[CuTe DSL] Add modular FMHA prefill and MLA decode attention kernels#2805

Merged

saltyminty merged 41 commits intoflashinfer-ai:mainfrom

pgera:cutedsl-fmha-prefill

Apr 14, 2026

Commits on Apr 1, 2026

[cutedsl] Add modular attention package with FMHA and MLA kernels
pgera
committed
Remove monolithic kernels from original PR, separate benchmarks
pgera
committed
Fix causal mask boundary + PV accumulate bugs in modular attention kernels
pgera
committed
Fix attention sink precision bugs: wrapper dtype mismatch and M_D_update domain conversion
pgera
committed
Fix sliding window mask, sink M_D_update scaling, and add comprehensive tests (AI-assisted)
pgera
committed
Remove MLA/decode files from prefill PR
pgera
committed
Replace patch/pipeline.py with upstream cutlass.pipeline
pgera
committed
Address review feedback: add @cute.jit to get_trip_count, remove unused tmem_utils.py
pgera
committed
Simplify RESIDUAL_MASK check and remove redundant qo_head_idx guard (AI-assisted)
pgera
committed
Unify variant runtime data into single params mechanism with score_mod hook (AI-assisted)
pgera
committed
Update copyright year to 2026 for new CuTe DSL attention files
pgera
committed
Update copyright year to 2026 for benchmark file
pgera
committed
Standardize license headers to abbreviated SPDX BSD-3-Clause
pgera
committed
Add validation and robustness improvements to CuTe DSL attention (AI-assisted)
pgera
committed
Fix SigmoidAttention bias not converted to log-base-2 domain (AI-assisted)
pgera
committed
Add transform_logits support: coordinate-free API, composable with score_mod, SigmoidTanhAttention (AI-assisted)
pgera
committed

Commits on Apr 3, 2026

Port monolithic MLA decode kernel to modular role-based architecture (AI-assisted)
pgera
committed

Commits on Apr 8, 2026

Commits on Apr 9, 2026

Commits on Apr 10, 2026

Guard against cute-dsl backend in paged KV cache wrapper (AI-assisted)
pgera
committed