[Bug]: aiter.ops.triton.attention.pa_mqa_logits.deepgemm_fp8_paged_mqa_logits_stage1` returns random topk for `context_len > 2048` on ROCm (gfx950), breaks GLM-5.1-FP8 decode
April 8, 2026 ยท #39303
Python
Difficulty: Easy
Labels
bug rocm
Parent Repository
vllm-project/vllm
Python repository
75,721 15,332