Black image (NaN) with int8_convrot diffusion models on gfx1201 / ROCm, but int8_convrot text encoder works fine

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Your question

Environment

• GPU: AMD Radeon RX 9070 XT (gfx1201, RDNA4)
• OS: Windows 11
• PyTorch: 2.9.1+rocm7.2.1, ROCm 7.2
• ComfyUI: v0.28.0-42-gf966a2b3
• comfy-kitchen: 0.2.22
• Python 3.12.10 (portable AMD build)
• Args: --windows-standalone-build --enable-triton-backend

Description

Loading a Krea2 Turbo int8_convrot model in UNETLoader produces an all-black image. Sampling completes without error; the failure surfaces at save time:

nodes.py:1682: RuntimeWarning: invalid value encountered in cast
  img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))

which indicates NaN/inf in the decoded tensor.

Notably, the same quantization format works correctly for the text encoder. Only the diffusion model path is affected.

Reproduction matrix (1024x1024, 16 steps, denoise 0.55, er_sde/simple)

UNET	Text encoder	Result	s/it
int8_convrot	int8_convrot	black	14.8
int8_convrot	fp8_scaled	black	16.6
fp8_scaled	int8_convrot	OK	36.7
fp8_scaled	fp8_scaled	OK	33.6

Models: Comfy-Org/Krea-2 (krea2_turbo_*, qwen3vl_4b_*), qwen_image_vae.

Backend info

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

comfy_kitchen cuda backend reports convrot capabilities but is disabled on ROCm; only the eager backend is active for the convrot path.

Note

int8_convrot is ~2.2x faster than fp8_scaled here, so this path is worth fixing for RDNA4. Possibly related to #14864 (black image on Turing with int4 models).

Logs

Other

No response