffer = []
+        return bufferString
+    }
+    return nil
+}

examples/batched/batched.cpp

@@ -66,7 +66,7 @@ int main(int argc, char ** argv) {
     ctx_params.seed  = 1234;
     ctx_params.n_ctx = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_threads = params.n_threads;
+    ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);

examples/embd-input/embd-input-lib.cpp

@@ -128,21 +128,22 @@ bool eval_string(struct MyModel * mymodel,const char* str){
 llama_token sampling_id(struct MyModel* mymodel) {
     llama_context* ctx = mymodel->ctx;
     gpt_params params = mymodel->params;
+    llama_sampling_params & sparams = params.sampling_params;
     // int n_ctx = llama_n_ctx(ctx);
 
     // out of user input, sample next token
-    const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : params.top_k;
-    const float   top_p           = params.top_p;
-    const float   tfs_z           = params.tfs_z;
-    const float   typical_p       = params.typical_p;
+    const float   temp            = sparams.temp;
+    const int32_t top_k           = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : sparams.top_k;
+    const float   top_p           = sparams.top_p;
+    const float   tfs_z           = sparams.tfs_z;
+    const float   typical_p       = sparams.typical_p;
     // const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
     // const float   repeat_penalty  = params.repeat_penalty;
     // const float   alpha_presence  = params.presence_penalty;
     // const float   alpha_frequency = params.frequency_penalty;
-    const int     mirostat        = params.mirostat;
-    const float   mirostat_tau    = params.mirostat_tau;
-    const float   mirostat_eta    = params.mirostat_eta;
+    const int     mirostat        = sparams.mirostat;
+    const float   mirostat_tau    = sparams.mirostat_tau;
+    const float   mirostat_eta    = sparams.mirostat_eta;
     // const bool    penalize_nl     = params.penalize_nl;
 
     llama_token id = 0;
@@ -151,7 +152,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
         auto n_vocab = llama_n_vocab(llama_get_model(ctx));
 
         // Apply params.logit_bias map
-        for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
             logits[it->first] += it->second;
         }
 

examples/infill/infill.cpp

@@ -0,0 +1,800 @@
+#include "common.h"
+
+#include "console.h"
+#include "llama.h"
+#include "build-info.h"
+#include "grammar-parser.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static gpt_params               * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+static bool is_interacting = false;
+
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = get_sortable_timestamp();
+
+    const bool success = create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: infill\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting) {
+            is_interacting = true;
+        } else {
+            console::cleanup();
+            printf("\n");
+            llama_print_timings(*g_ctx);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    llama_sampling_params & sparams = params.sampling_params;
+    g_params = &params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("infill", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.logits_all) {
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+    if (params.instruct) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for instruct mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.antiprompt.empty()) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
+        printf("\n************\n");
+        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (params.random_prompt) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.path_prompt_cache.empty()) {
+        printf("\n************\n");
+        printf("%s: infill does not support prompt caching\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    LOG("%s: llama backend init\n", __func__);
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+    llama_context * ctx_guidance = NULL;
+    g_model = &model;
+    g_ctx = &ctx;
+
+    // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (sparams.cfg_scale > 1.f) {
+        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
+        ctx_guidance = llama_new_context_with_model(model, lparams);
+    }
+
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
+
+    bool suff_rm_leading_spc = params.escape;
+    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+        params.input_suffix.erase(0, 1);
+        suff_rm_leading_spc = false;
+    }
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+    const int space_token = 29871;
+    if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
+        inp_sfx.erase(inp_sfx.begin());
+    }
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    if (add_bos) {
+        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+    }
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    embd_inp = inp_pfx;
+    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+    embd_inp.push_back(llama_token_middle(ctx));
+
+    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
+    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    }
+
+    // Tokenize negative prompt
+    std::vector<llama_token> guidance_inp;
+    int guidance_offset = 0;
+    int original_prompt_len = 0;
+    if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
+
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
+        original_prompt_len = original_inp.size();
+        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
+    }
+
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    }
+
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (ctx_guidance) {
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            for (int i = 0; i < (int) guidance_inp.size(); i++) {
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
+            }
+        }
+
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_TEE("'\n");
+        }
+        LOG_TEE("\n");
+    }
+
+    if (params.interactive) {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+        LOG_TEE("%s: interactive mode on.\n", __func__);
+
+        if (params.input_prefix_bos) {
+            LOG_TEE("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
+    }
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
+
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = sparams.logit_bias.find(llama_token_eos(ctx));
+            if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+    LOG_TEE("\n#####  Infill mode  #####\n\n");
+    if (params.infill) {
+        printf("\n************\n");
+        printf("no need to specify '--infill', always running infill\n");
+        printf("************\n\n");
+    }
+    if (params.interactive) {
+        const char *control_message;
+        if (params.multiline_input) {
+            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n";
+        } else {
+            control_message = " - Press Return to return control to LLaMa.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n"
+                              " - If you want to submit another line, end your input with '\\'.\n";
+        }
+        LOG_TEE("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_TEE(       "%s\n", control_message);
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool input_echo           = true;
+
+    int n_past             = 0;
+    int n_remain           = params.n_predict;
+    int n_consumed         = 0;
+    int n_past_guidance    = 0;
+
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(console::prompt);
+
+    std::vector<llama_token> embd;
+    std::vector<llama_token> embd_guidance;
+
+    const int n_vocab = llama_n_vocab(model);
+
+    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    while (n_remain != 0 || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(console::error);
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(console::reset);
+                fflush(stdout);
+            }
+
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                if (params.n_predict == -2) {
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    break;
+                }
+
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
+
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+
+            if (ctx_guidance) {
+                int input_size = 0;
+                llama_token * input_buf = NULL;
+
+                if (n_past_guidance < (int) guidance_inp.size()) {
+                    // Guidance context should have the same data with these modifications:
+                    //
+                    // * Replace the initial prompt
+                    // * Shift everything by guidance_offset
+                    embd_guidance = guidance_inp;
+                    if (embd.begin() + original_prompt_len < embd.end()) {
+                        embd_guidance.insert(
+                            embd_guidance.end(),
+                            embd.begin() + original_prompt_len,
+                            embd.end()
+                        );
+                    }
+
+                    input_buf  = embd_guidance.data();
+                    input_size = embd_guidance.size();
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                } else {
+                    input_buf  = embd.data();
+                    input_size = embd.size();
+                }
+
+                for (int i = 0; i < input_size; i += params.n_batch) {
+                    int n_eval = std::min(input_size - i, params.n_batch);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
+                        return 1;
+                    }
+
+                    n_past_guidance += n_eval;
+                }
+            }
+
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
+            }
+
+        }
+
+        embd.clear();
+        embd_guidance.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+
+            const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
+
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+
+            embd.push_back(id);
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo) {
+            for (auto id : embd) {
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                printf("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+            fflush(stdout);
+        }
+        // reset color to default if we there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(console::reset);
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+
+            // deal with eot token in infill mode
+            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+                if(is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                }
+                fflush(stdout);
+                printf("\n");
+                console::set_display(console::user_input);
+                std::string buffer;
+                std::string line;
+                bool another_line=true;
+                // set a new prefix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line, if so we use the old input
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_prefix = buffer;
+                }
+                buffer.clear();
+                // set a new suffix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_suffix = buffer;
+                }
+                buffer.clear();
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                if (params.escape) {
+                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
+                    process_escapes(params.input_prefix);
+                    process_escapes(params.input_suffix);
+                }
+                suff_rm_leading_spc = params.escape;
+                if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+                    params.input_suffix.erase(0, 1);
+                    suff_rm_leading_spc = false;
+                }
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+                if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
+                    inp_sfx.erase(inp_sfx.begin());
+                }
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                if (add_bos) {
+                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+                }
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                embd_inp = inp_pfx;
+                embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                embd_inp.push_back(llama_token_middle(ctx));
+                embd.clear();
+                embd_guidance.clear();
+                n_remain = params.n_predict;
+                n_past = 0;
+                n_consumed = 0;
+                // LOG_TEE("took new input\n");
+                is_interacting = false;
+            }
+            // deal with end of text token in interactive mode
+            else if (last_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
+                if (params.interactive) {
+
+                    is_interacting = true;
+                    printf("\n");
+                    console::set_display(console::user_input);
+                    fflush(stdout);
+               }
+            }
+
+            if (n_past > 0 && is_interacting && !params.interactive) {
+                LOG("waiting for user input\n");
+
+                if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_token_bos(ctx));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    printf("%s", buffer.c_str());
+                }
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
+                        printf("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << llama_token_to_piece(ctx, token);
+                    }
+
+                    n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0) {
+                if (is_interacting) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+    if (!params.interactive && n_remain <= 0) {
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        fflush(stdout);
+    }
+
+    llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
+    if (ctx_guidance) { llama_free(ctx_guidance); }
+    llama_free(ctx);
+    llama_free_model(model);
+
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
+    llama_backend_free();
+
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
+    return 0;
+}
+

examples/main/main.cpp

@@ -109,6 +109,7 @@ int main(int argc, char ** argv) {
     if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
+    llama_sampling_params & sparams = params.sampling_params;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("main", "log"));
@@ -179,7 +180,7 @@ int main(int argc, char ** argv) {
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (params.cfg_scale > 1.f) {
+    if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
     }
@@ -257,9 +258,9 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
         LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
@@ -296,6 +297,9 @@ int main(int argc, char ** argv) {
             LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         }
+
+        // remove any "future" tokens that we might have inherited from the previous session
+        llama_kv_cache_tokens_rm(ctx, n_matching_session_tokens, -1);
     }
 
     LOGLN(
@@ -343,7 +347,7 @@ int main(int argc, char ** argv) {
 
         if (ctx_guidance) {
             LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
             LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
                 LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
@@ -395,7 +399,7 @@ int main(int argc, char ** argv) {
         }
     }
     LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
-            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
+            sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
@@ -413,8 +417,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
 
         {
-            auto it = params.logit_bias.find(llama_token_eos(ctx));
-            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+            auto it = sparams.logit_bias.find(llama_token_eos(ctx));
+            if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
                 LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
             }
         }
@@ -469,6 +473,7 @@ int main(int argc, char ** argv) {
 
     const int n_vocab = llama_n_vocab(model);
 
+    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
 
@@ -622,7 +627,7 @@ int main(int argc, char ** argv) {
                 LOG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
+            const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
 
             last_tokens.erase(last_tokens.begin());
             last_tokens.push_back(id);
@@ -667,7 +672,7 @@ int main(int argc, char ** argv) {
             }
             fflush(stdout);
         }
-        // reset color to default if we there is no pending user input
+        // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
         }
@@ -694,10 +699,8 @@ int main(int argc, char ** argv) {
                     if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
-                            console::set_display(console::user_input);
                         }
                         is_antiprompt = true;
-                        fflush(stdout);
                         break;
                     }
                 }
@@ -721,8 +724,6 @@ int main(int argc, char ** argv) {
 
                     is_interacting = true;
                     printf("\n");
-                    console::set_display(console::user_input);
-                    fflush(stdout);
                 } else if (params.instruct) {
                     is_interacting = true;
                 }
@@ -747,6 +748,9 @@ int main(int argc, char ** argv) {
                     printf("%s", buffer.c_str());
                 }
 
+                // color user input only
+                console::set_display(console::user_input);
+
                 std::string line;
                 bool another_line = true;
                 do {

examples/parallel/parallel.cpp

@@ -10,6 +10,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <ctime>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {
@@ -70,6 +71,26 @@ struct client {
     std::vector<llama_token> tokens_prev;
 };
 
+static void print_date_time() {
+    std::time_t current_time = std::time(nullptr);
+    std::tm* local_time = std::localtime(&current_time);
+    char buffer[80];
+    strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
+
+    printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
+}
+
+// Define a split string function to ...
+static std::vector<std::string> split_string(const std::string& input, char delimiter) {
+    std::vector<std::string> tokens;
+    std::istringstream stream(input);
+    std::string token;
+    while (std::getline(stream, token, delimiter)) {
+        tokens.push_back(token);
+    }
+    return tokens;
+}
+
 int main(int argc, char ** argv) {
     srand(1234);
 
@@ -104,6 +125,25 @@ int main(int argc, char ** argv) {
     params.logits_all = true;
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
+    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, NULL);
+
+    // load the prompts from an external file if there are any
+    if (params.prompt.empty()) {
+        printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
+    } else {
+        // Output each line of the input params.prompts vector and copy to k_prompts
+        int index = 0;
+        printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
+
+        std::vector<std::string> prompts = split_string(params.prompt, '\n');
+        for (const auto& prompt : prompts) {
+            k_prompts.resize(index + 1);
+            k_prompts[index] = prompt;
+            index++;
+            printf("%3d prompt: %s\n", index, prompt.c_str());
+        }
+    }
+
     fprintf(stderr, "\n\n");
     fflush(stderr);
 
@@ -129,7 +169,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(params.n_ctx, 0);
+    llama_batch batch = llama_batch_init(n_ctx, 0);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -233,7 +273,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_TEE("\033[1mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+                    LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
 
                     g_seq_id += 1;
 
@@ -301,7 +341,7 @@ int main(int argc, char ** argv) {
                 //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
                 //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
 
-                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, client.tokens_prev, candidates, client.i_batch - i);
+                const llama_token id = llama_sampling_sample(ctx, NULL, ctx_sampling, client.tokens_prev, candidates, client.i_batch - i, client.seq_id);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -332,12 +372,12 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, n_ctx);
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
 
                     const auto t_main_end = ggml_time_us();
 
-                    LOG_TEE("\033[1mClient %3d, seq %4d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput:    %s\nResponse: %s\n\n",
-                            client.id, client.seq_id, client.n_prompt, client.n_decoded,
+                    LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput:    %s\n\033[35mResponse: %s\033[0m\n\n",
+                            client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
                             (t_main_end - client.t_start_prompt) / 1e6,
                             (double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
                             n_cache_miss,
@@ -346,7 +386,7 @@ int main(int argc, char ** argv) {
 
                     n_total_prompt += client.n_prompt;
                     n_total_gen    += client.n_decoded;
-
+                    llama_sampling_context_reset(ctx_sampling, client.seq_id);
                     client.seq_id = -1;
                 }
 
@@ -357,13 +397,21 @@ int main(int argc, char ** argv) {
 
     const auto t_main_end = ggml_time_us();
 
-    LOG_TEE("\n\n");
+    print_date_time();
+
+    LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    if (params.prompt_file.empty()) {
+        params.prompt_file = "used built-in defaults";
+    }
+    LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
+    LOG_TEE("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
+
     LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Total speed (AVG):   %6s  speed: %5.2f t/s\n", "",             (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Cache misses:        %6d\n", n_cache_miss);
 
-    LOG_TEE("\n\n");
+    LOG_TEE("\n");
 
     llama_print_timings(ctx);
 

examples/save-load-state/save-load-state.cpp

@@ -8,9 +8,10 @@
 
 int main(int argc, char ** argv) {
     gpt_params params;
+    llama_sampling_params & sparams = params.sampling_params;
     params.seed = 42;
     params.n_threads = 4;
-    params.repeat_last_n = 64;
+    sparams.repeat_last_n = 64;
     params.prompt = "The quick brown fox";
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -24,7 +25,7 @@ int main(int argc, char ** argv) {
     }
 
     auto n_past = 0;
-    auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
+    auto last_n_tokens_data = std::vector<llama_token>(sparams.repeat_last_n, 0);
 
     // init
     llama_model * model;

examples/server/public/index.html

@@ -136,6 +136,11 @@
       display: block;
     }
 
+    fieldset label.slim {
+      margin: 0 0.5em;
+      display: inline;
+    }
+
     header, footer {
       text-align: center;
     }
@@ -145,6 +150,14 @@
       color: #888;
     }
 
+    .mode-chat textarea[name=prompt] {
+      height: 4.5em;
+    }
+
+    .mode-completion textarea[name=prompt] {
+      height: 10em;
+    }
+
 
     @keyframes loading-bg-wipe {
       0% {
@@ -187,7 +200,7 @@
       template: "{{prompt}}\n\n{{history}}\n{{char}}:",
       historyTemplate: "{{name}}: {{message}}",
       transcript: [],
-      type: "chat",
+      type: "chat",  // "chat" | "completion"
       char: "Llama",
       user: "User",
     })
@@ -365,13 +378,44 @@
       return String(str).replaceAll(/\{\{(.*?)\}\}/g, (_, key) => template(settings[key]));
     }
 
+    async function runLlama(prompt, llamaParams, char) {
+      const currentMessages = [];
+      const history = session.value.transcript;
+      if (controller.value) {
+        throw new Error("already running");
+      }
+      controller.value = new AbortController();
+      for await (const chunk of llama(prompt, llamaParams, {controller: controller.value})) {
+        const data = chunk.data;
+
+        if (data.stop) {
+          while (
+            currentMessages.length > 0 &&
+            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
+            ) {
+            currentMessages.pop();
+          }
+          transcriptUpdate([...history, [char, currentMessages]])
+          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
+        } else {
+          currentMessages.push(data);
+          transcriptUpdate([...history, [char, currentMessages]])
+        }
+
+        if (data.timings) {
+          llamaStats.value = data.timings;
+        }
+      }
+
+      controller.value = null;
+    }
+
     // send message to server
     const chat = async (msg) => {
       if (controller.value) {
         console.log('already running...');
         return;
       }
-      controller.value = new AbortController();
 
       transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
 
@@ -391,55 +435,41 @@
         ).join("\n"),
       });
 
-      const currentMessages = [];
-      const history = session.value.transcript
-
-      const llamaParams = {
+      await runLlama(prompt, {
         ...params.value,
         stop: ["</s>", template("{{char}}:"), template("{{user}}:")],
-      }
-
-      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
-        const data = chunk.data;
+      }, "{{char}}");
+    }
 
-        if (data.stop) {
-          while (
-            currentMessages.length > 0 &&
-            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
-          ) {
-            currentMessages.pop();
-          }
-          transcriptUpdate([...history, ["{{char}}", currentMessages]])
-          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
-        } else {
-          currentMessages.push(data);
-          transcriptUpdate([...history, ["{{char}}", currentMessages]])
-        }
+    const runCompletion = async () => {
+      if (controller.value) {
+        console.log('already running...');
+        return;
+      }
+      const {prompt} = session.value;
+      transcriptUpdate([...session.value.transcript, ["", prompt]]);
+      await runLlama(prompt, {
+        ...params.value,
+        stop: [],
+      }, "");
+    }
 
-        if (data.timings) {
-          llamaStats.value = data.timings;
-        }
+    const stop = (e) => {
+      e.preventDefault();
+      if (controller.value) {
+        controller.value.abort();
+        controller.value = null;
       }
+    }
 
-      controller.value = null;
+    const reset = (e) => {
+      stop(e);
+      transcriptUpdate([]);
     }
 
     function MessageInput() {
       const message = useSignal("")
 
-      const stop = (e) => {
-        e.preventDefault();
-        if (controller.value) {
-          controller.value.abort();
-          controller.value = null;
-        }
-      }
-
-      const reset = (e) => {
-        stop(e);
-        transcriptUpdate([]);
-      }
-
       const submit = (e) => {
         stop(e);
         chat(message.value);
@@ -474,6 +504,19 @@
       `
     }
 
+    function CompletionControls() {
+      const submit = (e) => {
+        stop(e);
+        runCompletion();
+      }
+      return html`
+        <div>
+          <button onclick=${submit} type="button" disabled=${generating.value}>Start</button>
+          <button onclick=${stop} disabled=${!generating.value}>Stop</button>
+          <button onclick=${reset}>Reset</button>
+        </div>`;
+    }
+
     const ChatLog = (props) => {
       const messages = session.value.transcript;
       const container = useRef(null)
@@ -497,7 +540,11 @@
             data;
           message = html`<${Markdownish} text=${template(text)} />`
         }
-        return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
+        if(user) {
+          return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
+        } else {
+          return html`<p key=${index}>${message}</p>`
+        }
       };
 
       return html`
@@ -574,18 +621,31 @@
         userTemplateAutosave()
       }, [session.value, params.value])
 
-      return html`
-        <form>
-          <fieldset>
-            <${UserTemplateResetButton}/>
-          </fieldset>
+      const GrammarControl = () => (
+        html`
+          <div>
+            <label for="template">Grammar</label>
+            <textarea id="grammar" name="grammar" placeholder="Use gbnf or JSON Schema+convert" value="${params.value.grammar}" rows=4 oninput=${updateParams}/>
+            <input type="text" name="prop-order" placeholder="order: prop1,prop2,prop3" oninput=${updateGrammarJsonSchemaPropOrder} />
+            <button type="button" onclick=${convertJSONSchemaGrammar}>Convert JSON Schema</button>
+          </div>
+          `
+      );
 
-          <fieldset>
-            <div>
-              <label for="prompt">Prompt</label>
-              <textarea type="text" name="prompt" value="${session.value.prompt}" rows=4 oninput=${updateSession}/>
-            </div>
-          </fieldset>
+      const PromptControlFieldSet = () => (
+        html`
+        <fieldset>
+          <div>
+            <label htmlFor="prompt">Prompt</label>
+            <textarea type="text" name="prompt" value="${session.value.prompt}" oninput=${updateSession}/>
+          </div>
+        </fieldset>
+        `
+      );
+
+      const ChatConfigForm = () => (
+        html`
+          ${PromptControlFieldSet()}
 
           <fieldset class="two">
             <div>
@@ -609,15 +669,30 @@
               <label for="template">Chat history template</label>
               <textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/>
             </div>
+            ${GrammarControl()}
+          </fieldset>
+      `
+    );
+
+      const CompletionConfigForm = () => (
+        html`
+          ${PromptControlFieldSet()}
+          <fieldset>${GrammarControl()}</fieldset>
+        `
+      );
 
+      return html`
+        <form>
+          <fieldset class="two">
+            <${UserTemplateResetButton}/>
             <div>
-              <label for="template">Grammar</label>
-              <textarea id="grammar" name="grammar" placeholder="Use gbnf or JSON Schema+convert" value="${params.value.grammar}" rows=4 oninput=${updateParams}/>
-              <input type="text" name="prop-order" placeholder="order: prop1,prop2,prop3" oninput=${updateGrammarJsonSchemaPropOrder} />
-              <button type="button" onclick=${convertJSONSchemaGrammar}>Convert JSON Schema</button>
+              <label class="slim"><input type="radio" name="type" value="chat" checked=${session.value.type === "chat"} oninput=${updateSession} /> Chat</label>
+              <label class="slim"><input type="radio" name="type" value="completion" checked=${session.value.type === "completion"} oninput=${updateSession} /> Completion</label>
             </div>
           </fieldset>
 
+          ${session.value.type === 'chat' ? ChatConfigForm() : CompletionConfigForm()}
+
           <fieldset class="two">
             ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})}
             ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
@@ -851,7 +926,7 @@
     function App(props) {
 
       return html`
-        <div>
+        <div class="mode-${session.value.type}">
           <header>
             <h1>llama.cpp</h1>
           </header>
@@ -861,7 +936,7 @@
           </main>
 
           <section id="write">
-            <${MessageInput} />
+            <${session.value.type === 'chat' ? MessageInput : CompletionControls} />
           </section>
 
           <footer>

examples/server/server.cpp

@@ -200,6 +200,7 @@ struct llama_server_context
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
     gpt_params params;
+    llama_sampling_context ctx_sampling;
     int n_ctx;
 
     grammar_parser::parse_state parsed_grammar;
@@ -254,6 +255,7 @@ struct llama_server_context
         if (grammar != nullptr) {
             llama_grammar_free(grammar);
             grammar = nullptr;
+            ctx_sampling = llama_sampling_context_init(params, NULL);
         }
     }
 
@@ -329,8 +331,8 @@ struct llama_server_context
             grammar_parser::print_grammar(stderr, parsed_grammar);
 
             {
-                auto it = params.logit_bias.find(llama_token_eos(ctx));
-                if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                auto it = params.sampling_params.logit_bias.find(llama_token_eos(ctx));
+                if (it != params.sampling_params.logit_bias.end() && it->second == -INFINITY) {
                     LOG_WARNING("EOS token is disabled, which will cause most grammars to fail", {});
                 }
             }
@@ -339,9 +341,89 @@ struct llama_server_context
             grammar = llama_grammar_init(
                 grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
         }
+        ctx_sampling = llama_sampling_context_init(params, grammar);
         return true;
     }
 
+    void loadInfill()
+    {
+        bool suff_rm_leading_spc = true;
+        if (params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+            params.input_suffix.erase(0, 1);
+            suff_rm_leading_spc = false;
+        }
+
+        auto prefix_tokens = tokenize(params.input_prefix, false);
+        auto suffix_tokens = tokenize(params.input_suffix, false);
+        const int space_token = 29871;
+        if (suff_rm_leading_spc  && suffix_tokens[0] == space_token) {
+            suffix_tokens.erase(suffix_tokens.begin());
+        }
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+        prefix_tokens.push_back(llama_token_middle(ctx));
+        auto prompt_tokens = prefix_tokens;
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t)params.n_ctx)
+        {
+            printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
+            // todo we probably want to cut from both sides
+            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
+            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
+            std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+
+            LOG_VERBOSE("input truncated", {
+                                               {"n_ctx", params.n_ctx},
+                                               {"n_keep", params.n_keep},
+                                               {"n_left", n_left},
+                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                                           });
+
+            truncated = true;
+            prompt_tokens = new_tokens;
+        }
+        else
+        {
+            const size_t ps = num_prompt_tokens;
+            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
+            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            printf("we have to evaluate at least 1 token to generate logits\n");
+            n_past--;
+        }
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        LOG_VERBOSE("prompt ingested", {
+                                           {"n_past", n_past},
+                                           {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
+                                           {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
+                                       });
+
+        has_next_token = true;
+    }
     void loadPrompt()
     {
         auto prompt_tokens = tokenize(prompt, true);  // always add BOS
@@ -383,9 +465,6 @@ struct llama_server_context
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
 
-        // since #3228 we now have to manually manage the KV cache
-        llama_kv_cache_seq_rm(ctx, 0, n_past, params.n_ctx);
-
         embd = prompt_tokens;
         if (n_past == num_prompt_tokens)
         {
@@ -393,6 +472,9 @@ struct llama_server_context
             n_past--;
         }
 
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         LOG_VERBOSE("prompt ingested", {
                                            {"n_past", n_past},
                                            {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
@@ -440,9 +522,11 @@ struct llama_server_context
                                            });
         }
 
+        bool tg = true;
         while (n_past < embd.size())
         {
             int n_eval = (int)embd.size() - n_past;
+            tg = n_eval == 1;
             if (n_eval > params.n_batch)
             {
                 n_eval = params.n_batch;
@@ -468,98 +552,20 @@ struct llama_server_context
             return result;
         }
 
-        // out of user input, sample next token
-        const float temp = params.temp;
-        const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(model) : params.top_k;
-        const float top_p = params.top_p;
-        const float tfs_z = params.tfs_z;
-        const float typical_p = params.typical_p;
-        const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-        const float repeat_penalty = params.repeat_penalty;
-        const float alpha_presence = params.presence_penalty;
-        const float alpha_frequency = params.frequency_penalty;
-        const int mirostat = params.mirostat;
-        const float mirostat_tau = params.mirostat_tau;
-        const float mirostat_eta = params.mirostat_eta;
-        const bool penalize_nl = params.penalize_nl;
-        const int32_t n_probs = params.n_probs;
-
         {
-            auto *logits = llama_get_logits(ctx);
-            auto n_vocab = llama_n_vocab(model);
-
-            // Apply params.logit_bias map
-            for (const auto &it : params.logit_bias)
-            {
-                logits[it.first] += it.second;
-            }
-
+            // out of user input, sample next token
             std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++)
-            {
-                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-            }
+            candidates.reserve(llama_n_vocab(model));
 
-            llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
-
-            // Apply penalties
-            float nl_logit = logits[llama_token_nl(ctx)];
-            auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-            llama_sample_repetition_penalty(ctx, &candidates_p,
-                                            last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                            last_n_repeat, repeat_penalty);
-            llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-                                                          last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                                          last_n_repeat, alpha_frequency, alpha_presence);
-            if (!penalize_nl)
-            {
-                logits[llama_token_nl(ctx)] = nl_logit;
-            }
+            result.tok = llama_sampling_sample(ctx, NULL, ctx_sampling, last_n_tokens, candidates);
 
-            if (grammar != nullptr) {
-                llama_sample_grammar(ctx, &candidates_p, grammar);
-            }
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
-            if (temp <= 0)
+            const int32_t n_probs = params.sampling_params.n_probs;
+            if (params.sampling_params.temp <= 0 && n_probs > 0)
             {
-                // Greedy sampling
-                result.tok = llama_sample_token_greedy(ctx, &candidates_p);
-                if (n_probs > 0)
-                {
-                    llama_sample_softmax(ctx, &candidates_p);
-                }
-            }
-            else
-            {
-                if (mirostat == 1)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    const int mirostat_m = 100;
-                    llama_sample_temp(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-                }
-                else if (mirostat == 2)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    llama_sample_temp(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                }
-                else
-                {
-                    // Temperature sampling
-                    size_t min_keep = std::max(1, n_probs);
-                    llama_sample_top_k(ctx, &candidates_p, top_k, min_keep);
-                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
-                    llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
-                    llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
-                    llama_sample_temp(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token(ctx, &candidates_p);
-                }
-            }
-
-            if (grammar != nullptr) {
-                llama_grammar_accept_token(ctx, grammar, result.tok);
+                // For llama_sample_token_greedy we need to sort candidates
+                llama_sample_softmax(ctx, &candidates_p);
             }
 
             for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
@@ -569,7 +575,9 @@ struct llama_server_context
 
             last_n_tokens.erase(last_n_tokens.begin());
             last_n_tokens.push_back(result.tok);
-            num_tokens_predicted++;
+            if (tg) {
+                num_tokens_predicted++;
+            }
         }
 
         // add it to the context
@@ -629,7 +637,7 @@ struct llama_server_context
         const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
         generated_text += token_text;
 
-        if (params.n_probs > 0)
+        if (params.sampling_params.n_probs > 0)
         {
             generated_token_probs.push_back(token_with_probs);
         }
@@ -710,15 +718,16 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("usage: %s [options]\n", argv0);
     printf("\n");
     printf("options:\n");
-    printf("  -h, --help            show this help message and exit\n");
-    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    printf("  --rope-freq-base N    RoPE base frequency (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: loaded from model)\n");
-    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("  -h, --help                show this help message and exit\n");
+    printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    printf("  -t N,  --threads N        number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
+    printf("  -c N,  --ctx-size N       size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  --rope-freq-base N        RoPE base frequency (default: loaded from model)\n");
+    printf("  --rope-freq-scale N       RoPE frequency scaling factor (default: loaded from model)\n");
+    printf("  -b N,  --batch-size N     batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
     {
         printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
@@ -863,6 +872,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_threads = std::stoi(argv[i]);
         }
+        else if (arg == "--threads-batch" || arg == "-tb")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+        }
         else if (arg == "-b" || arg == "--batch-size")
         {
             if (++i >= argc)
@@ -947,7 +965,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.push_back({argv[i], 1.0f});
+            params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f));
             params.use_mmap = false;
         }
         else if (arg == "--lora-scaled")
@@ -963,7 +981,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.push_back({lora_adapter, std::stof(argv[i])});
+            params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i])));
             params.use_mmap = false;
         }
         else if (arg == "--lora-base")
@@ -1017,34 +1035,35 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
 
 static json format_generation_settings(llama_server_context &llama)
 {
-    const auto eos_bias = llama.params.logit_bias.find(llama_token_eos(llama.ctx));
-    const bool ignore_eos = eos_bias != llama.params.logit_bias.end() &&
+    const auto & sparams = llama.params.sampling_params;
+    const auto eos_bias = sparams.logit_bias.find(llama_token_eos(llama.ctx));
+    const bool ignore_eos = eos_bias != sparams.logit_bias.end() &&
                             eos_bias->second < 0.0f && std::isinf(eos_bias->second);
 
     return json{
         {"n_ctx", llama.n_ctx},
         {"model", llama.params.model_alias},
         {"seed", llama.params.seed},
-        {"temp", llama.params.temp},
-        {"top_k", llama.params.top_k},
-        {"top_p", llama.params.top_p},
-        {"tfs_z", llama.params.tfs_z},
-        {"typical_p", llama.params.typical_p},
-        {"repeat_last_n", llama.params.repeat_last_n},
-        {"repeat_penalty", llama.params.repeat_penalty},
-        {"presence_penalty", llama.params.presence_penalty},
-        {"frequency_penalty", llama.params.frequency_penalty},
-        {"mirostat", llama.params.mirostat},
-        {"mirostat_tau", llama.params.mirostat_tau},
-        {"mirostat_eta", llama.params.mirostat_eta},
-        {"penalize_nl", llama.params.penalize_nl},
+        {"temp", sparams.temp},
+        {"top_k", sparams.top_k},
+        {"top_p", sparams.top_p},
+        {"tfs_z", sparams.tfs_z},
+        {"typical_p", sparams.typical_p},
+        {"repeat_last_n", sparams.repeat_last_n},
+        {"repeat_penalty", sparams.repeat_penalty},
+        {"presence_penalty", sparams.presence_penalty},
+        {"frequency_penalty", sparams.frequency_penalty},
+        {"mirostat", sparams.mirostat},
+        {"mirostat_tau", sparams.mirostat_tau},
+        {"mirostat_eta", sparams.mirostat_eta},
+        {"penalize_nl", sparams.penalize_nl},
         {"stop", llama.params.antiprompt},
         {"n_predict", llama.params.n_predict},
         {"n_keep", llama.params.n_keep},
         {"ignore_eos", ignore_eos},
         {"stream", llama.stream},
-        {"logit_bias", llama.params.logit_bias},
-        {"n_probs", llama.params.n_probs},
+        {"logit_bias", sparams.logit_bias},
+        {"n_probs", sparams.n_probs},
         {"grammar", llama.params.grammar},
     };
 }
@@ -1060,8 +1079,6 @@ static json format_timings(llama_server_context &llama)
 {
     const auto timings = llama_get_timings(llama.ctx);
 
-    assert(timings.n_eval == ptrdiff_t(llama.num_tokens_predicted));
-
     return json{
         {"prompt_n", timings.n_p_eval},
         {"prompt_ms", timings.t_p_eval_ms},
@@ -1095,7 +1112,7 @@ static json format_final_response(llama_server_context &llama, const std::string
         {"timings", format_timings(llama)},
     };
 
-    if (llama.params.n_probs > 0)
+    if (llama.params.sampling_params.n_probs > 0)
     {
         res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
     }
@@ -1111,7 +1128,7 @@ static json format_partial_response(
         {"stop", false},
     };
 
-    if (llama.params.n_probs > 0)
+    if (llama.params.sampling_params.n_probs > 0)
     {
         res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
     }
@@ -1143,26 +1160,28 @@ static T json_value(const json &body, const std::string &key, const T &default_v
 static void parse_options_completion(const json &body, llama_server_context &llama)
 {
     gpt_params default_params;
+    const auto & default_sparams = default_params.sampling_params;
+    auto & sparams = llama.params.sampling_params;
 
     llama.stream = json_value(body, "stream", false);
     llama.params.n_predict = json_value(body, "n_predict", default_params.n_predict);
-    llama.params.top_k = json_value(body, "top_k", default_params.top_k);
-    llama.params.top_p = json_value(body, "top_p", default_params.top_p);
-    llama.params.tfs_z = json_value(body, "tfs_z", default_params.tfs_z);
-    llama.params.typical_p = json_value(body, "typical_p", default_params.typical_p);
-    llama.params.repeat_last_n = json_value(body, "repeat_last_n", default_params.repeat_last_n);
-    llama.params.temp = json_value(body, "temperature", default_params.temp);
-    llama.params.repeat_penalty = json_value(body, "repeat_penalty", default_params.repeat_penalty);
-    llama.params.presence_penalty = json_value(body, "presence_penalty", default_params.presence_penalty);
-    llama.params.frequency_penalty = json_value(body, "frequency_penalty", default_params.frequency_penalty);
-    llama.params.mirostat = json_value(body, "mirostat", default_params.mirostat);
-    llama.params.mirostat_tau = json_value(body, "mirostat_tau", default_params.mirostat_tau);
-    llama.params.mirostat_eta = json_value(body, "mirostat_eta", default_params.mirostat_eta);
-    llama.params.penalize_nl = json_value(body, "penalize_nl", default_params.penalize_nl);
+    sparams.top_k = json_value(body, "top_k", default_sparams.top_k);
+    sparams.top_p = json_value(body, "top_p", default_sparams.top_p);
+    sparams.tfs_z = json_value(body, "tfs_z", default_sparams.tfs_z);
+    sparams.typical_p = json_value(body, "typical_p", default_sparams.typical_p);
+    sparams.repeat_last_n = json_value(body, "repeat_last_n", default_sparams.repeat_last_n);
+    sparams.temp = json_value(body, "temperature", default_sparams.temp);
+    sparams.repeat_penalty = json_value(body, "repeat_penalty", default_sparams.repeat_penalty);
+    sparams.presence_penalty = json_value(body, "presence_penalty", default_sparams.presence_penalty);
+    sparams.frequency_penalty = json_value(body, "frequency_penalty", default_sparams.frequency_penalty);
+    sparams.mirostat = json_value(body, "mirostat", default_sparams.mirostat);
+    sparams.mirostat_tau = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    sparams.mirostat_eta = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    sparams.penalize_nl = json_value(body, "penalize_nl", default_sparams.penalize_nl);
     llama.params.n_keep = json_value(body, "n_keep", default_params.n_keep);
     llama.params.seed = json_value(body, "seed", default_params.seed);
     llama.params.grammar = json_value(body, "grammar", default_params.grammar);
-    llama.params.n_probs = json_value(body, "n_probs", default_params.n_probs);
+    sparams.n_probs = json_value(body, "n_probs", default_sparams.n_probs);
 
     if (body.count("prompt") != 0)
     {
@@ -1173,10 +1192,10 @@ static void parse_options_completion(const json &body, llama_server_context &lla
         llama.prompt = "";
     }
 
-    llama.params.logit_bias.clear();
+    sparams.logit_bias.clear();
     if (json_value(body, "ignore_eos", false))
     {
-        llama.params.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
+        sparams.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
     }
 
     const auto &logit_bias = body.find("logit_bias");
@@ -1192,11 +1211,11 @@ static void parse_options_completion(const json &body, llama_server_context &lla
                 {
                     if (el[1].is_number())
                     {
-                        llama.params.logit_bias[tok] = el[1].get<float>();
+                        sparams.logit_bias[tok] = el[1].get<float>();
                     }
                     else if (el[1].is_boolean() && !el[1].get<bool>())
                     {
-                        llama.params.logit_bias[tok] = -INFINITY;
+                        sparams.logit_bias[tok] = -INFINITY;
                     }
                 }
             }
@@ -1216,9 +1235,32 @@ static void parse_options_completion(const json &body, llama_server_context &lla
         }
     }
 
+    llama.ctx_sampling = llama_sampling_context_init(llama.params, llama.grammar);
+
     LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
 }
 
+static void parse_options_infill(const json &body, llama_server_context &llama)
+{
+    if (body.count("input_prefix") != 0)
+    {
+        llama.params.input_prefix = body["input_prefix"];
+    }
+    else
+    {
+        llama.params.input_prefix = "";
+    }
+    if (body.count("input_suffix") != 0)
+    {
+        llama.params.input_suffix = body["input_suffix"];
+    }
+    else
+    {
+        llama.params.input_suffix = "";
+    }
+    parse_options_completion(body, llama);
+}
+
 static void log_server_request(const Request &req, const Response &res)
 {
     LOG_INFO("request", {
@@ -1403,7 +1445,7 @@ int main(int argc, char **argv)
             }
 
             auto probs = llama.generated_token_probs;
-            if (llama.params.n_probs > 0 && llama.stopped_word) {
+            if (llama.params.sampling_params.n_probs > 0 && llama.stopped_word) {
                 const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
                 probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
             }
@@ -1455,7 +1497,7 @@ int main(int argc, char **argv)
 
                         std::vector<completion_token_output> probs_output = {};
 
-                        if (llama.params.n_probs > 0) {
+                        if (llama.params.sampling_params.n_probs > 0) {
                             const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
                             size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
                             size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
@@ -1519,6 +1561,127 @@ int main(int argc, char **argv)
             res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         } });
 
+    svr.Post("/infill", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        llama.rewind();
+
+        llama_reset_timings(llama.ctx);
+
+        parse_options_infill(json::parse(req.body), llama);
+
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+        llama.loadInfill();
+        llama.beginCompletion();
+        const auto chunked_content_provider = [&](size_t, DataSink & sink) {
+            size_t sent_count = 0;
+            size_t sent_token_probs_index = 0;
+
+            while (llama.has_next_token) {
+                const completion_token_output token_with_probs = llama.doCompletion();
+                if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
+                    continue;
+                }
+                const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+                size_t pos = std::min(sent_count, llama.generated_text.size());
+
+                const std::string str_test = llama.generated_text.substr(pos);
+                bool is_stop_full = false;
+                size_t stop_pos =
+                    llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+                if (stop_pos != std::string::npos) {
+                    is_stop_full = true;
+                    llama.generated_text.erase(
+                        llama.generated_text.begin() + pos + stop_pos,
+                        llama.generated_text.end());
+                    pos = std::min(sent_count, llama.generated_text.size());
+                } else {
+                    is_stop_full = false;
+                    stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                        STOP_PARTIAL);
+                }
+
+                if (
+                    stop_pos == std::string::npos ||
+                    // Send rest of the text if we are at the end of the generation
+                    (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                ) {
+                    const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
+                    sent_count += to_send.size();
+
+                    std::vector<completion_token_output> probs_output = {};
+
+                    if (llama.params.sampling_params.n_probs > 0) {
+                        const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                        size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                        size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                        if (probs_pos < probs_stop_pos) {
+                            probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                        }
+                        sent_token_probs_index = probs_stop_pos;
+                    }
+
+                    const json data = format_partial_response(llama, to_send, probs_output);
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+
+                if (!llama.has_next_token) {
+                    // Generation is done, send extra information.
+                    const json data = format_final_response(
+                        llama,
+                        "",
+                        std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                    );
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+            }
+
+            llama_print_timings(llama.ctx);
+            sink.done();
+            return true;
+        };
+        const auto on_complete = [&](bool) {
+            llama.mutex.unlock();
+        };
+        lock.release();
+        res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+        });
+
     svr.Get("/model.json", [&llama](const Request &, Response &res)
             {
         const json data = format_generation_settings(llama);

examples/speculative/speculative.cpp

@@ -125,6 +125,8 @@ int main(int argc, char ** argv) {
         grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
     }
 
+    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar_tgt);
+
     const auto t_dec_start = ggml_time_us();
 
     while (true) {
@@ -134,7 +136,7 @@ int main(int argc, char ** argv) {
 
         while (true) {
             // sample from the target model
-            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
+            llama_token id = llama_sampling_sample(ctx_tgt, NULL, ctx_sampling, last_tokens, candidates, i_dft);
 
             // remember which tokens were sampled - used for repetition penalties during sampling
             last_tokens.erase(last_tokens.begin());
@@ -172,7 +174,7 @@ int main(int argc, char ** argv) {
                 LOG("out of drafted tokens\n");
             }
 
-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, n_ctx);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
             llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
             ++n_past_dft;
 
@@ -211,7 +213,13 @@ int main(int argc, char ** argv) {
             if (grammar_dft) {
                 llama_grammar_free(grammar_dft);
             }
-            grammar_dft = llama_grammar_copy(grammar_tgt);
+            // Note: Hardcoded to sequence id 0, if this ever supports parallel generation
+            //       that will need to change.
+            auto it = ctx_sampling.sequence_contexts.find(0);
+            GGML_ASSERT(it != ctx_sampling.sequence_contexts.end());
+            // This is necessary because each sequence id in sequence_contexts
+            // uses a copy of the original grammar.
+            grammar_dft = llama_grammar_copy(it->second.grammar);
 
             LOG("copied target grammar to draft grammar\n");
         }
@@ -257,7 +265,7 @@ int main(int argc, char ** argv) {
             }
 
             // evaluate the drafted token on the draft model
-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, n_ctx);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
             llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
             ++n_past_cur;
 
@@ -267,7 +275,7 @@ int main(int argc, char ** argv) {
         }
 
         // evaluate the target model on the drafted tokens
-        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, n_ctx);
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
         llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
         ++n_past_tgt;
 

ggml-alloc.c

@@ -1,4 +1,5 @@
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #include "ggml.h"
 #include <assert.h>
 #include <stdarg.h>
@@ -6,25 +7,6 @@
 #include <stdlib.h>
 #include <string.h>
 
-#ifdef __has_include
-    #if __has_include(<unistd.h>)
-        #include <unistd.h>
-        #if defined(_POSIX_MAPPED_FILES)
-            #include <sys/types.h>
-            #include <sys/mman.h>
-        #endif
-    #endif
-#endif
-
-#if defined(_WIN32)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <memoryapi.h>
-#endif
-
 
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -80,8 +62,9 @@ struct free_block {
 #define MAX_FREE_BLOCKS 256
 
 struct ggml_allocr {
+    struct ggml_backend_buffer * buffer;
+    bool buffer_owned;
     void * data;
-    size_t size;
     size_t alignment;
     int n_free_blocks;
     struct free_block free_blocks[MAX_FREE_BLOCKS];
@@ -119,16 +102,9 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    return ggml_nbytes(tensor);
-
-    UNUSED(alloc);
-}
-
 // check if a tensor is allocated by this buffer
 static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+    return tensor->buffer == alloc->buffer;
 }
 
 static bool ggml_is_view(struct ggml_tensor * t) {
@@ -136,11 +112,10 @@ static bool ggml_is_view(struct ggml_tensor * t) {
 }
 
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-#ifdef GGML_ALLOCATOR_DEBUG
     GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
     GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
-#endif
-    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
+
+    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -188,6 +163,8 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 
     tensor->data = addr;
     AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data);
+    tensor->buffer = alloc->buffer;
+    ggml_backend_buffer_init_tensor(alloc->buffer, tensor);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
@@ -208,19 +185,21 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
 static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-
     if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
+        AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer);
         return;
     }
 
-    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
+    void * ptr = tensor->data;
+
+    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
-    AT_PRINTF("%s: alloc->data = %p alloc->data+alloc->size = %p alloc->data+alloc->max_size = %p\n", __func__, alloc->data, (char*)alloc->data + alloc->size, (char*)alloc->data + alloc->max_size);
+
+    ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
@@ -285,15 +264,18 @@ void ggml_allocr_reset(struct ggml_allocr * alloc) {
     alloc->n_free_blocks = 1;
     size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
     alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
-    alloc->free_blocks[0].size = alloc->size - align_offset;
+    alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
 }
 
 struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
+
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr));
 
     *alloc = (struct ggml_allocr){
-        /*.data          = */ data,
-        /*.size          = */ size,
+        /*.buffer        = */ buffer,
+        /*.buffer_owned  = */ true,
+        /*.base          = */ ggml_backend_buffer_get_base(buffer),
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -312,74 +294,26 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// OS specific functions to allocate and free uncommitted virtual memory
-static void * alloc_vmem(size_t size) {
-#if defined(_WIN32)
-    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
-#elif defined(_POSIX_MAPPED_FILES)
-    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
-    if (ptr == MAP_FAILED) {
-        return NULL;
-    }
-    return ptr;
-#else
-    // use a fixed address for other platforms
-    uintptr_t base_addr = (uintptr_t)-size - 0x100;
-    return (void *)base_addr;
-#endif
-}
-
-static void free_vmem(void * base_addr, size_t size) {
-#if defined(_WIN32)
-    VirtualFree(base_addr, 0, MEM_RELEASE);
-    UNUSED(size);
-#elif defined(_POSIX_MAPPED_FILES)
-    munmap(base_addr, size);
-#else
-    // nothing to do
-    UNUSED(base_addr);
-    UNUSED(size);
-#endif
-}
-
-// allocate uncommitted virtual memory to measure the size of the graph
-static void alloc_measure_vmem(void ** base_addr, size_t * size) {
-    // 128GB for 64-bit, 1GB for 32-bit
-    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
-    do {
-        *base_addr = alloc_vmem(*size);
-        if (*base_addr != NULL) {
-            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
-            return;
-        }
-        // try again with half the size
-        *size /= 2;
-    } while (*size > 0);
-
-    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
-}
-
-static void free_measure_vmem(void * base_addr, size_t size) {
-    free_vmem(base_addr, size);
-}
-
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+    struct ggml_allocr * alloc = ggml_allocr_new((void *)0x1000, (size_t)-0x1001, alignment);
+    alloc->measure = true;
 
-    void * base_addr;
-    size_t size;
+    return alloc;
+}
 
-    alloc_measure_vmem(&base_addr, &size);
+struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr));
 
     *alloc = (struct ggml_allocr){
-        /*.data          = */ base_addr,
-        /*.size          = */ size,
-        /*.alignment     = */ alignment,
+        /*.buffer        = */ buffer,
+        /*.buffer_owned  = */ false,
+        /*.base          = */ ggml_backend_buffer_get_base(buffer),
+        /*.alignment     = */ ggml_backend_buffer_get_alignment(buffer),
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
         /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
-        /*.measure       = */ true,
+        /*.measure       = */ false,
         /*.parse_seq     = */ {0},
         /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
@@ -393,8 +327,8 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
-    if (alloc->measure) {
-        free_measure_vmem(alloc->data, alloc->size);
+    if (alloc->buffer_owned) {
+        ggml_backend_buffer_free(alloc->buffer);
     }
     free(alloc);
 }
@@ -437,7 +371,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_CONT:
             return true;
 
         default:
@@ -445,12 +378,23 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
+static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) {
+    assert(view->view_src != NULL && view->view_src->data != NULL);
+    view->backend = view->view_src->backend;
+    view->buffer  = view->view_src->buffer;
+    view->data    = (char *)view->view_src->data + view->view_offs;
+
+    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
+    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
+    assert(ggml_allocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
+    ggml_backend_buffer_init_tensor(alloc->buffer, view);
+}
+
 static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            assert(node->view_src->data != NULL);
-            node->data = (char *)node->view_src->data + node->view_offs;
+            init_view(alloc, node);
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -478,13 +422,17 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                                 // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
                                 // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
                                 AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                                node->data = parent->data;
+                                node->view_src = view_src;
+                                view_src_hn->n_views += 1;
+                                init_view(alloc, node);
                                 return;
                             }
                         }
                         else {
                             AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                            node->data = parent->data;
+                            node->view_src = parent;
+                            p_hn->n_views += 1;
+                            init_view(alloc, node);
                             return;
                         }
                     }
@@ -495,7 +443,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocr_alloc_graph_tensors_n(
+size_t ggml_allocr_alloc_graph_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -513,6 +461,10 @@ static size_t ggml_allocr_alloc_graph_tensors_n(
             if (ggml_is_view(node)) {
                 struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
+                if (node->buffer == NULL && node->data != NULL) {
+                    // view of a pre-allocated tensor, didn't call init_view() yet
+                    init_view(alloc, node);
+                }
             }
 
             for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -521,6 +473,9 @@ static size_t ggml_allocr_alloc_graph_tensors_n(
                     break;
                 }
                 hash_get(ht, parent)->n_children += 1;
+                if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
+                    init_view(alloc, parent);
+                }
             }
         }
     }
@@ -631,7 +586,7 @@ static size_t ggml_allocr_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_n(alloc, &graph, 1, NULL, NULL);
 }
 
 size_t ggml_allocr_max_size(struct ggml_allocr * alloc) {

ggml-alloc.h

@@ -6,21 +6,27 @@
 extern "C" {
 #endif
 
+struct ggml_backend_buffer;
 
 GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
 GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
 GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
 
-GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
-GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API void   ggml_allocr_free       (struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure (struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset      (struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc      (struct ggml_allocr * alloc, struct ggml_tensor * tensor);
 GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
-GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
+GGML_API size_t ggml_allocr_max_size   (struct ggml_allocr * alloc);
 
+GGML_API size_t ggml_allocr_alloc_graph_n(
+                    struct ggml_allocr * alloc,
+                    struct ggml_cgraph ** graphs, int n_graphs,
+                    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs);
 
 #ifdef  __cplusplus
 }

ggml-backend.c

@@ -0,0 +1,385 @@
+#include "ggml-backend.h"
+#include "ggml-alloc.h"
+
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED GGML_UNUSED
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+        struct ggml_backend                  * backend,
+        struct ggml_backend_buffer_i           iface,
+               ggml_backend_buffer_context_t   context,
+               size_t                          size) {
+    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
+
+    GGML_ASSERT(iface.get_base != NULL);
+
+    (*buffer) = (struct ggml_backend_buffer) {
+        /* .interface = */ iface,
+        /* .backend   = */ backend,
+        /* .context   = */ context,
+        /* .size      = */ size,
+    };
+
+    return buffer;
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    free(buffer);
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_get_alignment(buffer->backend);
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_base(buffer);
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    if (buffer->iface.get_alloc_size) {
+        return buffer->iface.get_alloc_size(buffer, tensor);
+    }
+    return ggml_nbytes(tensor);
+}
+
+void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    if (buffer->iface.init_tensor) {
+        buffer->iface.init_tensor(buffer, tensor);
+    }
+}
+
+void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    if (buffer->iface.free_tensor) {
+        buffer->iface.free_tensor(buffer, tensor);
+    }
+}
+
+// backend
+
+ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
+    return tensor->buffer->backend;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return backend->iface.alloc_buffer(backend, size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return backend->iface.get_alignment(backend);
+}
+
+void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+    ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+    ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_compute(backend, plan);
+}
+
+void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return backend->iface.supports_op(backend, op);
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
+    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    // printf("cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
+
+    if (src == dst) {
+        return;
+    }
+
+    // TODO: allow backends to support copy to/from same backend
+
+    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
+        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
+    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
+        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
+    } else {
+        // shouldn't be hit when copying from/to CPU
+        #ifndef NDEBUG
+        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
+        #endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+// backend CPU
+
+struct ggml_backend_cpu_context {
+    int n_threads;
+    void * work_data;
+    size_t work_size;
+};
+
+static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+    return "CPU";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    free(cpu_ctx->work_data);
+    free(cpu_ctx);
+    free(backend);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
+    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL,
+    /* .free_tensor    = */ NULL,
+};
+
+static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+
+static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
+
+    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
+    return TENSOR_ALIGNMENT;
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy((char *)tensor->data + offset, data, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    // for a backend such as CUDA that can queue async calls, it is ok to do this asynchronously, but it may not be the case for other backends
+    ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    cpu_plan->cgraph = *cgraph;
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+    }
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    free(cpu_plan->cplan.work_data);
+    free(cpu_plan);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        // TODO: may be faster to free and use malloc to avoid the copy
+        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        cpu_ctx->work_size = cplan.work_size;
+    }
+
+    cplan.work_data = cpu_ctx->work_data;
+
+    ggml_graph_compute(cgraph, &cplan);
+}
+
+static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i cpu_backend_i = {
+    /* .get_name            = */ ggml_backend_cpu_name,
+    /* .free                = */ ggml_backend_cpu_free,
+    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_cpu_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
+    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op         = */ ggml_backend_cpu_supports_op,
+};
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
+
+    ctx->n_threads = GGML_DEFAULT_N_THREADS;
+    ctx->work_data = NULL;
+    ctx->work_size = 0;
+
+    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
+
+    *cpu_backend = (struct ggml_backend) {
+        /* .interface = */ cpu_backend_i,
+        /* .context   = */ ctx
+    };
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_cpu_name;
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
+    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
+}

ggml-backend.h

@@ -0,0 +1,143 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+    struct ggml_backend;
+    struct ggml_backend_buffer;
+
+    // type-erased backend-specific types / wrappers
+    typedef void * ggml_backend_context_t;
+    typedef void * ggml_backend_graph_plan_t;
+    typedef void * ggml_backend_buffer_context_t;
+
+    // avoid accessing internals of these types
+    typedef struct ggml_backend        * ggml_backend_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+
+    //
+    // backend buffer
+    //
+
+    struct ggml_backend_buffer_i {
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+    };
+
+    // TODO: hide behind API
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
+        ggml_backend_buffer_context_t context;
+
+        size_t size;
+    };
+
+    // backend buffer functions
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+    //
+    // backend
+    //
+
+    struct ggml_backend_i {
+        const char * (*get_name)(ggml_backend_t backend);
+
+        void (*free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);
+
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);
+
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // compute graph with a plan
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend supports an operation
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+    };
+
+    // TODO: hide behind API
+    struct ggml_backend {
+        struct ggml_backend_i iface;
+
+        ggml_backend_context_t context;
+    };
+
+    // backend helper functions
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-cuda.cu

@@ -62,6 +62,7 @@
 #define cudaMemcpyHostToDevice hipMemcpyHostToDevice
 #define cudaMemcpyKind hipMemcpyKind
 #define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
 #define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
 #define cudaSetDevice hipSetDevice
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
@@ -414,11 +415,13 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_CPY_BLOCK_SIZE 32
 #define CUDA_SCALE_BLOCK_SIZE 256
+#define CUDA_CLAMP_BLOCK_SIZE 256
 #define CUDA_ROPE_BLOCK_SIZE 256
 #define CUDA_ALIBI_BLOCK_SIZE 32
 #define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
 #define CUDA_QUANTIZE_BLOCK_SIZE 256
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
+#define CUDA_GET_ROWS_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_CUDA_DMMV_X
@@ -1574,6 +1577,34 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
     reinterpret_cast<half&>(y[ib].ds.y) = sum;
 }
 
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void k_get_rows(const void * x, const int32_t * y, dst_t * dst, const int ncols) {
+    const int col = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int row = blockDim.y*blockIdx.y + threadIdx.y;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int r = y[row];
+
+    // copy x[r*ncols + col] to dst[row*ncols + col]
+    const int xi = r*ncols + col;
+    const int di = row*ncols + col;
+
+    const int ib = xi/qk; // block index
+    const int iqs = (xi%qk)/qr; // quant index
+    const int iybs = di - di%qk; // y block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(x, ib, iqs, v);
+
+    dst[iybs + iqs + 0]        = v.x;
+    dst[iybs + iqs + y_offset] = v.y;
+}
+
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
     const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
@@ -4555,6 +4586,24 @@ static __global__ void scale_f32(const float * x, float * dst, const float scale
     dst[i] = scale * x[i];
 }
 
+static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
+}
+
+template<int qk, int qr, dequantize_kernel_t dq>
+static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) {
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ncols + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, nrows, 1);
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols);
+}
+
 static void add_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
     const int num_blocks = (kx + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
     add_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
@@ -5436,6 +5485,11 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons
     scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
 }
 
+static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
+    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
+}
+
 template<typename T>
 static void rope_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
                           const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
@@ -5699,7 +5753,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
         GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
         kind = cudaMemcpyDeviceToDevice;
-        struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
+        ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
         int id;
         CUDA_CHECK(cudaGetDevice(&id));
         src_ptr = (char *) extra->data_device[id];
@@ -5735,6 +5789,107 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     }
 }
 
+static void ggml_cuda_op_repeat(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & stream) {
+    // guaranteed to be an integer due to the check in ggml_can_repeat
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const int64_t ne3 = dst->ne[3];
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const size_t nb0 = dst->nb[0];
+    const size_t nb1 = dst->nb[1];
+    const size_t nb2 = dst->nb[2];
+    const size_t nb3 = dst->nb[3];
+
+    const size_t nb00 = src0->nb[0];
+    const size_t nb01 = src0->nb[1];
+    const size_t nb02 = src0->nb[2];
+    const size_t nb03 = src0->nb[3];
+
+    const int nr0 = (int)(ne0/ne00);
+    const int nr1 = (int)(ne1/ne01);
+    const int nr2 = (int)(ne2/ne02);
+    const int nr3 = (int)(ne3/ne03);
+
+    // TODO: support for transposed / permuted tensors
+    GGML_ASSERT(nb0  == sizeof(float));
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // TODO: very inefficient, implement in a kernel, or fewer cudaMemcpyAsync calls for contiguous tensors
+    for                         (int i3 = 0; i3 < nr3;  i3++) {
+        for                     (int k3 = 0; k3 < ne03; k3++) {
+            for                 (int i2 = 0; i2 < nr2;  i2++) {
+                for             (int k2 = 0; k2 < ne02; k2++) {
+                    for         (int i1 = 0; i1 < nr1;  i1++) {
+                        for     (int k1 = 0; k1 < ne01; k1++) {
+                            for (int i0 = 0; i0 < nr0;  i0++) {
+                                CUDA_CHECK(cudaMemcpyAsync(
+                                              (char *)  dst_d + (i3*ne03 + k3)*nb3  + (i2*ne02 + k2)*nb2  + (i1*ne01 + k1)*nb1  + (i0*ne00)*nb0,
+                                        (const char *) src0_d + (          k3)*nb03 + (          k2)*nb02 + (          k1)*nb01,
+                                        ne00*nb0, cudaMemcpyDeviceToDevice, stream));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    (void) src1;
+    (void) src1_d;
+}
+
+static void ggml_cuda_op_get_rows(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int ncols = src0->ne[0];
+    const int nrows = ggml_nelements(src1);
+
+    const int32_t * src1_i32 = (const int32_t *) src1_d;
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda<1, 1, convert_f16>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda<1, 1, convert_f32>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
 inline void ggml_cuda_op_add(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6275,12 +6430,12 @@ inline void ggml_cuda_op_alibi(
     const int64_t ne02 = src0->ne[2];
     const int64_t nrows = ggml_nrows(src0);
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    GGML_ASSERT(ne01 + n_past == ne00);
+    //GGML_ASSERT(ne01 + n_past == ne00);
     GGML_ASSERT(n_head == ne02);
 
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -6339,7 +6494,14 @@ inline void ggml_cuda_op_scale(
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const float scale = ((float *) src1->data)[0];
+    float scale;
+    // HACK: support for ggml backend interface
+    if (src1->backend == GGML_BACKEND_CPU) {
+        scale = ((float *) src1->data)[0];
+    } else {
+        // TODO: pass pointer to kernel instead of copying to host
+        CUDA_CHECK(cudaMemcpy(&scale, src1->data, sizeof(float), cudaMemcpyDeviceToHost));
+    }
 
     scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
     CUDA_CHECK(cudaGetLastError());
@@ -6349,6 +6511,24 @@ inline void ggml_cuda_op_scale(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_clamp(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const float min = ((float *) dst->op_params)[0];
+    const float max = ((float *) dst->op_params)[1];
+
+    clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) {
     const int64_t nrows0 = ggml_nrows(src0);
 
@@ -6358,9 +6538,9 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(              dst->backend != GGML_BACKEND_GPU_SPLIT);
 
-    struct ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    struct ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
+    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
+    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
 
     const bool src0_on_device =             src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
     const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
@@ -6501,9 +6681,9 @@ static void ggml_cuda_op_mul_mat(
     const size_t q8_1_ts = sizeof(block_q8_1);
     const size_t q8_1_bs = QK8_1;
 
-    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    struct ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
     const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
@@ -6581,7 +6761,7 @@ static void ggml_cuda_op_mul_mat(
         if (convert_src1_to_q8_1) {
             src1_ddq[id] = (char *) ggml_cuda_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]);
 
-            if (split && src1_on_device && src1_is_contiguous) {
+            if (src1_on_device && src1_is_contiguous) {
                 quantize_row_q8_1_cuda(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream);
                 CUDA_CHECK(cudaGetLastError());
             }
@@ -6663,7 +6843,7 @@ static void ggml_cuda_op_mul_mat(
                     GGML_ASSERT(false);
                 }
 
-                if (convert_src1_to_q8_1 && src1->backend == GGML_BACKEND_CPU) {
+                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) {
                     quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
@@ -6754,6 +6934,14 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
+static void ggml_cuda_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_repeat);
+}
+
+static void ggml_cuda_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_get_rows);
+}
+
 static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
 }
@@ -6808,13 +6996,13 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
-    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
 
-    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
     float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
 
-    struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
     ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
@@ -6839,13 +7027,13 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
-    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
 
-    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
     float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
 
-    struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
     const int64_t row_stride_x = nb01 / sizeof(half);
@@ -6866,11 +7054,11 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         }
     }
 
-    if (all_on_device && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+    if (all_on_device && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
     } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) {
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    }else if (src0->type == GGML_TYPE_F32) {
+    } else if (src0->type == GGML_TYPE_F32) {
         ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
         if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) {
@@ -6902,6 +7090,10 @@ static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1,
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
 }
 
+static void ggml_cuda_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_clamp);
+}
+
 static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -6931,8 +7123,8 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
-    const struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    const struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
 
     char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
     char * src1_ddc = (char *) src1_extra->data_device[g_main_device];
@@ -6987,8 +7179,8 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 
     const size_t nb1 = tensor->nb[1];
 
-    ggml_backend backend = tensor->backend;
-    struct ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu;
+    ggml_backend_type backend = tensor->backend;
+    ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu;
     memset(extra, 0, sizeof(*extra));
 
     for (int64_t id = 0; id < g_device_count; ++id) {
@@ -7042,7 +7234,6 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
             CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
         }
 
-
         CUDA_CHECK(cudaMemcpy(buf, buf_host, original_size, cudaMemcpyHostToDevice));
 
         extra->data_device[id] = buf;
@@ -7081,17 +7272,17 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) {
     delete extra;
 }
 
-static struct ggml_tensor_extra_gpu * g_temp_tensor_extras = nullptr;
+static ggml_tensor_extra_gpu * g_temp_tensor_extras = nullptr;
 static size_t g_temp_tensor_extra_index = 0;
 
-static struct ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
+static ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
     if (g_temp_tensor_extras == nullptr) {
         g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_MAX_NODES];
     }
 
     size_t alloc_index = g_temp_tensor_extra_index;
     g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_MAX_NODES;
-    struct ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index];
+    ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index];
     memset(extra, 0, sizeof(*extra));
 
     return extra;
@@ -7119,7 +7310,7 @@ static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scra
         return;
     }
 
-    struct ggml_tensor_extra_gpu * extra;
+    ggml_tensor_extra_gpu * extra;
 
     const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
         tensor->op == GGML_OP_VIEW ||
@@ -7128,7 +7319,7 @@ static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scra
 
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
-        struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
+        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
@@ -7137,7 +7328,7 @@ static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scra
         extra = ggml_cuda_alloc_temp_tensor_extra();
         extra->data_device[g_main_device] = src0_ddc + offset;
     } else if (tensor->op == GGML_OP_CPY) {
-        struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra;
+        ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra;
         void * src1_ddv = src1_extra->data_device[g_main_device];
         extra = ggml_cuda_alloc_temp_tensor_extra();
         extra->data_device[g_main_device] = src1_ddv;
@@ -7179,13 +7370,13 @@ void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset)
         CUDA_CHECK(cudaMalloc(&g_scratch_buffer, g_scratch_size));
     }
 
-    struct ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
+    ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
 
     const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
         tensor->op == GGML_OP_VIEW;
 
     if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
-        struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
+        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t view_offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
@@ -7203,7 +7394,7 @@ void ggml_cuda_copy_to_device(struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
     GGML_ASSERT(ggml_is_contiguous(tensor));
 
-    struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     CUDA_CHECK(cudaMemcpy(extra->data_device[g_main_device], tensor->data, ggml_nbytes(tensor), cudaMemcpyHostToDevice));
 }
@@ -7260,58 +7451,47 @@ void ggml_cuda_free_scratch() {
     g_scratch_buffer = nullptr;
 }
 
-bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor){
+bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     ggml_cuda_func_t func;
     const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
         || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
         || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
 
+    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT) {
+        return false;
+    }
+
     switch (tensor->op) {
+        case GGML_OP_REPEAT:
+            func = ggml_cuda_repeat;
+            break;
+        case GGML_OP_GET_ROWS:
+            func = ggml_cuda_get_rows;
+            break;
         case GGML_OP_DUP:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_dup;
             break;
         case GGML_OP_ADD:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_add;
             break;
         case GGML_OP_MUL:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_mul;
             break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(tensor)) {
                 case GGML_UNARY_OP_GELU:
-                    if (!any_on_device) {
-                        return false;
-                    }
                     func = ggml_cuda_gelu;
                     break;
                 case GGML_UNARY_OP_SILU:
-                    if (!any_on_device) {
-                        return false;
-                    }
                     func = ggml_cuda_silu;
                     break;
                 default:
                     return false;
             } break;
         case GGML_OP_NORM:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_norm;
             break;
         case GGML_OP_RMS_NORM:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_rms_norm;
             break;
         case GGML_OP_MUL_MAT:
@@ -7321,54 +7501,36 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             func = ggml_cuda_mul_mat;
             break;
         case GGML_OP_SCALE:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_scale;
             break;
-        case GGML_OP_CPY:
+        case GGML_OP_CLAMP:
             if (!any_on_device) {
                 return false;
             }
+            func = ggml_cuda_clamp;
+            break;
+        case GGML_OP_CPY:
             func = ggml_cuda_cpy;
             break;
         case GGML_OP_CONT:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_dup;
             break;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_nop;
             break;
         case GGML_OP_DIAG_MASK_INF:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_diag_mask_inf;
             break;
         case GGML_OP_SOFT_MAX:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_soft_max;
             break;
         case GGML_OP_ROPE:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_rope;
             break;
         case GGML_OP_ALIBI:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_alibi;
             break;
         default:
@@ -7396,3 +7558,263 @@ void ggml_cuda_get_device_description(int device, char * description, size_t des
     CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
     snprintf(description, description_size, "%s", prop.name);
 }
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+#define UNUSED GGML_UNUSED
+
+struct ggml_backend_context_cuda {
+};
+
+static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
+    return GGML_CUDA_NAME;
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cuda_free(ggml_backend_t backend) {
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+    delete cuda_ctx;
+    delete backend;
+}
+
+struct ggml_backend_buffer_context_cuda {
+    void * device;
+
+    ggml_tensor_extra_gpu * temp_tensor_extras = nullptr;
+    size_t temp_tensor_extra_index = 0;
+
+    ~ggml_backend_buffer_context_cuda() {
+        delete[] temp_tensor_extras;
+    }
+
+    ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
+        if (temp_tensor_extras == nullptr) {
+            temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_MAX_NODES];
+        }
+
+        size_t alloc_index = temp_tensor_extra_index;
+        temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_MAX_NODES;
+        ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index];
+        memset(extra, 0, sizeof(*extra));
+
+        return extra;
+    }
+};
+
+static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+    CUDA_CHECK(cudaFree(ctx->device));
+    delete ctx;
+}
+
+static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+    return ctx->device;
+}
+
+static size_t ggml_backend_cuda_buffer_get_alloc_size(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    int64_t row_low = 0;
+    int64_t row_high = ggml_nrows(tensor);
+    int64_t nrows_split = row_high - row_low;
+
+    size_t size = ggml_nbytes_split(tensor, nrows_split);
+
+    int64_t ne0 = tensor->ne[0];
+
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
+                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+        }
+    }
+
+    return size;
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        assert(tensor->view_src->buffer->backend == buffer->backend);
+        tensor->backend = tensor->view_src->backend;
+        tensor->extra = tensor->view_src->extra;
+        return;
+    }
+
+    ggml_tensor_extra_gpu * extra = ctx->ggml_cuda_alloc_temp_tensor_extra();
+
+    extra->data_device[g_main_device] = tensor->data;
+
+    tensor->backend = GGML_BACKEND_GPU;
+    tensor->extra = extra;
+
+    if (ggml_is_quantized(tensor->type)) {
+        // initialize padding to 0 to avoid possible NaN values
+        int64_t row_low = 0;
+        int64_t row_high = ggml_nrows(tensor);
+        int64_t nrows_split = row_high - row_low;
+
+        size_t original_size = ggml_nbytes_split(tensor, nrows_split);
+        size_t padded_size = ggml_backend_cuda_buffer_get_alloc_size(tensor->buffer, tensor);
+
+        if (padded_size > original_size && tensor->view_src == nullptr) {
+            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[g_main_device][0]));
+        }
+    }
+
+    UNUSED(buffer);
+}
+
+static struct ggml_backend_buffer_i cuda_backend_buffer_interface = {
+    /* .free_buffer    = */ ggml_backend_cuda_buffer_free_buffer,
+    /* .get_base       = */ ggml_backend_cuda_buffer_get_base,
+    /* .get_alloc_size = */ ggml_backend_cuda_buffer_get_alloc_size,
+    /* .init_tensor    = */ ggml_backend_cuda_buffer_init_tensor,
+    /* .free_tensor    = */ NULL,
+};
+
+static ggml_backend_buffer_t ggml_backend_cuda_alloc_buffer(ggml_backend_t backend, size_t size) {
+    ggml_cuda_set_device(g_main_device);
+
+    ggml_backend_buffer_context_cuda * ctx = new ggml_backend_buffer_context_cuda;
+    CUDA_CHECK(cudaMalloc(&ctx->device, size));
+    return ggml_backend_buffer_init(backend, cuda_backend_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_cuda_get_alignment(ggml_backend_t backend) {
+    return 128;
+    UNUSED(backend);
+}
+
+static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[g_main_device][0]));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+    CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+
+    UNUSED(backend);
+}
+
+static ggml_backend_graph_plan_t ggml_backend_cuda_graph_plan_create(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    GGML_ASSERT(!"not implemented");
+
+    return nullptr;
+
+    UNUSED(backend);
+    UNUSED(cgraph);
+}
+
+static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(!"not implemented");
+
+    UNUSED(backend);
+    UNUSED(plan);
+}
+
+static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(!"not implemented");
+
+    UNUSED(backend);
+    UNUSED(plan);
+}
+
+static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_cuda_set_device(g_main_device);
+
+    ggml_compute_params params = {};
+    params.type = GGML_TASK_COMPUTE;
+    params.ith = 0;
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor * node = cgraph->nodes[i];
+
+        assert(node->backend == GGML_BACKEND_GPU);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            if (node->src[j] != nullptr) {
+                assert(node->src[j]->backend == GGML_BACKEND_GPU);
+            }
+        }
+
+        bool ok = ggml_cuda_compute_forward(&params, node);
+        if (!ok) {
+            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+
+#if 0
+        if (node->type == GGML_TYPE_F32) {
+            cudaDeviceSynchronize();
+            std::vector<float> tmp(ggml_nelements(node), 0.0f);
+            cudaMemcpy(tmp.data(), node->data, ggml_nelements(node)*sizeof(float), cudaMemcpyDeviceToHost);
+            printf("\n%s (%s) (%s %s) (%s %s): ", node->name, ggml_op_name(node->op),
+                ggml_type_name(node->src[0]->type),
+                node->src[1] ? ggml_type_name(node->src[1]->type) : "none",
+                node->src[0]->name,
+                node->src[1] ? node->src[1]->name : "none");
+            double sum = 0.0;
+            double sq_sum = 0.0;
+            for (int i = 0; i < ggml_nelements(node); i++) {
+                printf("%f ", tmp[i]);
+                sum += tmp[i];
+                sq_sum += tmp[i]*tmp[i];
+            }
+            printf("\n");
+            printf("sum: %f, ", sum);
+            printf("sq_sum: %f\n", sq_sum);
+        }
+#endif
+    }
+
+    UNUSED(backend);
+}
+
+static ggml_backend_i cuda_backend_i = {
+    /* .get_name            = */ ggml_backend_cuda_name,
+    /* .free                = */ ggml_backend_cuda_free,
+    /* .alloc_buffer        = */ ggml_backend_cuda_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_cuda_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_cuda_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_cuda_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_cuda_synchronize,
+    /* .cpy_tensor_from     = */ nullptr,
+    /* .cpy_tensor_to       = */ nullptr,
+    /* .graph_plan_create   = */ ggml_backend_cuda_graph_plan_create,
+    /* .graph_plan_free     = */ ggml_backend_cuda_graph_plan_free,
+    /* .graph_plan_compute  = */ ggml_backend_cuda_graph_plan_compute,
+    /* .graph_compute       = */ ggml_backend_cuda_graph_compute,
+    /* .supports_op         = */ nullptr,
+};
+
+ggml_backend_t ggml_backend_cuda_init() {
+    ggml_init_cublas(); // TODO: remove from ggml.c
+
+    ggml_backend_context_cuda * ctx = new ggml_backend_context_cuda;
+
+    ggml_backend_t cuda_backend = new ggml_backend {
+        /* .interface = */ cuda_backend_i,
+        /* .context   = */ ctx
+    };
+
+    return cuda_backend;
+}

ggml-cuda.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
@@ -42,6 +43,9 @@ GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, s
 GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
+// backend API
+GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+
 #ifdef  __cplusplus
 }
 #endif

ggml-metal.h

@@ -20,6 +20,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #include <stddef.h>
 #include <stdbool.h>
@@ -35,10 +36,15 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+//
+// internal API
+// temporary exposed to user-code
+//
 
 struct ggml_metal_context;
 
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 // number of command buffers to use
 struct ggml_metal_context * ggml_metal_init(int n_cb);
 void ggml_metal_free(struct ggml_metal_context * ctx);
@@ -83,6 +89,17 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
 #ifdef __cplusplus
 }
 #endif

ggml-metal.m

@@ -779,8 +779,8 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_CONCAT:
                         {
+                            const int64_t nb = ne00;
 
-                            int64_t nb = ne00;
                             [encoder setComputePipelineState:ctx->pipeline_concat];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
@@ -812,6 +812,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
 
                             const int nth = MIN(1024, ne0);
+
                             [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ADD:
@@ -909,9 +910,10 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst)/4;
+                            const int64_t n = ggml_nelements(dst);
+                            GGML_ASSERT(n % 4 == 0);
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {
@@ -921,9 +923,10 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst)/4;
+                                    const int64_t n = ggml_nelements(dst);
+                                    GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             case GGML_UNARY_OP_RELU:
                                 {
@@ -941,9 +944,10 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst)/4;
+                                    const int64_t n = ggml_nelements(dst);
+                                    GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             default:
                                 {
@@ -1040,7 +1044,7 @@ void ggml_metal_graph_compute(
                                 !ggml_is_transposed(src0) &&
                                 !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
-                                ne00 % 32 == 0 &&
+                                ne00 % 32 == 0 && ne00 >= 64 &&
                                 ne11 > ne11_mm_min) {
                                 //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
                                 switch (src0->type) {
@@ -1251,6 +1255,8 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_RMS_NORM:
                         {
+                            GGML_ASSERT(ne00 % 4 == 0);
+
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
@@ -1293,7 +1299,7 @@ void ggml_metal_graph_compute(
 
                             const int nth = MIN(1024, ne00);
 
-                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            //const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_head = ((int32_t *) dst->op_params)[1];
                             float max_bias;
                             memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
@@ -1471,3 +1477,140 @@ preferably one under the recommended max working set size, or else fall back to
 
     }
 }
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+static const char * ggml_backend_metal_name(ggml_backend_t backend) {
+    return "Metal";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_free(ggml_backend_t backend) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    ggml_metal_free(ctx);
+    free(backend);
+}
+
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+static struct ggml_backend_buffer_i metal_backend_buffer_i = {
+    /* .free_buffer    = */ ggml_backend_metal_buffer_free_buffer,
+    /* .get_base       = */ ggml_backend_metal_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+static ggml_backend_buffer_t ggml_backend_metal_alloc_buffer(ggml_backend_t backend, size_t size) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+
+    void * data = ggml_metal_host_malloc(size);
+
+    // TODO: set proper name of the buffers
+    ggml_metal_add_buffer(ctx, "backend", data, size, 0);
+
+    return ggml_backend_buffer_init(backend, metal_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_metal_get_alignment(ggml_backend_t backend) {
+    return 32;
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy((char *)tensor->data + offset, data, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+
+    ggml_metal_graph_compute(metal_ctx, cgraph);
+}
+
+static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i metal_backend_i = {
+    /* .get_name            = */ ggml_backend_metal_name,
+    /* .free                = */ ggml_backend_metal_free,
+    /* .alloc_buffer        = */ ggml_backend_metal_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_metal_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_metal_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_metal_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_metal_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_metal_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_metal_cpy_tensor_to,
+    /* .graph_plan_create   = */ NULL, // the metal implementation does not require creating graph plans atm
+    /* .graph_plan_free     = */ NULL,
+    /* .graph_plan_compute  = */ NULL,
+    /* .graph_compute       = */ ggml_backend_metal_graph_compute,
+    /* .supports_op         = */ ggml_backend_metal_supports_op,
+};
+
+ggml_backend_t ggml_backend_metal_init(void) {
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
+
+    ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
+
+    *metal_backend = (struct ggml_backend) {
+        /* .interface = */ metal_backend_i,
+        /* .context   = */ ctx,
+    };
+
+    return metal_backend;
+}
+
+bool ggml_backend_is_metal(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_metal_name;
+}
+
+void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+
+    ggml_metal_set_n_cb(ctx, n_cb);
+}

ggml-metal.metal

@@ -345,10 +345,11 @@ kernel void kernel_rms_norm(
         uint sgitg[[simdgroup_index_in_threadgroup]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
-    device const float * x_scalar = (device const float *) x;
-    float4 sumf=0;
-    float all_sum=0;
+    device const float4 * x        = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+    device const float  * x_scalar = (device const float  *) x;
+
+    float4 sumf = 0;
+    float all_sum = 0;
 
     // parallel sum
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
@@ -361,6 +362,7 @@ kernel void kernel_rms_norm(
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // broadcast, simd group number is ntg / 32
     for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
        if (tpitg < i) {
@@ -368,7 +370,9 @@ kernel void kernel_rms_norm(
        }
     }
     if (tpitg == 0) {
-        for (int i = 4 * (ne00 / 4); i < ne00; i++) {sum[0] += x_scalar[i];}
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
         sum[0] /= ne00;
     }
 
@@ -383,7 +387,9 @@ kernel void kernel_rms_norm(
         y[i00] = x[i00] * scale;
     }
     if (tpitg == 0) {
-        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {y_scalar[i00] = x_scalar[i00] * scale;}
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
     }
 }
 

ggml.c

@@ -162,40 +162,16 @@ typedef void * thread_ret_t;
 
 #define GGML_PRINT(...) printf(__VA_ARGS__)
 
+//
+// end of logging block
+//
+
 #ifdef GGML_USE_ACCELERATE
 // uncomment to use vDSP for soft max computation
 // note: not sure if it is actually faster
 //#define GGML_SOFT_MAX_ACCELERATE
 #endif
 
-//
-// logging
-//
-
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_10(...)
-#endif
-
-#define GGML_PRINT(...) printf(__VA_ARGS__)
-
-//
-// end of logging block
-//
-
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
 #define GGML_ALIGNED_FREE(ptr)    _aligned_free(ptr)
@@ -4952,6 +4928,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     *result = (struct ggml_tensor) {
         /*.type         =*/ type,
         /*.backend      =*/ GGML_BACKEND_CPU,
+        /*.buffer       =*/ NULL,
         /*.n_dims       =*/ n_dims,
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
@@ -11257,7 +11234,7 @@ static void ggml_compute_forward_silu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
             UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
@@ -13083,24 +13060,22 @@ static void ggml_compute_forward_alibi_f32(
         return;
     }
 
-    const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    assert(n_past >= 0);
-
-    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int ne1 = src0->ne[1]; // seq_len_without_past
-    const int ne2 = src0->ne[2]; // n_head -> this is k
-    //const int ne3 = src0->ne[3]; // 1 -> bsz
+    const int64_t ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
+    const int64_t ne1 = src0->ne[1]; // seq_len_without_past
+    const int64_t ne2 = src0->ne[2]; // n_head -> this is k
+    //const int64_t ne3 = src0->ne[3]; // 1 -> bsz
 
-    const int n  = ggml_nrows(src0);
-    const int ne2_ne3 = n/ne1; // ne2*ne3
+    const int64_t n  = ggml_nrows(src0);
+    const int64_t ne2_ne3 = n/ne1; // ne2*ne3
 
-    const int nb0 = src0->nb[0];
-    const int nb1 = src0->nb[1];
-    const int nb2 = src0->nb[2];
+    const size_t nb0 = src0->nb[0];
+    const size_t nb1 = src0->nb[1];
+    const size_t nb2 = src0->nb[2];
     //const int nb3 = src0->nb[3];
 
     GGML_ASSERT(nb0 == sizeof(float));
@@ -13112,9 +13087,9 @@ static void ggml_compute_forward_alibi_f32(
     const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
-    for (int i = 0; i < ne0; i++) {
-        for (int j = 0; j < ne1; j++) {
-            for (int k = 0; k < ne2_ne3; k++) {
+    for (int64_t i = 0; i < ne0; i++) {
+        for (int64_t j = 0; j < ne1; j++) {
+            for (int64_t k = 0; k < ne2_ne3; k++) {
                 float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
                 float *      pdst = (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
 
@@ -13129,7 +13104,6 @@ static void ggml_compute_forward_alibi_f32(
                 }
 
                 pdst[0] = i * m_k + src[0];
-
             }
         }
     }
@@ -20200,6 +20174,10 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         ggml_vec_cpy_f32(nx, xp, x);
         ggml_vec_cpy_f32(nx, gp, g);
 
+        // TODO: instead of passing &cancel here, use the return code of the linesearch
+        //       to determine if the optimization should be cancelled
+        //       this is a simple change, but not doing this atm, since I don't have a nice
+        //       way to test and don't want to break something with so many changes lined up
         ls = linesearch_backtracking(&params, nx, x, &fx, g, d, step, xp, f, gb, &cplan, np, ps, &cancel, callback, callback_data);
         if (cancel) {
             return GGML_OPT_CANCEL;

ggml.h

@@ -326,7 +326,7 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
-    enum ggml_backend {
+    enum ggml_backend_type {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_GPU = 10,
         GGML_BACKEND_GPU_SPLIT = 20,
@@ -479,8 +479,10 @@ extern "C" {
 
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type    type;
-        enum ggml_backend backend;
+        enum ggml_type         type;
+        enum ggml_backend_type backend;
+
+        struct ggml_backend_buffer * buffer;
 
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
@@ -514,7 +516,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[4];
+        char padding[12];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -1358,7 +1360,7 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_alibi(
+    GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
@@ -1367,7 +1369,7 @@ extern "C" {
 
     // clamp
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_clamp(
+    GGML_API struct ggml_tensor * ggml_clamp(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             float                 min,
@@ -2102,7 +2104,7 @@ extern "C" {
         enum ggml_type    vec_dot_type;
     } ggml_type_traits_t;
 
-    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }

gguf-py/gguf/gguf.py

@@ -88,29 +88,31 @@ class MODEL_ARCH(IntEnum):
     PERSIMMON     : int = auto()
     REFACT        : int = auto()
     BERT          : int = auto()
+    BLOOM         : int = auto()
 
 
 class MODEL_TENSOR(IntEnum):
-    TOKEN_EMBD   : int = auto()
-    TOKEN_TYPES  : int = auto()
-    POS_EMBD     : int = auto()
-    OUTPUT       : int = auto()
-    OUTPUT_NORM  : int = auto()
-    ROPE_FREQS   : int = auto()
-    ATTN_Q       : int = auto()
-    ATTN_K       : int = auto()
-    ATTN_V       : int = auto()
-    ATTN_QKV     : int = auto()
-    ATTN_OUT     : int = auto()
-    ATTN_NORM    : int = auto()
-    ATTN_NORM_2  : int = auto()
-    ATTN_ROT_EMBD: int = auto()
-    FFN_GATE     : int = auto()
-    FFN_DOWN     : int = auto()
-    FFN_UP       : int = auto()
-    FFN_NORM     : int = auto()
-    ATTN_Q_NORM  : int = auto()
-    ATTN_K_NORM  : int = auto()
+    TOKEN_EMBD      : int = auto()
+    TOKEN_EMBD_NORM : int = auto()
+    TOKEN_TYPES     : int = auto()
+    POS_EMBD        : int = auto()
+    OUTPUT          : int = auto()
+    OUTPUT_NORM     : int = auto()
+    ROPE_FREQS      : int = auto()
+    ATTN_Q          : int = auto()
+    ATTN_K          : int = auto()
+    ATTN_V          : int = auto()
+    ATTN_QKV        : int = auto()
+    ATTN_OUT        : int = auto()
+    ATTN_NORM       : int = auto()
+    ATTN_NORM_2     : int = auto()
+    ATTN_ROT_EMBD   : int = auto()
+    FFN_GATE        : int = auto()
+    FFN_DOWN        : int = auto()
+    FFN_UP          : int = auto()
+    FFN_NORM        : int = auto()
+    ATTN_Q_NORM     : int = auto()
+    ATTN_K_NORM     : int = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -125,29 +127,31 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PERSIMMON:      "persimmon",
     MODEL_ARCH.REFACT:         "refact",
     MODEL_ARCH.BERT:           "bert",
+    MODEL_ARCH.BLOOM:          "bloom",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
-    MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
-    MODEL_TENSOR.TOKEN_TYPES:   "token_types",
-    MODEL_TENSOR.POS_EMBD:      "position_embd",
-    MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
-    MODEL_TENSOR.OUTPUT:        "output",
-    MODEL_TENSOR.ROPE_FREQS:    "rope_freqs",
-    MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
-    MODEL_TENSOR.ATTN_NORM_2:   "blk.{bid}.attn_norm_2",
-    MODEL_TENSOR.ATTN_QKV:      "blk.{bid}.attn_qkv",
-    MODEL_TENSOR.ATTN_Q:        "blk.{bid}.attn_q",
-    MODEL_TENSOR.ATTN_K:        "blk.{bid}.attn_k",
-    MODEL_TENSOR.ATTN_V:        "blk.{bid}.attn_v",
-    MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
-    MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
-    MODEL_TENSOR.ATTN_Q_NORM:   "blk.{bid}.attn_q_norm",
-    MODEL_TENSOR.ATTN_K_NORM:   "blk.{bid}.attn_k_norm",
-    MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
-    MODEL_TENSOR.FFN_GATE:      "blk.{bid}.ffn_gate",
-    MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
-    MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+    MODEL_TENSOR.TOKEN_EMBD:      "token_embd",
+    MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm",
+    MODEL_TENSOR.TOKEN_TYPES:     "token_types",
+    MODEL_TENSOR.POS_EMBD:        "position_embd",
+    MODEL_TENSOR.OUTPUT_NORM:     "output_norm",
+    MODEL_TENSOR.OUTPUT:          "output",
+    MODEL_TENSOR.ROPE_FREQS:      "rope_freqs",
+    MODEL_TENSOR.ATTN_NORM:       "blk.{bid}.attn_norm",
+    MODEL_TENSOR.ATTN_NORM_2:     "blk.{bid}.attn_norm_2",
+    MODEL_TENSOR.ATTN_QKV:        "blk.{bid}.attn_qkv",
+    MODEL_TENSOR.ATTN_Q:          "blk.{bid}.attn_q",
+    MODEL_TENSOR.ATTN_K:          "blk.{bid}.attn_k",
+    MODEL_TENSOR.ATTN_V:          "blk.{bid}.attn_v",
+    MODEL_TENSOR.ATTN_OUT:        "blk.{bid}.attn_output",
+    MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
+    MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
+    MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
+    MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
+    MODEL_TENSOR.FFN_DOWN:        "blk.{bid}.ffn_down",
+    MODEL_TENSOR.FFN_UP:          "blk.{bid}.ffn_up",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -282,6 +286,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.BLOOM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.GPT2: [
         # TODO
     ],
@@ -311,6 +327,7 @@ class TensorNameMap:
             "gpt_neox.embed_in",                        # gptneox
             "transformer.wte",                          # gpt2 gpt-j mpt refact
             "transformer.word_embeddings",              # falcon
+            "word_embeddings",                          # bloom
             "model.embed_tokens",                       # llama-hf
             "tok_embeddings",                           # llama-pth
             "embeddings.word_embeddings",               # bert
@@ -322,6 +339,11 @@ class TensorNameMap:
             "embeddings.token_type_embeddings",  # bert
         ),
 
+        # Normalization of token embeddings
+        MODEL_TENSOR.TOKEN_EMBD_NORM: (
+            "word_embeddings_layernorm",  # bloom
+        ),
+
         # Position embeddings
         MODEL_TENSOR.POS_EMBD: (
             "transformer.wpe",                 # gpt2
@@ -332,7 +354,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                # gptneox
             "lm_head",                  # gpt2 mpt falcon llama-hf baichuan
-            "output",                   # llama-pth
+            "output",                   # llama-pth bloom
             "word_embeddings_for_head", # persimmon
         ),
 
@@ -344,7 +366,7 @@ class TensorNameMap:
             "norm",                                   # llama-pth
             "embeddings.LayerNorm",                   # bert
             "transformer.norm_f",                     # mpt
-            "ln_f",                                   # refact
+            "ln_f",                                   # refact bloom
             "language_model.encoder.final_layernorm", # persimmon
         ),
 
@@ -361,6 +383,7 @@ class TensorNameMap:
             "transformer.h.{bid}.ln_1",                            # gpt2 gpt-j refact
             "transformer.blocks.{bid}.norm_1",                     # mpt
             "transformer.h.{bid}.input_layernorm",                 # falcon7b
+            "h.{bid}.input_layernorm",                             # bloom
             "transformer.h.{bid}.ln_mlp",                          # falcon40b
             "model.layers.{bid}.input_layernorm",                  # llama-hf
             "layers.{bid}.attention_norm",                         # llama-pth
@@ -379,6 +402,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.c_attn",                                    # gpt2
             "transformer.blocks.{bid}.attn.Wqkv",                                 # mpt
             "transformer.h.{bid}.self_attention.query_key_value",                 # falcon
+            "h.{bid}.self_attention.query_key_value",                             # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
         ),
 
@@ -412,6 +436,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.c_proj",                         # gpt2 refact
             "transformer.blocks.{bid}.attn.out_proj",                  # mpt
             "transformer.h.{bid}.self_attention.dense",                # falcon
+            "h.{bid}.self_attention.dense",                            # bloom
             "model.layers.{bid}.self_attn.o_proj",                     # llama-hf
             "layers.{bid}.attention.wo",                               # llama-pth
             "encoder.layer.{bid}.attention.output.dense",              # bert
@@ -429,6 +454,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_NORM: (
             "gpt_neox.layers.{bid}.post_attention_layernorm",               # gptneox
             "transformer.h.{bid}.ln_2",                                     # gpt2 refact
+            "h.{bid}.post_attention_layernorm",                             # bloom
             "transformer.blocks.{bid}.norm_2",                              # mpt
             "model.layers.{bid}.post_attention_layernorm",                  # llama-hf
             "layers.{bid}.ffn_norm",                                        # llama-pth
@@ -442,6 +468,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.c_fc",                          # gpt2
             "transformer.blocks.{bid}.ffn.up_proj",                  # mpt
             "transformer.h.{bid}.mlp.dense_h_to_4h",                 # falcon
+            "h.{bid}.mlp.dense_h_to_4h",                             # bloom
             "model.layers.{bid}.mlp.up_proj",                        # llama-hf refact
             "layers.{bid}.feed_forward.w3",                          # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                # bert
@@ -461,6 +488,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.c_proj",                        # gpt2 refact
             "transformer.blocks.{bid}.ffn.down_proj",                # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                 # falcon
+            "h.{bid}.mlp.dense_4h_to_h",                             # bloom
             "model.layers.{bid}.mlp.down_proj",                      # llama-hf
             "layers.{bid}.feed_forward.w2",                          # llama-pth
             "encoder.layer.{bid}.output.dense",                      # bert

gpttype_adapter.cpp

@@ -1768,7 +1768,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
     int realnpredict = params.n_predict-stopper_unused_tokens;
     float pt2 = (time2*1000.0/(realnpredict==0?1:realnpredict));
     float tokens_per_second = (realnpredict == 0 ? 0 : realnpredict / (time1 + time2));
-    printf("\nTime Taken - Processing:%.1fs (%.0fms/T), Generation:%.1fs (%.0fms/T), Total:%.1fs (%.1fT/s)", time1, pt1, time2, pt2, (time1 + time2), tokens_per_second);
+    printf("\nContextLimit: %d/%d, Processing:%.1fs (%.0fms/T), Generation:%.1fs (%.0fms/T), Total:%.1fs (%.1fT/s)",current_context_tokens.size(),nctx, time1, pt1, time2, pt2, (time1 + time2), tokens_per_second);
     fflush(stdout);
     output.status = 1;
     generation_finished = true;

koboldcpp.py

@@ -184,6 +184,10 @@ def init_library():
         os.add_dll_directory(dir_path)
         os.add_dll_directory(abs_path)
         os.add_dll_directory(os.getcwd())
+        if libname == lib_hipblas and "HIP_PATH" in os.environ:
+            os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin"))
+            if args.debugmode == 1:
+                print(f"HIP/ROCm SDK at {os.environ['HIP_PATH']} included in .DLL load path")
     handle = ctypes.CDLL(os.path.join(dir_path, libname))
 
     handle.load_model.argtypes = [load_model_inputs]
@@ -361,7 +365,7 @@ maxhordelen = 256
 modelbusy = threading.Lock()
 requestsinqueue = 0
 defaultport = 5001
-KcppVersion = "1.46.1"
+KcppVersion = "1.47"
 showdebug = True
 showsamplerwarning = True
 showmaxctxwarning = True
@@ -369,6 +373,8 @@ session_kudos_earned = 0
 session_jobs = 0
 session_starttime = None
 exitcounter = 0
+punishcounter = 0 #causes a timeout if too many errors
+rewardcounter = 0 #reduces error counts for successful jobs
 totalgens = 0
 currentusergenkey = "" #store a special key so polled streaming works even in multiuser
 args = None #global args
@@ -412,16 +418,34 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
             elif api_format==4:
                 # translate openai chat completion messages format into one big string.
                 messages_array = genparams.get('messages', [])
+                adapter_obj = genparams.get('adapter', {})
                 messages_string = ""
+                system_message_start = adapter_obj.get("system_start", "\n### Instruction:\n")
+                system_message_end = adapter_obj.get("system_end", "")
+                user_message_start = adapter_obj.get("user_start", "\n### Instruction:\n")
+                user_message_end = adapter_obj.get("user_end", "")
+                assistant_message_start = adapter_obj.get("assistant_start", "\n### Response:\n")
+                assistant_message_end = adapter_obj.get("assistant_end", "")
+
                 for message in messages_array:
                     if message['role'] == "system":
-                        messages_string+="\n### Instruction:\n"
+                        messages_string += system_message_start
                     elif message['role'] == "user":
-                        messages_string+="\n### Instruction:\n"
+                        messages_string += user_message_start
                     elif message['role'] == "assistant":
-                        messages_string+="\n### Response:\n"
-                    messages_string+=message['content']
-                messages_string += "\n### Response:\n"
+                        messages_string += assistant_message_start
+
+                    messages_string += message['content']
+
+                    if message['role'] == "system":
+                        messages_string += system_message_end
+                    elif message['role'] == "user":
+                        messages_string += user_message_end
+                    elif message['role'] == "assistant":
+                        messages_string += assistant_message_end
+
+                messages_string += assistant_message_start
+
                 genparams["prompt"] = messages_string
                 frqp = genparams.get('frequency_penalty', 0.1)
                 scaled_rep_pen = genparams.get('presence_penalty', frqp) + 1
@@ -497,9 +521,9 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
     async def handle_sse_stream(self, api_format):
         global friendlymodelname
         self.send_response(200)
-        self.send_header("Cache-Control", "no-cache")
-        self.send_header("Connection", "keep-alive")
-        self.end_headers(force_json=True, sse_stream_flag=True)
+        self.send_header("cache-control", "no-cache")
+        self.send_header("connection", "keep-alive")
+        self.end_headers(content_type='text/event-stream')
 
         current_token = 0
         incomplete_token_buffer = bytearray()
@@ -566,10 +590,10 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
         global maxctx, maxhordelen, friendlymodelname, KcppVersion, totalgens
         self.path = self.path.rstrip('/')
         response_body = None
-        force_json = False
+        content_type = 'application/json'
 
         if self.path in ["", "/?"] or self.path.startswith(('/?','?')): #it's possible for the root url to have ?params without /
-
+            content_type = 'text/html'
             if self.embedded_kailite is None:
                 response_body = (f"Embedded Kobold Lite is not found.<br>You will have to connect via the main KoboldAI client, or <a href='https://lite.koboldai.net?local=1&port={self.port}'>use this URL</a> to connect.").encode()
             else:
@@ -615,9 +639,9 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
 
         elif self.path.endswith('/v1/models'):
             response_body = (json.dumps({"object":"list","data":[{"id":friendlymodelname,"object":"model","created":1,"owned_by":"koboldcpp","permission":[],"root":"koboldcpp"}]}).encode())
-            force_json = True
 
         elif self.path=="/api":
+            content_type = 'text/html'
             if self.embedded_kcpp_docs is None:
                 response_body = (f"KoboldCpp partial API reference can be found at the wiki: https://github.com/LostRuins/koboldcpp/wiki").encode()
             else:
@@ -625,41 +649,40 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
         elif self.path.endswith(('/api')) or self.path.endswith(('/api/v1')):
             self.path = "/api"
             self.send_response(302)
-            self.send_header("Location", self.path)
-            self.end_headers()
+            self.send_header("location", self.path)
+            self.end_headers(content_type='text/html')
             return None
 
         if response_body is None:
             self.send_response(404)
-            self.end_headers()
+            self.end_headers(content_type='text/html')
             rp = 'Error: HTTP Server is running, but this endpoint does not exist. Please check the URL.'
             self.wfile.write(rp.encode())
         else:
             self.send_response(200)
-            self.send_header('Content-Length', str(len(response_body)))
-            self.end_headers(force_json=force_json)
+            self.send_header('content-length', str(len(response_body)))
+            self.end_headers(content_type=content_type)
             self.wfile.write(response_body)
         return
 
     def do_POST(self):
         global modelbusy, requestsinqueue, currentusergenkey, totalgens
-        content_length = int(self.headers['Content-Length'])
+        content_length = int(self.headers['content-length'])
         body = self.rfile.read(content_length)
         self.path = self.path.rstrip('/')
-        force_json = False
         if self.path.endswith(('/api/extra/tokencount')):
             try:
                 genparams = json.loads(body)
                 countprompt = genparams.get('prompt', "")
                 count = handle.token_count(countprompt.encode("UTF-8"))
                 self.send_response(200)
-                self.end_headers()
+                self.end_headers(content_type='application/json')
                 self.wfile.write(json.dumps({"value": count}).encode())
 
             except ValueError as e:
                 utfprint("Count Tokens - Body Error: " + str(e))
                 self.send_response(400)
-                self.end_headers()
+                self.end_headers(content_type='application/json')
                 self.wfile.write(json.dumps({"value": -1}).encode())
             return
 
@@ -672,11 +695,11 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
                 multiuserkey = ""
                 pass
 
-            if (multiuserkey!="" and multiuserkey==currentusergenkey) or requestsinqueue==0:
+            if (multiuserkey=="" and requestsinqueue==0) or (multiuserkey!="" and multiuserkey==currentusergenkey):
                 ag = handle.abort_generate()
                 time.sleep(0.3) #short delay before replying
                 self.send_response(200)
-                self.end_headers()
+                self.end_headers(content_type='application/json')
                 self.wfile.write(json.dumps({"success": ("true" if ag else "false")}).encode())
                 print("\nGeneration Aborted")
             else:
@@ -694,11 +717,11 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
                 pass
 
             if totalgens>0:
-                if (multiuserkey!="" and multiuserkey==currentusergenkey) or requestsinqueue==0:
+                if (multiuserkey=="" and requestsinqueue==0) or (multiuserkey!="" and multiuserkey==currentusergenkey):
                     pendtxt = handle.get_pending_output()
                     pendtxtStr = ctypes.string_at(pendtxt).decode("UTF-8","ignore")
             self.send_response(200)
-            self.end_headers()
+            self.end_headers(content_type='application/json')
             self.wfile.write(json.dumps({"results": [{"text": pendtxtStr}]}).encode())
             return
 
@@ -708,7 +731,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
             requestsinqueue += 1
         if not modelbusy.acquire(blocking=reqblocking):
             self.send_response(503)
-            self.end_headers()
+            self.end_headers(content_type='application/json')
             self.wfile.write(json.dumps({"detail": {
                     "msg": "Server is busy; please try again later.",
                     "type": "service_unavailable",
@@ -734,11 +757,9 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
 
             if self.path.endswith('/v1/completions'):
                 api_format = 3
-                force_json = True
 
             if self.path.endswith('/v1/chat/completions'):
                 api_format = 4
-                force_json = True
 
             if api_format > 0:
                 genparams = None
@@ -764,8 +785,8 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
                     # Headers are already sent when streaming
                     if not sse_stream_flag:
                         self.send_response(200)
-                        self.end_headers(force_json=force_json)
-                    self.wfile.write(json.dumps(gen).encode())
+                        self.end_headers(content_type='application/json')
+                        self.wfile.write(json.dumps(gen).encode())
                 except:
                     print("Generate: The response could not be sent, maybe connection was terminated?")
                 return
@@ -773,27 +794,23 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
             modelbusy.release()
 
         self.send_response(404)
-        self.end_headers()
+        self.end_headers(content_type='text/html')
 
 
     def do_OPTIONS(self):
         self.send_response(200)
-        self.end_headers()
+        self.end_headers(content_type='text/html')
 
     def do_HEAD(self):
         self.send_response(200)
-        self.end_headers()
-
-    def end_headers(self, force_json=False, sse_stream_flag=False):
-        self.send_header('Access-Control-Allow-Origin', '*')
-        self.send_header('Access-Control-Allow-Methods', '*')
-        self.send_header('Access-Control-Allow-Headers', '*')
-        if ("/api" in self.path and self.path!="/api") or force_json:
-            if sse_stream_flag:
-                self.send_header('Content-type', 'text/event-stream')
-            self.send_header('Content-type', 'application/json')
-        else:
-            self.send_header('Content-type', 'text/html')
+        self.end_headers(content_type='text/html')
+
+    def end_headers(self, content_type=None):
+        self.send_header('access-control-allow-origin', '*')
+        self.send_header('access-control-allow-methods', '*')
+        self.send_header('access-control-allow-headers', '*, Accept, Content-Type, Content-Length, Accept-Encoding, X-CSRF-Token, Client-Agent, X-Fields, Content-Type, Authorization, X-Requested-With, X-HTTP-Method-Override, apikey, genkey')
+        if content_type is not None:
+            self.send_header('content-type', content_type)
         return super(ServerRequestHandler, self).end_headers()
 
 
@@ -1017,7 +1034,8 @@ def show_new_gui():
     mmq_var = ctk.IntVar(value=1)
     blas_threads_var = ctk.StringVar()
     blas_size_var = ctk.IntVar()
-    version_var =ctk.StringVar(value="0")
+    version_var = ctk.StringVar(value="0")
+    tensor_split_str_vars = ctk.StringVar(value="")
 
     smartcontext = ctk.IntVar()
     context_var = ctk.IntVar()
@@ -1069,11 +1087,15 @@ def show_new_gui():
             quick_lowvram_box.grid(row=4, column=0, padx=8, pady=1,  stick="nw")
             mmq_box.grid(row=4, column=1, padx=8, pady=1,  stick="nw")
             quick_mmq_box.grid(row=4, column=1, padx=8, pady=1,  stick="nw")
+            tensor_split_label.grid(row=6, column=0, padx = 8, pady=1, stick="nw")
+            tensor_split_entry.grid(row=6, column=1, padx=8, pady=1, stick="nw")
         else:
             lowvram_box.grid_forget()
             quick_lowvram_box.grid_forget()
             mmq_box.grid_forget()
             quick_mmq_box.grid_forget()
+            tensor_split_label.grid_forget()
+            tensor_split_entry.grid_forget()
 
         if index == "Use CLBlast" or index == "Use CuBLAS" or index == "Use hipBLAS (ROCm)":
             gpu_layers_label.grid(row=5, column=0, padx = 8, pady=1, stick="nw")
@@ -1086,6 +1108,7 @@ def show_new_gui():
             quick_gpu_layers_label.grid_forget()
             quick_gpu_layers_entry.grid_forget()
 
+
     # presets selector
     makelabel(quick_tab, "Presets:", 1)
 
@@ -1118,7 +1141,7 @@ def show_new_gui():
     makeslider(quick_tab, "Context Size:", contextsize_text, context_var, 0, len(contextsize_text)-1, 30, set=2)
 
     # load model
-    makefileentry(quick_tab, "Model:", "Select GGML Model File", model_var, 40, 170,filetypes=[("GGML Model Files", "*.gguf;*.bin;*.ggml")])
+    makefileentry(quick_tab, "Model:", "Select GGML Model File", model_var, 40, 170)
 
     # Hardware Tab
     hardware_tab = tabcontent["Hardware"]
@@ -1137,6 +1160,7 @@ def show_new_gui():
     gpu_selector_box = ctk.CTkComboBox(hardware_tab, values=["1","2","3","4"], width=60, variable=gpu_choice_var, state="readonly")
     CUDA_gpu_selector_box = ctk.CTkComboBox(hardware_tab, values=["1","2","3","4", "All"], width=60, variable=gpu_choice_var, state="readonly")
     gpu_layers_entry,gpu_layers_label = makelabelentry(hardware_tab,"GPU Layers:", gpulayers_var, 5, 50)
+    tensor_split_entry,tensor_split_label = makelabelentry(hardware_tab, "Tensor Split:", tensor_split_str_vars, 6, 80)
     lowvram_box = makecheckbox(hardware_tab,  "Low VRAM", lowvram_var, 4,0)
     mmq_box = makecheckbox(hardware_tab,  "Use QuantMatMul (mmq)", mmq_var, 4,1)
 
@@ -1185,7 +1209,7 @@ def show_new_gui():
     # Model Tab
     model_tab = tabcontent["Model"]
 
-    makefileentry(model_tab, "Model:", "Select GGML Model File", model_var, 1, filetypes=[("GGML Model Files", "*.gguf;*.bin;*.ggml")])
+    makefileentry(model_tab, "Model:", "Select GGML Model File", model_var, 1)
     makefileentry(model_tab, "Lora:", "Select Lora File",lora_var, 3)
     makefileentry(model_tab, "Lora Base:", "Select Lora Base File", lora_base_var, 5)
 
@@ -1265,6 +1289,12 @@ def show_new_gui():
             args.noavx2 = True
             args.noblas = True
             args.nommap = True
+        if tensor_split_str_vars.get()!="":
+            tssv = tensor_split_str_vars.get()
+            if "," in tssv:
+                args.tensor_split = [float(x) for x in tssv.split(",")]
+            else:
+                args.tensor_split = [float(x) for x in tssv.split(" ")]
 
         args.blasthreads = None if blas_threads_var.get()=="" else int(blas_threads_var.get())
 
@@ -1329,6 +1359,9 @@ def show_new_gui():
                 runopts_var.set(openblas_option)
         if "gpulayers" in dict and dict["gpulayers"]:
             gpulayers_var.set(dict["gpulayers"])
+        if "tensor_split" in dict and dict["tensor_split"]:
+            tssep = ','.join(map(str, dict["tensor_split"]))
+            tensor_split_str_vars.set(tssep)
         if "blasthreads" in dict and dict["blasthreads"]:
             blas_threads_var.set(str(dict["blasthreads"]))
         else:
@@ -1447,7 +1480,7 @@ def show_gui_msgbox(title,message):
 def run_horde_worker(args, api_key, worker_name):
     import urllib.request
     from datetime import datetime
-    global friendlymodelname, maxhordectx, maxhordelen, exitcounter, modelbusy, session_starttime
+    global friendlymodelname, maxhordectx, maxhordelen, exitcounter, punishcounter, modelbusy, session_starttime
     epurl = f"http://localhost:{args.port}"
     if args.host!="":
         epurl = f"http://{args.host}:{args.port}"
@@ -1456,10 +1489,11 @@ def run_horde_worker(args, api_key, worker_name):
         print(f"{datetime.now().strftime('[%H:%M:%S]')} " + txt)
 
     def submit_completed_generation(url, jobid, sessionstart, submit_dict):
-        global exitcounter, session_kudos_earned, session_jobs
+        global exitcounter, punishcounter, session_kudos_earned, session_jobs, rewardcounter
         reply = make_url_request(url, submit_dict)
         if not reply:
             exitcounter += 1
+            punishcounter += 1
             print_with_time(f"Error, Job submit failed.")
         else:
             reward = reply["reward"]
@@ -1473,6 +1507,11 @@ def run_horde_worker(args, api_key, worker_name):
             elapsedtimestr = f"{hrs:03d}h:{mins:02d}m:{secs:02d}s"
             earnrate = session_kudos_earned/(elapsedtime.seconds/3600)
             print_with_time(f'Submitted {jobid} and earned {reward:.0f} kudos\n[Total:{session_kudos_earned:.0f} kudos, Time:{elapsedtimestr}, Jobs:{session_jobs}, EarnRate:{earnrate:.0f} kudos/hr]')
+            rewardcounter += 1
+            if rewardcounter > 50:
+                rewardcounter = 0
+                if exitcounter > 5:
+                    exitcounter -= 1
 
     def make_url_request(url, data, method='POST'):
         try:
@@ -1481,7 +1520,7 @@ def run_horde_worker(args, api_key, worker_name):
             if method=='POST':
                 json_payload = json.dumps(data).encode('utf-8')
                 request = urllib.request.Request(url, data=json_payload, headers=headers, method=method)
-                request.add_header('Content-Type', 'application/json')
+                request.add_header('content-type', 'application/json')
             else:
                 request = urllib.request.Request(url, headers=headers, method=method)
             response_data = ""
@@ -1508,17 +1547,23 @@ def run_horde_worker(args, api_key, worker_name):
     print(f"===\nEmbedded Horde Worker '{worker_name}' Starting...\n(To use your own KAI Bridge/Scribe worker instead, don't set your API key)")
     BRIDGE_AGENT = f"KoboldCppEmbedWorker:2:https://github.com/LostRuins/koboldcpp"
     cluster = "https://horde.koboldai.net"
-    while exitcounter < 10:
+    while exitcounter < 35:
         time.sleep(3)
         readygo = make_url_request(f'{epurl}/api/v1/info/version', None,'GET')
         if readygo:
             print_with_time(f"Embedded Horde Worker '{worker_name}' is started.")
             break
 
-    while exitcounter < 10:
+    while exitcounter < 35:
         currentjob_attempts = 0
         current_generation = None
 
+        if punishcounter >= 10:
+            punishcounter = 0
+            print_with_time(f"Horde Worker Paused for 10 min - Too many errors. It will resume automatically.")
+            print_with_time(f"Caution: Too many failed jobs may lead to entering maintenance mode.")
+            time.sleep(600)
+
         #first, make sure we are not generating
         if modelbusy.locked():
             time.sleep(0.2)
@@ -1537,6 +1582,7 @@ def run_horde_worker(args, api_key, worker_name):
         pop = make_url_request(f'{cluster}/api/v2/generate/text/pop',gen_dict)
         if not pop:
             exitcounter += 1
+            punishcounter += 1
             print_with_time(f"Failed to fetch job from {cluster}. Waiting 5 seconds...")
             time.sleep(5)
             continue
@@ -1555,7 +1601,7 @@ def run_horde_worker(args, api_key, worker_name):
         print_with_time(f"Job received from {cluster} for {current_payload.get('max_length',80)} tokens and {current_payload.get('max_context_length',1024)} max context. Starting generation...")
 
         #do gen
-        while exitcounter < 10:
+        while exitcounter < 35:
             if not modelbusy.locked():
                 current_generation = make_url_request(f'{epurl}/api/v1/generate', current_payload)
                 if current_generation:
@@ -1880,4 +1926,10 @@ if __name__ == '__main__':
     parser.add_argument("--multiuser", help="Runs in multiuser mode, which queues incoming requests instead of blocking them.", action='store_true')
     parser.add_argument("--foreground", help="Windows only. Sends the terminal to the foreground every time a new prompt is generated. This helps avoid some idle slowdown issues.", action='store_true')
 
+    # #deprecated hidden args. they do nothing. do not use
+    # parser.add_argument("--psutil_set_threads", action='store_true', help=argparse.SUPPRESS)
+    # parser.add_argument("--stream", action='store_true', help=argparse.SUPPRESS)
+    # parser.add_argument("--unbantokens", action='store_true', help=argparse.SUPPRESS)
+    # parser.add_argument("--usemirostat", action='store_true', help=argparse.SUPPRESS)
+
     main(parser.parse_args(),start_server=True)

llama.cpp

@@ -189,6 +189,7 @@ enum llm_arch {
     LLM_ARCH_STARCODER,
     LLM_ARCH_PERSIMMON,
     LLM_ARCH_REFACT,
+    LLM_ARCH_BLOOM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -202,7 +203,8 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
     { LLM_ARCH_BAICHUAN,        "baichuan"  },
     { LLM_ARCH_STARCODER,       "starcoder" },
     { LLM_ARCH_PERSIMMON,       "persimmon" },
-    { LLM_ARCH_REFACT,          "refact" },
+    { LLM_ARCH_REFACT,          "refact"    },
+    { LLM_ARCH_BLOOM,           "bloom"     },
 };
 
 enum llm_kv {
@@ -305,6 +307,7 @@ struct LLM_KV {
 
 enum llm_tensor {
     LLM_TENSOR_TOKEN_EMBD,
+    LLM_TENSOR_TOKEN_EMBD_NORM,
     LLM_TENSOR_POS_EMBD,
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
@@ -425,6 +428,14 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
         LLM_ARCH_MPT,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
     {
@@ -459,6 +470,21 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_BLOOM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1016,6 +1042,9 @@ struct llama_hparams {
     float rope_freq_base_train;
     float rope_freq_scale_train;
 
+    float f_clamp_kqv;
+    float f_max_alibi_bias;
+
     bool operator!=(const llama_hparams & other) const {
         if (this->vocab_only != other.vocab_only) return true;
         if (this->n_vocab != other.n_vocab) return true;
@@ -1201,6 +1230,8 @@ struct llama_model {
 
     struct ggml_tensor * tok_embeddings;
     struct ggml_tensor * pos_embeddings;
+    struct ggml_tensor * tok_norm;
+    struct ggml_tensor * tok_norm_b;
 
     struct ggml_tensor * output_norm;
     struct ggml_tensor * output_norm_b;
@@ -1330,7 +1361,11 @@ static bool llama_kv_cache_init(
     cache.cells.clear();
     cache.cells.resize(n_ctx);
 
+    // TODO: this should be:
+    //       cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
+    //       change it and test that it works
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+    memset(cache.buf.data, 0, cache.buf.size);
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size;
@@ -1736,7 +1771,7 @@ struct llama_model_loader {
         }
     }
 
-    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend backend) {
+    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend_type backend) {
         if (backend != GGML_BACKEND_CPU) {
             ggml_set_no_alloc(ctx, true);
         }
@@ -1754,7 +1789,7 @@ struct llama_model_loader {
         return tensor;
     }
 
-    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, ggml_backend backend) {
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, ggml_backend_type backend) {
         struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str());
 
         if (cur == NULL) {
@@ -2047,13 +2082,13 @@ static void llm_load_hparams(
                 }
             } break;
         case LLM_ARCH_PERSIMMON:
-        {
-            GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
-            switch (hparams.n_layer) {
-                case 36: model.type = e_model::MODEL_8B; break;
-                default: model.type = e_model::MODEL_UNKNOWN;
-            }
-        } break;
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                switch (hparams.n_layer) {
+                    case 36: model.type = e_model::MODEL_8B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_REFACT:
             {
                 GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
@@ -2062,6 +2097,33 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_BLOOM:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+
+                switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
+                    case 30:
+                        switch (hparams.n_embd) {
+                            case 2560: model.type = e_model::MODEL_3B; break;
+                            case 4096: model.type = e_model::MODEL_7B; break;
+                        } break;
+                }
+            } break;
+        case LLM_ARCH_MPT:
+            {
+                hparams.f_clamp_kqv = 0.0f;
+
+                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_CLAMP_KQV));
+                GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS));
+
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 48: model.type = e_model::MODEL_30B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -2206,6 +2268,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: n_gqa            = %u\n",     __func__, hparams.n_gqa());
     LLAMA_LOG_INFO("%s: f_norm_eps       = %.1e\n",   __func__, hparams.f_norm_eps);
     LLAMA_LOG_INFO("%s: f_norm_rms_eps   = %.1e\n",   __func__, hparams.f_norm_rms_eps);
+    LLAMA_LOG_INFO("%s: f_clamp_kqv      = %.1e\n",   __func__, hparams.f_clamp_kqv);
+    LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n",   __func__, hparams.f_max_alibi_bias);
     LLAMA_LOG_INFO("%s: n_ff             = %u\n",     __func__, hparams.n_ff);
     LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
     LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
@@ -2305,8 +2369,8 @@ static void llm_load_tensors(
 
                     // output
                     {
-                        ggml_backend backend_norm;
-                        ggml_backend backend_output;
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
                             // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@@ -2341,8 +2405,8 @@ static void llm_load_tensors(
                     model.layers.resize(n_layer);
 
                     for (uint32_t i = 0; i < n_layer; ++i) {
-                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
-                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+                        const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
                         auto & layer = model.layers[i];
 
@@ -2371,8 +2435,8 @@ static void llm_load_tensors(
                 {
                     model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
                     {
-                        ggml_backend backend_norm;
-                        ggml_backend backend_output;
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
                             // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@@ -2407,8 +2471,8 @@ static void llm_load_tensors(
                     model.layers.resize(n_layer);
 
                     for (uint32_t i = 0; i < n_layer; ++i) {
-                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
-                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+                        const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
                         auto & layer = model.layers[i];
 
@@ -2441,8 +2505,8 @@ static void llm_load_tensors(
 
                     // output
                     {
-                        ggml_backend backend_norm;
-                        ggml_backend backend_output;
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
                             // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@@ -2479,8 +2543,8 @@ static void llm_load_tensors(
                     model.layers.resize(n_layer);
 
                     for (uint32_t i = 0; i < n_layer; ++i) {
-                        const ggml_backend backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
-                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+                        const ggml_backend_type backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
                         auto & layer = model.layers[i];
 
@@ -2518,8 +2582,8 @@ static void llm_load_tensors(
 
                     // output
                     {
-                        ggml_backend backend_norm;
-                        ggml_backend backend_output;
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
                             // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@@ -2556,8 +2620,8 @@ static void llm_load_tensors(
                     model.layers.resize(n_layer);
 
                     for (uint32_t i = 0; i < n_layer; ++i) {
-                        const ggml_backend backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
-                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+                        const ggml_backend_type backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
                         auto & layer = model.layers[i];
 
@@ -2595,8 +2659,8 @@ static void llm_load_tensors(
                     model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"),  {n_embd, n_vocab}, GGML_BACKEND_CPU);
 
                     {
-                        ggml_backend backend_norm;
-                        ggml_backend backend_output;
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
                             // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@@ -2630,8 +2694,8 @@ static void llm_load_tensors(
                     const int i_gpu_start = n_layer - n_gpu_layers;
                     model.layers.resize(n_layer);
                     for (uint32_t i = 0; i < n_layer; ++i) {
-                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
-                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT;
                         auto & layer = model.layers[i];
                         layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
                         layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, backend);
@@ -2651,6 +2715,155 @@ static void llm_load_tensors(
                         layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i),   {64}, backend);
                     }
                 } break;
+            case LLM_ARCH_BLOOM:
+                {
+                    // TODO: CPU-only for now
+
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+                    model.tok_norm       = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd},          GGML_BACKEND_CPU);
+                    model.tok_norm_b     = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd},          GGML_BACKEND_CPU);
+
+                    // output
+                    {
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd},          backend_norm);
+                        model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                            vram_weights += ggml_nbytes(model.output_norm_b);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend_type backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, backend);
+                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, backend);
+
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend_split);
+
+                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},                backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},                        backend_split);
+
+                        layer.ffn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+                        layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, backend);
+
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
+                        layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend_split);
+
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) +
+                                ggml_nbytes(layer.wqkv)      + ggml_nbytes(layer.bqkv)        +
+                                ggml_nbytes(layer.wo)        + ggml_nbytes(layer.bo)          +
+                                ggml_nbytes(layer.ffn_norm)  + ggml_nbytes(layer.ffn_norm_b)  +
+                                ggml_nbytes(layer.w3)        + ggml_nbytes(layer.b3)          +
+                                ggml_nbytes(layer.w2)        + ggml_nbytes(layer.b2);
+                        }
+                    }
+                } break;
+            case LLM_ARCH_MPT:
+                {
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+
+                    // output
+                    {
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, 3*n_embd}, backend_split);
+                        layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},     backend_split);
+
+                        layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) +
+                                ggml_nbytes(layer.wqkv)      +
+                                ggml_nbytes(layer.wo)        +
+                                ggml_nbytes(layer.ffn_norm)  +
+                                ggml_nbytes(layer.w2)        +
+                                ggml_nbytes(layer.w3);
+                        }
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -4507,7 +4720,6 @@ static struct ggml_cgraph * llm_build_starcoder(
     return gf;
 }
 
-
 static struct ggml_cgraph * llm_build_persimmon(
          llama_context & lctx,
      const llama_batch & batch) {
@@ -4905,37 +5117,604 @@ static struct ggml_cgraph * llm_build_persimmon(
     return gf;
 }
 
-static struct ggml_cgraph * llama_build_graph(
+static struct ggml_cgraph * llm_build_bloom(
          llama_context & lctx,
      const llama_batch & batch) {
-    const auto & model = lctx.model;
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
 
-    struct ggml_cgraph * result = NULL;
+    const auto & kv_self = lctx.kv_self;
 
-    switch (model.arch) {
-        case LLM_ARCH_LLAMA:
-            {
-                result = llm_build_llama(lctx, batch);
-            } break;
-        case LLM_ARCH_BAICHUAN:
-            {
-                result = llm_build_baichaun(lctx, batch);
-            } break;
-        case LLM_ARCH_FALCON:
-            {
-                result = llm_build_falcon(lctx, batch);
-            } break;
-        case LLM_ARCH_STARCODER:
-            {
-                result = llm_build_starcoder(lctx, batch);
-            } break;
-        case LLM_ARCH_PERSIMMON:
-            {
-                result = llm_build_persimmon(lctx, batch);
-            } break;
-        case LLM_ARCH_REFACT:
-            {
-                result = llm_build_refact(lctx, batch);
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    const float norm_eps = hparams.f_norm_eps;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ false,
+    };
+
+    params.no_alloc = true;
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * token;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, token);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token));
+        }
+    }
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, token, norm_eps);
+        inpL = ggml_add(ctx0, ggml_mul(ctx0, inpL, model.tok_norm), model.tok_norm_b);
+    }
+
+    ggml_set_name(inpL, "inpL");
+
+    for (int il = 0; il < n_layer; ++il) {
+        {
+            // Norm
+            cur = ggml_norm(ctx0, inpL, norm_eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b);
+        }
+
+        {
+            // Self Attention
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv);
+
+            struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd);
+            struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa));
+
+            struct ggml_tensor * Qcur = tmpq;
+            struct ggml_tensor * Kcur = tmpk;
+
+            // store key and value to memory
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens));
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)),
+                        0, 2, 1, 3);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            ggml_set_name(K, "K");
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            ggml_set_name(KQ, "KQ");
+
+            // KQ_scaled = KQ / sqrt(n_embd_head)
+            // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ kv_head, n_head, 8);
+            ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            // split cached V into n_head heads
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            ggml_set_name(KQV, "KQV");
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+        }
+
+        // Projection
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo);
+
+        // Add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        // FF
+        {
+            // Norm
+            {
+                cur = ggml_norm(ctx0, inpFF, norm_eps);
+                cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b);
+            }
+
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // Projection
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2);
+        }
+
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    // Output Norm
+    {
+        cur = ggml_norm(ctx0, inpL, norm_eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b);
+    }
+    ggml_set_name(cur, "result_norm");
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_cgraph * llm_build_mpt(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv; // == n_head for MPT, as there's no MQA/GQA
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    const float norm_eps       = hparams.f_norm_eps;
+    const float clamp_kqv      = hparams.f_clamp_kqv;
+    const float max_alibi_bias = hparams.f_max_alibi_bias;
+
+    const int n_gpu_layers = model.n_gpu_layers;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ false,
+    };
+
+    params.no_alloc = true;
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * inpL;
+
+    //int warmup = 0;
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+            //warmup = ((uint32_t*) inp_tokens->data)[0] == 0;
+        }
+
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
+        }
+    }
+
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
+    }
+#endif // GGML_USE_CUBLAS
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * attn_norm;
+
+        offload_func_t offload_func = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (il >= i_gpu_start) {
+            offload_func = ggml_cuda_assign_buffers_no_alloc;
+        }
+#endif // GGML_USE_CUBLAS
+
+        // self-attention
+        // TODO: refactor into common function (shared with LLaMA)
+        {
+            attn_norm = ggml_norm(ctx0, inpL, norm_eps);
+            offload_func(attn_norm);
+
+            attn_norm = ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm);
+            offload_func(attn_norm);
+
+            if (1) {
+                cur = attn_norm;
+            }
+
+            // compute QKV
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+            offload_func_kq(cur);
+
+            if (clamp_kqv > 0.0f) {
+                cur = ggml_clamp(ctx0, cur, -clamp_kqv, clamp_kqv);
+                offload_func_kq(cur);
+            }
+
+            const size_t wsize = ggml_type_size(cur->type);
+
+            struct ggml_tensor * Qcur = ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                0);
+            offload_func_kq(Qcur);
+
+            struct ggml_tensor * Kcur = ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head_kv, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                wsize * n_embd_head *  n_head);
+            offload_func_kq(Kcur);
+
+            struct ggml_tensor * tmpv = ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head_kv, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                wsize * n_embd_head * (n_head +     n_head_kv));
+            offload_func_kq(Kcur);
+
+            ggml_set_name(Qcur, "Qcur");
+            ggml_set_name(Kcur, "Kcur");
+
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens));
+                offload_func_v(Vcur);
+                offload_func_v(Vcur->src[0]->src[0]);
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                offload_func_kq(k);
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            offload_func_kq(Q);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            offload_func_kq(K);
+            ggml_set_name(K, "K");
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
+            ggml_set_name(KQ, "KQ");
+
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            // TODO: replace with ggml_add()
+            struct ggml_tensor * KQ_scaled_alibi =
+                ggml_alibi(ctx0, KQ_scaled, 0, n_head, max_alibi_bias);
+            offload_func_kq(KQ_scaled_alibi);
+            ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");
+
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask);
+            offload_func_kq(KQ_masked);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            offload_func_v(V);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
+            ggml_set_name(KQV, "KQV");
+
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_wo");
+        }
+
+        // Add the input
+        cur = ggml_add(ctx0, cur, inpL);
+        offload_func(cur);
+
+        struct ggml_tensor * attn_out = cur;
+
+        // feed forward
+        {
+            // Norm
+            {
+                cur = ggml_norm(ctx0, attn_out, norm_eps);
+                offload_func(cur);
+
+                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm);
+                offload_func(cur);
+            }
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur);
+            offload_func(cur);
+
+            cur = ggml_gelu(ctx0, cur);
+            offload_func(cur);
+            cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur);
+            offload_func(cur);
+        }
+
+        cur = ggml_add(ctx0, cur, attn_out);
+        offload_func(cur);
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    {
+        cur = ggml_norm(ctx0, cur, norm_eps);
+        offload_func_nr(cur);
+
+        cur = ggml_mul(ctx0, cur, model.output_norm);
+        ggml_set_name(cur, "result_norm");
+    }
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_cgraph * llama_build_graph(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model = lctx.model;
+
+    struct ggml_cgraph * result = NULL;
+
+    switch (model.arch) {
+        case LLM_ARCH_LLAMA:
+            {
+                result = llm_build_llama(lctx, batch);
+            } break;
+        case LLM_ARCH_BAICHUAN:
+            {
+                result = llm_build_baichaun(lctx, batch);
+            } break;
+        case LLM_ARCH_FALCON:
+            {
+                result = llm_build_falcon(lctx, batch);
+            } break;
+        case LLM_ARCH_STARCODER:
+            {
+                result = llm_build_starcoder(lctx, batch);
+            } break;
+        case LLM_ARCH_PERSIMMON:
+            {
+                result = llm_build_persimmon(lctx, batch);
+            } break;
+        case LLM_ARCH_REFACT:
+            {
+                result = llm_build_refact(lctx, batch);
+            } break;
+        case LLM_ARCH_BLOOM:
+            {
+                result = llm_build_bloom(lctx, batch);
+            } break;
+        case LLM_ARCH_MPT:
+            {
+                result = llm_build_mpt(lctx, batch);
             } break;
         default:
             GGML_ASSERT(false);
@@ -5067,7 +5846,8 @@ static int llama_decode_internal(
     const bool full_offload_supported = model.arch == LLM_ARCH_LLAMA ||
         model.arch == LLM_ARCH_BAICHUAN ||
         model.arch == LLM_ARCH_FALCON ||
-        model.arch == LLM_ARCH_REFACT;
+        model.arch == LLM_ARCH_REFACT ||
+        model.arch == LLM_ARCH_MPT;
     const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
     if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {
         n_threads = 1;
@@ -5568,7 +6348,6 @@ private:
         for (int i = 0; i < (int)text_utf.size(); i++) {
             const std::string & utf_char = text_utf[i];
             bool split_condition = false;
-            // const char* text_pos = raw_text_p + utf_char.seq_offset_bytes;
             int bytes_remain = text_utf.size() - i;
             // forward backward lookups
             const std::string & utf_char_next = (i + 1 < (int)text_utf.size()) ? text_utf[i + 1] : "";
@@ -5594,9 +6373,9 @@ private:
             if (!split_condition && bytes_remain >= 3) {
                 // 're|'ve|'ll
                 if (utf_char == "\'" && (
-                    (utf_char_next == "r" || utf_char_next_next == "e") ||
-                    (utf_char_next == "v" || utf_char_next_next == "e") ||
-                    (utf_char_next == "l" || utf_char_next_next == "l"))
+                    (utf_char_next == "r" && utf_char_next_next == "e") ||
+                    (utf_char_next == "v" && utf_char_next_next == "e") ||
+                    (utf_char_next == "l" && utf_char_next_next == "l"))
                     ) {
                     split_condition = true;
                 }
@@ -5647,7 +6426,7 @@ private:
                 else if (collecting_special && (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
                     split_condition = true;
                 }
-                else if (collecting_whitespace_lookahead && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) {
+                else if (collecting_whitespace_lookahead && (codepoint_type(utf_char_next) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
                     split_condition = true;
                 }
             }
@@ -7166,7 +7945,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         const std::string name = ggml_get_name(meta);
 
         // TODO: avoid hardcoded tensor names - use the TN_* constants
-        if (name.find("attn_v.weight") != std::string::npos) {
+        if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
             ++n_attention_wv;
         }
         else if (name.find("ffn_down.weight") != std::string::npos) {

otherarch/llama_v3.cpp

@@ -63,9 +63,8 @@ static void llama_v3_log_callback_default(llama_v3_log_level level, const char *
 #define LLAMA_V3_LOG_WARN(...)  llama_v3_log_internal(LLAMA_V3_LOG_LEVEL_WARN , __VA_ARGS__)
 #define LLAMA_V3_LOG_ERROR(...) llama_v3_log_internal(LLAMA_V3_LOG_LEVEL_ERROR, __VA_ARGS__)
 
-
-#if !defined(GGML_USE_CUBLAS)
 #include "ggml-alloc.h"
+#if !defined(GGML_USE_CUBLAS)
 #define LLAMA_V3_USE_ALLOCATOR
 #else
 #define LLAMA_V3_USE_SCRATCH
@@ -725,7 +724,7 @@ struct llama_v3_model_loader {
         }
     }
 
-    struct ggml_tensor * get_tensor(const std::string & name, const std::vector<uint32_t> & ne, ggml_backend backend) {
+    struct ggml_tensor * get_tensor(const std::string & name, const std::vector<uint32_t> & ne, ggml_backend_type backend) {
         auto it = tensors_map.name_to_idx.find(name);
         if (it == tensors_map.name_to_idx.end()) {
             throw std::runtime_error(std::runtime_error(format_old("llama.cpp: tensor '%s' is missing from model", name.c_str())));
@@ -739,7 +738,7 @@ struct llama_v3_model_loader {
         return get_tensor_for(lt, backend);
     }
 
-    struct ggml_tensor * get_tensor_for(llama_v3_load_tensor & lt, ggml_backend backend) {
+    struct ggml_tensor * get_tensor_for(llama_v3_load_tensor & lt, ggml_backend_type backend) {
         struct ggml_tensor * tensor;
         if (backend != GGML_BACKEND_CPU) {
             ggml_set_no_alloc(ggml_ctx, true);
@@ -1230,8 +1229,8 @@ static void llama_v3_model_load_internal(
 
         // "output" tensor
         {
-            ggml_backend backend_norm;
-            ggml_backend backend_output;
+            ggml_backend_type backend_norm;
+            ggml_backend_type backend_output;
             if (n_gpu_layers > int(n_layer)) { // NOLINT
                 // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
                 // on Windows however this is detrimental unless everything is on the GPU
@@ -1261,8 +1260,8 @@ static void llama_v3_model_load_internal(
 
         model.layers.resize(n_layer);
         for (uint32_t i = 0; i < n_layer; ++i) {
-            const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_V3_BACKEND_OFFLOAD; // NOLINT
-            const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_V3_BACKEND_OFFLOAD_SPLIT; // NOLINT
+            const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_V3_BACKEND_OFFLOAD; // NOLINT
+            const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_V3_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
             auto & layer = model.layers[i];
 

prompts/mnemonics.txt

@@ -0,0 +1,93 @@
+For each kanji character, write a Markdown‐formatted mnemonic that uses its keyword and the keyword of all its components.
+
+Kanji: 欠 (lack of)
+Components: 𠂊 (hook claw), 人 (person)
+Mnemonic: This **person** is a pirate. He lost his hand to a crocodile many years ago. Nowadays, the ***lack of*** a hand does not bother him too much. In fact, the **hook claw** that replaces it is the mark of a true pirate, so he is quite proud of it!
+
+Kanji: 類 (kind (of something))
+Components: 米 (rice), 大 (large), 頁 (page)
+Mnemonic: The waiter at a Chinese restaurant hands you a **large** menu. Each **page** has all ***kinds*** of **rice** on offer!
+
+Kanji: 燃 (burn)
+Components: 火 (fire), 然 (sort of thing)
+Mnemonic: ***Burning*** things up with **fire** is just my **sort of thing**. (Spoken like a true pyromaniac.)
+
+Kanji: 頂 (top of)
+Components: 丁 (street), 頁 (page)
+Mnemonic: To be at the ***top of*** your game, you need both practical knowledge (**street** smarts) and theoretical knowledge (having read many **pages**).
+
+Kanji: 険 (risky and steep)
+Components: 阝 (small village), 㑒 (consensus)
+Mnemonic: Everyone agrees (there is **consensus**) that the path to the **small village** is ***risky and steep***.
+
+Kanji: 困 (distressed)
+Components: 囗 (closed box), 木 (tree)
+Mnemonic: You would feel ***distressed*** too if you were a **tree** trapped in a **closed box**! I have no place to grow!
+
+Kanji: 頭 (head)
+Components: 豆 (bean), 頁 (page)
+Mnemonic: What do you have in that ***head*** of yours? A **bean** for a brain? Go read more **pages** and become more knowledgeable about the world!
+
+Kanji: 確 (certain)
+Components: 石 (stone), 冖 (roof without a chimney), 隹 (old bird)
+Mnemonic: An **old bird** has made a nest on your **roof**. What do you do? You call Misaka from a <cite>A ***Certain*** Scientific Railgun</cite> to get rid of it, of course! But she doesn’t really want to vaporize the poor thing, so she just throws a **stone** to scare it away. (What was the point of calling her, then‽)
+
+Kanji: 魚 (fish)
+Components: 𠂊 (hook claw), 田 (rice field), 灬 (fire sparks)
+Mnemonic: Catch ***fish*** with a **hook**, collect rice from the **rice field**, cook them with **fire**… And my meal is ready!
+
+Kanji: 警 (to police (something))
+Components: 敬 (respect), 言 (say)
+Mnemonic: ***To police something*** is to make people **respect** what the law **says**.
+
+Kanji: 筆 (writing brush)
+Components: 竹 (bamboo), 聿 (brush)
+Mnemonic: A traditional ***writing brush*** is a **brush** made of **bamboo**.
+
+Kanji: 獄 (prison)
+Components: 犭 (animal), 言 (say), 犬 (dog)
+Mnemonic: In ***prison***, like in the **animal** kingdom, only the toughest survive. You have to watch what you **say**. It’s a **dog**‐eat‐dog world.
+
+Kanji: 新 (new)
+Components: 立 (standing up), 木 (tree), 斤 (axe)
+Mnemonic: In order for a ***new*** construction to be made, an empty lot is needed. If there are any **trees** **standing up**, they must be cut down with an **axe**.
+
+Kanji: 怪 (suspicious)
+Components: 忄 (weak heart), 圣 (sacred)
+Mnemonic: That painting of the **Sacred** **Heart** of Jesus looks ***suspicious***. I think it might be a forgery.
+
+Kanji: 温 (warm (to the touch))
+Components: 氵 (water drops), 日 (sun), 皿 (dish)
+Mnemonic: If you leave **water** on a **dish** in the **sun**, it will get ***warm***.
+
+Kanji: 階 (floor (of a building))
+Components: 阝 (small village), 皆 (all)
+Mnemonic: It might be a **small village**, but, despite that, **all** of its buildings have many ***floors***. It’s a village of skyscrapers!
+
+Kanji: 多 (many)
+Components: 夕 (evening (before sunset)), 夕 (evening (before sunset))
+Mnemonic: Two **evenings** in a day would be one too ***many***.
+
+Kanji: 別 (separate)
+Components: 口 (mouth), 万 (ten thousand), 刂 (knife)
+Mnemonic: Tom Six is at it again. For his next flick, he wants to stitch together **ten thousand** people, **mouth**‐to‐anus. One of the most graphic and disturbing scenes will feature one of the victims using a **knife** to ***separate*** perself.
+
+Kanji: 並 (line up)
+Components: 䒑 (antlers on a wall), 业 (runway)
+Mnemonic: In order to land a plane you have to ***line up*** properly with the **runway**. The things that look like **antlers** at the end of the runway are the control towers; you should follow their instructions.
+
+Kanji: 姿 (figure)
+Components: 次 (next), 女 (woman)
+Mnemonic: The **next** **woman** that I date will have a perfect **figure**. Because I’m done with 3D women—it will *literally* be an anime figure!
+
+Kanji: 実 (real)
+Components: 宀 (roof with a chimney), 𡗗 (three people)
+Mnemonic: Living under a **roof with a chimney** with **three people** (a wife and two children)—a happy family life—is not something I could have ever imagined. It does not feel ***real***.
+
+Kanji: 謝 (apologize)
+Components: 言 (say), 射 (shoot)
+Mnemonic: **Shot** first, ***apologize*** (**say** you are sorry) later.
+
+Kanji: 提 (propose)
+Components: 扌 (left hand), 是 (go with)
+Mnemonic: