ace-step: patch upstream infer-api + missing runtime deps + cache mount

Three upstream gaps surfaced once /generate was actually exercised:

  1. infer-api.py builds an 18-arg positional tuple but the pipeline
     expects 24 — first missing arg is `format`, so audio_duration
     shifts into format's slot and the pipeline calls len() on an
     int. Ship a patched copy of infer-api.py and COPY over upstream's
     in the Dockerfile. Also handle empty lora_name_or_path -> "none"
     (empty string trips HF Hub's repo-id validator).
  2. torchcodec + ffmpeg are required by the WAV save path but neither
     is in upstream requirements.txt. Without them every /generate
     runs to completion and then 500s at write-time.
  3. ACE-Step caches checkpoints at /root/.cache/ace-step/checkpoints
     (HARDCODED, not honored by HF_HOME). Mount our persistent dir
     there so the ~7 GB model survives container recreates.

Bench on A6000 (cached model, lo-fi hip hop, 60-step euler/apg):
  10s @ 27 steps -> 9.4s  (0.94x)
  30s @ 60 steps -> 11.2s (0.37x, ~2.7x realtime)
  60s @ 60 steps -> 14.8s (0.24x, ~4x realtime)

stacks/ace-step/Dockerfile

@@ -35,9 +35,13 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         git \
         curl \
         ca-certificates \
+        ffmpeg \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/* \
     && ln -sf /usr/bin/python3 /usr/bin/python
+# ffmpeg is required by torchcodec at runtime — torchcodec dlopens
+# libavcodec/libavformat. Without it the WAV save step fails with
+# "Could not load libtorchcodec".
 
 RUN python -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
@@ -60,7 +64,17 @@ RUN pip install --no-cache-dir --upgrade pip \
         --index-url https://download.pytorch.org/whl/cu126 \
     && pip install --no-cache-dir hf_transfer peft \
     && pip install --no-cache-dir -r requirements.txt \
-    && pip install --no-cache-dir .
+    && pip install --no-cache-dir . \
+    && pip install --no-cache-dir torchcodec
+# torchcodec — required by torchaudio's save_with_torchcodec (the save
+# path ACE-Step uses on output). Not pulled in by upstream's
+# requirements.txt; without it, /generate runs to completion and then
+# 500s at the WAV write step.
+
+# Replace upstream's infer-api.py with our patched copy. Upstream's
+# version builds an 18-arg positional tuple but the pipeline expects
+# 24 — see infer-api.py header comment for the fix.
+COPY infer-api.py /app/infer-api.py
 
 EXPOSE 8000
 CMD ["python3", "infer-api.py"]