Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

fix(longcat): install onnxruntime-gpu so the vocal separator imports at job start

nirholase10f35e10069
diff --git a/workers/longcat/check_pins.py b/workers/longcat/check_pins.py
index 33dbce12b..7160c487a 100644
--- a/workers/longcat/check_pins.py
+++ b/workers/longcat/check_pins.py
@@ -5,6 +5,14 @@ own torch and numpy requirements, and without this gate a transitive resolution
that swapped the cu124 wheel for a CPU-only build would only surface at
inference time, as "Torch not compiled with CUDA enabled", after a 40-minute
build had already reported success.
+
+It also proves the avatar script's import chain resolves. LongCat's
+run_demo_avatar_single_audio_to_video.py imports
+`audio_separator.separator.Separator` at module scope, and that module imports
+`onnxruntime`, which audio-separator only declares behind an optional extra.
+Without the explicit onnxruntime-gpu pin, the image built green and every
+avatar job died with ModuleNotFoundError before rendering a frame. Listing the
+execution providers needs no GPU, so this runs on the CPU-only builder.
"""
from __future__ import annotations
@@ -50,6 +58,34 @@ if hub_parts < HUB_FLOOR:
f"{'.'.join(str(n) for n in HUB_FLOOR)} floor diffusers {diffusers.__version__} requires"
)
+# onnxruntime-gpu 1.20.x is the last line built for any CUDA 12.x with cuDNN 9;
+# 1.21 and later need CUDA 12.8, newer than this image's 12.4 runtime. A CPU-only
+# `onnxruntime` wheel arriving through some later extra would replace the module
+# and silently drop the CUDA provider, so check both.
+try:
+ import onnxruntime
+except ImportError as exc:
+ providers = []
+ failures.append(f"onnxruntime is not installed ({exc}); pin onnxruntime-gpu in requirements-model.txt")
+else:
+ providers = onnxruntime.get_available_providers()
+ if not onnxruntime.__version__.startswith("1.20."):
+ failures.append(
+ f"onnxruntime drifted to {onnxruntime.__version__}, expected 1.20.x (CUDA 12.x + cuDNN 9 build)"
+ )
+ if "CUDAExecutionProvider" not in providers:
+ failures.append(
+ f"onnxruntime {onnxruntime.__version__} has no CUDAExecutionProvider ({providers}); "
+ "a CPU-only onnxruntime wheel replaced onnxruntime-gpu"
+ )
+
+# The exact import the avatar script makes at module scope. A failure here is the
+# one every job would otherwise hit first.
+try:
+ from audio_separator.separator import Separator # noqa: F401
+except Exception as exc: # noqa: BLE001
+ failures.append(f"audio_separator.separator does not import: {type(exc).__name__}: {exc}")
+
if failures:
for line in failures:
print(f"PIN DRIFT {line}", file=sys.stderr)
@@ -63,4 +99,6 @@ print(
f"transformers {transformers.__version__}",
f"diffusers {diffusers.__version__}",
f"huggingface-hub {huggingface_hub.__version__}",
+ f"onnxruntime {onnxruntime.__version__} ({', '.join(providers)})",
+ "audio_separator.separator ok",
)
diff --git a/workers/longcat/requirements-model.txt b/workers/longcat/requirements-model.txt
index 15358ba62..e09e2d683 100644
--- a/workers/longcat/requirements-model.txt
+++ b/workers/longcat/requirements-model.txt
@@ -40,3 +40,13 @@ scipy==1.15.3
pyloudnorm==0.1.1
onnx==1.22.0
audio-separator==0.30.2
+# The runtime that executes the separator's Kim_Vocal_2.onnx. audio-separator
+# only pulls one in through its [cpu]/[gpu] extras, so a plain install leaves
+# `import onnxruntime` (separator.py, module scope) unresolvable and every
+# avatar job dies with ModuleNotFoundError before the first frame. The GPU build
+# provides the same `onnxruntime` module. 1.20.x specifically: it is the last
+# onnxruntime-gpu line built for any CUDA 12.x with cuDNN 9 (1.21 and later
+# require CUDA 12.8), which is what this image's CUDA 12.4.1 runtime base and the
+# cuDNN 9 that torch 2.6.0+cu124 loads provide. check_pins.py asserts the version,
+# the CUDA execution provider, and that the separator imports.
+onnxruntime-gpu==1.20.2