[TRTLLM-14597][perf] Fuse the DSv4 MLA prologue: kv_a_layernorm, q_nope FP8 quant and Q RoPE - #17273

Draft

dc3671 wants to merge 18 commits into

NVIDIA:mainfrom

dc3671:user/zhenhuanc/dsv4-kv-norm-fusion-main

Commits on Aug 5, 2026

[None][perf] Fuse kv_a_layernorm into the DSv4 context RoPE kernel
dc3671
committed
d8972bc
[None][perf] Fuse kv_a_layernorm into the DSv4 generation KV path
dc3671
committed
8715871
[None][fix] Keep the fused kv-norm flag alive for the generation half
dc3671
committed
6c3d15c
[None][perf] Rehome the DSv4 FMHA scheduler prologue, and fix a PDL race
dc3671
committed
711e949
[None][perf] Pick the fused kv-norm block size by measurement, not occupancy
dc3671
committed
976f4fc
[None][perf] Drop the q_nope FP8 requantize from the DSv4 decode RoPE kernel
dc3671
committed
df46a4f
[None][perf] Fold the Q RoPE into q_b_layernorm for DSv4 decode
dc3671
committed
b9529b8
[None][perf] Fold the Q RoPE into q_b_layernorm for DSv4 context too
dc3671
committed
51561b9
[None][perf] Split the DSv4 context KV prologue into its own kernel
dc3671
committed
660e831
[None][perf] Tighten the DSv4 fused kv-norm and q-norm kernels
dc3671
committed
e60e81e
[None][perf] Pre-launch the DSv4 compressor and hoist the fused KV kernel by default
dc3671
committed
84c95bf
[None][chore] Drop the DSv4 fusion debug gates and experiment modes
dc3671
committed
a771f0f
[None][perf] Keep the DSv4 prologue off the shared MLA RoPE kernels
dc3671
committed
369daeb

Commits on Aug 6, 2026