File size: 8,281 Bytes
54f9196 |
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 |
import os
import torch
import torch.nn as nn
from transformers import (
AutoModelForCausalLM,
CLIPVisionModel,
PreTrainedModel,
PretrainedConfig,
AutoConfig,
AutoModel
)
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING
from typing import Optional
class MultimodalLFM2Config(PretrainedConfig):
model_type = "multimodal_lfm2"
def __init__(
self,
lfm2_model_name="LiquidAI/LFM2-1.2B",
clip_model_name="openai/clip-vit-base-patch32",
vision_projection_dim=512,
**kwargs
):
super().__init__(**kwargs)
self.lfm2_model_name = lfm2_model_name
self.clip_model_name = clip_model_name
self.vision_projection_dim = vision_projection_dim
class MultimodalLFM2Model(PreTrainedModel):
config_class = MultimodalLFM2Config
def __init__(self, config):
super().__init__(config)
# --- Language Model ---
self.language_model = AutoModelForCausalLM.from_pretrained(
config.lfm2_model_name,
torch_dtype=torch.bfloat16,
trust_remote_code=True
)
# --- Vision Encoder ---
self.vision_encoder = CLIPVisionModel.from_pretrained(config.clip_model_name)
for param in self.vision_encoder.parameters():
param.requires_grad = False
# --- Projection Layer ---
self.language_hidden_size = self.language_model.config.hidden_size
self.vision_hidden_size = self.vision_encoder.config.hidden_size
self.vision_projection = nn.Sequential(
nn.Linear(self.vision_hidden_size, config.vision_projection_dim),
nn.ReLU(),
nn.Dropout(0.1),
nn.Linear(config.vision_projection_dim, self.language_hidden_size),
nn.LayerNorm(self.language_hidden_size)
)
self.image_token_id = None
def gradient_checkpointing_enable(self, **kwargs):
"""Delegates gradient checkpointing to the language model."""
self.language_model.gradient_checkpointing_enable(**kwargs)
def _prepare_multimodal_inputs(
self,
input_ids: torch.Tensor,
images: torch.Tensor
) -> torch.Tensor:
"""
Prepares input embeddings by combining text and image features.
"""
inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
vision_outputs = self.vision_encoder(pixel_values=images)
image_features = vision_outputs.last_hidden_state
projected_image_features = self.vision_projection(image_features).to(self.language_model.dtype)
batch_size = input_ids.shape[0]
image_token_mask = (input_ids == self.image_token_id)
for i in range(batch_size):
image_positions = torch.where(image_token_mask[i])[0]
if len(image_positions) > 0:
img_feat = projected_image_features[i]
# match length
if len(image_positions) > img_feat.shape[0]:
repeat_times = (len(image_positions) + img_feat.shape[0] - 1) // img_feat.shape[0]
img_feat = img_feat.repeat(repeat_times, 1)[:len(image_positions)]
elif len(image_positions) < img_feat.shape[0]:
img_feat = img_feat[:len(image_positions)]
inputs_embeds[i, image_positions] = img_feat
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
images: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
**kwargs
):
"""
Forward pass for training.
"""
if images is not None and self.image_token_id is not None:
inputs_embeds = self._prepare_multimodal_inputs(input_ids, images)
final_input_ids = None
else:
inputs_embeds = None
final_input_ids = input_ids
return self.language_model(
input_ids=final_input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
labels=labels,
return_dict=True
)
def generate(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
images: Optional[torch.Tensor] = None,
**kwargs
):
"""
Generation method for inference.
"""
if images is not None and self.image_token_id is not None:
inputs_embeds = self._prepare_multimodal_inputs(input_ids, images)
final_input_ids = None
else:
inputs_embeds = None
final_input_ids = input_ids
return self.language_model.generate(
input_ids=final_input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
**kwargs
)
def save_pretrained(self, save_directory, **kwargs):
"""
Custom save method - saves everything in one directory.
"""
os.makedirs(save_directory, exist_ok=True)
# Save config
self.config.save_pretrained(save_directory)
# Save language model state dict directly
torch.save(
self.language_model.state_dict(),
os.path.join(save_directory, "language_model.bin")
)
# Save language model config
self.language_model.config.save_pretrained(save_directory, config_file_name="language_model_config.json")
# Save vision projection
torch.save(
self.vision_projection.state_dict(),
os.path.join(save_directory, "vision_projection.bin")
)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
"""
Custom loading method - works with your current structure.
"""
config = cls.config_class.from_pretrained(pretrained_model_name_or_path)
model = cls(config)
# Try to load from pytorch_model.bin (your current structure)
main_model_path = os.path.join(pretrained_model_name_or_path, "pytorch_model.bin")
if os.path.exists(main_model_path):
# Load the full model state dict
full_state_dict = torch.load(main_model_path, map_location="cpu")
# Separate language model and vision projection weights
language_state_dict = {}
projection_state_dict = {}
for key, value in full_state_dict.items():
if key.startswith("language_model."):
# Remove the "language_model." prefix
new_key = key[len("language_model."):]
language_state_dict[new_key] = value
elif key.startswith("vision_projection."):
# Remove the "vision_projection." prefix
new_key = key[len("vision_projection."):]
projection_state_dict[new_key] = value
# Load the separated state dicts
if language_state_dict:
model.language_model.load_state_dict(language_state_dict)
if projection_state_dict:
model.vision_projection.load_state_dict(projection_state_dict)
else:
# Fallback to separate files
language_model_path = os.path.join(pretrained_model_name_or_path, "language_model.bin")
if os.path.exists(language_model_path):
language_state_dict = torch.load(language_model_path, map_location="cpu")
model.language_model.load_state_dict(language_state_dict)
projection_path = os.path.join(pretrained_model_name_or_path, "vision_projection.bin")
if os.path.exists(projection_path):
projection_state_dict = torch.load(projection_path, map_location="cpu")
model.vision_projection.load_state_dict(projection_state_dict)
return model
# Register the model with transformers
AutoConfig.register("multimodal_lfm2", MultimodalLFM2Config)
AutoModelForCausalLM.register(MultimodalLFM2Config, MultimodalLFM2Model) |