Add fastchat api server tutorial

.gitignore

@@ -33,4 +33,4 @@ tests/state_of_the_union.txt
 
 # Build
 build
-!dummy_file
+!dummy_file

data_process/redis_data_filtering/prompt_reverse_engineering.py

@@ -2,7 +2,7 @@
 import os
 from collections import defaultdict
 from typing import Any, Dict, List, Tuple, Union, cast
-
+import transformers
 import pandas as pd
 import rich
 from rich.console import Console
@@ -15,33 +15,52 @@
 import enum
 
 #PROMPT_PREFIX = "Prompt after formatting:\n"
-
+MAX_TOKEN = 2048
 PROMPT_TEMPLATE="""Prompt after formatting:\nImagine you are {agent}, your task is to act/speak as {agent} would, keeping in mind {agent}'s social goal.
 You can find {agent}'s background and goal in the 'Here is the context of the interaction' field.
 Note that {agent}'s secret and goal is only visible to you.
 You should try your best to achieve {agent}'s goal in a way that align with their character traits.
 Additionally, maintaining the conversation's naturalness and realism is essential (e.g., do not repeat what other people has already said before).
 {history}.
-You are at Turn #{turn_number}. Your available action types are
-{action_list}.
-Note: You can "leave" this conversation if 1. you have achieved your social goals, 2. this conversation makes you uncomfortable, 3. you find it uninteresting/you lose your patience, 4. or for other reasons you want to leave.
-
-Please only generate a JSON string including the action type and the argument.
-Your action should follow the given format:
-{format_instructions}
-"""
+You are at Turn #{turn_number}."""
 
 #PYDANTIC_FORMAT_INSTRUCTIONS.format(schema=schema_str)
 FORMAT_TEMPLATE = """\nAs an example, for the schema {\"properties\": {\"foo\": {\"title\": \"Foo\", \"description\": \"a list of strings\", \"type\": \"array\", \"items\": {\"type\": \"string\"}}}, \"required\": [\"foo\"]}
 the object {\"foo\": [\"bar\", \"baz\"]} is a well-formatted instance of the schema. The object {\"properties\": {\"foo\": [\"bar\", \"baz\"]}} is not well-formatted.
 \nHere is the output schema:\n```\n{\"description\": \"An interface for messages.\\nThere is only one required method: to_natural_language\", \"properties\": {\"action_type\": {\"title\": \"Action Type\", \"description\": \"whether to speak at this turn or choose to not do anything\", \"enum\": [\"none\", \"speak\", \"non-verbal communication\", \"action\", \"leave\"], \"type\": \"string\"}, \"argument\": {\"title\": \"Argument\", \"description\": \"the utterance if choose to speak, the expression or gesture if choose non-verbal communication, or the physical action if choose action\", \"type\": \"string\"}}, \"required\": [\"action_type\", \"argument\"]}\n```\u001b[0m"""
 
 
+PROMPT_TEMPLATE_W_FORMAT="""Prompt after formatting:\nImagine you are {agent}, your task is to act/speak as {agent} would, keeping in mind {agent}'s social goal.
+You can find {agent}'s background and goal in the 'Here is the context of the interaction' field.
+Note that {agent}'s secret and goal is only visible to you.
+You should try your best to achieve {agent}'s goal in a way that align with their character traits.
+Additionally, maintaining the conversation's naturalness and realism is essential (e.g., do not repeat what other people has already said before).
+{history}.
+You are at Turn #{turn_number}. Your available action types are
+"none action speak non-verbal communication leave".
+Note: You can "leave" this conversation if 1. you have achieved your social goals, 2. this conversation makes you uncomfortable, 3. you find it uninteresting/you lose your patience, 4. or for other reasons you want to leave.
+
+Please only generate a JSON string including the action type and the argument.
+Your action should follow the given format:
+\nAs an example, for the schema {\"properties\": {\"foo\": {\"title\": \"Foo\", \"description\": \"a list of strings\", \"type\": \"array\", \"items\": {\"type\": \"string\"}}}, \"required\": [\"foo\"]}
+the object {\"foo\": [\"bar\", \"baz\"]} is a well-formatted instance of the schema. The object {\"properties\": {\"foo\": [\"bar\", \"baz\"]}} is not well-formatted.
+\nHere is the output schema:\n```\n{\"description\": \"An interface for messages.\\nThere is only one required method: to_natural_language\", \"properties\": {\"action_type\": {\"title\": \"Action Type\", \"description\": \"whether to speak at this turn or choose to not do anything\", \"enum\": [\"none\", \"speak\", \"non-verbal communication\", \"action\", \"leave\"], \"type\": \"string\"}, \"argument\": {\"title\": \"Argument\", \"description\": \"the utterance if choose to speak, the expression or gesture if choose non-verbal communication, or the physical action if choose action\", \"type\": \"string\"}}, \"required\": [\"action_type\", \"argument\"]}\n```\u001b[0m
+"""
 # static
 ACTION_LIST = "none action speak non-verbal communication leave" #" ".join(ActionType)
 
 ACTION_REVERSE_MAP = {"left ": "leave", 'did n': 'none', 'said:': 'speak'}
 
+MODEL_CHECKPOINT = "meta-llama/Llama-2-13b-chat-hf"
+HF_TOKEN = "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"
+
+
+TOKENIZER = transformers.AutoTokenizer.from_pretrained(
+    MODEL_CHECKPOINT,
+    padding = False,
+    truncation = False,
+    token=HF_TOKEN,
+    )
 
 def to_natural_language(self) -> str:
         match self.action_type:
@@ -101,10 +120,27 @@ def generate_result(msg):
 
     return str_result
 
-def reverse_episode_log(epilog, later_speak=False):
+def surpass_max_token_check(string, max_token=MAX_TOKEN, tokenizer=TOKENIZER):
+    prompt_tokens = len(tokenizer(string)['input_ids'])
+    return max(prompt_tokens - max_token, 0)
+
+def truncate_prompt_to_length(dia_his, surpass_num, tokenizer=TOKENIZER):
+    # context_len = len(tokenizer(context)['input_ids'])
+    dia_sen = dia_his.split("\n")
+    remove_len = 0
+    i = 0
+    while remove_len < surpass_num:
+        remove_len+=len(tokenizer(dia_sen[i])['input_ids'])
+        i+=1
+    trunc_dia = "\n".join(p for p in dia_sen[i:])
+    return trunc_dia
+
+
+def reverse_episode_log(epilog, later_speak=False, include_format=False, max_token=MAX_TOKEN):
     episode_msg = epilog.messages
     # per episode
     agent_model = epilog.models[1]
+    promt_template = PROMPT_TEMPLATE_W_FORMAT if include_format else PROMPT_TEMPLATE
 
     if len(episode_msg) > 0:
         init_loop = episode_msg[0]
@@ -131,23 +167,31 @@ def reverse_episode_log(epilog, later_speak=False):
                     dial_history += "\n"+tpl[2]
                 else:
                     # for the first context, we don't need \n 
-                    dial_history += tpl[2]
+                    context = tpl[2]
+                    dial_history += context
 
             if tpl[0] == speaker: # if speaker is the agent, use what he said as result
                 str_result = generate_result(tpl[2])
                 # check if this is the end 
         if i%2 == turn_div:  
             # take alternative turns as we always want to predict one agent, not both
             next_turn = i
-            prompt = PROMPT_TEMPLATE.format(
-                    agent=speaker, history=dial_history, turn_number=next_turn, 
-                    action_list=ACTION_LIST, format_instructions=FORMAT_TEMPLATE)
+            prompt = promt_template.format(
+                    agent=speaker, history=dial_history, turn_number=next_turn)
+            over_tokens = surpass_max_token_check(prompt)
+            if over_tokens > 0:
+                all_dial = dial_history[len(context):]
+                #print(all_dial)
+                trun_dial = truncate_prompt_to_length(all_dial, over_tokens)
+                prompt = promt_template.format(
+                    agent=speaker, history=context+"\n"+trun_dial, turn_number=next_turn)
             turn_dic["prompt"] = prompt   
             turn_dic['result'] = str_result
             prompt_result_instances.append(turn_dic)
 
     return prompt_result_instances
 
+
 def parse_prompt_to_json(episode, dir, init_speak):
     prompt_result_instances = reverse_episode_log(episode, init_speak)
 

data_process/redis_data_filtering/redis_filtering.py

@@ -95,7 +95,7 @@ def goal_filter_per_env_agent(episodes):
             env_tpls.append((episodes[agent1_rank[i]], 0))
             env_tpls.append((episodes[agent2_rank[i]], 1))
         else:
-            if goal_score['agent1'][agent1_rank[i]] >= min(GOAL_KEEP_THRESHOD, agent1_avg) and (goal_score['agent2'][agent2_rank[i]] >= min(KEEP_THRESHOD, agent2_avg)):
+            if goal_score['agent1'][agent1_rank[i]] >= min(GOAL_KEEP_THRESHOD, agent1_avg) and (goal_score['agent2'][agent2_rank[i]] >= min(GOAL_KEEP_THRESHOD, agent2_avg)):
                 env_tpls.append((episodes[agent1_rank[i]], 0))
                 env_tpls.append((episodes[agent1_rank[i]], 1))
 

llm_deploy/README.md

@@ -7,6 +7,7 @@ Go to the vllm dir and pip install -e .
 To notice https://github.com/vllm-project/vllm/issues/1283, need to modify the config file to "== 2.0.1" and the pytorch version if facing with CUDA version error.
 
 
+
 ## Deploy finetuned model on babel via FastChat API server
 ### Login with SSH key
 1. Add public ed25519 key to server
@@ -161,6 +162,7 @@ If the above command runs successfully, you should be able to use REST API on yo
 
 
 
+
 ### Userful resource links for babel
 1. https://hpc.lti.cs.cmu.edu/wiki/index.php?title=BABEL#Cluster_Architecture
 2. https://hpc.lti.cs.cmu.edu/wiki/index.php?title=VSCode

llm_ft/data/create_dummy.py

@@ -0,0 +1,12 @@
+import json
+
+dummy_qa = {"id": "", "conversations": [{"from": "human", "value": "How old is Haofei?"}, {"from": "gpt", "value": "He is one year old."}]}
+
+res = []
+for i in range(1000):
+    new_qa = dict(dummy_qa)
+    new_qa["id"] = f"identity_{i}"
+    res.append(new_qa)
+
+with open("./dummy_convs.json", "w") as f:
+    json.dump(res, f, indent=4)

llm_ft/data/data_filter_out_long.py

@@ -0,0 +1,28 @@
+import json
+import transformers
+
+INPUT_PATH = "fastchat-ft-gpt4-gpt4-easy-2-side-partial-speak.json"
+OUTPUT_PATH = "fastchat-ft-gpt4-gpt4-easy-2-side-partial-speak-no-long.json"
+MODEL_CHECKPOINT = "meta-llama/Llama-2-13b-chat-hf"
+HF_TOKEN = "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"
+
+with open(INPUT_PATH, 'r') as f:
+    data = json.load(f)
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(
+    MODEL_CHECKPOINT,
+    padding = False,
+    truncation = False,
+    token=HF_TOKEN,
+    )
+
+res = []
+for d in data:
+    for conv in d['conversations']:
+        if conv['from'] == "human":
+            input_ids = tokenizer(conv['value'])
+            if len(input_ids) <= 2048:
+                res.append(d)
+
+with open(OUTPUT_PATH, 'w') as f:
+    json.dump(res, f, indent=4)

llm_ft/data/data_keep_only_speak.py

@@ -0,0 +1,16 @@
+import json
+
+INPUT_PATH = "fastchat-ft-gpt4-gpt4-easy-2-side-partial.json"
+OUTPUT_PATH = "fastchat-ft-gpt4-gpt4-easy-2-side-partial-speak.json"
+
+with open(INPUT_PATH, 'r') as f:
+    data = json.load(f)
+
+res = []
+for d in data:
+    for conv in d['conversations']:
+        if conv['from'] == "gpt" and "'action_type': 'speak'" in conv['value']:
+            res.append(d)
+
+with open(OUTPUT_PATH, 'w') as f:
+    json.dump(res, f, indent=4)

llm_ft/fastchat/model/model_adapter.py

@@ -60,10 +60,12 @@ def load_model(self, model_path: str, from_pretrained_kwargs: dict):
                 use_fast=self.use_fast_tokenizer,
                 revision=revision,
                 trust_remote_code=True,
+                token=None if not "token" in from_pretrained_kwargs else from_pretrained_kwargs["token"]
             )
         except TypeError:
             tokenizer = AutoTokenizer.from_pretrained(
-                model_path, use_fast=False, revision=revision, trust_remote_code=True
+                model_path, use_fast=False, revision=revision, trust_remote_code=True,
+                token=None if not "token" in from_pretrained_kwargs else from_pretrained_kwargs["token"]
             )
         try:
             model = AutoModelForCausalLM.from_pretrained(
@@ -154,6 +156,7 @@ def load_model(
     awq_config: Optional[AWQConfig] = None,
     revision: str = "main",
     debug: bool = False,
+    hf_access_token: Optional[str|None] = None,
 ):
     """Load a model from Hugging Face."""
     # get model adapter
@@ -280,6 +283,9 @@ def load_model(
 
     if dtype is not None:  # Overwrite dtype if it is provided in the arguments.
         kwargs["torch_dtype"] = dtype
+
+    if hf_access_token:
+        kwargs["token"] = hf_access_token
 
     # Load model
     model, tokenizer = adapter.load_model(model_path, kwargs)

llm_ft/fastchat/serve/cli.py

@@ -236,6 +236,7 @@ def main(args):
             judge_sent_end=args.judge_sent_end,
             debug=args.debug,
             history=not args.no_history,
+            hf_access_token = args.hf_access_token,
         )
     except KeyboardInterrupt:
         print("exit...")
@@ -281,5 +282,8 @@ def main(args):
         action="store_true",
         help="Print useful debug information (e.g., prompts)",
     )
+    parser.add_argument(
+        "--hf-access-token", type=str, default=None, help="Optional access token for Hugging Face."
+    )
     args = parser.parse_args()
     main(args)

llm_ft/fastchat/serve/inference.py

@@ -308,6 +308,7 @@ def chat_loop(
     judge_sent_end: bool = True,
     debug: bool = True,
     history: bool = True,
+    hf_access_token: Optional[str|None] = None,
 ):
     # Model
     model, tokenizer = load_model(
@@ -322,6 +323,7 @@ def chat_loop(
         awq_config=awq_config,
         revision=revision,
         debug=debug,
+        hf_access_token=hf_access_token,
     )
     generate_stream_func = get_generate_stream_function(model, model_path)