File size: 9,056 Bytes
b8f4d94 8adbcfe b8f4d94 8adbcfe b8f4d94 8adbcfe b8f4d94 8adbcfe b8f4d94 8adbcfe b8f4d94 8adbcfe b8f4d94 8adbcfe 6ba2d75 b8f4d94 8adbcfe b8f4d94 8adbcfe 6ba2d75 8adbcfe 6ba2d75 8adbcfe 6ba2d75 8adbcfe 6ba2d75 8adbcfe b8f4d94 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 |
import os
from abc import ABC, abstractmethod
import torch
from huggingface_hub import InferenceClient
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel, PeftConfig
from safetensors.torch import load_file, save_file
from huggingface_hub import hf_hub_download
import tempfile
import shutil
from pathlib import Path
HF_TOKEN = os.environ.get("HF_TOKEN")
# Cache model instances to avoid reloading
_model_cache = {}
model_configs = {
"baseline": {
"id": "meta-llama/Llama-3.2-1B-Instruct",
"type": "endpoint"
},
"fine_tuned": {
"base_model_id": "meta-llama/Llama-3.2-1B-Instruct",
"adapter_id": "aracape/teaching-assistant-1B-dpo",
"type": "pipeline"
},
"prompted": {
"id": "meta-llama/Llama-3.2-1B-Instruct",
"type": "endpoint"
},
}
class ModelWrapper(ABC):
"""Abstract base class for model wrappers"""
def __init__(self, model_id: str):
self.model_id = model_id
@abstractmethod
def generate(self, messages: list[dict], max_tokens: int, temperature: float) -> str:
"""Generate a response given messages"""
pass
class InferenceEndpointModel(ModelWrapper):
"""Wrapper for models deployed as HF inference endpoints"""
def __init__(self, model_id: str):
super().__init__(model_id)
self.client = InferenceClient(model_id, token=HF_TOKEN)
def generate(self, messages: list[dict], max_tokens: int, temperature: float) -> str:
"""Generate a complete response"""
result = self.client.chat_completion(
messages,
max_tokens=max_tokens,
temperature=temperature
)
return result.choices[0].message["content"]
def generate(self, messages: list[dict], max_tokens: int, temperature: float):
"""Generate a streaming response"""
response = ""
for message_chunk in self.client.chat_completion(
messages,
max_tokens=max_tokens,
stream=True,
temperature=temperature,
):
if message_chunk.choices and message_chunk.choices[0].delta.content:
token = message_chunk.choices[0].delta.content
response += token
return response
class PipelineModel(ModelWrapper):
"""Wrapper for models loaded locally with transformers pipeline"""
def __init__(self, base_model_id: str, adapter_id: str):
super().__init__(adapter_id)
# Load pipeline lazily on first use to save memory
self._pipeline = None
self.base_model_id = base_model_id
self.adapter_id = adapter_id
def _load_model(self):
if torch.cuda.is_available():
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype="bfloat16",
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
)
else:
quantization_config = None
print(f"Loading base model: {self.base_model_id}")
base_model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
dtype="auto",
device_map="auto",
quantization_config=quantization_config,
token=HF_TOKEN
)
# Load the LoRA adapter
print(f"Loading adapter: {self.adapter_id}")
model = self._load_adapter_robust(base_model)
return model
def _load_adapter_robust(self, base_model):
"""
Robustly load the adapter, handling potential key mismatches
(common with TRL/DPO trained models having extra nesting).
"""
# Create offload folder if needed for accelerate
offload_folder = Path("offload")
offload_folder.mkdir(exist_ok=True)
try:
return PeftModel.from_pretrained(
base_model,
self.adapter_id,
offload_folder=str(offload_folder),
token=HF_TOKEN
)
except KeyError as e:
# Check for common key errors indicating extra nesting
# The error is usually like "KeyError: 'base_model.model.model.model.embed_tokens'"
# or similar, implying the adapter keys have an extra .model
if "base_model.model.model" in str(e) or "embed_tokens" in str(e):
print(f"Encountered KeyError loading adapter: {e}. Attempting to fix keys...")
return self._load_adapter_with_key_fix(base_model, offload_folder=str(offload_folder))
raise e
def _load_adapter_with_key_fix(self, base_model, offload_folder=None):
"""
Download adapter, fix keys by removing extra 'model' nesting, and load.
"""
print("Patching adapter weights to fix key mismatch...")
with tempfile.TemporaryDirectory() as temp_dir:
temp_path = Path(temp_dir)
try:
config_path = hf_hub_download(self.adapter_id, "adapter_config.json", token=HF_TOKEN)
shutil.copy(config_path, temp_path / "adapter_config.json")
except Exception as e:
raise RuntimeError(f"Failed to download adapter config: {e}")
# 2. Download and load weights
is_safetensors = True
try:
weights_path = hf_hub_download(self.adapter_id, "adapter_model.safetensors", token=HF_TOKEN)
weights = load_file(weights_path)
except Exception:
try:
weights_path = hf_hub_download(self.adapter_id, "adapter_model.bin", token=HF_TOKEN)
weights = torch.load(weights_path, map_location="cpu")
is_safetensors = False
except Exception as e:
raise RuntimeError(f"Failed to download adapter weights: {e}")
# 3. Fix keys: replace 'base_model.model.model.' with 'base_model.model.'
new_weights = {}
fixed_count = 0
for k, v in weights.items():
if "base_model.model.model." in k:
new_k = k.replace("base_model.model.model.", "base_model.model.")
new_weights[new_k] = v
fixed_count += 1
else:
new_weights[k] = v
print(f"Fixed {fixed_count} keys in adapter weights.")
# 4. Save fixed weights
if is_safetensors:
save_file(new_weights, temp_path / "adapter_model.safetensors")
else:
torch.save(new_weights, temp_path / "adapter_model.bin")
# 5. Load PeftModel manually to handle meta tensors
print("Loading PeftModel with manual state_dict assignment...")
config = PeftConfig.from_pretrained(temp_dir)
model = PeftModel(base_model, config)
# Load state dict with assign=True to handle meta tensors
# This replaces the meta tensors with the loaded CPU tensors
try:
model.load_state_dict(new_weights, strict=False, assign=True)
except TypeError:
print("Warning: load_state_dict does not support assign=True. Loading might fail for meta tensors.")
model.load_state_dict(new_weights, strict=False)
return model
@property
def pipe(self):
if self._pipeline is None:
model = self._load_model()
tokenizer = AutoTokenizer.from_pretrained(self.adapter_id, token=HF_TOKEN)
self._pipeline = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
)
return self._pipeline
def generate(self, messages: list[dict], max_tokens: int, temperature: float) -> str:
"""Generate a complete response"""
result = self.pipe(
messages,
max_new_tokens=max_tokens,
temperature=temperature,
do_sample=temperature > 0,
)
return result[0]["generated_text"][-1]["content"]
def create_model(model_key: str) -> ModelWrapper:
"""Factory function to create the appropriate model wrapper"""
config = model_configs[model_key]
if config["type"] == "endpoint":
return InferenceEndpointModel(config["id"])
elif config["type"] == "pipeline":
return PipelineModel(config["base_model_id"], config["adapter_id"])
else:
raise ValueError(f"Unknown model type: {config['type']}")
def get_model(model_key: str) -> ModelWrapper:
"""Get or create a model instance"""
if model_key not in _model_cache:
_model_cache[model_key] = create_model(model_key)
return _model_cache[model_key]
|