add recipes and perplexity graph

README.md

@@ -34,6 +34,7 @@ Perplexity computed against *wiki.test.raw*.
 These first two are just test quants for baseline perplexity comparison:
 * `BF16` 205.811 GiB (16.004 BPW)
   - Final estimate: PPL = 4.5704 +/- 0.02796
+
 * `Q8_0` 109.381 GiB (8.505 BPW)
   - Final estimate: PPL = 4.5798 +/- 0.02804
 
@@ -45,7 +46,56 @@ Final estimate: PPL = 4.5867 +/- 0.02806
 <summary>👈 Secret Recipe</summary>
 
 ```bash
-echo TODO
+#!/usr/bin/env bash
+
+custom="
+# 47 Repeating Layers [0-46]
+# Note: All ffn_down.* layers are not divisible by 256 so have limited quantization options.
+
+# Attention
+blk\..*\.attn_q.*=q8_0
+blk\..*\.attn_k.*=q8_0
+blk\..*\.attn_v.*=q8_0
+blk\..*\.attn_output.*=q8_0
+
+# First 1 Dense Layers [0]
+blk\..*\.ffn_down\.weight=q8_0
+blk\..*\.ffn_(gate|up)\.weight=q8_0
+
+# Shared Expert Layers [1-46]
+blk\..*\.ffn_down_shexp\.weight=q8_0
+blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0
+
+# Routed Experts Layers [1-46]
+blk\.(1)\.ffn_down_exps\.weight=q8_0
+blk\.(1)\.ffn_(gate|up)_exps\.weight=q8_0
+
+blk\..*\.ffn_down_exps\.weight=q6_0
+blk\..*\.ffn_(gate|up)_exps\.weight=iq5_k
+
+# NextN MTP Layer [46]
+blk\..*\.nextn\.embed_tokens\.weight=iq5_ks
+blk\..*\.nextn\.shared_head_head\.weight=iq5_ks
+blk\..*\.nextn\.eh_proj\.weight=q8_0
+
+# Non-Repeating Layers
+token_embd\.weight=iq6_k
+output\.weight=iq6_k
+"
+
+custom=$(
+  echo "$custom" | grep -v '^#' | \
+  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
+)
+
+numactl -N 0 -m 0 \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/imatrix-GLM-4.5-Air-BF16.dat \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x9.4B-BF16-00001-of-00005.gguf \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-IQ5_K.gguf \
+    IQ5_K \
+    192
 ```
 
 </details>
@@ -59,7 +109,53 @@ Final estimate: PPL = 4.7056 +/- 0.02909
 <summary>👈 Secret Recipe</summary>
 
 ```bash
-echo TODO
+#!/usr/bin/env bash
+
+custom="
+# 47 Repeating Layers [0-46]
+# Note: All ffn_down.* layers are not divisible by 256 so have limited quantization options.
+
+# Attention
+blk\.(0|1)\.attn_q.*=q8_0
+blk\.(0|1)\.attn_k.*=q8_0
+blk\.(0|1)\.attn_v.*=q8_0
+blk\.(0|1)\.attn_output.*=q8_0
+
+blk\..*\.attn_q.*=iq5_ks
+blk\..*\.attn_k.*=iq5_ks
+blk\..*\.attn_v.*=iq5_ks
+blk\..*\.attn_output.*=iq5_ks
+
+# First 1 Dense Layers [0]
+blk\..*\.ffn_down\.weight=q6_0
+blk\..*\.ffn_(gate|up)\.weight=iq5_ks
+
+# Shared Expert Layers [1-46]
+blk\..*\.ffn_down_shexp\.weight=q6_0
+blk\..*\.ffn_(gate|up)_shexp\.weight=iq5_ks
+
+# Routed Experts Layers [1-46]
+blk\..*\.ffn_down_exps\.weight=iq4_xs
+blk\..*\.ffn_(gate|up)_exps\.weight=iq4_kss
+
+# Non-Repeating Layers
+token_embd\.weight=iq4_k
+output\.weight=iq6_k
+"
+
+custom=$(
+  echo "$custom" | grep -v '^#' | \
+  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
+)
+
+numactl -N 1 -m 1 \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/imatrix-GLM-4.5-Air-BF16.dat \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x8.1B-BF16-00001-of-00005.gguf \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-IQ4_KSS.gguf \
+    IQ4_KSS \
+    192
 ```
 
 </details>
@@ -72,7 +168,56 @@ Final estimate: PPL = 5.0697 +/- 0.03166
 <summary>👈 Secret Recipe</summary>
 
 ```bash
-echo TODO
+#!/usr/bin/env bash
+
+custom="
+# 47 Repeating Layers [0-46]
+# Note: All ffn_down.* layers are not divisible by 256 so have limited quantization options.
+
+# Attention
+blk\..*\.attn_q.*=iq4_ks
+blk\..*\.attn_k.*=iq5_ks
+blk\..*\.attn_v.*=iq5_ks
+blk\..*\.attn_output.*=iq4_ks
+
+# First 1 Dense Layers [0]
+blk\..*\.ffn_down\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)\.weight=iq4_kss
+
+# Shared Expert Layers [1-46]
+blk\..*\.ffn_down_shexp\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)_shexp\.weight=iq4_kss
+
+# Routed Experts Layers [1-46]
+blk\.(1)\.ffn_down_exps\.weight=iq4_nl
+blk\.(1)\.ffn_(gate|up)_exps\.weight=iq4_kss
+
+blk\..*\.ffn_down_exps\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)_exps\.weight=iq2_kl
+
+# NextN MTP Layer [46]
+blk\..*\.nextn\.embed_tokens\.weight=iq4_ks
+blk\..*\.nextn\.shared_head_head\.weight=iq4_ks
+blk\..*\.nextn\.eh_proj\.weight=q6_0
+
+# Non-Repeating Layers
+token_embd\.weight=iq4_k
+output\.weight=iq6_k
+"
+
+custom=$(
+  echo "$custom" | grep -v '^#' | \
+  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
+)
+
+numactl -N 0 -m 0 \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/imatrix-GLM-4.5-Air-BF16.dat \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x9.4B-BF16-00001-of-00005.gguf \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-IQ2_KL.gguf \
+    IQ2_KL \
+    192
 ```
 
 </details>