RDNA1/RDNA2: comfy-kitchen Triton backend auto-enabled by #14862 hangs INT8 models

[liminfei-amd]

Problem

#14862 auto-enables the comfy-kitchen Triton backend on all ROCm GPUs (Triton >= 3.7). The INT8 matmul kernels need matrix cores — WMMA on RDNA3+/RDNA4 (gfx11xx/gfx12xx), MFMA on CDNA (gfx9xx). RDNA1/RDNA2 (gfx10xx) have neither, so the INT8 path hangs the GPU instead of falling back to eager.

Report

RDNA2 + triton-windows 3.7.1: both native "load diffusion model" and the custom-node INT8 path freeze until a reset (Comfy-Org/comfy-kitchen#69).

Fix

Gate the ROCm auto-enable on GPU arch (only WMMA/MFMA archs) so RDNA1/RDNA2 stay on eager; add --disable-triton-backend. Workaround today: run the v0.27.0 release (predates #14862).

[coderabbitai]

⚠️ Possible Duplicate Issue(s)

• Using the now mandated "triton-backend" completely freezes PC with AMD rx 6800 on windows #14866

🔗 Related PRs

#14862 - Enable comfy-kitchen Triton backend by default on ROCm/AMD [merged]

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[0xDELUXA]

@liminfei-amd As I mentioned in #14866 (comment), on my setup (RDNA4, Windows), enabling comfy-kitchen's Triton backend by default (introduced in 099522f) consistently triggers an "unspecified launch failure" error at the first step of generation, no matter the workflow. I've tried updating triton-windows (tested 3.7.1 and 3.8.0), and staying on an outdated version isn't a real solution. At the moment, the only way I can use ComfyUI is by launching it with --disable-triton-backend.

[Jules-A]

@liminfei-amd As I mentioned in #14866 (comment), on my setup (RDNA4, Windows), enabling comfy-kitchen's Triton backend by default (introduced in 099522f) consistently triggers an "unspecified launch failure" error at the first step of generation, no matter the workflow. I've tried updating triton-windows (tested 3.7.1 and 3.8.0), and staying on an outdated version isn't a real solution. At the moment, the only way I can use ComfyUI is by launching it with --disable-triton-backend.

I assume you mean using the official Comfyui? It's working fine for me on the ROCm fork with 9070 XT and Triton 3.7.1 with Anima Int8/BF16, SDXL BF16, ZImage Turbo Int8, PiD Int8. I haven't tried any heavier models though. I did update a lot of the other dependencies though and am using flash-attention instead of sage-attention.

EDIT: Krea2-Turbo instantly crashes, BF16 and a downloaded Int8 Convrot. It doesn't seem to be Triton though as it still occurs with it disabled.