File size: 14,435 Bytes
1bcbdb8 41881bf 1bcbdb8 41881bf 1bcbdb8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 | from __future__ import annotations
import json
import os
from pathlib import Path
from threading import Lock
import gradio as gr
import torch
from huggingface_hub import snapshot_download
from transformers import AutoModelForMultimodalLM, AutoProcessor
MODEL_TITLE = "LumynaX Infused Gemma E4B Model"
DEFAULT_MODEL_REPO_ID = "AbteeXAILab/lumynax-infused-gemma-e4b"
MODEL_REPO_ENV_VAR = "LUMYNAX_MODEL_REPO_ID"
HF_TOKEN_ENV_VARS = ("HF_TOKEN", "HUGGING_FACE_HUB_TOKEN", "HUGGINGFACE_HUB_TOKEN")
DEFAULT_IMAGE_URL = "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/GoldenGate.png"
DEFAULT_AUDIO_URL = "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/journal1.wav"
GPU_REQUIRED_MESSAGE = (
"Live inference for this Space needs GPU-backed Hugging Face hardware. "
"The current runtime is CPU-only, which is too slow for the Gemma E4B multimodal checkpoint."
)
SHOWCASE_MESSAGE = (
"This Space is running in showcase mode on CPU hardware. "
"The examples below were captured during package validation so people can still see how the model behaves. "
"If GPU hardware is attached later, this same Space will switch back to live inference automatically."
)
SHOWCASE_SAMPLES = {
"text": {
"prompt": "Who are you? Reply in one short sentence.",
"response": "I am LumynaX, operating from the LumynaX Infused Gemma E4B Model package.",
"parsed_output": {
"role": "assistant",
"content": "I am LumynaX, operating from the LumynaX Infused Gemma E4B Model package.",
},
},
"image": {
"prompt": "What is shown in this image? Reply in under 12 words.",
"response": "The iconic Golden Gate Bridge spans the water under a clear sky. I am LumynaX.",
"parsed_output": {
"role": "assistant",
"content": "The iconic Golden Gate Bridge spans the water under a clear sky. I am LumynaX.",
},
},
"audio": {
"prompt": "Transcribe the speech in one line only.",
"response": 'A local validation run transcribed the bundled sample audio and included: "My name is LumynaX."',
"parsed_output": {
"validation_summary": 'A local validation run transcribed the bundled sample audio and included: "My name is LumynaX."',
},
},
"reasoning": {
"prompt": "Explain what this package is in one short sentence.",
"response": "Reasoning mode was verified locally and returned a non-empty structured thinking field.",
"parsed_output": {
"validation_summary": "Reasoning mode was verified locally and returned a non-empty structured thinking field.",
},
},
}
_MODEL = None
_PROCESSOR = None
_LOAD_ERROR = None
_LOAD_LOCK = Lock()
def _resolve_hf_token() -> str | None:
for env_var in HF_TOKEN_ENV_VARS:
raw_value = os.environ.get(env_var, "").strip()
if raw_value:
return raw_value
return None
def _has_supported_gpu_runtime() -> bool:
return bool(torch.cuda.is_available())
def _load_runtime() -> tuple[object, object]:
global _MODEL, _PROCESSOR, _LOAD_ERROR
if _MODEL is not None and _PROCESSOR is not None:
return _MODEL, _PROCESSOR
if _LOAD_ERROR is not None:
raise RuntimeError(_LOAD_ERROR)
with _LOAD_LOCK:
if _MODEL is not None and _PROCESSOR is not None:
return _MODEL, _PROCESSOR
if _LOAD_ERROR is not None:
raise RuntimeError(_LOAD_ERROR)
try:
if not _has_supported_gpu_runtime():
raise RuntimeError(GPU_REQUIRED_MESSAGE)
repo_id = os.environ.get(MODEL_REPO_ENV_VAR, "").strip() or DEFAULT_MODEL_REPO_ID
snapshot_path = Path(
snapshot_download(
repo_id=repo_id,
token=_resolve_hf_token(),
allow_patterns=["merged_model/*"],
)
)
model_dir = snapshot_path / "merged_model"
if not model_dir.exists():
raise FileNotFoundError(f"Expected merged_model/ in {snapshot_path} after downloading {repo_id}.")
processor = AutoProcessor.from_pretrained(str(model_dir))
model = AutoModelForMultimodalLM.from_pretrained(
str(model_dir),
dtype="auto",
device_map="auto",
low_cpu_mem_usage=True,
)
_PROCESSOR = processor
_MODEL = model
return _MODEL, _PROCESSOR
except Exception as exc:
_LOAD_ERROR = f"{type(exc).__name__}: {exc}"
raise
def _resolve_media_reference(upload_value: str | None, url_value: str | None) -> str | None:
if isinstance(url_value, str) and url_value.strip():
return url_value.strip()
if isinstance(upload_value, str) and upload_value.strip():
return upload_value.strip()
return None
def _extract_response_text(parsed: object) -> str:
if isinstance(parsed, dict):
content = parsed.get("content")
if isinstance(content, str) and content.strip():
return content.strip()
if isinstance(parsed, str):
return parsed.strip()
return json.dumps(parsed, indent=2, ensure_ascii=False, default=str)
def _format_json(value: object) -> str:
return json.dumps(value, indent=2, ensure_ascii=False, default=str)
def run_request(
*,
prompt: str,
thinking: bool,
max_new_tokens: int,
image_upload: str | None = None,
image_url: str = "",
audio_upload: str | None = None,
audio_url: str = "",
) -> tuple[str, str]:
if not prompt.strip():
raise gr.Error("A prompt is required.")
if not _has_supported_gpu_runtime():
return GPU_REQUIRED_MESSAGE, _format_json({"error": GPU_REQUIRED_MESSAGE})
image_ref = _resolve_media_reference(image_upload, image_url)
audio_ref = _resolve_media_reference(audio_upload, audio_url)
content: list[dict[str, str]] = []
if image_ref:
content.append({"type": "image", "url": image_ref})
if audio_ref:
content.append({"type": "audio", "audio": audio_ref})
content.append({"type": "text", "text": prompt.strip()})
messages = [
{
"role": "user",
"content": content,
},
]
model, processor = _load_runtime()
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
enable_thinking=thinking,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
with torch.inference_mode():
outputs = model.generate(
**inputs,
max_new_tokens=int(max_new_tokens),
do_sample=False,
)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
parsed = processor.parse_response(response) if hasattr(processor, "parse_response") else response
return _extract_response_text(parsed), _format_json(parsed)
def run_text(prompt: str, thinking: bool, max_new_tokens: int) -> tuple[str, str]:
return run_request(
prompt=prompt,
thinking=thinking,
max_new_tokens=max_new_tokens,
)
def run_image(
prompt: str,
image_upload: str | None,
image_url: str,
thinking: bool,
max_new_tokens: int,
) -> tuple[str, str]:
return run_request(
prompt=prompt,
thinking=thinking,
max_new_tokens=max_new_tokens,
image_upload=image_upload,
image_url=image_url,
)
def run_audio(
prompt: str,
audio_upload: str | None,
audio_url: str,
thinking: bool,
max_new_tokens: int,
) -> tuple[str, str]:
return run_request(
prompt=prompt,
thinking=thinking,
max_new_tokens=max_new_tokens,
audio_upload=audio_upload,
audio_url=audio_url,
)
def _render_showcase_sample(
*,
prompt: str,
response: str,
parsed_output: object,
media_markdown: str | None = None,
media_url: str | None = None,
) -> None:
if media_markdown:
gr.Markdown(media_markdown)
if media_url:
gr.Textbox(label="Sample Asset URL", value=media_url, interactive=False, lines=1)
gr.Textbox(label="Example Prompt", value=prompt, interactive=False, lines=3)
gr.Textbox(label="Example Response", value=response, interactive=False, lines=6)
gr.Code(label="Example Parsed Output", value=_format_json(parsed_output), language="json")
def _build_live_ui() -> None:
gr.Markdown(
f"# {MODEL_TITLE}\n\n"
"Live multimodal demo mode is active because GPU hardware is available. "
"The LumynaX identity comes from the packaged model template and is not user-editable here."
)
with gr.Tab("Text"):
text_prompt = gr.Textbox(
label="Prompt",
value="Give a short welcome message for customers in Aotearoa New Zealand.",
lines=4,
)
with gr.Row():
text_thinking = gr.Checkbox(label="Enable Reasoning", value=False)
text_max_tokens = gr.Slider(label="Max New Tokens", minimum=16, maximum=256, value=64, step=16)
text_run = gr.Button("Run Text Demo", variant="primary")
text_answer = gr.Textbox(label="Response", lines=8)
text_debug = gr.Code(label="Parsed Output", language="json")
text_run.click(
run_text,
inputs=[text_prompt, text_thinking, text_max_tokens],
outputs=[text_answer, text_debug],
)
with gr.Tab("Image"):
image_prompt = gr.Textbox(
label="Prompt",
value="What is shown in this image? Reply in under 12 words.",
lines=3,
)
image_upload = gr.Image(label="Upload Image", type="filepath")
image_url = gr.Textbox(label="Or Image URL", value=DEFAULT_IMAGE_URL)
with gr.Row():
image_thinking = gr.Checkbox(label="Enable Reasoning", value=False)
image_max_tokens = gr.Slider(label="Max New Tokens", minimum=16, maximum=256, value=64, step=16)
image_run = gr.Button("Run Image Demo", variant="primary")
image_answer = gr.Textbox(label="Response", lines=8)
image_debug = gr.Code(label="Parsed Output", language="json")
image_run.click(
run_image,
inputs=[image_prompt, image_upload, image_url, image_thinking, image_max_tokens],
outputs=[image_answer, image_debug],
)
with gr.Tab("Audio"):
audio_prompt = gr.Textbox(
label="Prompt",
value="Transcribe the speech in one line only.",
lines=3,
)
audio_upload = gr.Audio(label="Upload Audio", type="filepath")
audio_url = gr.Textbox(label="Or Audio URL", value=DEFAULT_AUDIO_URL)
with gr.Row():
audio_thinking = gr.Checkbox(label="Enable Reasoning", value=False)
audio_max_tokens = gr.Slider(label="Max New Tokens", minimum=16, maximum=256, value=64, step=16)
audio_run = gr.Button("Run Audio Demo", variant="primary")
audio_answer = gr.Textbox(label="Response", lines=8)
audio_debug = gr.Code(label="Parsed Output", language="json")
audio_run.click(
run_audio,
inputs=[audio_prompt, audio_upload, audio_url, audio_thinking, audio_max_tokens],
outputs=[audio_answer, audio_debug],
)
def _build_showcase_ui() -> None:
gr.Markdown(
f"# {MODEL_TITLE}\n\n"
f"{SHOWCASE_MESSAGE}\n\n"
"This is still the real package identity and real package structure, but not live inference on this CPU-only Space."
)
with gr.Tab("Overview"):
gr.Markdown(
"### What this Space is showing\n"
"- verified text, image, audio, and reasoning examples from package validation\n"
"- the real packaged Gemma E4B release structure and LumynaX identity behavior\n"
"- honest provenance: packaged upstream Gemma weights under a LumynaX runtime identity\n\n"
"### Why this is showcase mode\n"
"- Hugging Face `cpu-basic` cannot serve this checkpoint interactively\n"
"- the same Space will switch to live inference automatically if GPU hardware is added later"
)
with gr.Tab("Text Sample"):
sample = SHOWCASE_SAMPLES["text"]
_render_showcase_sample(
prompt=sample["prompt"],
response=sample["response"],
parsed_output=sample["parsed_output"],
)
with gr.Tab("Image Sample"):
sample = SHOWCASE_SAMPLES["image"]
_render_showcase_sample(
prompt=sample["prompt"],
response=sample["response"],
parsed_output=sample["parsed_output"],
media_markdown=f"",
media_url=DEFAULT_IMAGE_URL,
)
with gr.Tab("Audio Sample"):
sample = SHOWCASE_SAMPLES["audio"]
_render_showcase_sample(
prompt=sample["prompt"],
response=sample["response"],
parsed_output=sample["parsed_output"],
media_url=DEFAULT_AUDIO_URL,
)
with gr.Tab("Reasoning Note"):
sample = SHOWCASE_SAMPLES["reasoning"]
_render_showcase_sample(
prompt=sample["prompt"],
response=sample["response"],
parsed_output=sample["parsed_output"],
)
with gr.Tab("Run It"):
gr.Markdown(
"### Local or GPU-backed run\n"
"Use the packaged files directly for a real interactive run, or attach GPU hardware to this Space."
)
gr.Textbox(
label="Quickstart",
interactive=False,
lines=4,
value=(
"pip install -r requirements.txt\n"
"python quickstart.py\n"
"python quickstart.py --mode image --image path-or-url\n"
"python quickstart.py --mode audio --audio path-or-url"
),
)
with gr.Blocks() as demo:
if _has_supported_gpu_runtime():
_build_live_ui()
else:
_build_showcase_ui()
if __name__ == "__main__":
demo.queue().launch(show_error=True)
|