supports_int8_compute() disables native INT8 on MPS unconditionally — torch._int_mm has been implemented for MPS since Aug 14 2026

[designis1gt-alt]

Summary

comfy/model_management.py::supports_int8_compute() returns False for MPS
unconditionally. The inline comment states that PyTorch does not implement
torch._int_mm for MPS and links pytorch/pytorch#141287.

That premise is out of date. Native _int_mm for MPS landed in PyTorch on
2026-08-14 (PR pytorch/pytorch#193153, commit 65366750d9, "[MPS] Add native
_int_mm support"). On a current nightly it works and is bit-exact against CPU.

The commit that introduced the gate — 313a76fb (#16130) — is titled "Disable
int8 weight-only quantization on devices without torch._int_mm", but the
implementation disables it on all MPS devices without probing the capability.

Effect

On macOS 26 + PyTorch nightly ≥ 2.15.0.dev20260815, every INT8 path silently
falls back to emulation. There is no error and no warning — only the model-load
line changes:

Before (ComfyUI 0.33.1):

Native ops: convrot_w4a4, int8_tensorwise, asym_w4a8_int8 , emulated ops: nvfp4, mxfp8, float8_e4m3fn, float8_e5m2

After (ComfyUI 0.37.0):

Native ops: , emulated ops: convrot_w4a4, mxfp8, int8_tensorwise, nvfp4, float8_e4m3fn, float8_e5m2, asym_w4a8_int8

Measured on a MacBook Pro 16" M5 Pro, 48 GB unified, macOS 26.6.2, with
torch 2.15.0.dev20260816 + torchvision 0.30.0.dev20260816, started with
env -u PYTORCH_ENABLE_MPS_FALLBACK, reference workflow Z-Image Turbo INT8
(z_image_turbo_int8_convrot.safetensors), 1024x1024, 8 steps,
res_multistep/simple, cfg 1.0, warm runs only:

Build | Warmed reference time | Native ops -- | -- | -- 0.33.1 | 30.09 s | three ops 0.37.0, unpatched | 32.09 s (+6.6%) | none 0.37.0 + patch below | 30.08 s | three ops

Verification that _int_mm really works on MPS here

python

import torch
for M, K, N in [(64,128,64), (256,512,256), (1024,1024,1024)]:
    a = torch.randint(-128, 127, (M, K), dtype=torch.int8)
    b = torch.randint(-128, 127, (K, N), dtype=torch.int8)
    assert torch.equal(torch._int_mm(a, b), torch._int_mm(a.to("mps"), b.to("mps")).cpu())
print("bit-exact on all three sizes")

All three pass on the configuration above.

Suggested fix

Probe the capability instead of denying by device type — this matches the stated
intent of #16130 and keeps the crash fix for builds where _int_mm is genuinely
missing:

python

_MPS_INT_MM_WORKS = None
def _mps_int_mm_works():

global _MPS_INT_MM_WORKS

if _MPS_INT_MM_WORKS is None:

try:

a = torch.randint(-8, 8, (32, 32), dtype=torch.int8, device="mps")

b = torch.randint(-8, 8, (32, 32), dtype=torch.int8, device="mps")

torch._int_mm(a, b)

_MPS_INT_MM_WORKS = True

except Exception:

_MPS_INT_MM_WORKS = False

return _MPS_INT_MM_WORKS
def supports_int8_compute(device=None):

if (device is not None and is_device_mps(device)) or mps_mode():

return _mps_int_mm_works()

...

Happy to open a PR if that shape is acceptable.

Environment

• ComfyUI 0.37.0 (73c9bad4)
• macOS 26.6.2 (build 25G83), MacBook Pro 16" M5 Pro, 48 GB unified
• Python 3.12.13, torch 2.15.0.dev20260816, torchvision 0.30.0.dev20260816
• comfy-kitchen 0.2.35, comfy-aimdo 0.5.5
• Launched with env -u PYTORCH_ENABLE_MPS_FALLBACK python main.py --use-pytorch-cross-attention

[Isalia20]

I wasn't aware this issue existed. I submitted PR yesterday here:
#16451