Spaces:
Running on Zero
Running on Zero
File size: 11,261 Bytes
7d03019 | 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 | from __future__ import annotations
import os
import json
import atexit
from pathlib import Path
import time
from datetime import datetime, timezone
# ZeroGPU v2 serializes global virtual CUDA tensors before the replica starts.
# Keep the service-managed NVMe location, matching the reference Space.
os.environ.setdefault("HF_XET_HIGH_PERFORMANCE", "1")
os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
# PyTorch's default CUDA availability check invokes the CUDA device-count API,
# which poisons forked children. NVML discovery is explicitly fork-safe.
os.environ.setdefault("PYTORCH_NVML_BASED_CUDA_CHECK", "1")
import spaces
import torch
_ZERO_GPU = os.environ.get("SPACES_ZERO_GPU", "").strip().lower() in {"1", "true", "yes", "on"}
def _is_zerogpu_parent() -> bool:
if not _ZERO_GPU:
return False
try:
from spaces.zero import wrappers as zero_wrappers
return not zero_wrappers.forked
except Exception:
return True
def _assert_cuda_fork_clean(event: str) -> None:
"""Reproduce the CUDA bad-fork check used by the ZeroGPU worker."""
read_fd, write_fd = os.pipe()
pid = os.fork()
if pid == 0:
os.close(read_fd)
try:
verdict = b"1" if torch.cuda._is_in_bad_fork() else b"0"
os.write(write_fd, verdict)
finally:
os.close(write_fd)
os._exit(0)
os.close(write_fd)
try:
verdict = os.read(read_fd, 1)
finally:
os.close(read_fd)
_, status = os.waitpid(pid, 0)
if status != 0 or verdict != b"0":
raise RuntimeError("ZeroGPU parent CUDA state poisons the worker fork")
print(f'[WAN_SERVICE] {{"bad_fork": false, "event": "{event}"}}', flush=True)
def _install_parent_cuda_firewall() -> None:
"""Keep discovery fork-safe in the parent and restore real CUDA in workers."""
original_is_available = torch.cuda.is_available
original_device_count = torch.cuda.device_count
original_is_bf16_supported = torch.cuda.is_bf16_supported
original_get_allocator_backend = torch.cuda.get_allocator_backend
original_lazy_init = torch.cuda._lazy_init
original_c_device_count = torch._C._cuda_getDeviceCount
original_c_init = torch._C._cuda_init
def dispatch(original, parent_value):
def wrapped(*args, **kwargs):
if _is_zerogpu_parent():
return parent_value
return original(*args, **kwargs)
return wrapped
def guarded_lazy_init(*args, **kwargs):
if _is_zerogpu_parent():
raise RuntimeError("CUDA context initialization attempted in ZeroGPU parent")
return original_lazy_init(*args, **kwargs)
def guarded_c_init(*args, **kwargs):
if _is_zerogpu_parent():
raise RuntimeError("Direct CUDA initialization attempted in ZeroGPU parent")
return original_c_init(*args, **kwargs)
torch.cuda.is_available = dispatch(original_is_available, False)
torch.cuda.device_count = dispatch(original_device_count, 0)
torch.cuda.is_bf16_supported = dispatch(original_is_bf16_supported, True)
torch.cuda.get_allocator_backend = dispatch(original_get_allocator_backend, "zerogpu")
torch.cuda._lazy_init = guarded_lazy_init
torch.cuda.init = guarded_lazy_init
torch._C._cuda_getDeviceCount = dispatch(original_c_device_count, 0)
torch._C._cuda_init = guarded_c_init
if _ZERO_GPU:
_assert_cuda_fork_clean("parent.cuda_import_clean")
_install_parent_cuda_firewall()
import gradio as gr
if _ZERO_GPU:
_assert_cuda_fork_clean("parent.cuda_gradio_clean")
from core_anim.space_config import load_space_config
from core_anim.space_postprocess import encode_video_with_preview
from core_anim.space_runtime import LoopGeneratorService
if _ZERO_GPU:
_assert_cuda_fork_clean("parent.cuda_runtime_imports_clean")
CONFIG = load_space_config()
_SERVICE = None
if _ZERO_GPU:
_assert_cuda_fork_clean("parent.cuda_config_clean")
def _shutdown_global_service() -> None:
global _SERVICE
service, _SERVICE = _SERVICE, None
if service is not None:
service.close()
def get_service() -> LoopGeneratorService:
global _SERVICE
if _SERVICE is None:
_SERVICE = LoopGeneratorService(CONFIG)
return _SERVICE
# ZeroGPU optimizes CUDA placement performed during module initialization.
# Tests and manifest validation can explicitly skip the multi-GB model setup.
if os.environ.get("SPACE_SKIP_MODEL_LOAD") != "1":
_SERVICE = LoopGeneratorService(CONFIG)
if _ZERO_GPU:
if torch.cuda.is_initialized():
raise RuntimeError("ZeroGPU parent initialized CUDA before worker fork")
print('[WAN_SERVICE] {"cuda_initialized": false, "event": "parent.cuda_clean"}', flush=True)
_assert_cuda_fork_clean("parent.cuda_fork_clean")
atexit.register(_shutdown_global_service)
def _diag(event: str, **fields) -> None:
payload = {
"ts": datetime.now(timezone.utc).isoformat(timespec="milliseconds"),
"event": event,
**fields,
}
print(f"[WAN_JOB] {json.dumps(payload, sort_keys=True)}", flush=True)
def validate_request(image, prompt: str):
if image is None:
raise gr.Error("Upload an image.")
if not (prompt or "").strip():
raise gr.Error("Enter a prompt.")
return image, prompt
def finish_video(frame_bundle: str, output_format: str) -> tuple[str, str]:
return encode_video_with_preview(
frame_bundle,
fps=CONFIG.fps,
output_format=output_format,
crossfade=CONFIG.use_crossfade,
)
@spaces.GPU(duration=CONFIG.gpu_duration, size=CONFIG.gpu_size)
def generate_loop(image, prompt: str, output_format: str, progress=gr.Progress(track_tqdm=False)):
image, prompt = validate_request(image, prompt)
if os.environ.get("SPACE_SKIP_MODEL_LOAD") == "1":
raise RuntimeError("Model runtime disabled by SPACE_SKIP_MODEL_LOAD.")
service = None
frame_bundle = None
started = time.perf_counter()
outcome = "failed"
_diag(
"job.accepted",
input_width=getattr(image, "width", None),
input_height=getattr(image, "height", None),
)
try:
progress(0.0, desc="Requesting GPU…")
yield gr.skip(), gr.skip(), "Requesting GPU…"
progress(0.03, desc="Initializing runtime…")
yield gr.skip(), gr.skip(), "Initializing runtime…"
init_started = time.perf_counter()
service = get_service()
_diag("runtime.init.done", elapsed_s=round(time.perf_counter() - init_started, 3))
for stage_result in service.generate_iter(
image,
prompt,
progress_callback=progress,
):
if stage_result is None: # Compatibility and cancellation checkpoint.
yield gr.skip(), gr.skip(), "Processing…"
elif isinstance(stage_result, dict) and "stage" in stage_result:
yield gr.skip(), gr.skip(), stage_result["stage"]
elif isinstance(stage_result, dict) and "bundle" in stage_result:
frame_bundle = stage_result["bundle"]
else: # Compatibility with simple test doubles.
frame_bundle = stage_result
# Release per-job diffusion/VAE objects before CPU-only video encoding;
# reusable process-global components remain prepared for the next job.
service.cleanup_job()
format_name = "MKV" if output_format == "mkv" else "MP4"
encoding_label = f"Encoding {format_name} and preview…" if output_format == "mkv" else "Encoding MP4…"
progress(0.97, desc=encoding_label)
yield gr.skip(), gr.skip(), encoding_label
encode_started = time.perf_counter()
_diag("video.encode.start", output_format=output_format)
preview, download = finish_video(frame_bundle, output_format)
_diag(
"video.encode.done",
output_format=output_format,
elapsed_s=round(time.perf_counter() - encode_started, 3),
)
frame_bundle = None # encode_video_with_preview removes its source bundle.
progress(1.0, desc="Complete")
outcome = "complete"
yield preview, download, "Complete"
except GeneratorExit:
outcome = "cancelled"
_diag("job.cancelled", elapsed_s=round(time.perf_counter() - started, 3))
raise
except BaseException as exc:
_diag(
"job.failed",
elapsed_s=round(time.perf_counter() - started, 3),
error_type=type(exc).__name__,
error=str(exc)[:500],
)
raise
finally:
_diag("job.cleanup.start", outcome=outcome)
if service is not None:
service.cleanup_job()
if frame_bundle is not None:
Path(frame_bundle).unlink(missing_ok=True)
_diag(
"job.cleanup.done",
outcome=outcome,
elapsed_s=round(time.perf_counter() - started, 3),
)
with gr.Blocks(title="Wan Loop Generator", delete_cache=(86400, 86400)) as demo:
gr.Markdown(
"# Wan Loop Generator\n"
"This experimental Space is a test for generating one-second video loops from a single image."
)
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### Image Upload")
image_input = gr.Image(
type="pil",
label="Image",
sources=["upload"],
height=420,
)
prompt_input = gr.Textbox(label="Prompt", placeholder="Describe the motion of the loop…", lines=4)
format_input = gr.Dropdown(
choices=[
("MKV (High Quality)", "mkv"),
("MP4 (Compressed)", "mp4"),
],
value="mkv",
label="Output format",
interactive=True,
)
with gr.Row():
generate_button = gr.Button("Generate loop", variant="primary")
cancel_button = gr.Button("Cancel", variant="stop")
with gr.Column(scale=1):
gr.Markdown("### Progress / Output")
status_output = gr.Textbox(
label="Status",
value="Ready",
interactive=False,
)
preview_output = gr.Video(
label="Preview",
format="mp4",
height=360,
autoplay=True,
loop=True,
)
download_output = gr.File(
label="Download generated loop",
)
generation_event = generate_button.click(
generate_loop,
inputs=[image_input, prompt_input, format_input],
outputs=[preview_output, download_output, status_output],
concurrency_limit=CONFIG.concurrency_limit,
api_name="generate_loop",
show_progress="full",
)
cancel_button.click(
fn=None,
cancels=[generation_event],
queue=False,
api_visibility="private",
)
if __name__ == "__main__":
demo.queue(default_concurrency_limit=CONFIG.concurrency_limit, max_size=8).launch(ssr_mode=False)
|