inference/research
verifiedhigh · 0.90updated 2026-10-03

vLLM issue #50264: hybrid-Mamba models fall back to Triton paged attention on RDNA and decode collapses

forum

Fields

id
sources/gotcha-vllm-issue-50264-rdna-triton-paged-attn
URL
https://github.com/vllm-project/vllm/issues/50264
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-07-29
Accessed
2026-10-03
notes
Kernel-level profile table attributes the whole regression to kernel_paged_attention_2d; reporter notes the upstream fix widened the gate inside the PR.

Notes

Kernel-level profile table attributes the whole regression to kernel_paged_attention_2d; reporter notes the upstream fix widened the gate inside the PR.

Cross-references

This record references no other records.