[TRTLLM-12807][feat] add CuteDSL FP8/FP16 MLA decode attention fmha lib by brnguyen2 · Pull Request #17266 · NVIDIA/TensorRT-LLM

[None][feat] add CuteDSL FP8/FP16 MLA decode attention backend

haow-nv

committed

e37ef7f

[None][feat] CuteDSL MLA decode: fp8 scaling fix, BF16 support, per-layer paged-KV, multi-layer test

haow-nv

committed

4c01205

[None][fix] CuteDSL MLA decode: v1 paged-KV layout + CUDA-graph-safe fp8 scale

haow-nv

committed

7678862

[TRTLLM-12807][feat] Wrap CuteDSL MLA decode as FMHA lib

yuxianq

authored and

haow-nv

committed

748ffeb

[None][feat] CuteDSL MLA decode: MTP causal mask + can_implement gate

haow-nv

committed

66159f1

[None][feat] CuteDSL MLA decode: persistent + split-KV alignment (flashinfer #2743/#3309)

haow-nv

committed

0fe7413

[None][feat] CuteDSL MLA decode: mixed prefill+decode page-table offset, cache-key fix, TRTLLM-Gen mixed warmup

haow-nv

committed

7aed06f

[None][feat] CuteDSL MLA decode: autotuned is_persistent tactic, perf whitelist gate, op schema cleanup

haow-nv

committed

cb20915

[None][chore] CuteDSL MLA: drop stale is_available comment

haow-nv

committed

0960936

[None][chore] CuteDSL MLA decode: comment cleanup, drop default_is_persistent, gate-aware test asserts

haow-nv

committed

722a7de

[None][chore] CuteDSL MLA decode: apply pre-commit formatting, fix comment typos

haow-nv

committed

e396dc1

[None][chore] CuteDSL MLA decode: drop standalone decode unit test

haow-nv

committed

5d6bc50

[None][fix] CuteDSL MLA: reject sparse attention in FMHA gate

haow-nv

committed

9bc6bc1

[None][chore] CuteDSL MLA decode: trim comments, drop unused runner init params

haow-nv

committed

b082608

[None][perf] CuteDSL MLA: gate (128,1) decode on spec-decode off

haow-nv

committed

cb08644

[None][perf] CuteDSL MLA: batch-aware perf gate, autotuner decode warmup, max-batch plumbing

haow-nv

committed

3968be1

[None][chore] CuteDSL MLA: apply pre-commit formatting

haow-nv

committed

de72d57

[None][perf] CuteDSL MLA: restructure decode perf gate, drop (128,1) spec-decode special-case

haow-nv

committed

d0949c6

[None][fix] stabilize CuteDSL MLA split-KV workspace for CUDA graphs

haow-nv

committed

66b8e90

[None][chore] CuteDSL MLA decode: drop debug kernel-arg dumps, fix autotuner free dim, trim comments

haow-nv

committed

a9d21f5

[None][fix] CuteDSL MLA decode: reject Helix, slice AutoTuner workspace, trim gate

haow-nv

committed

60b0425

[None][fix] CuteDSL MLA decode: tighten split_kv and batch gating

haow-nv

committed

2915a9e

[None][chore] CuteDSL MLA decode: rename fmha/cute_dsl.py to cute_dsl_mla.py

haow-nv

committed

a575f97

[None][fix] CuteDSL MLA decode: keep dtype/heads gate during autotuning

haow-nv

committed

faf5aba

[None][fix] CuteDSL MLA decode: drop (128,4)/(128,8) from fp8 perf allowlist

haow-nv

committed

346fb8c

[None][fix] CuteDSL MLA decode: bucket the fallback tactic's batch size

brnguyen2

committed

827cc9e

[None][test] CuteDSL MLA decode: cover the AutoTuner tuning path

brnguyen2

committed

9f9e319

[None][test] Add disagg decode-only smoke for the CuTe DSL MLA FMHA lib

brnguyen2

committed

ee0319b

[None][chore] Rename CuteDslMlaFmha FMHA lib to CuteDslFmha

brnguyen2

committed

780fa9b

Navigation Menu

[TRTLLM-12807][feat] add CuteDSL FP8/FP16 MLA decode attention fmha lib - #17266

[TRTLLM-12807][feat] add CuteDSL FP8/FP16 MLA decode attention fmha lib#17266
brnguyen2 wants to merge 29 commits into
NVIDIA:mainfrom
brnguyen2:prbs/mirror/pr-15138

Commits on Aug 4, 2026

[None][feat] add CuteDSL FP8/FP16 MLA decode attention backend

[None][feat] CuteDSL MLA decode: fp8 scaling fix, BF16 support, per-layer paged-KV, multi-layer test

[None][fix] CuteDSL MLA decode: v1 paged-KV layout + CUDA-graph-safe fp8 scale

[TRTLLM-12807][feat] Wrap CuteDSL MLA decode as FMHA lib

[None][feat] CuteDSL MLA decode: MTP causal mask + can_implement gate

[None][feat] CuteDSL MLA decode: persistent + split-KV alignment (flashinfer #2743/#3309)

[None][feat] CuteDSL MLA decode: mixed prefill+decode page-table offset, cache-key fix, TRTLLM-Gen mixed warmup

[None][feat] CuteDSL MLA decode: autotuned is_persistent tactic, perf whitelist gate, op schema cleanup

[None][chore] CuteDSL MLA: drop stale is_available comment

[None][chore] CuteDSL MLA decode: comment cleanup, drop default_is_persistent, gate-aware test asserts

[None][chore] CuteDSL MLA decode: apply pre-commit formatting, fix comment typos

[None][chore] CuteDSL MLA decode: drop standalone decode unit test

[None][fix] CuteDSL MLA: reject sparse attention in FMHA gate

[None][chore] CuteDSL MLA decode: trim comments, drop unused runner init params

[None][perf] CuteDSL MLA: gate (128,1) decode on spec-decode off

[None][perf] CuteDSL MLA: batch-aware perf gate, autotuner decode warmup, max-batch plumbing

[None][chore] CuteDSL MLA: apply pre-commit formatting

[None][perf] CuteDSL MLA: restructure decode perf gate, drop (128,1) spec-decode special-case

[None][fix] stabilize CuteDSL MLA split-KV workspace for CUDA graphs

[None][chore] CuteDSL MLA decode: drop debug kernel-arg dumps, fix autotuner free dim, trim comments

[None][fix] CuteDSL MLA decode: reject Helix, slice AutoTuner workspace, trim gate

[None][fix] CuteDSL MLA decode: tighten split_kv and batch gating

[None][chore] CuteDSL MLA decode: rename fmha/cute_dsl.py to cute_dsl_mla.py

[None][fix] CuteDSL MLA decode: keep dtype/heads gate during autotuning

[None][fix] CuteDSL MLA decode: drop (128,4)/(128,8) from fp8 perf allowlist

[None][fix] CuteDSL MLA decode: bucket the fallback tactic's batch size

[None][test] CuteDSL MLA decode: cover the AutoTuner tuning path

[None][test] Add disagg decode-only smoke for the CuTe DSL MLA FMHA lib

[None][chore] Rename CuteDslMlaFmha FMHA lib to CuteDslFmha