add dtype and seed (#2430)

fastchat/llm_judge/README.md

@@ -10,7 +10,7 @@ To automate the evaluation process, we prompt strong LLMs like GPT-4 to act as j
 - [Review Pre-Generated Model Answers and Judgments](#review-pre-generated-model-answers-and-judgments)
 - [MT-Bench](#mt-bench)
 - [Agreement Computation](#agreement-computation)
-- [Dataset](#dataset)
+- [Datasets](#datasets)
 - [Citation](#citation)
 
 ## Install
@@ -133,7 +133,7 @@ We released 3.3K human annotations for model responses generated by 6 models in
 
 This Colab [notebook](https://colab.research.google.com/drive/1ctgygDRJhVGUJTQy8-bRZCl1WNcT8De6?usp=sharing) shows how to compute the agreement between humans and GPT-4 judge with the dataset. Our results show that humans and GPT-4 judge achieve over 80\% agreement, the same level of agreement between humans.
 
-## Dataset
+## Datasets
 - [Chatbot Arena Conversation Dataset](https://huggingface.co/datasets/lmsys/chatbot_arena_conversations)
 - [MT-bench Human Annotation Dataset](https://huggingface.co/datasets/lmsys/mt_bench_human_judgments)
 

fastchat/llm_judge/gen_model_answer.py

@@ -15,6 +15,7 @@
 
 from fastchat.llm_judge.common import load_questions, temperature_config
 from fastchat.model import load_model, get_conversation_template
+from fastchat.utils import str_to_torch_dtype
 
 
 def run_eval(
@@ -29,6 +30,7 @@ def run_eval(
     num_gpus_per_model,
     num_gpus_total,
     max_gpu_memory,
+    dtype,
 ):
     questions = load_questions(question_file, question_begin, question_end)
     # random shuffle the questions to balance the loading
@@ -45,7 +47,7 @@ def run_eval(
     else:
         get_answers_func = get_model_answers
 
-    chunk_size = len(questions) // (num_gpus_total // num_gpus_per_model) // 2
+    chunk_size = len(questions) // (num_gpus_total // num_gpus_per_model)
     ans_handles = []
     for i in range(0, len(questions), chunk_size):
         ans_handles.append(
@@ -58,6 +60,7 @@ def run_eval(
                 num_choices,
                 num_gpus_per_model,
                 max_gpu_memory,
+                dtype=dtype,
             )
         )
 
@@ -75,12 +78,14 @@ def get_model_answers(
     num_choices,
     num_gpus_per_model,
     max_gpu_memory,
+    dtype,
 ):
     model, tokenizer = load_model(
         model_path,
         device="cuda",
         num_gpus=num_gpus_per_model,
         max_gpu_memory=max_gpu_memory,
+        dtype=dtype,
         load_8bit=False,
         cpu_offloading=False,
         debug=False,
@@ -192,7 +197,9 @@ def reorg_answer_file(answer_file):
         required=True,
         help="The path to the weights. This can be a local folder or a Hugging Face repo ID.",
     )
-    parser.add_argument("--model-id", type=str, required=True)
+    parser.add_argument(
+        "--model-id", type=str, required=True, help="A custom name for the model."
+    )
     parser.add_argument(
         "--bench-name",
         type=str,
@@ -234,6 +241,14 @@ def reorg_answer_file(answer_file):
         type=str,
         help="Maxmum GPU memory used for model weights per GPU.",
     )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16"],
+        help="Override the default dtype. If not set, it will use float16 on GPU and float32 on CPU.",
+        default=None,
+    )
+
     args = parser.parse_args()
 
     if args.num_gpus_total // args.num_gpus_per_model > 1:
@@ -250,17 +265,18 @@ def reorg_answer_file(answer_file):
     print(f"Output to {answer_file}")
 
     run_eval(
-        args.model_path,
-        args.model_id,
-        question_file,
-        args.question_begin,
-        args.question_end,
-        answer_file,
-        args.max_new_token,
-        args.num_choices,
-        args.num_gpus_per_model,
-        args.num_gpus_total,
-        args.max_gpu_memory,
+        model_path=args.model_path,
+        model_id=args.model_id,
+        question_file=question_file,
+        question_begin=args.question_begin,
+        question_end=args.question_end,
+        answer_file=answer_file,
+        max_new_token=args.max_new_token,
+        num_choices=args.num_choices,
+        num_gpus_per_model=args.num_gpus_per_model,
+        num_gpus_total=args.num_gpus_total,
+        max_gpu_memory=args.max_gpu_memory,
+        dtype=str_to_torch_dtype(args.dtype),
     )
 
     reorg_answer_file(answer_file)

fastchat/model/model_adapter.py

@@ -152,6 +152,7 @@ def load_model(
     device: str = "cuda",
     num_gpus: int = 1,
     max_gpu_memory: Optional[str] = None,
+    dtype: Optional[torch.dtype] = None,
     load_8bit: bool = False,
     cpu_offloading: bool = False,
     gptq_config: Optional[GptqConfig] = None,
@@ -282,6 +283,9 @@ def load_model(
         return model, tokenizer
     kwargs["revision"] = revision
 
+    if dtype is not None:  # Overwrite dtype if it is provided in the arguments.
+        kwargs["torch_dtype"] = dtype
+
     # Load model
     model, tokenizer = adapter.load_model(model_path, kwargs)
 
@@ -393,6 +397,13 @@ def add_model_args(parser):
         type=str,
         help="The maximum memory per GPU for storing model weights. Use a string like '13Gib'",
     )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16"],
+        help="Override the default dtype. If not set, it will use float16 on GPU and float32 on CPU.",
+        default=None,
+    )
     parser.add_argument(
         "--load-8bit", action="store_true", help="Use 8-bit quantization"
     )

fastchat/serve/cli.py

@@ -26,11 +26,13 @@
 from rich.console import Console
 from rich.live import Live
 from rich.markdown import Markdown
+import torch
 
 from fastchat.model.model_adapter import add_model_args
 from fastchat.modules.gptq import GptqConfig
 from fastchat.modules.awq import AWQConfig
 from fastchat.serve.inference import ChatIO, chat_loop
+from fastchat.utils import str_to_torch_dtype
 
 
 class SimpleChatIO(ChatIO):
@@ -208,6 +210,7 @@ def main(args):
             args.device,
             args.num_gpus,
             args.max_gpu_memory,
+            str_to_torch_dtype(args.dtype),
             args.load_8bit,
             args.cpu_offloading,
             args.conv_template,

fastchat/serve/inference.py

@@ -291,6 +291,7 @@ def chat_loop(
     device: str,
     num_gpus: int,
     max_gpu_memory: str,
+    dtype: Optional[torch.dtype],
     load_8bit: bool,
     cpu_offloading: bool,
     conv_template: Optional[str],
@@ -312,6 +313,7 @@ def chat_loop(
         device=device,
         num_gpus=num_gpus,
         max_gpu_memory=max_gpu_memory,
+        dtype=dtype,
         load_8bit=load_8bit,
         cpu_offloading=cpu_offloading,
         gptq_config=gptq_config,

fastchat/serve/model_worker.py

@@ -34,6 +34,7 @@
     )
 import torch
 import torch.nn.functional as F
+from transformers import set_seed
 import uvicorn
 
 from fastchat.constants import WORKER_HEART_BEAT_INTERVAL, ErrorCode, SERVER_ERROR_MSG
@@ -46,7 +47,12 @@
 )
 from fastchat.modules.gptq import GptqConfig
 from fastchat.modules.awq import AWQConfig
-from fastchat.utils import build_logger, pretty_print_semaphore, get_context_length
+from fastchat.utils import (
+    build_logger,
+    pretty_print_semaphore,
+    get_context_length,
+    str_to_torch_dtype,
+)
 
 
 worker_id = str(uuid.uuid4())[:8]
@@ -190,13 +196,15 @@ def __init__(
         device: str,
         num_gpus: int,
         max_gpu_memory: str,
+        dtype: Optional[torch.dtype] = None,
         load_8bit: bool = False,
         cpu_offloading: bool = False,
         gptq_config: Optional[GptqConfig] = None,
         awq_config: Optional[AWQConfig] = None,
         stream_interval: int = 2,
-        conv_template: str = None,
+        conv_template: Optional[str] = None,
         embed_in_truncate: bool = False,
+        seed: Optional[int] = None,
         **kwargs,
     ):
         super().__init__(
@@ -215,6 +223,7 @@ def __init__(
             device=device,
             num_gpus=num_gpus,
             max_gpu_memory=max_gpu_memory,
+            dtype=dtype,
             load_8bit=load_8bit,
             cpu_offloading=cpu_offloading,
             gptq_config=gptq_config,
@@ -227,6 +236,7 @@ def __init__(
         self.generate_stream_func = get_generate_stream_function(self.model, model_path)
         self.stream_interval = stream_interval
         self.embed_in_truncate = embed_in_truncate
+        self.seed = seed
 
         if not no_register:
             self.init_heart_beat()
@@ -235,6 +245,8 @@ def generate_stream_gate(self, params):
         self.call_ct += 1
 
         try:
+            if self.seed is not None:
+                set_seed(self.seed)
             for output in self.generate_stream_func(
                 self.model,
                 self.tokenizer,
@@ -475,6 +487,12 @@ def create_model_worker():
     )
     parser.add_argument("--stream-interval", type=int, default=2)
     parser.add_argument("--no-register", action="store_true")
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=None,
+        help="Overwrite the random seed for each generation.",
+    )
     args = parser.parse_args()
     logger.info(f"args: {args}")
 
@@ -508,13 +526,15 @@ def create_model_worker():
         device=args.device,
         num_gpus=args.num_gpus,
         max_gpu_memory=args.max_gpu_memory,
+        dtype=str_to_torch_dtype(args.dtype),
         load_8bit=args.load_8bit,
         cpu_offloading=args.cpu_offloading,
         gptq_config=gptq_config,
         awq_config=awq_config,
         stream_interval=args.stream_interval,
         conv_template=args.conv_template,
         embed_in_truncate=args.embed_in_truncate,
+        seed=args.seed,
     )
     return args, worker
 

fastchat/utils.py

@@ -302,3 +302,18 @@ def get_context_length(config):
         if val is not None:
             return int(rope_scaling_factor * val)
     return 2048
+
+
+def str_to_torch_dtype(dtype: str):
+    import torch
+
+    if dtype is None:
+        return None
+    elif dtype == "float32":
+        return torch.float32
+    elif dtype == "float16":
+        return torch.float16
+    elif dtype == "bfloat16":
+        return torch.bfloat16
+    else:
+        raise ValueError(f"Unrecognized dtype: {dtype}")