Update with h2oGPT hash 6e04392f877303e49bab527df50aac557f4b9d14

client_test.py

@@ -100,7 +100,7 @@ def get_args(prompt, prompt_type, chat=False, stream_output=False,
                          chunk_size=512,
                          document_choice=[DocumentChoices.All_Relevant.name],
                          )
-    from gen import eval_func_param_names
+    from evaluate_params import eval_func_param_names
     assert len(set(eval_func_param_names).difference(set(list(kwargs.keys())))) == 0
     if chat:
         # add chatbot output on end.  Assumes serialize=False

gen.py

@@ -20,6 +20,8 @@ from urllib3.exceptions import ConnectTimeoutError, MaxRetryError, ConnectionErr
 from requests.exceptions import ConnectionError as ConnectionError2
 from requests.exceptions import ReadTimeout as ReadTimeout2
 
+from evaluate_params import eval_func_param_names, no_default_param_names
+
 if os.path.dirname(os.path.abspath(__file__)) not in sys.path:
     sys.path.append(os.path.dirname(os.path.abspath(__file__)))
 
@@ -48,8 +50,6 @@ from transformers import GenerationConfig, AutoModel, TextIteratorStreamer
 from prompter import Prompter, inv_prompt_type_to_model_lower, non_hf_types, PromptType, get_prompt, generate_prompt
 from stopping import get_stopping
 
-eval_extra_columns = ['prompt', 'response', 'score']
-
 langchain_modes = [x.value for x in list(LangChainMode)]
 
 langchain_actions = [x.value for x in list(LangChainAction)]
@@ -144,7 +144,7 @@ def main(
         force_langchain_evaluate: bool = False,
         visible_langchain_modes: list = ['UserData', 'MyData'],
         # WIP:
-        #visible_langchain_actions: list = langchain_actions.copy(),
+        # visible_langchain_actions: list = langchain_actions.copy(),
         visible_langchain_actions: list = [LangChainAction.QUERY.value, LangChainAction.SUMMARIZE_MAP.value],
         document_choice: list = [DocumentChoices.All_Relevant.name],
         user_path: str = None,
@@ -354,6 +354,7 @@ def main(
         # assert not prompt_type, "Don't specify model_lock and prompt_type"
         # assert not prompt_dict, "Don't specify model_lock and prompt_dict"
 
+    n_jobs = int(os.getenv('n_jobs', str(n_jobs)))
     is_hf = bool(int(os.getenv("HUGGINGFACE_SPACES", '0')))
     is_gpth2oai = bool(int(os.getenv("GPT_H2O_AI", '0')))
     is_public = is_hf or is_gpth2oai  # multi-user case with fixed model and disclaimer
@@ -361,6 +362,8 @@ def main(
         memory_restriction_level = 2 if is_hf else 0  # 2 assumes run on 24GB consumer GPU
     else:
         assert 0 <= memory_restriction_level <= 3, "Bad memory_restriction_level=%s" % memory_restriction_level
+    if is_public and os.getenv('n_jobs') is None:
+        n_jobs = max(1, min(os.cpu_count() // 2, 8))
     admin_pass = os.getenv("ADMIN_PASS")
     # will sometimes appear in UI or sometimes actual generation, but maybe better than empty result
     # but becomes unrecoverable sometimes if raise, so just be silent for now
@@ -1188,52 +1191,6 @@ def get_score_model(score_model: str = None,
     return smodel, stokenizer, sdevice
 
 
-no_default_param_names = [
-    'instruction',
-    'iinput',
-    'context',
-    'instruction_nochat',
-    'iinput_nochat',
-]
-
-gen_hyper = ['temperature',
-             'top_p',
-             'top_k',
-             'num_beams',
-             'max_new_tokens',
-             'min_new_tokens',
-             'early_stopping',
-             'max_time',
-             'repetition_penalty',
-             'num_return_sequences',
-             'do_sample',
-             ]
-
-eval_func_param_names = ['instruction',
-                         'iinput',
-                         'context',
-                         'stream_output',
-                         'prompt_type',
-                         'prompt_dict'] + \
-                        gen_hyper + \
-                        ['chat',
-                         'instruction_nochat',
-                         'iinput_nochat',
-                         'langchain_mode',
-                         'langchain_action',
-                         'top_k_docs',
-                         'chunk',
-                         'chunk_size',
-                         'document_choice',
-                         ]
-
-# form evaluate defaults for submit_nochat_api
-eval_func_param_names_defaults = eval_func_param_names.copy()
-for k in no_default_param_names:
-    if k in eval_func_param_names_defaults:
-        eval_func_param_names_defaults.remove(k)
-
-
 def evaluate(
         model_state,
         my_db_state,

gpt_langchain.py

@@ -1182,7 +1182,7 @@ def file_to_doc(file, base_path=None, verbose=False, fail_any_exception=False,
             # so just extract in path where
             zip_ref.extractall(base_path)
             # recurse
-            doc1 = path_to_docs(base_path, verbose=verbose, fail_any_exception=fail_any_exception)
+            doc1 = path_to_docs(base_path, verbose=verbose, fail_any_exception=fail_any_exception, n_jobs=n_jobs)
     else:
         raise RuntimeError("No file handler for %s" % os.path.basename(file))
 

gradio_runner.py

@@ -57,9 +57,9 @@ from prompter import prompt_type_to_model_name, prompt_types_strings, inv_prompt
     get_prompt
 from utils import get_githash, flatten_list, zip_data, s3up, clear_torch_cache, get_torch_allocated, system_info_print, \
     ping, get_short_name, get_url, makedirs, get_kwargs, remove, system_info, ping_gpu
-from gen import get_model, languages_covered, evaluate, eval_func_param_names, score_qa, langchain_modes, \
-    inputs_kwargs_list, scratch_base_dir, no_default_param_names, \
-    eval_func_param_names_defaults, get_max_max_new_tokens, get_minmax_top_k_docs, history_to_context, langchain_actions
+from gen import get_model, languages_covered, evaluate, score_qa, langchain_modes, inputs_kwargs_list, scratch_base_dir, \
+    get_max_max_new_tokens, get_minmax_top_k_docs, history_to_context, langchain_actions
+from evaluate_params import eval_func_param_names, no_default_param_names, eval_func_param_names_defaults
 
 from apscheduler.schedulers.background import BackgroundScheduler
 
@@ -307,7 +307,7 @@ def go_gradio(**kwargs):
 
                     col_chat = gr.Column(visible=kwargs['chat'])
                     with col_chat:
-                        instruction, submit, stop_btn = make_prompt_form(kwargs)
+                        instruction, submit, stop_btn = make_prompt_form(kwargs, LangChainMode)
                         text_output, text_output2, text_outputs = make_chatbots(output_label0, output_label0_model2,
                                                                                 **kwargs)
 
@@ -756,6 +756,7 @@ def go_gradio(**kwargs):
                                                 caption_loader=caption_loader,
                                                 verbose=kwargs['verbose'],
                                                 user_path=kwargs['user_path'],
+                                                n_jobs=kwargs['n_jobs'],
                                                 )
         add_file_outputs = [fileup_output, langchain_mode, add_to_shared_db_btn, add_to_my_db_btn]
         add_file_kwargs = dict(fn=update_user_db_func,
@@ -834,6 +835,7 @@ def go_gradio(**kwargs):
                                               caption_loader=caption_loader,
                                               verbose=kwargs['verbose'],
                                               user_path=kwargs['user_path'],
+                                              n_jobs=kwargs['n_jobs'],
                                               )
 
         add_my_file_outputs = [fileup_output, langchain_mode, my_db_state, add_to_shared_db_btn, add_to_my_db_btn]
@@ -2269,7 +2271,8 @@ def _update_user_db(file, db1, x, y, chunk, chunk_size, dbs=None, db_type=None,
                     captions_model=None,
                     enable_ocr=None,
                     verbose=None,
-                    is_url=None, is_txt=None):
+                    is_url=None, is_txt=None,
+                    n_jobs=-1):
     assert use_openai_embedding is not None
     assert hf_embedding_model is not None
     assert caption_loader is not None
@@ -2310,6 +2313,7 @@ def _update_user_db(file, db1, x, y, chunk, chunk_size, dbs=None, db_type=None,
         print("Adding %s" % file, flush=True)
     sources = path_to_docs(file if not is_url and not is_txt else None,
                            verbose=verbose,
+                           n_jobs=n_jobs,
                            chunk=chunk, chunk_size=chunk_size,
                            url=file if is_url else None,
                            text=file if is_txt else None,

gradio_utils/prompt_form.py

@@ -3,8 +3,6 @@ import math
 
 import gradio as gr
 
-from enums import LangChainMode
-
 
 def make_chatbots(output_label0, output_label0_model2, **kwargs):
     text_outputs = []
@@ -97,7 +95,7 @@ def make_chatbots(output_label0, output_label0_model2, **kwargs):
     return text_output, text_output2, text_outputs
 
 
-def make_prompt_form(kwargs):
+def make_prompt_form(kwargs, LangChainMode):
     if kwargs['langchain_mode'] != LangChainMode.DISABLED.value:
         extra_prompt_form = ".  For summarization, empty submission uses first top_k_docs documents."
     else: