拓展功能检查依赖提示安装

models/loader/loader.py

@@ -4,18 +4,11 @@ import os
 import re
 import time
 from pathlib import Path
-from peft import PeftModel
 from typing import Optional, List, Dict, Tuple, Union
 import torch
 import transformers
-
 from transformers import (AutoConfig, AutoModel, AutoModelForCausalLM,
-                          AutoTokenizer, BitsAndBytesConfig, LlamaTokenizer)
-from transformers.dynamic_module_utils import get_class_from_dynamic_module
-from transformers.modeling_utils import no_init_weights
-from transformers.utils import ContextManagers
-from accelerate import init_empty_weights
-from accelerate.utils import get_balanced_memory, infer_auto_device_map
+                          AutoTokenizer, LlamaTokenizer)
 from configs.model_config import LLM_DEVICE
 
 
@@ -150,16 +143,35 @@ class LoaderCheckPoint:
                 )
 
         elif self.is_llamacpp:
+
+            try:
                 from models.extensions.llamacpp_model_alternative import LlamaCppModel
 
+            except ImportError as exc:
+                raise ValueError(
+                    "Could not import depend python package "
+                    "Please install it with `pip install llama-cpp-python`."
+                ) from exc
+
             model_file = list(checkpoint.glob('ggml*.bin'))[0]
             print(f"llama.cpp weights detected: {model_file}\n")
 
             model, tokenizer = LlamaCppModel.from_pretrained(model_file)
             return model, tokenizer
 
-        # Custom
-        else:
+        elif self.load_in_8bit:
+            try:
+                from accelerate import init_empty_weights
+                from accelerate.utils import get_balanced_memory, infer_auto_device_map
+                from transformers import BitsAndBytesConfig
+
+            except ImportError as exc:
+                raise ValueError(
+                    "Could not import depend python package "
+                    "Please install it with `pip install transformers` "
+                    "`pip install bitsandbytes``pip install accelerate`."
+                ) from exc
+
             params = {"low_cpu_mem_usage": True}
 
             if not self.llm_device.lower().startswith("cuda"):
@@ -167,18 +179,11 @@ class LoaderCheckPoint:
             else:
                 params["device_map"] = 'auto'
                 params["trust_remote_code"] = True
-                if self.load_in_8bit:
                 params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True,
                                                                    llm_int8_enable_fp32_cpu_offload=False)
-                elif self.bf16:
-                    params["torch_dtype"] = torch.bfloat16
-                else:
-                    params["torch_dtype"] = torch.float16
 
-            if self.load_in_8bit and params.get('max_memory', None) is not None and params['device_map'] == 'auto':
-                config = AutoConfig.from_pretrained(checkpoint)
             with init_empty_weights():
-                    model = LoaderClass.from_config(config)
+                model = LoaderClass.from_config(self.model_config,trust_remote_code = True)
             model.tie_weights()
             if self.device_map is not None:
                 params['device_map'] = self.device_map
@@ -186,11 +191,18 @@ class LoaderCheckPoint:
                 params['device_map'] = infer_auto_device_map(
                     model,
                     dtype=torch.int8,
-                        max_memory=params['max_memory'],
                     no_split_module_classes=model._no_split_modules
                 )
+            try:
 
                 model = LoaderClass.from_pretrained(checkpoint, **params)
+            except ImportError as exc:
+                raise ValueError(
+                    "如果开启了8bit量化加载,项目无法启动，参考此位置，选择合适的cuda版本，https://github.com/TimDettmers/bitsandbytes/issues/156"
+                ) from exc
+        # Custom
+        else:
+            pass
 
         # Loading the tokenizer
         if type(model) is transformers.LlamaForCausalLM:
@@ -247,6 +259,20 @@ class LoaderCheckPoint:
         return device_map
 
     def moss_auto_configure_device_map(self, num_gpus: int, model_name) -> Dict[str, int]:
+        try:
+
+            from accelerate import init_empty_weights
+            from accelerate.utils import get_balanced_memory, infer_auto_device_map
+            from transformers.dynamic_module_utils import get_class_from_dynamic_module
+            from transformers.modeling_utils import no_init_weights
+            from transformers.utils import ContextManagers
+        except ImportError as exc:
+            raise ValueError(
+                "Could not import depend python package "
+                "Please install it with `pip install transformers` "
+                "`pip install bitsandbytes``pip install accelerate`."
+            ) from exc
+
         checkpoint = Path(f'{self.model_dir}/{model_name}')
 
         if self.model_path:
@@ -271,6 +297,16 @@ class LoaderCheckPoint:
             return device_map
 
     def _add_lora_to_model(self, lora_names):
+
+        try:
+
+            from peft import PeftModel
+
+        except ImportError as exc:
+            raise ValueError(
+                "Could not import depend python package. "
+                "Please install it with `pip install peft``pip install accelerate`."
+            ) from exc
         # 目前加载的lora
         prior_set = set(self.lora_names)
         # 需要加载的

requirements.txt

@@ -11,16 +11,16 @@ beautifulsoup4
 icetk
 cpm_kernels
 faiss-cpu
-accelerate~=0.18.0
 gradio==3.28.3
 fastapi~=0.95.0
 uvicorn~=0.21.1
-peft~=0.3.0
 pypinyin~=0.48.0
 click~=8.1.3
 tabulate
 azure-core
-bitsandbytes; platform_system != "Windows"
+#accelerate~=0.18.0
+#peft~=0.3.0
+#bitsandbytes; platform_system != "Windows"
 #llama-cpp-python==0.1.34; platform_system != "Windows"
 #https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.34/llama_cpp_python-0.1.34-cp310-cp310-win_amd64.whl; platform_system == "Windows"