Spaces:
Running on Zero
Running on Zero
Alexander Claude Opus 5.5 (1M context) commited on
Commit 路
6f89b98
1
Parent(s): a4eb9c1
Give up on a conditioner that never answers
Browse filesSince 2026-10-01 every encode logged `[gpu] encode` and then nothing: no
booking, no error, no GPU spent, while the same /encode called from
outside the Space finished in 12s. `predict` waits forever, so wait a
bounded time instead, log where the job stood and whether a token was
forwarded, cancel it, and drop the shared client in case its stream went
stale.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- h3_engine.py +27 -1
h3_engine.py
CHANGED
|
@@ -16,6 +16,7 @@ import tempfile
|
|
| 16 |
import time
|
| 17 |
import traceback
|
| 18 |
from collections import OrderedDict
|
|
|
|
| 19 |
from functools import cache
|
| 20 |
|
| 21 |
# `import spaces` patches `torch.cuda` before anything can initialize it, which is what lets the 77 GB load happen at
|
|
@@ -32,6 +33,9 @@ import lora_stack
|
|
| 32 |
|
| 33 |
MODEL_REPO = os.getenv("H3_MODEL_REPO", "MiniMaxAI/MiniMax-H3")
|
| 34 |
CONDITIONER_SPACE = os.getenv("H3_CONDITIONER", "multimodalart/qwen3vl-conditioner")
|
|
|
|
|
|
|
|
|
|
| 35 |
GPU_SIZE = os.getenv("H3_GPU_SIZE", "xlarge")
|
| 36 |
# cuDNN's fused attention needs nothing installed and beats the SDPA default on this pool.
|
| 37 |
ATTENTION = os.getenv("H3_ATTENTION", "_native_cudnn")
|
|
@@ -719,7 +723,11 @@ def encode(prompt, first, last, canvas, num_frames, ip_token=None):
|
|
| 719 |
print("[h3] conditioning served from cache", flush=True)
|
| 720 |
return _encode_cache[key]
|
| 721 |
|
| 722 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 723 |
prompt=prompt,
|
| 724 |
image_path=handle_file(first) if first else None,
|
| 725 |
last_image_path=handle_file(last) if last else None,
|
|
@@ -728,6 +736,24 @@ def encode(prompt, first, last, canvas, num_frames, ip_token=None):
|
|
| 728 |
rewrite_prompt=False,
|
| 729 |
api_name="/encode",
|
| 730 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 731 |
with safe_open(path, framework="pt") as handle:
|
| 732 |
conditioning = (handle.get_tensor("prompt_embeds"), handle.get_tensor("text_token_tags"), handle.metadata(), plan)
|
| 733 |
|
|
|
|
| 16 |
import time
|
| 17 |
import traceback
|
| 18 |
from collections import OrderedDict
|
| 19 |
+
from concurrent.futures import TimeoutError as FutureTimeout
|
| 20 |
from functools import cache
|
| 21 |
|
| 22 |
# `import spaces` patches `torch.cuda` before anything can initialize it, which is what lets the 77 GB load happen at
|
|
|
|
| 33 |
|
| 34 |
MODEL_REPO = os.getenv("H3_MODEL_REPO", "MiniMaxAI/MiniMax-H3")
|
| 35 |
CONDITIONER_SPACE = os.getenv("H3_CONDITIONER", "multimodalart/qwen3vl-conditioner")
|
| 36 |
+
# An encode takes about 40s with a cold ZeroGPU slot and 12s warm; this is room for a queue on the conditioner's
|
| 37 |
+
# side, not for a job nobody is running.
|
| 38 |
+
CONDITIONER_TIMEOUT = int(os.getenv("H3_CONDITIONER_TIMEOUT", "180"))
|
| 39 |
GPU_SIZE = os.getenv("H3_GPU_SIZE", "xlarge")
|
| 40 |
# cuDNN's fused attention needs nothing installed and beats the SDPA default on this pool.
|
| 41 |
ATTENTION = os.getenv("H3_ATTENTION", "_native_cudnn")
|
|
|
|
| 723 |
print("[h3] conditioning served from cache", flush=True)
|
| 724 |
return _encode_cache[key]
|
| 725 |
|
| 726 |
+
# `predict` has no timeout, and a job the conditioner never answers held the request forever: from 2026-10-01
|
| 727 |
+
# every encode logged `[gpu] encode` and then nothing - no booking, no error, no GPU spent - while the same call
|
| 728 |
+
# from outside finished in 12s. So wait a bounded time, say where the job stood, and let the request fail.
|
| 729 |
+
client = _conditioner(ip_token)
|
| 730 |
+
job = client.submit(
|
| 731 |
prompt=prompt,
|
| 732 |
image_path=handle_file(first) if first else None,
|
| 733 |
last_image_path=handle_file(last) if last else None,
|
|
|
|
| 736 |
rewrite_prompt=False,
|
| 737 |
api_name="/encode",
|
| 738 |
)
|
| 739 |
+
try:
|
| 740 |
+
path, plan = job.result(timeout=CONDITIONER_TIMEOUT)
|
| 741 |
+
except FutureTimeout:
|
| 742 |
+
status = job.status()
|
| 743 |
+
print(
|
| 744 |
+
f"[h3] conditioner silent for {CONDITIONER_TIMEOUT}s 路 job {status.code.name}"
|
| 745 |
+
f" 路 rank {status.rank} of {status.queue_size} 路 token {'forwarded' if ip_token else 'none'}",
|
| 746 |
+
flush=True,
|
| 747 |
+
)
|
| 748 |
+
job.cancel()
|
| 749 |
+
# The shared client is held for the life of the process; if its stream to the conditioner went stale,
|
| 750 |
+
# keeping it would hang every request that has no token of its own.
|
| 751 |
+
if not ip_token:
|
| 752 |
+
_shared_conditioner.cache_clear()
|
| 753 |
+
raise TimeoutError(
|
| 754 |
+
f"the conditioner `{CONDITIONER_SPACE}` did not answer in {CONDITIONER_TIMEOUT}s "
|
| 755 |
+
f"(job {status.code.name.lower()}) - no GPU time was booked, try again"
|
| 756 |
+
) from None
|
| 757 |
with safe_open(path, framework="pt") as handle:
|
| 758 |
conditioning = (handle.get_tensor("prompt_embeds"), handle.get_tensor("text_token_tags"), handle.metadata(), plan)
|
| 759 |
|