- bench/failures.py: per-GPU success rates and error-class breakdown by config
- bench/validate_outputs.py: ffprobe every recorded output via /view
- finding: 24GB cards fail on multi-ref (TE vision OOM), not duration/size;
100% of successful runs produce valid videos
- outputs now written to /workspace (survives pod resets)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- comfyui/: extra_model_paths.yaml + the three MiniMax H3 workflow JSONs
- push-to-pod.sh stages everything onto the persistent volume
- install-minimax-h3.sh rebuilds a wiped container from that stage
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The network volume was resized to 1TB (which reset the pod and wiped the
container disk). All weights now live persistently on /workspace; the
/dev/shm overflow scheme is no longer needed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- refheavy suite: ref2v saturation with 1/4/8 refs, the priority modality
- family suite: reduced grid for int8/bf16 family comparisons
- 8 distinct 1344x768 reference scenes generated into /ComfyUI/input
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Hanashi