fix(asr): CrisperWhisper ct2 venv needs torch+transformers for HF→CT2 conversion

[wilke0818]

Problem

On Linux x86_64, the CrisperWhisper backend builds its subprocess venv from crisperwhisper[ct2] only. CT2 inference is torch-free, but the first-run HF→CT2 conversion (model._init_v2 → converter.ensure_ct2_model → ctranslate2.converters.transformers._load) needs torch and transformers: that module does try: import huggingface_hub, torch, transformers and, when they're absent, leaves the names unbound; _load() then calls torch.no_grad() and the transformers loader → NameError: name 'torch' / 'transformers' is not defined.

Impact: the venv "builds" but every clip's CrisperWhisper ASR fails in ~0.7 s with that NameError, silently dropping the whisper model from multi-ASR runs (found on a CUDA-13.1 GPU batch; canary-qwen + Qwen3-ASR still succeed). macOS is unaffected — it uses the transformers backend (torch already declared) and never invokes the CT2 converter.

Fix

Add the [transformers] extra (== [all] alongside ct2: transformers + torch + accelerate) to the Linux-x86 ct2 requirements, and pin torch/torchaudio so ensure_venv routes them through the CUDA-aware index. The conversion result is cached, so torch/transformers are installed but never loaded at CT2 inference time.

# before
_CRISPER_REQUIREMENTS = ["crisperwhisper[ct2]==2.0.1"]
# after
_CRISPER_REQUIREMENTS = ["crisperwhisper[ct2,transformers]==2.0.1", "torch>=2.4", "torchaudio>=2.4"]

Why CI didn't catch it

crisperwhisper_test.py:

• the assembly test mocks ensure_venv / venv_python / subprocess → the ct2 venv is never built;
• test_crisperwhisper_transcribes_when_venv_present is @skipif(not CRISPER_VENV.exists()) → skipped in CI, and where a dev has the venv (macOS) it runs the torch-having transformers backend.

So no test builds the [ct2] venv on Linux-x86 and runs a conversion — the only path that hits this. A GPU-gated regression test mirroring brouhaha_venv_gpu_test.py (build ct2 venv + transcribe) would cover it; happy to add as a follow-up.

Validation

Rebuilt the venv on a GPU node: CrisperWhisper now transcribes (e.g. "blue or maybe green blue red blue blue" on a Word-color-Stroop clip) and all three ASR models return.

Notes

Base: alpha. Independent of #539 (HF-cache) — the two touch different regions of crisperwhisper.py (requirements list vs hf_subprocess_env model staging) and merge cleanly.

🤖 Generated with Claude Code

Version

Published prerelease version: 1.3.1-alpha.39

Changelog

🐛 Bug Fix

• fix(asr): CrisperWhisper ct2 venv needs torch+transformers for HF→CT2 conversion #540 (@wilke0818)
• hf-cache: stop HF Hub 429s under parallel batch (resolve-once + SHA-pin) #539 (@wilke0818)
• fix(adaptive): guard None history in convergence improvement delta #541 (@wilke0818)
• refactor(cache): cached-inference layer out of analyze_audio.py (T051, parts 1-2) #538 (@satra)
• Scene-aware presence/quality axis + ASR transcription overhaul (CrisperWhisper 2.0, Qwen) #536 (@satra @claude)
• feat(canary): split long audio at pauses to avoid the ~40s truncation #534 (@wilke0818)
• ScriptLine: accept empty text as "model produced no transcript" signal #520 (@wilke0818)
• Feat/pii subprocess venv #519 (@wilke0818)
• subprocess_venv: route torch/torchaudio via CUDA-aware PyTorch index #516 (@satra @wilke0818)
• subprocess_venv: two-stage install to fix torch/torchaudio CUDA-tag s… #518 (@wilke0818)
• Fix/ser wav2vec2 regression head #511 (@satra @wilke0818 @claude)
• comparison & uncertainty stage for analyze_audio.py #512 (@satra)
• audio analysis script + multilingual MMS aligner + 4 new ASR backends #510 (@satra)

Authors: 3

• Claude (@claude)
• Jordan Wilke (@wilke0818)
• Satrajit Ghosh (@satra)

[wilke0818]

the other option, which requires a bit more logic, would be to create a one time venv for the conversion and then another for the inference. The challenge is that it is more work + if the cache directory for the model gets changed/moved, the venv needs to be reconstituted