File size: 12,347 Bytes
31112ad |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 |
from pathlib import Path
import torch
import torch.nn as nn
from einops import rearrange
from PIL import Image
from transformers import AutoProcessor, Mistral3ForConditionalGeneration, pipeline
from shared.utils import files_locator as fl
from .sampling import cap_pixels, concatenate_images
from .system_messages import (
PROMPT_IMAGE_INTEGRITY,
PROMPT_IMAGE_INTEGRITY_FOLLOW_UP,
PROMPT_TEXT_INTEGRITY,
SYSTEM_MESSAGE,
SYSTEM_MESSAGE_UPSAMPLING_I2I,
SYSTEM_MESSAGE_UPSAMPLING_T2I,
SYSTEM_PROMPT_CONTENT_FILTER,
)
OUTPUT_LAYERS = [10, 20, 30]
MAX_LENGTH = 512
NSFW_THRESHOLD = 0.85
UPSAMPLING_MAX_IMAGE_SIZE = 768**2
from mmgp import offload
import os
class Mistral3SmallEmbedder(nn.Module):
def __init__(
self,
model_spec = None,
torch_dtype: str = "bfloat16",
):
super().__init__()
file_path = model_spec
self.model = offload.fast_load_transformers_model(file_path, writable_tensors= False, defaultConfigPath= os.path.join(os.path.dirname(file_path), "config.json"))
self.processor = AutoProcessor.from_pretrained(os.path.dirname(file_path), use_fast=False)
self.yes_token, self.no_token = self.processor.tokenizer.encode(
["yes", "no"], add_special_tokens=False
)
self.max_length = MAX_LENGTH
self.upsampling_max_image_size = UPSAMPLING_MAX_IMAGE_SIZE
self.nsfw_classifier = None
def _validate_and_process_images(
self, img: list[list[Image.Image]] | list[Image.Image]
) -> list[list[Image.Image]]:
# Simple validation: ensure it's a list of PIL images or list of lists of PIL images
if not img:
return []
# Check if it's a list of lists or a list of images
if isinstance(img[0], Image.Image):
# It's a list of images, convert to list of lists
img = [[im] for im in img]
# potentially concatenate multiple images to reduce the size
img = [[concatenate_images(img_i)] if len(img_i) > 1 else img_i for img_i in img]
# cap the pixels
img = [[cap_pixels(img_i, self.upsampling_max_image_size) for img_i in img_i] for img_i in img]
return img
def format_input(
self,
txt: list[str],
system_message: str = SYSTEM_MESSAGE,
img: list[Image.Image] | list[list[Image.Image]] | None = None,
) -> list[list[dict]]:
"""
Format a batch of text prompts into the conversation format expected by apply_chat_template.
Optionally, add images to the input.
Args:
txt: List of text prompts
system_message: System message to use (default: CREATIVE_SYSTEM_MESSAGE)
img: List of images to add to the input.
Returns:
List of conversations, where each conversation is a list of message dicts
"""
# Remove [IMG] tokens from prompts to avoid Pixtral validation issues
# when truncation is enabled. The processor counts [IMG] tokens and fails
# if the count changes after truncation.
cleaned_txt = [prompt.replace("[IMG]", "") for prompt in txt]
if img is None or len(img) == 0:
return [
[
{
"role": "system",
"content": [{"type": "text", "text": system_message}],
},
{"role": "user", "content": [{"type": "text", "text": prompt}]},
]
for prompt in cleaned_txt
]
else:
assert len(img) == len(txt), "Number of images must match number of prompts"
img = self._validate_and_process_images(img)
messages = [
[
{
"role": "system",
"content": [{"type": "text", "text": system_message}],
},
]
for _ in cleaned_txt
]
for i, (el, images) in enumerate(zip(messages, img)):
# optionally add the images per batch element.
if images is not None:
el.append(
{
"role": "user",
"content": [{"type": "image", "image": image_obj} for image_obj in images],
}
)
# add the text.
el.append(
{
"role": "user",
"content": [{"type": "text", "text": cleaned_txt[i]}],
}
)
return messages
@torch.no_grad()
def upsample_prompt(
self,
txt: list[str],
img: list[Image.Image] | list[list[Image.Image]] | None = None,
temperature: float = 0.15,
) -> list[str]:
"""
Upsample prompts using the model's generate method.
Args:
txt: List of input prompts to upsample
img: Optional list of images or list of lists of images. If None or all None, uses t2i mode, otherwise i2i mode.
Returns:
List of upsampled prompts
"""
# Set system message based on whether images are provided
if img is None or len(img) == 0 or img[0] is None:
system_message = SYSTEM_MESSAGE_UPSAMPLING_T2I
else:
system_message = SYSTEM_MESSAGE_UPSAMPLING_I2I
# Format input messages
messages_batch = self.format_input(txt=txt, system_message=system_message, img=img)
# Process all messages at once
# with image processing a too short max length can throw an error in here.
try:
inputs = self.processor.apply_chat_template(
messages_batch,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=2048,
)
except ValueError as e:
print(
f"Error processing input: {e}, your max length is probably too short, when you have images in the input."
)
raise e
# Move to device
inputs["input_ids"] = inputs["input_ids"].to(self.model.device)
inputs["attention_mask"] = inputs["attention_mask"].to(self.model.device)
if "pixel_values" in inputs:
inputs["pixel_values"] = inputs["pixel_values"].to(self.model.device, self.model.dtype)
# Generate text using the model's generate method
try:
generated_ids = self.model.generate(
**inputs,
max_new_tokens=512,
do_sample=True,
temperature=temperature,
use_cache=True,
)
# Decode only the newly generated tokens (skip input tokens)
# Extract only the generated portion
input_length = inputs["input_ids"].shape[1]
generated_tokens = generated_ids[:, input_length:]
raw_txt = self.processor.tokenizer.batch_decode(
generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
return raw_txt
except Exception as e:
print(f"Error generating upsampled prompt: {e}, returning original prompt")
return txt
@torch.no_grad()
def forward(self, txt: list[str]):
# Format input messages
messages_batch = self.format_input(txt=txt)
# Process all messages at once
# with image processing a too short max length can throw an error in here.
inputs = self.processor.apply_chat_template(
messages_batch,
add_generation_prompt=False,
tokenize=True,
return_dict=True,
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=self.max_length,
)
# Move to device
input_ids = inputs["input_ids"].to(self.model.device)
attention_mask = inputs["attention_mask"].to(self.model.device)
# Forward pass through the model
output = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
output_hidden_states=True,
use_cache=False,
)
out = torch.stack([output.hidden_states[k] for k in OUTPUT_LAYERS], dim=1)
return rearrange(out, "b c l d -> b l (c d)")
def yes_no_logit_processor(
self, input_ids: torch.LongTensor, scores: torch.FloatTensor
) -> torch.FloatTensor:
"""
Sets all tokens but yes/no to the minimum.
"""
scores_yes_token = scores[:, self.yes_token].clone()
scores_no_token = scores[:, self.no_token].clone()
scores_min = scores.min()
scores[:, :] = scores_min - 1
scores[:, self.yes_token] = scores_yes_token
scores[:, self.no_token] = scores_no_token
return scores
def test_image(self, image: Image.Image | str | Path | torch.Tensor) -> bool:
if isinstance(image, torch.Tensor):
image = rearrange(image[0].clamp(-1.0, 1.0), "c h w -> h w c")
image = Image.fromarray((127.5 * (image + 1.0)).cpu().byte().numpy())
elif isinstance(image, (str, Path)):
image = Image.open(image)
classification = next(c for c in self.nsfw_classifier(image) if c["label"] == "nsfw")
if classification["score"] > NSFW_THRESHOLD:
return True
# 512^2 pixels are enough for checking
w, h = image.size
f = (512**2 / (w * h)) ** 0.5
image = image.resize((int(f * w), int(f * h)))
chat = [
{
"role": "system",
"content": [
{
"type": "text",
"text": SYSTEM_PROMPT_CONTENT_FILTER,
},
],
},
{
"role": "user",
"content": [
{
"type": "text",
"text": PROMPT_IMAGE_INTEGRITY,
},
{
"type": "image",
"image": image,
},
{
"type": "text",
"text": PROMPT_IMAGE_INTEGRITY_FOLLOW_UP,
},
],
},
]
inputs = self.processor.apply_chat_template(
chat,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(self.model.device)
inputs["pixel_values"] = inputs["pixel_values"].to(dtype=self.model.dtype)
generate_ids = self.model.generate(
**inputs,
max_new_tokens=1,
logits_processor=[self.yes_no_logit_processor],
do_sample=False,
)
return generate_ids[0, -1].item() == self.yes_token
def test_txt(self, txt: str) -> bool:
chat = [
{
"role": "system",
"content": [
{
"type": "text",
"text": SYSTEM_PROMPT_CONTENT_FILTER,
},
],
},
{
"role": "user",
"content": [
{
"type": "text",
"text": PROMPT_TEXT_INTEGRITY.format(prompt=txt),
},
],
},
]
inputs = self.processor.apply_chat_template(
chat,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(self.model.device)
generate_ids = self.model.generate(
**inputs,
max_new_tokens=1,
logits_processor=[self.yes_no_logit_processor],
do_sample=False,
)
return generate_ids[0, -1].item() == self.yes_token
|