grant/spark-control
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 3 Wiki Activity

v0.12.0:3 - hotfix: build torchaudio from source against NGC's torch

Browse Source 
NGC PyTorch (the only base with working torch on Spark's ARM64 + sm_120
Blackwell) doesn't ship torchaudio. Stock pip wheels are amd64-only AND
ABI-incompatible with NGC's custom torch 2.10.0a anyway. Pip install
just fails or crashes at runtime.

Real fix:
  - apt install git cmake build-essential ninja-build
  - pip install git+https://github.com/pytorch/audio.git@v2.5.1
    with TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0" (sm_120 for Blackwell GB10)
  - this compiles torchaudio against the torch already in the image, so
    ABI matches by construction

Then constraints.txt locks torch + torchvision + torchaudio so the later
`pip install whisperx` can't swap any of them.

Cost: +3-5 min to the first install. Docker layer cache reuses the
built torchaudio on every subsequent rebuild.

Torchaudio v2.5.1 is the last tag that builds cleanly against
torch 2.5-2.10 — main branch is too volatile against NGC's alpha torch.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Keysat
2026-05-18 21:40:50 -05:00
parent 98aeef8779
commit 09a1d3590d
2 changed files with 33 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
package/startos/versions/v0_1_0.ts
+2 -2
View File
@@ -1,10 +1,10 @@
import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
export const v0_1_0 = VersionInfo.of({
version: '0.12.0:2',
version: '0.12.0:3',
releaseNotes: {
en_US:
'v0.12.0:2 — hotfix: WhisperX docker build was failing at the model-prewarm step with "undefined symbol: torch_library_impl". Root cause: the NGC PyTorch base image ships custom builds of torch + torchaudio + torchvision compiled together for Blackwell (sm_120); pip pulled a stock torchaudio wheel as a transitive dep of whisperx/pyannote, the ABIs didn\'t match, and the resulting .so file refused to load. Fix: generate a constraints.txt at build time from the NGC base\'s installed torch versions, and pass it to every pip install so pip can\'t swap torch/torchaudio/torchvision out from under us. Build should now finish through the model-prewarm step. No other changes vs 0.12.0:1.',
'v0.12.0:3 — hotfix: deeper torchaudio fix. The Spark is ARM64 (Grace + GB10 Blackwell), and the NGC PyTorch container — the only base with a working torch for sm_120 ARM64 — does NOT ship torchaudio at all. Stock pip wheels are amd64-only and ABI-incompatible with NGC\'s custom torch anyway. Real fix: build torchaudio from source against NGC\'s torch (v2.5.1, the last torchaudio tag that compiles cleanly against torch 2.5–2.10) with TORCH_CUDA_ARCH_LIST set for Blackwell sm_120. Build adds ~3-5 min to the first WhisperX install (only first time — Docker layer cache reuses it after). Plus the constraints.txt approach from 0.12.0:2 to lock torch + torchvision + torchaudio against any later pip swap-out.',
},
migrations: {
up: async ({ effects }) => {},