JPeetz/agent-skills / skills/marketing/local-video-production-m5
local-video-production-m5
VerifiedUse this skill when producing local videos on the 16GB M5 executor.
| File | Kind | Size |
|---|---|---|
| CHANGELOG.md | markdown | 586 B |
| SKILL.md | skill md | 9.6 KB |
| evals/evals.json | config | 2.4 KB |
| references/comfyui-mps-custom-node-pitfalls.md | markdown | 3.4 KB |
| references/kokoro-tts-mature-voice.md | markdown | 2.1 KB |
| references/ltx-mlx-a2v-setup.md | markdown | 3.9 KB |
| scripts/m5_env_check.shexecutable | script | 665 B |
references/comfyui-mps-custom-node-pitfalls.md
View on GitHub# ComfyUI custom-node pitfalls on Apple Silicon (MPS)
Debugging walls when a CUDA-written custom node refuses to run on the M5's
MPS build. Order they tend to appear, roughly top to bottom:
## 1. Node fails to import (custom nodes never show in /object_info)
Check "IMPORT FAILED" / traceback in the headless server log, not just GUI.
Common causes, each fixed by a small patch (back up the file first):
- **Legacy mediapipe `solutions`** — mediapipe >=1.0 removed `mp.solutions.*`.
Nodes still do `from mediapipe import solutions` or `mp.solutions.face_mesh`
at module level. Patch options: (a) comment out a truly dead import; (b)
install a `.pth` shim in site-packages that installs a permissive dummy on
`mp.solutions` whose `__getattr__` returns nested dummies, `__iter__` yields
empty, and `__call__` raises a clear error — module-level references load,
only actual calls fail.
- **`torchaudio.save` → torchcodec wall** — torchcodec's dylib links
`libavcodec.58` (FFmpeg 4), which the M5 lacks (FFmpeg 9 provides
`libavcodec.63`), and installing FFmpeg@4 needs Xcode CLT 27. The clean fix is
the same as cozui Wav2Lip: patch `torchaudio.save(...)` to write with
`soundfile` or the stdlib `wave` module (mono int16 PCM). See
`creative/comfyui` references/local-lip-sync-wav2lip.md for the model.
- **Missing vendored module** (e.g. a custom node that `from .dist import`
xFuser shims that the git clone didn't include) — obtain it from the
upstream/main repo and copy the file into the node's src tree. The module
degrades gracefully (None fallbacks) once present.
- **`pip install` of a git subpackage + pyproject `readme`** fails with
`Readme path must be within the project directory` when `readme` points
outside the subdir and the PEP 517 hook can't find it. Fix: clone at the
pinned tag, delete the `readme = ...` line from each subpackage's pyproject,
pip install the local clones.
- Missing pip deps surface one at a time (`pyloudnorm`, `decord` has no py3.13
wheel — comment out its import if only used by an unused code path). Iterate:
install → restart → read next error.
## 2. Loads but model files "not found" / empty dropdowns
Custom nodes often resolve model paths relative to the install tree
(`ComfyUI/models/...`) instead of the shared dir that `extra_model_paths.yaml`
maps. Fix: `ln -sfn` the shared model dirs into the install tree
(remove any pre-existing real dir first, or the symlink lands INSIDE it). Also
add missing keys to extra_model_paths.yaml (e.g. `clip_vision:`) or the
loader's option list comes back empty.
## 3. Runs, then a config / device wall during inference
- Omegaconf `Missing key` for a field the node reads from its own config.yaml —
append the missing key with a sensible default (`quantize_transformer: false`).
- mmgp offload profiler calls `torch.cuda` → on MPS set the node's `use_mmgp`
to `None`/`"off"` and rely on block_offload + lowvram + teacache_offload
instead of the CUDA memory profiler.
- Any `torch.cuda.*` call → hard `AssertionError`. Runtime-detect the device in
the node's constructors/defaults.
## Watch the shared flag through every restart
Any rebuild of headless ComfyUI must re-export `PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0`
(and keep the mediapipe `.pth` shim in site-packages). Env like DYLD_LIBRARY_PATH
often doesn't survive the launch + macOS strips it for spawned processes — set
it inline on the launch command (`env DLL=... nohup ... main.py`).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |