packaging: build a profile's extra venvs generically, and make audiocraft's installable

The audio build failed with 'No matching distribution found for torch==2.1.0':
audiocraft's pin is unsatisfiable on Python 3.12 at all, because torch support
for 3.12 starts at 2.2. So the venv uses the oldest torch this Python can run
and installs audiocraft --no-deps, with the deps it actually uses listed
explicitly. --no-deps is reckless in a shared venv and routine in an isolated
one: nothing in there can reach the server's stack, and the build's import check
fails the image if the version judgement turns out to be wrong.

The step is now generic — any <profile>.venv-<name>.txt builds
/opt/coderai/venvs/<name>, with an optional .nodeps beside it — rather than
hardcoding audiocraft into the Dockerfile. The next library that pins an
impossible torch will not need a new build step.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0182DhR1eFNPQyGmDrpHbedY

packaging/runpod/Dockerfile.capability

@@ -57,12 +57,18 @@ RUN --mount=type=cache,target=/root/.cache/pip set -eux; \
 ARG PROFILE
 COPY packaging/runpod/profiles/ /tmp/venv-profiles/
 RUN set -eux; \
-    extra="/tmp/venv-profiles/${PROFILE}.venv-audiocraft.txt"; \
-    if [ -f "$extra" ]; then \
-        python -m venv /opt/coderai/venvs/audiocraft; \
-        /opt/coderai/venvs/audiocraft/bin/pip install --no-cache-dir -r "$extra"; \
-        /opt/coderai/venvs/audiocraft/bin/python -c "import audiocraft; print('audiocraft', audiocraft.__version__)"; \
-    fi; \
+    for extra in /tmp/venv-profiles/${PROFILE}.venv-*.txt; do \
+        [ -f "$extra" ] || continue; \
+        name="$(basename "$extra" .txt)"; name="${name#${PROFILE}.venv-}"; \
+        vdir="/opt/coderai/venvs/${name}"; \
+        python -m venv "$vdir"; \
+        "$vdir/bin/pip" install --no-cache-dir -r "$extra"; \
+        nodeps="${extra%.txt}.nodeps"; \
+        if [ -f "$nodeps" ]; then \
+            "$vdir/bin/pip" install --no-cache-dir --no-deps -r "$nodeps"; \
+        fi; \
+        "$vdir/bin/python" -c "import ${name}; print('${name}', getattr(${name}, '__version__', 'ok'), 'in its own venv')"; \
+    done; \
     rm -rf /tmp/venv-profiles
 
 COPY . /opt/coderai/app

packaging/runpod/profiles/audio.venv-audiocraft.nodeps

+# Installed with --no-deps into the venv described by the .txt beside this file.
+# audiocraft's declared pins (torch==2.1.0, torchvision==0.16.0,
+# torchtext==0.16.0) cannot all be satisfied on this Python, and its inference
+# code does not need those exact versions — the pin is the library being
+# conservative. The import check in the build fails the image if that judgement
+# turns out to be wrong.
+audiocraft==1.3.0

packaging/runpod/profiles/audio.venv-audiocraft.txt

 # A SECOND venv inside the audio image, for audiocraft alone.
 #
-# audiocraft pins torch==2.1.0, torchvision==0.16.0 and torchtext==0.16.0. The
-# server runs torch 2.11 built for a specific CUDA, and torch 2.1.0 has no build
-# for that CUDA at all — so installing audiocraft beside the server does not
-# merely risk a conflict, it leaves the pod with no working GPU torch. Its own
-# venv is the only way both can be right.
+# audiocraft pins torch==2.1.0, torchvision==0.16.0 and torchtext==0.16.0. Two
+# separate problems with that:
 #
-# Baked at build time, not on first use: a pod is disposable and rented by the
-# second, so paying minutes of GPU rental to pip-install a library on every cold
-# start would cost more than the image layer does.
+#  1. torch 2.1.0 has no build for Python 3.12 at all — support starts at 2.2 —
+#     so the pin is unsatisfiable in this image whatever else we do.
+#  2. Even if it existed, installing it into the server's venv would downgrade
+#     torch five majors and leave the pod with no working GPU torch, because
+#     2.1.0 has no build for this image's CUDA either.
+#
+# So: its own venv, on the oldest torch this Python can run, with audiocraft
+# installed --no-deps (see the .nodeps file beside this one) and the deps it
+# actually uses listed here. --no-deps is reckless in a shared venv and routine
+# in an isolated one — nothing here can reach the server's stack.
 --extra-index-url https://download.pytorch.org/whl/cu121
-torch==2.1.0
-torchaudio==2.1.0
-audiocraft==1.3.0
+torch==2.2.2
+torchaudio==2.2.2
+torchvision==0.17.2
+
+# audiocraft's real runtime dependencies. Its dev-only ones (pytest, flake8,
+# mypy, coverage, pdoc3) and gradio are deliberately absent: a pod serves an
+# API, it does not run the library's test suite or its demo UI. torchtext is
+# absent too — it is abandoned, unbuildable here, and unused at inference.
+av>=11.0.0
+einops
+flashy>=0.0.1
+hydra-core>=1.1
+hydra_colorlog
+julius
+num2words
+sentencepiece
+spacy>=3.6.1
+huggingface_hub
+tqdm
+librosa
+torchmetrics
+encodec
+protobuf
+numpy<2