Added google/flan models and fixed AutoModelForSeq2SeqLM when loading T5 compression model (lm-sys#2402)

Author: wangzhen263

fastchat/model/compression.py

@@ -11,7 +11,13 @@
 from torch.nn import functional as F
 import torch.nn as nn
 from tqdm import tqdm
-from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, AutoModel
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    AutoModel,
+    AutoModelForSeq2SeqLM,
+)
 
 
 @dataclasses.dataclass
@@ -123,7 +129,13 @@ def load_compress_model(model_path, device, torch_dtype, use_fast, revision="mai
         # some models are loaded by AutoModel but not AutoModelForCausalLM,
         # such as chatglm, chatglm2
         try:
-            model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
+            # google/flan-* models are based on an AutoModelForSeq2SeqLM.
+            if "T5Config" in str(type(config)):
+                model = AutoModelForSeq2SeqLM.from_config(
+                    config, trust_remote_code=True
+                )
+            else:
+                model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
         except NameError:
             model = AutoModel.from_config(config, trust_remote_code=True)
         linear_weights = get_compressed_list(model)

fastchat/model/model_adapter.py

@@ -649,6 +649,13 @@ def load_model(self, model_path: str, from_pretrained_kwargs: dict):
         return model, tokenizer
 
 
+class FlanAdapter(T5Adapter):
+    """The model adapter for flan-t5-*, flan-ul2"""
+
+    def match(self, model_path: str):
+        return "flan" in model_path.lower()
+
+
 class KoalaAdapter(BaseModelAdapter):
     """The model adapter for koala"""
 
@@ -1592,6 +1599,7 @@ def get_default_conv_template(self, model_path: str) -> Conversation:
 register_model_adapter(LongChatAdapter)
 register_model_adapter(CodeT5pAdapter)
 register_model_adapter(T5Adapter)
+register_model_adapter(FlanAdapter)
 register_model_adapter(KoalaAdapter)
 register_model_adapter(AlpacaAdapter)
 register_model_adapter(ChatGLMAdapter)

pyproject.toml

@@ -19,7 +19,7 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-model_worker = ["accelerate>=0.21", "peft", "sentencepiece", "torch", "transformers>=4.31.0"]
+model_worker = ["accelerate>=0.21", "peft", "sentencepiece", "torch", "transformers>=4.31.0", "protobuf"]
 webui = ["gradio"]
 train = ["einops", "flash-attn>=2.0", "wandb"]
 llm_judge = ["openai", "anthropic>=0.3", "ray"]