MOSS-Image int8_convrot: HIPBLAS_STATUS_INVALID_VALUE in torch._int_mm on RX 6900 XT (gfx1030, ROCm 7.2)
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Newbie friendliness
- 48/100
- Issue type
- Bug
- Clarity
- Mostly clear
- Activity status
- Active
- Domain
- backend, machine-learning
Research direction
Start by tracing the int8_convrot model path used by ComfyUI and how it reaches torch._int_mm on ROCm; the report does not name a source file or test. Compare that path with the working bf16 variant on the RX 6900 XT and reproduce the failure using the supplied MOSS-Image template at 1024×1024. Done means int8 sampling completes without the HIPBLAS_STATUS_INVALID_VALUE error and produces an image.
Written by the indexing model from the issue text.
Description
Custom Node Testing
- I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)
Expected Behavior
The MOSS-Image int8_convrot model should complete sampling on an RX 6900 XT, as the bf16 variant does.
Actual Behavior
Generation stops at the first sampling step.
The node is marked as failed, and the prompt is aborted. No image is produced. The same error appears on every attempt with the int8 models, including with --use-split-cross-attention, and no output file is written.
Steps to Reproduce
Launch ComfyUI with python main.py --listen
Load the MOSS-Image template with moss_image_sft_int8_convrot.safetensors
Queue a generation at 1024×1024
Debug Logs
RuntimeError: CUDA error: HIPBLAS_STATUS_INVALID_VALUE when calling
`hipblasLtMatmulAlgoGetHeuristic(ltHandle, computeDesc.descriptor(), Adesc.descriptor(),
Bdesc.descriptor(), Cdesc.descriptor(), Cdesc.descriptor(), preference.descriptor(), 1,
&heuristicResult, &returnedResult)`
Other
Enabling --use-split-cross-attention (log confirms “Using split optimization for attention”) does not change the error, so attention is not the cause.
The same template with the bf16 model runs correctly on this card.
Flux.1 Krea (fp8 scaled) also runs on this card.
- Dominant language
- Python
- Stars
- 146k
- Forks
- 17.4k
- Avg merge
- 19h 44m
- Merged PRs (30d)
- 178
Getting set up
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from Comfy-Org/ComfyUI
-
Difficulty 2/5 1-3 hours Newbie friendliness 75/100
Maintainers usually reply within 1 day
-
ROCm gfx1100: contiguous-k/v SDPA workaround is gfx1151-only, MiniMax H3 sampling 2.2x slower on RX 7900 XTXPossibly taken @TwelveNights claimed this 3 days ago. OpenPotential Bug
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
Maintainers usually reply within 1 day
-
Sub-quadratic attention is ~10x slower on long sequences on RDNA2 (gfx1030) because the key passed to bmm is non-contiguousPossibly taken @KornienkoDev claimed this 3 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 75/100
Maintainers usually reply within 1 day
-
MoGe-3 output is nondeterministic on CUDA: sparse_pool3d_mean sums with float index_add_ (atomics)Possibly taken @potoior claimed this 10 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
Comfy-Org/ComfyUI#17498 · 1 comment ·
Maintainers usually reply within 1 day
-
TRELLIS.2 texture bake fails under --bf16-vae: BakeTextureFromVoxel 'Got unsupported ScalarType BFloat16'Possibly taken @Shubham-Padkonde claimed this 11 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 65/100
Maintainers usually reply within 1 day
All issues in Comfy-Org/ComfyUI
Similar issues
-
add entry to blockOpen
Difficulty 1/5 Under an hour Newbie friendliness 65/100
StevenBlack/hosts#3299 ·
-
`garak --report` overwrites the input report when its name has no ".report"Possibly taken @0z5a claimed this today. Openneeds-triage
Difficulty 2/5 Half a day Newbie friendliness 72/100
Maintainers usually reply within 2 days
-
plot_graph crashes when a mesh node has no m2m edgePossibly taken @tygvi claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
mllam/neural-lam#883 · 1 comment ·
Maintainers usually reply within 2 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
future-agi/future-agi#4092 · 1 comment ·
Maintainers usually reply within 1 day
-
transcript analytics: a text whose only protected spans the prose pass redacts is withheld wholeOpen
Difficulty 1/5 Under an hour Newbie friendliness 82/100
WingedGuardian/GENesis-AGI#4641 ·
Maintainers usually reply within 1 day