Apple Silicon MPS: official FL2VA ComfyUI INT8 path fails on aten::_int_mm

Local validation report — H3 FL2VA on Apple Silicon

Status

Blocked for a functional local generation through the official quantized ComfyUI path. Installation and model loading succeed; sampling fails on the first iteration because an MPS operator is unavailable.

Machine

Mac Studio Mac14,14 — Apple M2 Ultra, 24 CPU, 60 GPU, 64 GB unified memory, Metal 4, macOS 26.5.2.

Tested backend

ComfyUI stock 0.33.0 + PyTorch 2.13.0 + MPS, with no additional H3 custom node and no Sage Attention. Outside the sandbox, mps_available=True and ComfyUI detects Device: mps.

Models

• minimax_h3_fl2va_pruned_int8_convrot.safetensors — ~20 GiB
• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors — ~15 GiB
• Video VAE — ~4.8 GiB
• Audio VAE — ~577 MiB

The FL2VA download total was confirmed at 42.5 GiB before transfer. Ref2VA and BF16/full variants were not downloaded.

First test

Native local T2VA workflow, audio enabled, 124 frames at 24 fps (~5 seconds), 608x352 preview, 20 steps, fixed seed and simple prompt.

The prompt was accepted locally (node_errors: {}). The text encoder loaded 14,960 MB and the DiT loaded 19,996 MB. Sampling failed after 47.92 seconds at 0/20 steps. No MP4 was produced.

Error

NotImplementedError: The operator 'aten::_int_mm' is not currently implemented for the MPS device

The error occurs in the INT8 quantized comfy_kitchen path used by the DiT ConvRot implementation. PyTorch suggests PYTORCH_ENABLE_MPS_FALLBACK=1, but that would move the operation to CPU and would not constitute a satisfactory native MPS baseline.

Memory compression and swap were also non-zero after the failed run.

Recommendation

Please clarify whether Apple Silicon/MPS is expected to support the official INT8/NVFP4/AWQ ComfyUI workflow. A Diffusers/MPS route or an explicit BF16/CPU route may be worth testing separately. Until aten::_int_mm is implemented for MPS, the current quantized ComfyUI path should not be treated as a working Apple Silicon baseline.

Start by reproducing the native T2VA workflow with the official INT8 ComfyUI path on MPS and capture the aten::_int_mm failure in the comfy_kitchen DiT ConvRot path. Check whether the limitation belongs to PyTorch MPS or the quantized integration, then establish the supported Apple Silicon route or document that it is unsupported. Done means a functional local generation or a clear, verified support limitation.