vllm
[ROCm][K3] Dequantize the fp8 decode query for MLA backends without quant-query support - TRITON_MLA
#51860
Merged
Merged
[ROCm][K3] Dequantize the fp8 decode query for MLA backends without quant-query support - TRITON_MLA #51860
hongxiayang
merged
4 commits into
vllm-project:main
from
hongxiayang:k3-dspark-fp8-query-fallback
[ROCm][K3] Dequantize the fp8 decode query for MLA backends without q…
8565b44e
claude
commented
on 2026-08-11
copilot-pull-request-reviewer
commented
on 2026-08-11
hongxiayang
changed the title 
njhill
commented
on 2026-08-11
dllehr-amd
approved these changes
on 2026-08-11
Address review: keep dequantized query in the query dtype; direct att…
ee76dd10
njhill
commented
on 2026-08-12
Lint: collapse cast to one line (ruff-format)
c77a2224
Merge branch 'main' into k3-dspark-fp8-query-fallback
5b387bc4
disabled auto-merge 10 days ago
Manually disabled by user
Manually disabled by user
Assignees
No one assigned
Labels
rocm
ready
quantization
kimi
k3
Milestone
No milestone
Login to write a write a comment.
Login via GitHub