build(whisper): pin model source and revision, pin Python dependencies

README.md

@@ -236,6 +236,16 @@ Measured on CPU: ~2.0 GiB RSS peak; roughly real-time speed on 2 threads (a
   VAD and decoder (`WHISPER_WARMUP`, default `true`) so the first real request
   does not pay the initialisation cost. `/health` returns 503 until model and
   warmup are done; a failed warmup is logged and does not block startup.
+- Model source and pins: the image bakes `large-v3-turbo` from
+  `dropbox-dash/faster-whisper-large-v3-turbo` at a fixed commit
+  (`WHISPER_MODEL_REVISION` in `deploy/Containerfile.whisper`) and verifies
+  `model.bin` against its SHA-256; the build fails if either changes. This is
+  the repository faster-whisper's `large-v3-turbo` alias
+  (`mobiuslabsgmbh/faster-whisper-large-v3-turbo`) redirects to (HTTP 307);
+  `Systran/faster-whisper-large-v3-turbo` answers HTTP 401. Model license: MIT.
+  Python dependencies are pinned in `deploy/whisper/requirements.txt` (`av`
+  must stay on 16.x, newer majors break faster-whisper 1.2.1) and the base
+  image by digest.
 - Per-request log line (JSON after the prefix `whisper_request`), e.g.
   `whisper_request {"status":"ok","path":"standard","audio_s":9.0,"prep_s":0.4,"infer_s":7.6,"total_s":8.1,"rtf":0.889,...}`:
   audio duration (`audio_s`, `audio_after_vad_s`), upload, queue wait, decode +

deploy/Containerfile.whisper

@@ -14,31 +14,48 @@
 # Model is baked into the image (large-v3-turbo, ~1.6 GB) and loaded from a
 # local path — the container runs fully offline and needs no PVC.
 #
+# Model source (pinned, see the ARGs below):
+#   faster-whisper maps "large-v3-turbo" to mobiuslabsgmbh/faster-whisper-large-v3-turbo.
+#   That repository now answers HTTP 307 and redirects to
+#   dropbox-dash/faster-whisper-large-v3-turbo, which is where the files are
+#   served from (checked 2026-10). Systran/faster-whisper-large-v3-turbo
+#   answers HTTP 401. Instead of relying on that alias and redirect, the build
+#   downloads the repository that actually serves the files, at a fixed commit,
+#   and verifies model.bin against its SHA-256. A moved or changed source fails
+#   the build instead of silently changing the image.
+#   License of the model: MIT (CTranslate2 conversion of openai/whisper-large-v3-turbo).
+#   To update: pick a new commit, set WHISPER_MODEL_REVISION and the new
+#   model.bin SHA-256 (Hugging Face file page, "Git LFS Details").
+#
 # Hardened: non-root (uid 1000), no network access needed at runtime,
 # ctranslate2 4.x ships a non-executable stack, so the default seccomp
 # profile and allowPrivilegeEscalation=false work.
 
-FROM python:3.12-slim
+FROM python:3.12-slim@sha256:02108f5d322dd89f1c9e552442c25acb0543dfdbc455693a5599624f20d9155d
 
+# WHISPER_MODEL names the model (image tag, MODEL_NAME); the release workflow
+# reads this line, keep it as is.
 ARG WHISPER_MODEL=large-v3-turbo
+ARG WHISPER_MODEL_REPO=dropbox-dash/faster-whisper-large-v3-turbo
+ARG WHISPER_MODEL_REVISION=0a363e9161cbc7ed1431c9597a8ceaf0c4f78fcf
+ARG WHISPER_MODEL_SHA256=e76620f83d5f5b69efd3d87e3dc180c1bd21df9fbebacfd4335e5e1efcc018da
 
-# av is pinned: faster-whisper 1.2.x breaks with newer PyAV major versions
-# ("open() got an unexpected keyword argument 'metadata_errors'").
-RUN pip install --no-cache-dir \
-    "faster-whisper==1.2.1" \
-    "av==16.1.0" \
-    "fastapi>=0.110.0" \
-    "uvicorn[standard]>=0.27.0" \
-    "python-multipart>=0.0.9"
+# All Python dependencies are pinned in deploy/whisper/requirements.txt
+# (including av==16.1.0, which must not move: newer PyAV majors break
+# faster-whisper 1.2.1). The base image is pinned by digest.
+COPY deploy/whisper/requirements.txt /tmp/requirements.txt
+RUN pip install --no-cache-dir -r /tmp/requirements.txt && rm /tmp/requirements.txt
 
 RUN groupadd -g 1000 whisper && \
     useradd -u 1000 -g whisper -m -s /bin/bash whisper && \
     mkdir -p /opt/model && chown whisper:whisper /opt/model
 
-# Bake the model (needs network at build time only).
+# Bake the model (needs network at build time only). Same file set as
+# faster-whisper's own download.
 USER 1000
-RUN python -c "from faster_whisper.utils import download_model; download_model('${WHISPER_MODEL}', output_dir='/opt/model')" && \
-    rm -rf /home/whisper/.cache
+RUN python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='${WHISPER_MODEL_REPO}', revision='${WHISPER_MODEL_REVISION}', local_dir='/opt/model', allow_patterns=['config.json', 'preprocessor_config.json', 'model.bin', 'tokenizer.json', 'vocabulary.*'])" && \
+    echo "${WHISPER_MODEL_SHA256}  /opt/model/model.bin" | sha256sum -c - && \
+    rm -rf /opt/model/.cache /home/whisper/.cache
 
 WORKDIR /app
 COPY --chown=1000:1000 deploy/whisper/server.py /app/server.py

deploy/whisper/requirements.txt

@@ -0,0 +1,24 @@
+# Python dependencies of the whisper image (deploy/Containerfile.whisper).
+# Everything the server imports is pinned; bump deliberately and re-run the
+# image smoke test (start the container, POST a short clip to
+# /v1/audio/transcriptions).
+
+faster-whisper==1.2.1
+
+# av must stay on 16.x: faster-whisper 1.2.1 breaks with newer PyAV majors
+# ("open() got an unexpected keyword argument 'metadata_errors'"). Renovate is
+# told to leave it alone (renovate.json).
+av==16.1.0
+
+# faster-whisper's runtime and model download, pinned to the versions the
+# image was tested with (faster-whisper itself only constrains them loosely).
+ctranslate2==4.8.2
+huggingface_hub==1.33.0
+numpy==2.5.3
+onnxruntime==1.30.0
+tokenizers==0.23.2
+
+# HTTP server (deploy/whisper/server.py).
+fastapi==0.142.2
+uvicorn[standard]==0.54.0
+python-multipart==0.0.32

renovate.json

@@ -11,6 +11,12 @@
       "matchPackagePatterns": ["*"],
       "groupName": "inference-stack deps",
       "groupSlug": "inference-stack-deps"
+    },
+    {
+      "description": "av must stay on 16.x: faster-whisper 1.2.1 breaks with newer PyAV majors",
+      "matchPackageNames": ["av"],
+      "matchManagers": ["pip_requirements"],
+      "enabled": false
     }
   ],