renpas22 commited on
Commit ·
e20135f
1
Parent(s): 61dbc34
Add VLM support to trainer with auto-detection
Browse files- src/reasoning/step_level_cot.py +75 -10
src/reasoning/step_level_cot.py
CHANGED
|
@@ -114,19 +114,84 @@ class StepLevelCoTTrainer:
|
|
| 114 |
logger.info("StepLevelCoTTrainer initialized")
|
| 115 |
|
| 116 |
def _load_model(self) -> None:
|
| 117 |
-
"""Load vision-language model."""
|
| 118 |
logger.info(f"Loading model: {self.config.pretrained_model_name_or_path}")
|
| 119 |
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
)
|
| 125 |
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
# Add special tokens for chain-of-thought reasoning
|
| 132 |
self._add_special_tokens()
|
|
|
|
| 114 |
logger.info("StepLevelCoTTrainer initialized")
|
| 115 |
|
| 116 |
def _load_model(self) -> None:
|
| 117 |
+
"""Load vision-language model (supports both VLM and diffusion models)."""
|
| 118 |
logger.info(f"Loading model: {self.config.pretrained_model_name_or_path}")
|
| 119 |
|
| 120 |
+
model_name = self.config.pretrained_model_name_or_path.lower()
|
| 121 |
+
|
| 122 |
+
# Detect model type
|
| 123 |
+
is_vlm = any(x in model_name for x in ["qwen2", "qwen-vl", "internvl", "llava", "cogvlm"])
|
| 124 |
+
is_diffusion = any(x in model_name for x in ["flux", "qwen-image", "stable-diffusion", "sd"])
|
| 125 |
|
| 126 |
+
if is_vlm:
|
| 127 |
+
# Load as Vision-Language Model
|
| 128 |
+
from transformers import AutoProcessor, AutoModelForVision2Seq, BitsAndBytesConfig
|
| 129 |
+
|
| 130 |
+
logger.info("Loading as Vision-Language Model...")
|
| 131 |
+
|
| 132 |
+
# Prepare quantization config if enabled
|
| 133 |
+
quantization_config = None
|
| 134 |
+
if self.config.get("quantize", False):
|
| 135 |
+
quantization_config = BitsAndBytesConfig(
|
| 136 |
+
load_in_8bit=True,
|
| 137 |
+
llm_int8_threshold=6.0,
|
| 138 |
+
llm_int8_has_fp16_weight=False,
|
| 139 |
+
)
|
| 140 |
+
|
| 141 |
+
# Load processor and model
|
| 142 |
+
self.tokenizer = AutoProcessor.from_pretrained(
|
| 143 |
+
self.config.pretrained_model_name_or_path,
|
| 144 |
+
trust_remote_code=True
|
| 145 |
+
)
|
| 146 |
+
|
| 147 |
+
torch_dtype = torch.float16 if self.config.mixed_precision == "fp16" else torch.bfloat16
|
| 148 |
+
|
| 149 |
+
self.model = AutoModelForVision2Seq.from_pretrained(
|
| 150 |
+
self.config.pretrained_model_name_or_path,
|
| 151 |
+
torch_dtype=torch_dtype,
|
| 152 |
+
quantization_config=quantization_config,
|
| 153 |
+
trust_remote_code=True,
|
| 154 |
+
device_map="auto"
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
# For VLMs, model itself is the main component
|
| 158 |
+
self.pipeline = None
|
| 159 |
+
self.vae = None
|
| 160 |
+
self.text_encoder = None
|
| 161 |
+
|
| 162 |
+
elif is_diffusion:
|
| 163 |
+
# Load as Diffusion Model
|
| 164 |
+
logger.info("Loading as Diffusion Model...")
|
| 165 |
+
|
| 166 |
+
self.pipeline = QwenImagePipeline.from_pretrained(
|
| 167 |
+
self.config.pretrained_model_name_or_path,
|
| 168 |
+
torch_dtype=torch.float16 if self.config.mixed_precision == "fp16" else torch.float32,
|
| 169 |
+
)
|
| 170 |
+
|
| 171 |
+
self.model = self.pipeline.transformer
|
| 172 |
+
self.vae = self.pipeline.vae
|
| 173 |
+
self.text_encoder = self.pipeline.text_encoder
|
| 174 |
+
self.tokenizer = self.pipeline.tokenizer
|
| 175 |
+
else:
|
| 176 |
+
# Try loading as VLM by default
|
| 177 |
+
logger.warning(f"Unknown model type, attempting to load as VLM...")
|
| 178 |
+
from transformers import AutoProcessor, AutoModelForVision2Seq
|
| 179 |
+
|
| 180 |
+
self.tokenizer = AutoProcessor.from_pretrained(
|
| 181 |
+
self.config.pretrained_model_name_or_path,
|
| 182 |
+
trust_remote_code=True
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
self.model = AutoModelForVision2Seq.from_pretrained(
|
| 186 |
+
self.config.pretrained_model_name_or_path,
|
| 187 |
+
torch_dtype=torch.bfloat16,
|
| 188 |
+
trust_remote_code=True,
|
| 189 |
+
device_map="auto"
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
self.pipeline = None
|
| 193 |
+
self.vae = None
|
| 194 |
+
self.text_encoder = None
|
| 195 |
|
| 196 |
# Add special tokens for chain-of-thought reasoning
|
| 197 |
self._add_special_tokens()
|