fix(scriberr): diarization works again: HF_TOKEN passthrough, pyannote env torchcodec pin (patch 0003); mirror retention; YaRN stays off
Scriberr: pyannote's diarization models are gated, and Scriberr had no Hugging Face token, so every diarized job died at the diarize step (WhisperX after a full CPU transcription). HF_TOKEN now passes through compose; the value is only in the host .env (vault fv-ml1/hf-token). The separate pyannote env had torchcodec 0.16 (needs torch >=2.11/CUDA 13) beside torch 2.8, so pyannote 4 had no AudioDecoder. Pinned to 0.7 by hand on the live env; patch 0003 makes it durable at the next rebuild. WhisperX's adapter defaults to cpu/small/float32, now documented. Both paths verified on GPU 3 with a real multi-speaker clip. pbs-nh3: the backups-mirror prune job now keeps the primary's history (last 3, daily 14, weekly 8, monthly 12, yearly 5). Dry-run: keep 738, remove 3,194 across 32 groups; first run at the job's 00:00.
This commit is contained in:
@@ -134,6 +134,16 @@ rebuild — a local patch to carry, which is why it was not done.
|
||||
- **This stack is a guest on GPU1**, which it shares with the `sec` seat. If
|
||||
VRAM gets tight, this is the thing that should yield.
|
||||
|
||||
## Diarization: what broke it and what fixed it (2026-10-06)
|
||||
|
||||
Prime: "whisper doesn't work at all, and I need diarized transcripts." There were three separate faults, and all three looked like "transcription is broken":
|
||||
|
||||
1. **No Hugging Face token.** pyannote's diarization models are gated. Every diarized job died at the diarize step: WhisperX with `GatedRepoError 401` on `pyannote/speaker-diarization-3.1`, but only after a full CPU transcription (12m45s for a 1 h file), and Parakeet+pyannote with "HuggingFace token is required". **Fix:** `HF_TOKEN` is passed through compose. Its value is in the host `.env` only (vault `fv-ml1/hf-token`, account lkraven, read role; access to `speaker-diarization-3.1`, `segmentation-3.0` and `speaker-diarization-community-1` was checked).
|
||||
2. **The separate pyannote env had a torchcodec that cannot load.** uv resolved torchcodec 0.16.0 beside torch 2.8.0+cu128. 0.16 needs torch ≥2.11 and CUDA 13 (`libnvrtc.so.13`), so pyannote.audio 4 has no decoder and fails with `name 'AudioDecoder' is not defined`. **Fix (live, by hand):** `torchcodec==0.7.*` (its table pairs 0.7 with torch 2.8), pinned in `/tank/scriberr/whisperx-env/pyannote/pyproject.toml`, then `uv lock && uv sync` as uid 10001; the old files are kept as `*.bak-20261006`. Scriberr only rewrites that pyproject if its import test fails at startup. **Durable fix:** `patches/0003-pyannote-env-pin-torchcodec.patch` pins torch/torchaudio 2.8 and torchcodec 0.7 in the embedded pyproject, so it lands at the next rebuild.
|
||||
3. **WhisperX defaults to CPU.** The adapter's defaults are `device cpu`, `model small`, `compute_type float32` (`whisperx_adapter.go`). A quick transcribe with no profile ran exactly that. **Use** device `cuda`, model `large-v3`, compute `float16`.
|
||||
|
||||
**Verified on a 4-min, multi-speaker clip of a real upload, GPU 3:** WhisperX large-v3/cuda/float16 + `pyannote/speaker-diarization-3.1` exited 0 in 65 s with 2 speakers. The pyannote env script with `speaker-diarization-community-1` (the adapter's only option, used by the Parakeet profile) exited 0 in 11 s with 3 speakers. The two pipelines disagree on the speaker count, so set min/max speakers when you know it. WhisperX's own env (pyannote.audio 4.0.4) loads 3.1 fine and feeds the pipeline in-memory audio, so it never needed torchcodec.
|
||||
|
||||
## Optional: summarisation via the LiteLLM gateway
|
||||
|
||||
Scriberr speaks the OpenAI API, so point it at the fleet gateway instead of a
|
||||
|
||||
@@ -72,6 +72,15 @@ services:
|
||||
# and takes out the Parakeet + Sortformer backends (WhisperX survives).
|
||||
# `copy` trades a little disk and time for it actually working.
|
||||
- UV_LINK_MODE=${SCRIBERR_UV_LINK_MODE:-copy}
|
||||
# Hugging Face token for the GATED pyannote models (speaker-diarization-3.1,
|
||||
# segmentation-3.0). Without it every diarized job dies at the diarize step:
|
||||
# WhisperX with a 401 GatedRepoError AFTER the whole transcription has run, and
|
||||
# Parakeet+pyannote with "HuggingFace token is required". So it looks as if
|
||||
# nothing works. Scriberr's config and both adapters read HF_TOKEN
|
||||
# (internal/config/config.go), and it reaches the uv subprocesses. Value: vault
|
||||
# `fv-ml1/hf-token` (account lkraven, read role; access to the gated repos was
|
||||
# checked 2026-10-06). It lives only in the host .env, never in git.
|
||||
- HF_TOKEN=${HF_TOKEN:-}
|
||||
# ── Memory settings, measured while Scriberr shared GPU 1 (2026-09-30; it moved to
|
||||
# GPU 3 at 1322 the same day, so the 5.5 GB budget no longer binds, but the values stand). Prime: Scriberr shared GPU 1 with
|
||||
# intern-decision, which holds ~9.7 GB resting / 10.3 GB peak). ──────────
|
||||
|
||||
@@ -0,0 +1,34 @@
|
||||
Subject: [PATCH 0003] pyannote env: pin torch/torchaudio 2.8 and torchcodec 0.7
|
||||
|
||||
The embedded pyannote env leaves torchcodec unpinned (and torch >=2.5). On fv-ml1
|
||||
(2026-10-06) uv resolved torch 2.8.0+cu128 with torchcodec 0.16.0, which needs
|
||||
torch >=2.11 and CUDA 13 (libnvrtc.so.13). torchcodec's import fails, so
|
||||
pyannote.audio 4 has no AudioDecoder, and every Parakeet+pyannote diarization
|
||||
fails with "name 'AudioDecoder' is not defined". torchcodec's own compatibility
|
||||
table pairs 0.7 with torch 2.8. Pin all three together, so a fresh env resolves
|
||||
the set that was verified: community-1 diarization on GPU, 3 speakers on a 4-min
|
||||
clip, 11 s.
|
||||
|
||||
---
|
||||
--- a/internal/transcription/adapters/py/pyannote/pyproject.toml
|
||||
+++ b/internal/transcription/adapters/py/pyannote/pyproject.toml
|
||||
@@ -4,8 +4,8 @@
|
||||
description = "Audio diarization using PyAnnote"
|
||||
requires-python = ">=3.10"
|
||||
dependencies = [
|
||||
- "torch>=2.5.0",
|
||||
- "torchaudio>=2.5.0",
|
||||
+ "torch==2.8.*",
|
||||
+ "torchaudio==2.8.*",
|
||||
"huggingface-hub>=0.28.1",
|
||||
"pyannote.audio==4.0.2"
|
||||
]
|
||||
@@ -18,7 +18,7 @@
|
||||
|
||||
[tool.uv]
|
||||
override-dependencies = [
|
||||
- "torchcodec ; sys_platform != 'linux' or platform_machine != 'aarch64'"
|
||||
+ "torchcodec==0.7.* ; sys_platform != 'linux' or platform_machine != 'aarch64'"
|
||||
]
|
||||
|
||||
[tool.uv.sources]
|
||||
@@ -9,6 +9,7 @@ distinctly tagged image, and proves it before anyone deploys it.
|
||||
|---|---|---|
|
||||
| `0001-parakeet-pause-aware-slicer.patch` | upstream `a353078` (HEAD 2026-09-20) | **LIVE on fv-ml1 since 2026-09-30 1211 PT** as `scriberr:local-blackwell-a353078-slicer1`; superseded in production by 0002 on top of it; **no upstream PR (Prime, 2026-10-01)**. Both patches are carried locally only |
|
||||
| `0002-parakeet-model-path-and-gap-retry.patch` | 0001 | **LIVE on fv-ml1 since 2026-09-30 1602 PT** as `scriberr:local-blackwell-a353078-dropout2` (Prime: "Scriberr gets the basic fix, no surgery for the new toolkit"; v3 stays). Live check: a 20-min file peaked at 5,502 MiB on GPU 3, `retried_gaps: 1`, +29 words against the slicer1 image. Why: `docs/pfi/parakeet-dropout-investigation-2026-09-30.md` |
|
||||
| `0003-pyannote-env-pin-torchcodec.patch` | 0002 | **NOT in an image yet (2026-10-06).** Pins torch/torchaudio 2.8 and torchcodec 0.7 in the embedded pyannote env pyproject. The live env was fixed by hand the same way (`../README.md` § Diarization), so this only matters when the env is rebuilt or Scriberr re-runs its setup. `git apply --check` passes on a353078+0001+0002. |
|
||||
|
||||
|
||||
Rollback for the live deploy: `SCRIBERR_IMAGE=scriberr:local-blackwell` (the
|
||||
|
||||
Reference in New Issue
Block a user