Repository navigation
v1.5.2
What's Changed
- [benchmarks] Compare the linear-attention examples across backends by @jansel in #3996
- [cutedsl] Stop seeding collective MMA for sites the lowering refuses by @jansel in #3997
- [runtime] Cut the per-call host cost: prepared guards, binding plans by @jansel in #3998
- [autotuner] Run the execution default as the accuracy baseline by @jansel in #3999
- [cutedsl] cute_mma_dag: warp-MMA lowering for single-loop dot programs by @jansel in #4000
- [cutedsl] cute_mma_dag="tcgen05": state recurrences and K loops on TMEM by @jansel in #4001
- [megakernel] Warp-specialized static task loops with readiness counters by @ethche in #3914
- [cutedsl] cute_mma_dag warp: region tree, reductions and the tf32 chain by @jansel in #4002
- [cutedsl] cute_mma_dag tcgen05: delta state kernels and region programs by @jansel in #4003
- [cutedsl] cute_mma_dag warp: hoisted loads, overlaid tiles, f32x2 pairs by @jansel in #4004
- [cutedsl] cute_mma_dag tcgen05: pass warps, TMA stores, MN-major tiles by @jansel in #4005
- [cutedsl] cute_mma_dag regions: aux ring tiles, wy chain, seed ranking by @jansel in #4006
- [cutedsl] launcher: C-level relaunch guards, alias and alignment checks by @jansel in #4007
- [compiler] Balance dispatch across root loops by @ZenAlexa in #3906
- [cutedsl] Fix a BF16 rounding flake in the top-k normalization test by @yushangdi in #4017
- [cutedsl] Fused tcgen05 backward for the FlexAttention envelope by @jansel in #4046
- [cutedsl] Flex backward: block order, SMEM budget, statistics MMAs, polynomial exp2 by @jansel in #4047
- [cutedsl] Flex backward: CLC persistent list walk (cute_flash_bwd_persistent) by @jansel in #4048
- [cutedsl] Backward flash search surfaces drop epilogue_subtile by @jansel in #4049
- [cutedsl] Flex backward: stage the hook's gathered aux tables in shared memory by @jansel in #4050
- [cutedsl] Flex backward: the dK MMA reads its dS^T operand from TMEM (head dim 64) by @jansel in #4051
- [cutedsl] Flex backward: review fixes and tests for the non-default knobs by @jansel in #4052
- Fix helion cute_backend test listing broken by CuTe DSL 4.7-only imports by @jansel in #4054
- Fix test_compiled_config_refuses_a_larger_extent_bucket on Triton backend (#4055) by @jansel in #4055
- Fix fatal tpx skip event in test_host_integer_power_keeps_runtime_scalar_origin (#4056) by @jansel in #4056
- Fix test_full_slice_of_block_sized_buffer_reduction after CuTe gained the lowering (#4057) by @jansel in #4057
- [cutedsl] Pack FP8 persistent reduction stores by @yushangdi in #4015
- [cutedsl] Add proven nc_l2_first packet loads and an l2_evict_last packet store policy by @yushangdi in #4016
- [cutedsl] Seed short persistent row reductions and admit them to proven bounds by @yushangdi in #4018
- [cutedsl] Add guarded dense tcgen05 schedules and output variants by @yushangdi in #4021
- Support safe affine one-trip consumer packet prefetch by @yushangdi in #4039
- Admit block-dimension geometry in consumer packet prefetch by @yushangdi in #4040
- Reject rolled-cluster padding beyond signed Int32 bounds by @yushangdi in #4041
- Seed bounded cluster and full-row register packets by @yushangdi in #4042
- [cutedsl] Improve narrow pointwise tile seeds by @coder-2011 in #4013
- Keep the benchmark phase in-process on MTIA (#4058) by @jansel in #4058
- Gate CuTe-only unittest tests with unittest-visible skips by @jansel in #4078
- [CI] Stop the flaky ref-eager shard from blocking the stack by @jansel in #4074
- [benchmarks] example comparison: harness fixes from the full example sweep by @jansel in #4059
- [benchmarks] Mamba2 chunk-state backend comparison harness by @jansel in #4060
- [examples] Treat an FLA install that fails to import as unavailable by @jansel in #4061
- [examples] nvfp4_gemm: build the FP4 decode table on the device by @jansel in #4062
- [examples] moe_matmul_ogs: mask padding rows out of the output store by @jansel in #4063
- [autotuner] Give a SIGKILLed benchmark worker 60s to exit before it counts as unreapable by @jansel in #4064
- [cutedsl] flash: let examples/attention.py autotune on the fused tcgen05 families by @jansel in #4065
- [cutedsl] tcgen05: stage an M-major A through TMA; refuse two configs that fault or hang by @jansel in #4066
- [cutedsl] computed fragments: resolve iota start/step nodes; leave grid-axis carries alone by @jansel in #4067
- [cutedsl] mma_dag: admit chunk-offset tiles, leading tile begins, torch.matmul dots, tile.index masks and int32 side loads by @jansel in #4068
- [cutedsl] mma_dag tcgen05: hoist invariant side loads, evaluate row-only pass values once per row, fold the staging tile into the smem estimate by @jansel in #4069
- [cutedsl] mma_dag tcgen05: persistent launch knob with cross-tile ring prefetch by @jansel in #4070
- [cutedsl] jagged: promote the nested-row seed to the autotune-off default; keep feature-axis loads scalar by @jansel in #4076
- [cutedsl] Search ordered phase launches for hl.barrier() kernels by @jansel in #4072
- Skip test_cute_unsigned_scalar_load_sdk_type without a supported CuTe DSL (#4081) by @jansel in #4081
- [cutedsl] Flex forward/backward: fused paths take strided Q/K/V/dO/O/dK/dV by @jansel in #4082
- [cutedsl] Fused flash forward at every head dim that is a multiple of 16 up to 128 by @jansel in #4083
- [cutedsl] Flex backward: head dims beyond 64 / 128, the MLA pair and a chunked step table by @jansel in #4084
- [autotuner] Streaming budgeted population, forked CuTe render pool, flash qualification floor by @jansel in #4085
- [cutedsl] Flex forward: the cursor walk over the item's bitmask words and one PASS2 per step by @jansel in #4086
- [cutedsl] Flash forward above head dim 128: the cooperative pair (ws_2cta) and the flex walk on it by @jansel in #4087
- [cutedsl] Flex list walk: shared softmax region, resident PASS2 regions, host flags, split-KV layout by @jansel in #4088
- [cutedsl] Two-stage bodies: issue-stream knobs; flex list-walk seeds and the all-kept fast path by @jansel in #4089
- [cutedsl] Flex backward at head dim 256: the two-role pair kernel with Q-major dQ clusters by @jansel in #4090
- [cutedsl] Flex backward: the 2-CTA cluster family and the 1-CTA walk's boundary knobs by @jansel in #4091
- [cutedsl] fa4 list walk: tail split, early next-item QK, one staged output tile, P-handoff fences by @jansel in #4092
- [cutedsl] Flex forward: compiler seeds reach the flex surface; a leaner host launch path by @jansel in #4093
- [cutedsl] Generic lowering: a full-slice store beside a rolled reduction (resolver rule, own chunk lane loop) by @jansel in #4094
- [runtime] Prepared-call dispatch: one TensorGuards check for plain tensor arguments by @jansel in #4095
- [metal] Support mixed accumulator/output dtype MPP matmul via scalar store by @aditvenk in #3800
- [cute] Vectorize broadcasted integer indices with aligned bounds by @yushangdi in #4079
- [cute] Defer radix selection output and narrow scan barriers by @yushangdi in #4080
- [cute] Add CTA int64 top-k and uniform sampling control flow by @yushangdi in #4102
- [tests] Clean up cache contexts when setup skips or fails by @jansel in #4110
- [cutedsl] extend native MXFP8 for packed scales and registered grids by @yushangdi in #4103
- [cutedsl] add TMA and in-place staging for packed MXFP8 scales by @yushangdi in #4104
- [cutedsl] add optional MXFP8 accumulator and TMA output modes by @yushangdi in #4105
- [cutedsl] measure joint MXFP8 output modes after the initial population cap by @yushangdi in #4106
- [cutedsl] Validate MXFP8 metadata and configuration choices by @yushangdi in #4107
- [TileIR] Add structural autotuner seed configurations by @yushangdi in #4116
- [cutedsl] Add proved native grouped worklists and tuning families by @yushangdi in #4112
- [cutedsl] Compose prefix-adaptive native grouped recipes by @yushangdi in #4113
- Make three tests independent of test order and the runner's environment (#4109) by @jansel in #4109
- [xpu] supports torch_compile template lowering by @etaf in #4118
- [benchmark] Bump tritonbench pin past the os.environ fix by @etaf in #4101
- [cutedsl] Fix heuristic axis and slot mapping by @coder-2011 in #4011
- Avoid retaining tensors in prepared argument guards (#4077) by @mengluy0125 in #4077
- [autotuner] L2 flush regimes for the benchmarks (call / off / replay), batch sizing, final selection in the replay regime, mutated-argument kernels in the killable worker by @jansel in #4126
- [bench] Example catalogue: sustained 16-call graph timer, pinned-config and ABAB tools, replay-regime autotunes by @jansel in #4127
- [autotuner] Grid block floors water-fill each root grid's program budget by @jansel in #4128
- [cutedsl] Exact fp32 SIMT GEMM (cute_simt_mma) and ordered-phase launches for integer-metadata preludes by @jansel in #4129
- [cutedsl] cute_mma_dag: full-slice GEMM + row epilogue on tcgen05 (matmul_layernorm), cluster / fixup row reductions, varying-bound K loops, side-load prefetch and aux ring, TMA warp route, exact fp32 widening, epilogue staging, corner seeds by @jansel in #4130
- [cutedsl] Block-scaled GEMMs: W4A16 scaled contractions, literal-offset load merge, prepare knobs, bn=64 and fused NVFP4 tiles, cute_tiles_per_block packing, two-barrier interleaved reductions by @jansel in #4131
- [cutedsl] GEMMs: fused int16 cast operands, cute_pdl on tcgen05 / warp / collective / grouped kernels, M-major A, collective epilogue-load hoist, TF32 register rounding, thread-layout repair, narrow row-union tiles by @jansel in #4132
- [cutedsl] Reductions: fragment-owner capacity decline, host zero fills, prefetched host sum tails, occupancy-filling resident seeds by @jansel in #4133
- [cutedsl] Lane loops: persistent slice vectors, V-loop sink over sibling lane loops, vectorized flat gathers, packet caches, fold-mask hoist, lane-unroll cap, pointwise / nested-row / reduction-tile seeds by @jansel in #4134
- [examples] addmm backward without unit scalings, grpo_loss gathers the completion logit in the kernels, swiglu under fast_math by @jansel in #4135
- [benchmarks] Example comparison: production linear-attention workloads, FLA baselines, fused_linear_jsd gradients by @jansel in #4137
- [examples] linear_attention_engine: unneeded dg skipped, cu_seqlens checked once; se_block forward tiles N and K by @jansel in #4138
- [autotuner] Search bookkeeping, compile deadline, accuracy-check fixes, mutated-arg kernels in the killable worker by @jansel in #4139
- [cutedsl] Full-slice stripmine, per-phase tcgen05 planning, two-output epilogues, rolled pair, layout/seed rules by @jansel in #4140
- [cutedsl] Flash fwd/bwd: fp8 fa4 body, TMA bias ring, dense flex plan, fa4_alt o_stage pin, row_mma small grids by @jansel in #4141
- [cutedsl] mma_dag admissions and derived pass, fragment tf32 mma.sync and transports, duplicate-reduction CSE by @jansel in #4142
- [cutedsl] mma_dag: region schedules, varlen windows, aux ring, SASS-checked L2 hints, persistent region tiles by @jansel in #4143
- [cutedsl] Register tiles: schedule knobs, an accumulate nest per reduction level, value budget, rolled fallback by @jansel in #4144
- [cutedsl] Exact loop-invariant fp32 division: hoisted Markstein sequence with a div.rn redo (cute_exact_division) by @jansel in #4145
- [examples] JSD budget and gradient, fast_math, se_block / squeeze backward rework, clone-free recurrent step by @jansel in #4146
- [autotuner] Defer benchmark worker pipe cleanup to the waiting thread by @jansel in #4147
- [megakernel] Defer a keyed peer producer's last release into the next root by @ethche in #4117
- [megakernel] Fix cross-loop publication counts and warp-specialized helper barriers by @ethche in #4119
- [triton] Guard tensor-descriptor base alignment and enable cases on XPU. by @etaf in #4098
- [pallas] canonicalize relative jagged row bounds for ordered carry by @cota in #3961
- [megakernel] Add Muse Spark speculative-decoding verify megakernels by @ethche in #4148
- [cute] Split exact division config tests to avoid CI timeouts by @jansel in #4177
- [cute] Keep reduction seed chunks legal by @coder-2011 in #3912
- [autotuner] Vectorize dynamic CuTe reduction seeds by @coder-2011 in #3922
- Fix test_fused_linear_jsd_fwd timeout: pin num_warps=16 by @jansel in #4174
- [cutedsl] Allow two scheduler stages for single-CTA grouped N,M worklists by @yushangdi in #4136
- Fix lint error on main branch: remove unused local variable by @etaf in #4178
- [benchmarks] Skip dead weakref proxies when enumerating live example kernels by @jansel in #4150
- torch_tpu: update pin to aaba9aa5e67f80916979c69770c77df9504f416c by @cota in #4179
- [autotuner] Benchmark in the steady CUDA graph-dispatch state; callable kernel arguments reach the benchmark subprocess by @jansel in #4163
- [bench] Catalogue readings in the steady graph-dispatch state; the CPU codegen test helper restores the probes it poisons by @jansel in #4164
- [cutedsl] Generic PDL kernels release their dependents (cute_pdl_release placements, auto rule, seeds); final selection prefers a releasing config by @jansel in #4165
- [cutedsl] Register-MMA GEMMs: M-contiguous A operands, grouped row union, epilogue-chain helpers, per-chunk TF32 base addresses by @jansel in #4166
- [cutedsl] tcgen05 GEMMs: two-CTA row-union prologue and ring depth, epilogue SMEM gates, streamed epilogue, CLC pacing, persistent grid cap by @jansel in #4167
- [cutedsl] Block-scaled NVFP4 fused consumer: spin waits before the transform warps' waits, early PDL trigger, cute_scaled_epilogue_n knob by @jansel in #4168
- [cutedsl] Elementwise lowering: exact 16-bit products, lane-loop reduction folds, affine lane vectorizer, fast_math epilogue math by @jansel in #4169
- [cutedsl] Host tails on the device: 1-D sum/mean after the launch as one CuTe CTA, remainder fills written by the kernel by @jansel in #4170
- [autotuner] Seeds: tcgen05 M-paired deep-ring tiles with cute_pdl, pointwise L2-resident inputs and memory-order walks, 32-byte row packets by @jansel in #4171
- [examples] grpo_loss: no materialized zero grads for unused outputs; epilogue_subtiling: fast_math GELU epilogue by @jansel in #4172
- [cutedsl] Host sum tails: refresh the ATen fingerprint pin to the 2026-10-26 nightly by @jansel in #4173
- Add H100 fbtriton test shard with legacy CuTe import coverage by @jansel in #4176
- [cutedsl] Two codegen fixes: in-loop TMA-store waits only for loop-region stores; a packed grid axis is a thread axis by @jansel in #4151
- [cutedsl] Barrier chains with one native GEMM phase, and materialized derived operands of seeded and transposed GEMMs by @jansel in #4152
- [autotuner] Final selection: per-call-flushed tiebreak column, per-family finalist quota, and the two-tile fa4 seed by @jansel in #4153
- [cutedsl] mma_dag: a CTA-shared side tile for scalar/row side loads and an MMA-warp pass mode for the region renderer by @jansel in #4154
- [cutedsl] Epilogue-issued TMA staging of exact-shape auxiliary tiles (tcgen05_aux_load_mode="epi_tma") by @jansel in #4155
- [cutedsl] Paired tcgen05: fanout by SMEM budget, 2x2 multicast clusters, co-resident cluster cap, seed carriers and twins by @jansel in #4156
- [cutedsl] Epilogue warp groups, a configured epilogue subtile and an auxiliary prefetch ring for the role-local TMA-store epilogue by @jansel in #4157
- [cutedsl] Register sub-block family (cute_register_subblocks): mma.sync tf32 from registers; per-row sums, rsqrt, repeated operands by @jansel in #4158
- [cutedsl] Flash forward: fa4_alt item-boundary fixes and resident value-graph softmax lowering on the dense LSE surface by @jansel in #4159
- [examples] fused_linear_jsd: no zero fills, in-place row GEMMs, one loss mean; se_block: shared backward operand once by @jansel in #4160
- [cutedsl] Materialized chains: exact-shape auxiliary prefetch on the GEMM launches, flat producer twin seeds first by @jansel in #4161
- [cutedsl] PDL default: cute_pdl pinned True on tcgen05 kernels (sm_90+) by @jansel in #4162
- [triton] fast_math: drop NaN propagation in max/min by @etaf in #4100
- [compiler] Fix PyTorch nightly compatibility and CuTe CI timeouts by @jansel in #4189
- [ROCm][CI] Update ROCm version to 10.1 by @akashveramd in #4188
- [compiler] fake_tensor: type the contiguous-cache replace loosely so pyrefly passes on every supported torch by @jansel in #4191
- [benchmarks] matmul_sweep: 96-cell GEMM sweep harness (CuTe vs Triton vs torch) with burst and sustained timers by @jansel in #4192
- [cutedsl] Stage TMA-unaligned matmul operands through a padded host copy by @jansel in #4193
- [cutedsl] Flat tcgen05 kernels run ragged tiles through TMA with vector stores; admit ragged batched and 2-63-row GEMMs by @jansel in #4194
- [cutedsl] Split-K for K-heavy GEMMs: in-kernel tcgen05_split_k and the workspace schedule on bf16 and plain stores by @jansel in #4195
- [cutedsl] Planner: two-CTA edge family on ragged GEMMs down to 256 with 128-row pair seeds by @jansel in #4196
- [examples] Second frontend versions for the residual GEMM cases: shared-weight grouped GEMM, unfused matmul+layernorm by @jansel in #4197
- [megakernel] Add the Muse Spark EP4 MoE verify megakernel by @ethche in #4198
- Let cute tests run on other CuTe DSL releases, Python 3.13+, and environments (#4180) by @jansel in #4180
- [cutedsl] Fix CuTe matmul seeds beside unrelated loops by @coder-2011 in #3972
- [language] hl.prefetch: L2 prefetch hint with hl.load indexing by @ethche in #4190
- [triton] Drop launch options the installed Triton build lacks; blackwell_attention pins its AutoWS config only where it applies by @jansel in #4200
- [cutedsl] Re-raise a codegen refusal as itself instead of chaining it from itself by @jansel in #4201
- [examples] CUDA-graph-capturable references for fp8_attention and gather_gemv; no flash SDPA baseline for biased attention by @jansel in #4202
- [autotuner] Fall back to a sampled config for the accuracy baseline when the reference config is refused by @jansel in #4203
- [cutedsl] Elide a split/join round trip over a register tile into the elementwise op on the source by @jansel in #4204
- [cutedsl] Affine vectorizer: the fast-math exp2/division call forms are pure by @jansel in #4205
- [examples] aot_example: fast_math for row_softmax and the rms_norm kernels by @jansel in #4206
- [compiler] FastMathUnused: warn when a kernel has operations fast_math would speed up but the setting is off by @jansel in #4207
- [examples] se_block: fast_math for the forward kernel by @jansel in #4208
- [CI] Install Python development headers for XPU by @jansel in #4210
New Contributors
Full Changelog: v1.5.1...v1.5.2