feat(deploy): vendor ace-step as submodule, skip nano-vllm via sys.path injection

ace-step's pyproject pins 'nano-vllm; sys_platform != darwin' which isn't
on PyPI (their README says manually 'cd acestep/third_parts/nano-vllm &&
pip install .'). All nano-vllm imports in ace-step are lazy (function-
scoped, try/except wrapped), so the package works without it.

Vendored at vendor/ace-step (git submodule of the apple-silicon fork
which has a looser gradio pin >=6.5.1). requirements.txt now lists
ace-step's transitive deps directly. app.py adds vendor/ace-step to
sys.path BEFORE importing ace_pipeline so 'from acestep import ...'
resolves to the vendored copy.

Upstream updates: 'git submodule update --remote vendor/ace-step'.

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "vendor/ace-step"]
+	path = vendor/ace-step
+	url = https://github.com/clockworksquirrel/ace-step-apple-silicon.git

app.py

@@ -46,11 +46,22 @@ os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1")
 # Don't pin HF download source — let HF default for both Spaces and local cache.
 os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
 
+# Vendored ace-step (git submodule at vendor/ace-step/) — added to sys.path
+# BEFORE any module that imports `from acestep import ...`. We vendor
+# instead of pip-installing because the upstream pyproject.toml declares
+# `nano-vllm; sys_platform != "darwin"`, a path-source dep not on PyPI
+# that breaks `pip install -r requirements.txt` on HF Spaces (Linux).
+import sys
+from pathlib import Path
+
+_VENDORED_ACE_STEP = Path(__file__).resolve().parent / "vendor" / "ace-step"
+if _VENDORED_ACE_STEP.exists() and str(_VENDORED_ACE_STEP) not in sys.path:
+    sys.path.insert(0, str(_VENDORED_ACE_STEP))
+
 import hashlib
 import random
 import shutil  # noqa: F401  (reserved for future cleanup paths)
 import subprocess
-from pathlib import Path
 
 import gradio as gr
 

pyproject.toml

@@ -13,12 +13,17 @@ target-version = "py311"
 # checkpoints/ holds downloaded model code shipped by upstream HF repos —
 # linting third-party drop-ins isn't our concern. output/ and research/
 # are runtime artefacts / scratch.
-extend-exclude = ["checkpoints", "output", "research", ".venv"]
+extend-exclude = ["checkpoints", "output", "research", ".venv", "vendor"]
 
 [tool.ruff.lint]
 select = ["E", "F", "I", "B", "UP", "RUF"]
 ignore = ["E501"]
 
+[tool.ruff.lint.per-file-ignores]
+# app.py needs sys.path injection (vendored ace-step submodule) BEFORE the
+# remaining imports — E402 (module-level-import-not-at-top) is intentional.
+"app.py" = ["E402"]
+
 [tool.pytest.ini_options]
 addopts = "-m 'not gpu' --tb=short"
 markers = ["gpu: requires a real GPU/MPS device; opt in with -m gpu"]

requirements.txt

@@ -1,21 +1,38 @@
-# gradio is injected by HF Spaces per README sdk_version. The apple-silicon
-# ace-step fork allows gradio>=6.5.1; we keep sdk_version at 6.14.0.
-transformers>=4.45,<5
+# gradio is injected by HF Spaces per README sdk_version (6.14.0).
+# transformers — pinned narrower to match ace-step's range (4.51.0–4.57.x).
+transformers>=4.51.0,<4.58.0
 torch>=2.4
 torchaudio>=2.4
 safetensors>=0.4
 peft>=0.13
 demucs>=4.0
 pyloudnorm>=0.1.1
-soundfile>=0.12
+soundfile>=0.13.1
 librosa>=0.10
 huggingface_hub>=0.25
 hf_transfer>=0.1.9
 numpy>=1.26,<2
-# Use the apple-silicon fork on all platforms — upstream ace-step pins
-# gradio==6.2.0 and requires nano-vllm (not on PyPI) on Linux. The fork
-# declares torch==2.10.0 for Linux + has a looser gradio pin.
-ace-step @ git+https://github.com/clockworksquirrel/ace-step-apple-silicon.git
+# --- ace-step's transitive deps (NOT installing ace-step itself; it's
+# vendored as a git submodule at vendor/ace-step. The submodule's
+# pyproject pins nano-vllm; sys_platform != "darwin" which isn't on PyPI,
+# so we list ace-step's deps directly and bypass the pip install of the
+# package. app.py adds vendor/ace-step to sys.path before importing. ---
+diffusers
+matplotlib>=3.7.5
+scipy>=1.10.1
+loguru>=0.7.3
+einops>=0.8.1
+accelerate>=1.12.0
+fastapi>=0.110.0
+diskcache
+uvicorn[standard]>=0.27.0
+numba>=0.63.1
+vector-quantize-pytorch>=1.27.15
+torchcodec>=0.9.1
+torchao
+toml
+lightning>=2.0.0
+modelscope
 # HF Spaces ZeroGPU. Do NOT pin — HF's ZeroGPU build injects its own version
 # and a pin causes pip-resolve failure. Locally the `import spaces` in
 # app._maybe_spaces_gpu() is wrapped in try/except so absence is fine.

vendor/ace-step

@@ -0,0 +1 @@
+Subproject commit 8637c8299f87faecae815565da5ef52d9a93eb4c