Spaces:
Running
on
Zero
Running
on
Zero
Vladyslav Humennyy
Claude
commited on
Commit
·
a113c8a
1
Parent(s):
9203469
Rewrite image handling to match app_chat_vllm.py format
Browse files- User function now converts images to base64 with image_url format
- Removed complex unused helper functions for message processing
- Bot function properly handles base64 images with processor
- Converts base64 back to PIL images when using processor
- Falls back to tokenizer for text-only messages
- Simplified and cleaner implementation matching app_chat_vllm.py
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <[email protected]>
app.py
CHANGED
|
@@ -60,44 +60,40 @@ def load_model():
|
|
| 60 |
model, tokenizer, processor, device = load_model()
|
| 61 |
|
| 62 |
|
| 63 |
-
def _ensure_image_path(image_data: Any) -> str | None:
|
| 64 |
-
"""Return a valid file path for the provided image data."""
|
| 65 |
-
if image_data is None:
|
| 66 |
-
return None
|
| 67 |
-
|
| 68 |
-
try:
|
| 69 |
-
from PIL import Image
|
| 70 |
-
except ImportError: # pragma: no cover - PIL is bundled with Gradio's image component
|
| 71 |
-
return None
|
| 72 |
-
|
| 73 |
-
# Already a path string
|
| 74 |
-
if isinstance(image_data, str) and os.path.exists(image_data):
|
| 75 |
-
return image_data
|
| 76 |
-
|
| 77 |
-
# PIL Image object - save to temp file
|
| 78 |
-
if isinstance(image_data, Image.Image):
|
| 79 |
-
fd, tmp_path = tempfile.mkstemp(suffix=".png")
|
| 80 |
-
os.close(fd)
|
| 81 |
-
image_data.save(tmp_path, format="PNG")
|
| 82 |
-
return tmp_path
|
| 83 |
-
|
| 84 |
-
return None
|
| 85 |
-
|
| 86 |
-
|
| 87 |
def user(user_message, image_data, history: list):
|
| 88 |
-
|
|
|
|
|
|
|
|
|
|
| 89 |
|
|
|
|
| 90 |
updated_history = list(history)
|
| 91 |
has_content = False
|
| 92 |
|
| 93 |
stripped_message = user_message.strip()
|
| 94 |
-
if stripped_message:
|
| 95 |
-
updated_history.append({"role": "user", "content": stripped_message})
|
| 96 |
-
has_content = True
|
| 97 |
|
| 98 |
-
|
| 99 |
-
if
|
| 100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
has_content = True
|
| 102 |
|
| 103 |
if not has_content:
|
|
@@ -117,257 +113,116 @@ def append_example_message(x: gr.SelectData, history):
|
|
| 117 |
return history
|
| 118 |
|
| 119 |
|
| 120 |
-
def
|
| 121 |
-
content
|
| 122 |
-
if isinstance(content,
|
| 123 |
-
|
| 124 |
-
return True
|
| 125 |
-
if content.get("type") in {"image", "image_url"}:
|
| 126 |
-
return True
|
| 127 |
if isinstance(content, list):
|
|
|
|
| 128 |
for item in content:
|
| 129 |
-
if isinstance(item, dict) and item.get("type")
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
def _content_to_text(content: Any) -> str:
|
| 135 |
-
if isinstance(content, dict):
|
| 136 |
-
if "text" in content:
|
| 137 |
-
return content.get("text", "")
|
| 138 |
-
if "path" in content:
|
| 139 |
-
alt_text = content.get("alt_text")
|
| 140 |
-
placeholder = alt_text or os.path.basename(content["path"]) or "image"
|
| 141 |
-
return f"[image: {placeholder}]"
|
| 142 |
-
if "image" in content:
|
| 143 |
-
return "[image]"
|
| 144 |
-
if content.get("type") == "image_url":
|
| 145 |
-
image_url = content.get("image_url")
|
| 146 |
-
if isinstance(image_url, dict):
|
| 147 |
-
image_url = image_url.get("url", "")
|
| 148 |
-
return f"[image: {image_url}]"
|
| 149 |
-
if content.get("type") == "text":
|
| 150 |
-
return content.get("text", "")
|
| 151 |
-
return str(content)
|
| 152 |
-
if isinstance(content, list):
|
| 153 |
-
text_parts: list[str] = []
|
| 154 |
-
for item in content:
|
| 155 |
-
if isinstance(item, dict):
|
| 156 |
-
item_type = item.get("type")
|
| 157 |
-
if item_type == "text":
|
| 158 |
-
text_parts.append(item.get("text", ""))
|
| 159 |
-
elif item_type == "image":
|
| 160 |
-
text_parts.append("[image]")
|
| 161 |
-
elif item_type == "image_url":
|
| 162 |
-
image_url = item.get("image_url")
|
| 163 |
-
if isinstance(image_url, dict):
|
| 164 |
-
image_url = image_url.get("url", "")
|
| 165 |
-
text_parts.append(f"[image: {image_url}]")
|
| 166 |
-
else:
|
| 167 |
-
text_parts.append(str(item))
|
| 168 |
-
else:
|
| 169 |
-
text_parts.append(str(item))
|
| 170 |
-
filtered = [part for part in text_parts if part]
|
| 171 |
-
return "\n".join(filtered) if filtered else "[image]"
|
| 172 |
return str(content)
|
| 173 |
|
| 174 |
|
| 175 |
-
def _collect_recent_user_contents(history: list[dict[str, Any]]) -> list[Any]:
|
| 176 |
-
"""Collect the trailing sequence of user messages prior to the assistant reply."""
|
| 177 |
-
chunks: list[Any] = []
|
| 178 |
-
for message in reversed(history):
|
| 179 |
-
if message.get("role") != "user":
|
| 180 |
-
break
|
| 181 |
-
chunks.append(message.get("content"))
|
| 182 |
-
chunks.reverse()
|
| 183 |
-
return chunks
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
def _prepare_text_history(history: list[dict[str, Any]]) -> list[dict[str, str]]:
|
| 187 |
-
text_history: list[dict[str, str]] = []
|
| 188 |
-
for message in history:
|
| 189 |
-
role = message.get("role", "user")
|
| 190 |
-
content_text = _content_to_text(message.get("content"))
|
| 191 |
-
if not content_text:
|
| 192 |
-
continue
|
| 193 |
-
if text_history and text_history[-1]["role"] == role:
|
| 194 |
-
text_history[-1]["content"] = text_history[-1]["content"] + "\n" + content_text
|
| 195 |
-
else:
|
| 196 |
-
text_history.append({"role": role, "content": content_text})
|
| 197 |
-
return text_history
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
def _prepare_processor_history(history: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
| 201 |
-
"""Prepare history for processor with proper image format."""
|
| 202 |
-
processor_history = []
|
| 203 |
-
|
| 204 |
-
for message in history:
|
| 205 |
-
role = message.get("role", "user")
|
| 206 |
-
content = message.get("content")
|
| 207 |
-
|
| 208 |
-
# Handle different content formats
|
| 209 |
-
if isinstance(content, str):
|
| 210 |
-
# Simple text message
|
| 211 |
-
processor_history.append({"role": role, "content": content})
|
| 212 |
-
elif isinstance(content, list):
|
| 213 |
-
# Multi-modal content (text + images)
|
| 214 |
-
formatted_content = []
|
| 215 |
-
for item in content:
|
| 216 |
-
if isinstance(item, dict):
|
| 217 |
-
item_type = item.get("type")
|
| 218 |
-
if item_type == "text":
|
| 219 |
-
formatted_content.append({"type": "text", "text": item.get("text", "")})
|
| 220 |
-
elif item_type == "image":
|
| 221 |
-
# Extract PIL Image from _pil_image field or load from path
|
| 222 |
-
pil_image = item.get("_pil_image")
|
| 223 |
-
if pil_image is None and "path" in item:
|
| 224 |
-
from PIL import Image
|
| 225 |
-
pil_image = Image.open(item["path"])
|
| 226 |
-
if pil_image is not None:
|
| 227 |
-
formatted_content.append({"type": "image", "image": pil_image})
|
| 228 |
-
if formatted_content:
|
| 229 |
-
processor_history.append({"role": role, "content": formatted_content})
|
| 230 |
-
elif isinstance(content, dict):
|
| 231 |
-
# Legacy format or single image
|
| 232 |
-
if "image" in content or "_pil_image" in content:
|
| 233 |
-
pil_image = content.get("_pil_image") or content.get("image")
|
| 234 |
-
if pil_image is None and "path" in content:
|
| 235 |
-
from PIL import Image
|
| 236 |
-
pil_image = Image.open(content["path"])
|
| 237 |
-
if pil_image is not None:
|
| 238 |
-
processor_history.append({
|
| 239 |
-
"role": role,
|
| 240 |
-
"content": [{"type": "image", "image": pil_image}]
|
| 241 |
-
})
|
| 242 |
-
else:
|
| 243 |
-
# Try to extract text
|
| 244 |
-
text = _content_to_text(content)
|
| 245 |
-
if text:
|
| 246 |
-
processor_history.append({"role": role, "content": text})
|
| 247 |
-
|
| 248 |
-
return processor_history
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
def _clean_history_for_display(history: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
| 252 |
-
"""Remove internal metadata fields like _pil_image before displaying in Gradio."""
|
| 253 |
-
cleaned = []
|
| 254 |
-
|
| 255 |
-
for message in history:
|
| 256 |
-
cleaned_message = {"role": message.get("role", "user")}
|
| 257 |
-
content = message.get("content")
|
| 258 |
-
|
| 259 |
-
if isinstance(content, str):
|
| 260 |
-
cleaned_message["content"] = content
|
| 261 |
-
elif isinstance(content, list):
|
| 262 |
-
cleaned_content = []
|
| 263 |
-
for item in content:
|
| 264 |
-
if isinstance(item, dict):
|
| 265 |
-
# Remove _pil_image and ensure alt_text is string or absent
|
| 266 |
-
cleaned_item = {}
|
| 267 |
-
for k, v in item.items():
|
| 268 |
-
if k == "_pil_image":
|
| 269 |
-
continue
|
| 270 |
-
if k == "alt_text":
|
| 271 |
-
# Ensure alt_text is a string
|
| 272 |
-
if isinstance(v, str):
|
| 273 |
-
cleaned_item[k] = v
|
| 274 |
-
# Skip non-string alt_text values
|
| 275 |
-
continue
|
| 276 |
-
cleaned_item[k] = v
|
| 277 |
-
# Ensure alt_text exists for image type
|
| 278 |
-
if cleaned_item.get("type") == "image" and "alt_text" not in cleaned_item:
|
| 279 |
-
cleaned_item["alt_text"] = "uploaded image"
|
| 280 |
-
cleaned_content.append(cleaned_item)
|
| 281 |
-
else:
|
| 282 |
-
cleaned_content.append(item)
|
| 283 |
-
cleaned_message["content"] = cleaned_content
|
| 284 |
-
elif isinstance(content, dict):
|
| 285 |
-
# Remove _pil_image and ensure alt_text is string or absent
|
| 286 |
-
cleaned_item = {}
|
| 287 |
-
for k, v in content.items():
|
| 288 |
-
if k == "_pil_image":
|
| 289 |
-
continue
|
| 290 |
-
if k == "alt_text":
|
| 291 |
-
# Ensure alt_text is a string
|
| 292 |
-
if isinstance(v, str):
|
| 293 |
-
cleaned_item[k] = v
|
| 294 |
-
# Skip non-string alt_text values
|
| 295 |
-
continue
|
| 296 |
-
cleaned_item[k] = v
|
| 297 |
-
# Ensure alt_text exists for image content
|
| 298 |
-
if "path" in cleaned_item and "alt_text" not in cleaned_item:
|
| 299 |
-
cleaned_item["alt_text"] = "uploaded image"
|
| 300 |
-
cleaned_message["content"] = cleaned_item
|
| 301 |
-
else:
|
| 302 |
-
cleaned_message["content"] = content
|
| 303 |
-
|
| 304 |
-
cleaned.append(cleaned_message)
|
| 305 |
-
|
| 306 |
-
return cleaned
|
| 307 |
-
|
| 308 |
-
|
| 309 |
@spaces.GPU
|
| 310 |
def bot(
|
| 311 |
history: list[dict[str, Any]]
|
| 312 |
-
# max_tokens,
|
| 313 |
-
# temperature,
|
| 314 |
-
# top_p,
|
| 315 |
):
|
| 316 |
-
|
| 317 |
-
if not user_chunks:
|
| 318 |
-
user_message_text = ""
|
| 319 |
-
else:
|
| 320 |
-
user_message_text = "\n".join(filter(None, (_content_to_text(chunk) for chunk in user_chunks)))
|
| 321 |
-
print('User message:', user_message_text)
|
| 322 |
-
# [{"role": "system", "content": system_message}] +
|
| 323 |
-
# Build conversation
|
| 324 |
max_tokens = 4096
|
| 325 |
temperature = 0.7
|
| 326 |
top_p = 0.95
|
| 327 |
|
| 328 |
-
text_history = _prepare_text_history(history)
|
| 329 |
-
|
| 330 |
-
# Handle empty history case
|
| 331 |
-
if not text_history:
|
| 332 |
-
input_text = ""
|
| 333 |
-
else:
|
| 334 |
-
input_text: str = tokenizer.apply_chat_template(
|
| 335 |
-
text_history,
|
| 336 |
-
tokenize=False,
|
| 337 |
-
add_generation_prompt=True,
|
| 338 |
-
# enable_thinking=True,
|
| 339 |
-
)
|
| 340 |
-
|
| 341 |
-
if input_text and tokenizer.bos_token:
|
| 342 |
-
input_text = input_text.replace(tokenizer.bos_token, "", 1)
|
| 343 |
-
print(input_text)
|
| 344 |
-
model_inputs = None
|
| 345 |
-
|
| 346 |
# Early return if no input
|
| 347 |
-
if not
|
| 348 |
return
|
| 349 |
|
| 350 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 351 |
try:
|
| 352 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 353 |
model_inputs = processor(
|
| 354 |
messages=processor_history,
|
| 355 |
return_tensors="pt",
|
| 356 |
add_generation_prompt=True,
|
| 357 |
).to(model.device)
|
| 358 |
-
|
| 359 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 360 |
|
| 361 |
if model_inputs is None:
|
| 362 |
-
|
| 363 |
|
| 364 |
-
decoded_input = tokenizer.decode(model_inputs["input_ids"][0])
|
| 365 |
-
print("Decoded input:", decoded_input)
|
| 366 |
-
print([{int(token_id.item()): tokenizer.decode([int(token_id.item())])} for token_id in model_inputs["input_ids"][0]])
|
| 367 |
# Streamer setup
|
| 368 |
-
streamer = TextIteratorStreamer(
|
| 369 |
-
tokenizer, skip_prompt=True # skip_special_tokens=True # ,
|
| 370 |
-
)
|
| 371 |
|
| 372 |
# Run model.generate in background thread
|
| 373 |
generation_kwargs = dict(
|
|
@@ -377,7 +232,6 @@ def bot(
|
|
| 377 |
top_p=top_p,
|
| 378 |
top_k=64,
|
| 379 |
do_sample=True,
|
| 380 |
-
# eos_token_id=tokenizer.eos_token_id,
|
| 381 |
streamer=streamer,
|
| 382 |
)
|
| 383 |
thread = threading.Thread(target=model.generate, kwargs=generation_kwargs)
|
|
@@ -387,7 +241,7 @@ def bot(
|
|
| 387 |
# Yield tokens as they come in
|
| 388 |
for new_text in streamer:
|
| 389 |
history[-1]["content"] += new_text
|
| 390 |
-
yield
|
| 391 |
|
| 392 |
assistant_message = history[-1]["content"]
|
| 393 |
logger.log_interaction(user=user_message_text, answer=assistant_message)
|
|
|
|
| 60 |
model, tokenizer, processor, device = load_model()
|
| 61 |
|
| 62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
def user(user_message, image_data, history: list):
|
| 64 |
+
"""Format user message with optional image (like app_chat_vllm.py)."""
|
| 65 |
+
import base64
|
| 66 |
+
import io
|
| 67 |
+
from PIL import Image
|
| 68 |
|
| 69 |
+
user_message = user_message or ""
|
| 70 |
updated_history = list(history)
|
| 71 |
has_content = False
|
| 72 |
|
| 73 |
stripped_message = user_message.strip()
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
+
# Format message with image in base64 format (matching app_chat_vllm.py)
|
| 76 |
+
if image_data is not None:
|
| 77 |
+
# Convert PIL image to base64
|
| 78 |
+
buffered = io.BytesIO()
|
| 79 |
+
image_data.save(buffered, format="JPEG")
|
| 80 |
+
img_base64 = base64.b64encode(buffered.getvalue()).decode()
|
| 81 |
+
|
| 82 |
+
text_content = stripped_message if stripped_message else "Describe this image"
|
| 83 |
+
|
| 84 |
+
updated_history.append({
|
| 85 |
+
"role": "user",
|
| 86 |
+
"content": [
|
| 87 |
+
{"type": "text", "text": text_content},
|
| 88 |
+
{
|
| 89 |
+
"type": "image_url",
|
| 90 |
+
"image_url": {"url": f"data:image/jpeg;base64,{img_base64}"},
|
| 91 |
+
},
|
| 92 |
+
],
|
| 93 |
+
})
|
| 94 |
+
has_content = True
|
| 95 |
+
elif stripped_message:
|
| 96 |
+
updated_history.append({"role": "user", "content": stripped_message})
|
| 97 |
has_content = True
|
| 98 |
|
| 99 |
if not has_content:
|
|
|
|
| 113 |
return history
|
| 114 |
|
| 115 |
|
| 116 |
+
def _extract_text_from_content(content: Any) -> str:
|
| 117 |
+
"""Extract text from message content for logging."""
|
| 118 |
+
if isinstance(content, str):
|
| 119 |
+
return content
|
|
|
|
|
|
|
|
|
|
| 120 |
if isinstance(content, list):
|
| 121 |
+
text_parts = []
|
| 122 |
for item in content:
|
| 123 |
+
if isinstance(item, dict) and item.get("type") == "text":
|
| 124 |
+
text_parts.append(item.get("text", ""))
|
| 125 |
+
return " ".join(text_parts) if text_parts else "[Image]"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
return str(content)
|
| 127 |
|
| 128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
@spaces.GPU
|
| 130 |
def bot(
|
| 131 |
history: list[dict[str, Any]]
|
|
|
|
|
|
|
|
|
|
| 132 |
):
|
| 133 |
+
"""Generate bot response with support for text and images."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
max_tokens = 4096
|
| 135 |
temperature = 0.7
|
| 136 |
top_p = 0.95
|
| 137 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
# Early return if no input
|
| 139 |
+
if not history:
|
| 140 |
return
|
| 141 |
|
| 142 |
+
# Extract last user message for logging
|
| 143 |
+
last_user_msg = next((msg for msg in reversed(history) if msg.get("role") == "user"), None)
|
| 144 |
+
user_message_text = _extract_text_from_content(last_user_msg.get("content")) if last_user_msg else ""
|
| 145 |
+
print('User message:', user_message_text)
|
| 146 |
+
|
| 147 |
+
# Check if any message contains images
|
| 148 |
+
has_images = any(
|
| 149 |
+
isinstance(msg.get("content"), list) and
|
| 150 |
+
any(item.get("type") == "image_url" for item in msg.get("content") if isinstance(item, dict))
|
| 151 |
+
for msg in history
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
model_inputs = None
|
| 155 |
+
|
| 156 |
+
# Use processor if images are present
|
| 157 |
+
if processor is not None and has_images:
|
| 158 |
try:
|
| 159 |
+
# Processor expects messages with PIL images, not base64
|
| 160 |
+
# We need to convert base64 back to PIL for the processor
|
| 161 |
+
from PIL import Image
|
| 162 |
+
import base64
|
| 163 |
+
import io
|
| 164 |
+
|
| 165 |
+
processor_history = []
|
| 166 |
+
for msg in history:
|
| 167 |
+
role = msg.get("role", "user")
|
| 168 |
+
content = msg.get("content")
|
| 169 |
+
|
| 170 |
+
if isinstance(content, str):
|
| 171 |
+
processor_history.append({"role": role, "content": content})
|
| 172 |
+
elif isinstance(content, list):
|
| 173 |
+
formatted_content = []
|
| 174 |
+
for item in content:
|
| 175 |
+
if isinstance(item, dict):
|
| 176 |
+
if item.get("type") == "text":
|
| 177 |
+
formatted_content.append({"type": "text", "text": item.get("text", "")})
|
| 178 |
+
elif item.get("type") == "image_url":
|
| 179 |
+
# Extract base64 and convert to PIL
|
| 180 |
+
img_url = item.get("image_url", {}).get("url", "")
|
| 181 |
+
if img_url.startswith("data:image"):
|
| 182 |
+
base64_data = img_url.split(",")[1]
|
| 183 |
+
img_data = base64.b64decode(base64_data)
|
| 184 |
+
pil_image = Image.open(io.BytesIO(img_data))
|
| 185 |
+
formatted_content.append({"type": "image", "image": pil_image})
|
| 186 |
+
if formatted_content:
|
| 187 |
+
processor_history.append({"role": role, "content": formatted_content})
|
| 188 |
+
|
| 189 |
model_inputs = processor(
|
| 190 |
messages=processor_history,
|
| 191 |
return_tensors="pt",
|
| 192 |
add_generation_prompt=True,
|
| 193 |
).to(model.device)
|
| 194 |
+
print("Using processor for vision input")
|
| 195 |
+
except Exception as exc:
|
| 196 |
+
print(f"Processor failed: {exc}")
|
| 197 |
+
model_inputs = None
|
| 198 |
+
|
| 199 |
+
# Fallback to tokenizer for text-only
|
| 200 |
+
if model_inputs is None:
|
| 201 |
+
# Convert to text-only format for tokenizer
|
| 202 |
+
text_history = []
|
| 203 |
+
for msg in history:
|
| 204 |
+
role = msg.get("role", "user")
|
| 205 |
+
content = msg.get("content")
|
| 206 |
+
text_content = _extract_text_from_content(content)
|
| 207 |
+
if text_content:
|
| 208 |
+
text_history.append({"role": role, "content": text_content})
|
| 209 |
+
|
| 210 |
+
if text_history:
|
| 211 |
+
input_text = tokenizer.apply_chat_template(
|
| 212 |
+
text_history,
|
| 213 |
+
tokenize=False,
|
| 214 |
+
add_generation_prompt=True,
|
| 215 |
+
)
|
| 216 |
+
if input_text and tokenizer.bos_token:
|
| 217 |
+
input_text = input_text.replace(tokenizer.bos_token, "", 1)
|
| 218 |
+
model_inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
|
| 219 |
+
print("Using tokenizer for text-only input")
|
| 220 |
|
| 221 |
if model_inputs is None:
|
| 222 |
+
return
|
| 223 |
|
|
|
|
|
|
|
|
|
|
| 224 |
# Streamer setup
|
| 225 |
+
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True)
|
|
|
|
|
|
|
| 226 |
|
| 227 |
# Run model.generate in background thread
|
| 228 |
generation_kwargs = dict(
|
|
|
|
| 232 |
top_p=top_p,
|
| 233 |
top_k=64,
|
| 234 |
do_sample=True,
|
|
|
|
| 235 |
streamer=streamer,
|
| 236 |
)
|
| 237 |
thread = threading.Thread(target=model.generate, kwargs=generation_kwargs)
|
|
|
|
| 241 |
# Yield tokens as they come in
|
| 242 |
for new_text in streamer:
|
| 243 |
history[-1]["content"] += new_text
|
| 244 |
+
yield history
|
| 245 |
|
| 246 |
assistant_message = history[-1]["content"]
|
| 247 |
logger.log_interaction(user=user_message_text, answer=assistant_message)
|