[None][fix] Forward layer_idx for VSWA + narrow FLASHINFER+spec rejection - #15692

Draft

mihai-chiorean wants to merge 18 commits into

NVIDIA:mainfrom

mihai-chiorean:fix/flashinfer-vswa-sm121

Commits on Jun 28, 2026

[None][fix] Forward layer_idx to KV cache for FlashInfer / Vanilla backends (VSWA)
mihai-chiorean
committed
631dc3d
[None][fix] Lift FLASHINFER + speculative decoding rejection (vLLM-style)
mihai-chiorean
committed
b504384
[None][fix] Handle V1 KVCacheManager VSWA in FlashInfer prepare() (follow-up)
mihai-chiorean
committed
464e91b
[None][fix] Fix VSWA detection in FlashInfer prepare() for mixed-window models
mihai-chiorean
committed
5ccb9f3
[None][fix] Set blocks_in_primary_pool in VSWA / linear-attention branch (closes FIXME)
mihai-chiorean
committed
a01b7d4
[None][fix] Populate Mamba2Metadata.query_start_loc on pure-decode path
mihai-chiorean
committed
cc0aa43
[None][fix] Add spec_decoding_* fields to FlashInferAttentionMetadata
mihai-chiorean
committed
8d90877

Commits on Jun 29, 2026