Repository navigation

Optimize Qwen CUDA sparse attention, indexer, and quantized MoE kernels - #33393

Closed

Akshay Sonawane (apsonawane) wants to merge 551 commits into

mainfrom

asonawane/qwen-cuda-kernel-optimizations

Closed

Optimize Qwen CUDA sparse attention, indexer, and quantized MoE kernels#33393
Akshay Sonawane (apsonawane) wants to merge 551 commits into
mainfrom
asonawane/qwen-cuda-kernel-optimizations

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 15, 2026
- Fix SparsePagedAttention WGSL generation
  
  Copilot
  and
  kunal-vaishnavi
  authored
  24f8399
- Fix WebGPU plugin API initialization mismatch
  
  Copilot
  and
  kunal-vaishnavi
  authored
  5c0301d
- Merge remote-tracking branch 'origin/main' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  a25b7cd
- Address SparseAttentionIndexer review feedback
  
  Copilot
  and
  kunal-vaishnavi
  committed
  a9b3b69
- Gate XQA dispatch tests by CUDA architecture
  
  Copilot
  and
  kunal-vaishnavi
  committed
  2a42087
- Address SparsePagedAttention review cleanup and coverage
  
  Copilot
  and
  kunal-vaishnavi
  committed
  4f7f255
- Ignore plugin EP NuGet build artifacts
  
  Copilot
  and
  kunal-vaishnavi
  committed
  2f6bc3a
- Optimize single-source sparse paged attention
  
  Copilot
  and
  kunal-vaishnavi
  committed
  4f8a7a5
- Strengthen direct auxiliary output test
  
  Copilot
  and
  kunal-vaishnavi
  committed
  408cc68
- Fix SparsePagedAttention WGSL generation
  
  Copilot
  and
  kunal-vaishnavi
  committed
  36821cb
- Sync target and fix formatting
  
  Copilot
  and
  kunal-vaishnavi
  committed
  a67b414
- Update generated contrib docs
  
  Copilot
  and
  kunal-vaishnavi
  committed
  f9bd45a
- Remove unused SparsePagedAttention binding limit
  
  Copilot
  and
  kunal-vaishnavi
  authored
  6efcefd
- Merge remote-tracking branch 'origin/copilot/copilotadd-cuda-sparsepagedattention' into copilot/copilotcopilotadd-cuda-sparsepagedattention-s2-non
  
  Copilot
  and
  kunal-vaishnavi
  authored
  1a9a9c4
- Fix WebGPU plugin logger definitions
  
  Copilot
  and
  kunal-vaishnavi
  authored
  c13bfc0
- Use valid head size in SparsePagedAttention rotary test
  
  Copilot
  and
  kunal-vaishnavi
  authored
  d94ab7e
- Initial plan
  Copilot
  authored
  dbaeb30
- Apply remaining changes
  
  Copilot
  and
  kunal-vaishnavi
  authored
  2e3693c
- Restore GatherBlockQuantized MLAS optimization
  
  Copilot
  and
  kunal-vaishnavi
  authored
  879017f
- Test GatherBlockQuantized batch boundary
  
  Copilot
  and
  kunal-vaishnavi
  authored
  0d86780
- Merge remote-tracking branch 'origin/main' into copilot/add-cuda-dynamicsparseattention
  Copilot
  authored
  3d2be52
- Fix SparsePagedAttention validation and indexing
  
  Copilot
  and
  kunal-vaishnavi
  authored
  e6e26ae
- Address DynamicSparseAttention review feedback
  
  Copilot
  and
  kunal-vaishnavi
  authored
  2f4024f
- Fix half-split rotary test alignment
  
  Copilot
  and
  kunal-vaishnavi
  authored
  52c497e
Commits on Sep 16, 2026
- Fix packed QKV validation test expectation
  
  Copilot
  and
  kunal-vaishnavi
  authored
  3520cfb
- Add CUDA host-pageable gather implementation
  
  Copilot
  and
  kunal-vaishnavi
  authored
  20ac107
- Harden CUDA gather fallback and tests
  
  Copilot
  and
  kunal-vaishnavi
  authored
  e879ff6
- Complete plugin and constant-input coverage
  
  Copilot
  and
  kunal-vaishnavi
  authored
  9d26367
- Support shared SparseAttentionIndexer caches
  
  Copilot
  and
  kunal-vaishnavi
  authored
  f039ff6
- Implement packed sparse attention indexer
  
  Copilot
  and
  kunal-vaishnavi
  authored
  cf54693
- Harden packed indexer state validation
  
  Copilot
  and
  kunal-vaishnavi
  authored
  6f8041c
- Validate packed indexer capacity bounds
  
  Copilot
  and
  kunal-vaishnavi
  authored
  5b9d10e
- Fix packed indexer output-count test
  
  Copilot
  and
  kunal-vaishnavi
  authored
  de0e587
- Consolidate SparseAttentionIndexer cache state
  
  Copilot
  and
  kunal-vaishnavi
  authored
  5a37cb6
- Enable host-pageable gather in CUDA Graphs
  
  Copilot
  and
  kunal-vaishnavi
  authored
  e2126e3
- Strengthen direct host CUDA Graph coverage
  
  Copilot
  and
  kunal-vaishnavi
  authored
  4d9c851
- Fix SparseAttentionIndexer test narrowing warning
  
  Copilot
  and
  kunal-vaishnavi
  authored
  76d6532
- Address CUDA gather review feedback
  
  Copilot
  and
  kunal-vaishnavi
  authored
  4eb6a08
- Handle disabled prepacking in direct mode
  
  Copilot
  and
  kunal-vaishnavi
  authored
  2651deb
- Fix invalid-width shape inference test
  
  Copilot
  and
  kunal-vaishnavi
  authored
  c5dcbd1
- Merge remote-tracking branch 'origin/main' into copilot/add-cuda-dynamicsparseattention
  Copilot
  authored
  00a7441
- Merge remote-tracking branch 'origin/copilot/add-cuda-dynamicsparseattention' into copilot/copilotadd-cuda-dynamicsparseattention
  
  Copilot
  and
  kunal-vaishnavi
  authored
  20776a3
- Fix GatedDeltaNet parameter dispatch
  
  Copilot
  and
  kunal-vaishnavi
  authored
  0cb82f8
- Fix WebGPU provider test build
  
  Copilot
  and
  kunal-vaishnavi
  authored
  3de7958
- Merge main into copilot/implement-gated-delta-net-webgpu
  kunal-vaishnavi
  committed
  97d8f00
- Optimize GatedDeltaNet recurrent WebGPU execution
  
  Copilot
  and
  kunal-vaishnavi
  committed
  bb02499
- Cover optimized GatedDeltaNet recurrent path
  
  Copilot
  and
  kunal-vaishnavi
  committed
  b62c0a4
- Exercise vectorized GatedDeltaNet path on baseline adapters
  
  Copilot
  and
  kunal-vaishnavi
  committed
  c823868
- Add effective WebGPU limit overrides
  
  Copilot
  and
  kunal-vaishnavi
  committed
  1174fc7
- Update generated SparseAttentionIndexer docs
  
  Copilot
  and
  kunal-vaishnavi
  authored
  daf73e3
- Merge remote-tracking branch 'origin/copilot/implement-sparse-attention-indexer' into copilot/implement-sparse-attention-indexer-again
  
  Copilot
  and
  kunal-vaishnavi
  authored
  2652363
- Adapt WebGPU indexer to shared cache ABI
  
  Copilot
  and
  kunal-vaishnavi
  authored
  3d328b2
- Merge remote-tracking branch 'origin/copilot/implement-sparse-attention-indexer-again' into copilot/copilotimplement-packed-sparse-attention-indexer
  Copilot
  authored
  2125b41
- Address packed indexer review feedback
  
  Copilot
  and
  kunal-vaishnavi
  authored
  df26284
- Optimize CUDA QSA scoring and selection
  
  Copilot
  and
  kunal-vaishnavi
  committed
  ebc6bdf
- Add single-pass TopK coverage
  
  Copilot
  and
  kunal-vaishnavi
  committed
  ecabf39
- Fix packed indexer CI failures
  
  Copilot
  and
  kunal-vaishnavi
  authored
  8ece210
- Merge remote-tracking branch 'origin/copilot/copilotimplement-packed-sparse-attention-indexer' into copilot/copilotcopilotimplement-packed-sparse-attention-in
  Copilot
  authored
  49ade0c
- Address CUDA QSA review feedback
  
  Copilot
  and
  kunal-vaishnavi
  authored
  8f506aa
- Test BF16 sparse mixer routing
  kunal-vaishnavi
  committed
  80f4f66
- Fix CUDA CI compilation and generated docs
  
  Copilot
  and
  kunal-vaishnavi
  authored
  2aa61ae
Commits on Sep 17, 2026
- Merge remote-tracking branch 'origin/main' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  b56a5d3
- Merge remote-tracking branch 'origin/pr/32358' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  8619cb3
- Merge remote-tracking branch 'origin/pr/32467' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  4ef561c
- Merge remote-tracking branch 'origin/pr/32480' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  cbaddd9
- Merge remote-tracking branch 'origin/pr/32512' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  cc5d018
- Merge remote-tracking branch 'origin/pr/32621' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  1cb5545
- Merge remote-tracking branch 'origin/pr/32524' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  7c05ac9
- Merge remote-tracking branch 'origin/pr/32525' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  17807cf
- Merge remote-tracking branch 'origin/pr/32526' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  9f1ff55
- Merge remote-tracking branch 'origin/pr/32626' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  cfe7428
- Merge remote-tracking branch 'origin/pr/32656' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  d131e8c
- Merge remote-tracking branch 'origin/pr/32659' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  e7ee810
- Merge remote-tracking branch 'origin/pr/32529' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  bea3fe7
- Merge remote-tracking branch 'origin/pr/32533' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  86da2b4
- Merge remote-tracking branch 'origin/pr/32613' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  783b40e
- Merge remote-tracking branch 'origin/pr/32645' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  0863883
- Restore wrapped Engram hash addition
  Copilot
  committed
  9950a03
- Fix packed indexer schema validation
  
  Copilot
  and
  kunal-vaishnavi
  authored
  d0b6faf
- Merge remote-tracking branch 'origin/copilot/copilotimplement-packed-sparse-attention-indexer' into copilot/copilotcopilotimplement-packed-sparse-attention-in
  Copilot
  authored
  51b6403
- Optimize SparsePagedAttention workgroup softmax
  
  Copilot
  and
  kunal-vaishnavi
  authored
  fbb15f9
- Merge remote-tracking branch 'origin/copilot/copilotadd-cuda-sparsepagedattention' into copilot/copilotcopilotadd-cuda-sparsepagedattention-s2-non
  
  Copilot
  and
  kunal-vaishnavi
  authored
  87c6740
- Reduce packed indexer WebGPU bindings
  
  Copilot
  and
  kunal-vaishnavi
  authored
  e81906b
- Merge remote-tracking branch 'origin/copilot/copilotadd-cuda-dynamicsparseattention' into copilot/optimize-dynamic-sparse-attention
  Copilot
  authored
  5b1566b
- Parallelize CUDA DynamicSparseAttention
  
  Copilot
  and
  kunal-vaishnavi
  authored
  cfacbec
- Harden fused attention synchronization
  
  Copilot
  and
  kunal-vaishnavi
  authored
  70fd72a
- Scale DynamicSparseAttention to large contexts
  
  Copilot
  and
  kunal-vaishnavi
  authored
  d158543
- Reduce large-context attention workspace
  
  Copilot
  and
  kunal-vaishnavi
  authored
  3837ccf
- Merge remote-tracking branch 'origin/copilot/optimize-dynamic-sparse-attention' into kvaishnavi/qwen-38-flash
  Copilot
  committed
  26b93ff
- Implement profile-driven QSA selection optimizations
  
  Copilot
  and
  kunal-vaishnavi
  authored
  24bccfd
- Fix and format QSA radix selection
  
  Copilot
  and
  kunal-vaishnavi
  authored
  1593b64
- Include CUDA kernel scratch API
  
  Copilot
  and
  kunal-vaishnavi
  authored
  5b77441
- Fix CUDA provider test option construction
  
  Copilot
  and
  kunal-vaishnavi
  committed
  9c88191
- Implement profile-driven QSA selection optimizations
  
  Copilot
  and
  kunal-vaishnavi
  committed
  9d906b1
- Fix and format QSA radix selection
  
  Copilot
  and
  kunal-vaishnavi
  committed
  2c03167
- Include CUDA kernel scratch API
  
  Copilot
  and
  kunal-vaishnavi
  committed
  7dd2fae
- Bound sparse attention split workspace
  
  Copilot
  and
  kunal-vaishnavi
  committed
  833faca
- Bound DynamicSparseAttention split work
  
  Copilot
  and
  kunal-vaishnavi
  committed
  b69c135
- Optimize SparsePagedAttention workgroup softmax
  
  Copilot
  and
  kunal-vaishnavi
  committed
  f8d960b
- Fix packed indexer schema validation
  
  Copilot
  and
  kunal-vaishnavi
  committed
  d3645c1
- Reduce packed indexer WebGPU bindings
  
  Copilot
  and
  kunal-vaishnavi
  committed
  32f7fce
- Fix GatedDeltaNet WebGPU CI compilation
  
  Copilot
  and
  kunal-vaishnavi
  committed
  7d6428a
- Fix packed GatedDeltaNet WebGPU test setup
  
  Copilot
  and
  kunal-vaishnavi
  committed
  7b6e195
- Fix generated contrib documentation whitespace
  Copilot
  committed
  6a2cbd1
- Fix sparse attention indexer CUDA kernel type
  kunal-vaishnavi
  committed
  7a76fff
- Fix sparse attention indexer CUDA kernel type
  kunal-vaishnavi
  committed
  1f12360
- Fix CUDA gather block quantized tests
  kunal-vaishnavi
  committed
  082cde9
- Implement bounded exact QSA selection
  
  Copilot
  and
  kunal-vaishnavi
  authored
  d2de258
- Finalize bounded QSA decode path
  
  Copilot
  and
  kunal-vaishnavi
  authored
  a6c7873
Commits on Sep 18, 2026
- Add HyperConnectionMix CUDA contrib operator
  
  Copilot
  and
  kunal-vaishnavi
  authored
  3b5660c
- Format HyperConnectionMix CUDA kernel
  
  Copilot
  and
  kunal-vaishnavi
  authored
  c12532d
- Remove incorrect HyperConnectionMix operator
  
  Copilot
  and
  kunal-vaishnavi
  authored
  0be2808
- Add weightless hyper-connection fusion operators
  
  Copilot
  and
  kunal-vaishnavi
  authored
  6ce8251
- Split hyper-connection kernels by operator
  
  Copilot
  and
  kunal-vaishnavi
  authored
  e31e472
- Add dedicated hyper-connection shape inference
  
  Copilot
  and
  kunal-vaishnavi
  authored
  f51ae49
- Merge remote-tracking branch 'origin/main' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  038a7ff
- Merge remote-tracking branch 'origin/pr/32687' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  f216d45
- Allow flattened HyperConnection pre-mix gates
  kunal-vaishnavi
  committed
  ccda237
- Merge remote-tracking branch 'origin/pr/32687' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  9e76e4c
- Export layer norm helpers for branchwise RMS norm
  kunal-vaishnavi
  committed
  abd4a3b
- Fix hyper-connection CI and review issues
  
  Copilot
  and
  kunal-vaishnavi
  authored
  853b48a
- Fix provider build for branchwise RMS norm
  kunal-vaishnavi
  committed
  6f00305
- Fix provider build for branchwise RMS norm
  kunal-vaishnavi
  committed
  b4eeea8
- Merge remote-tracking branch 'origin/pr/32687' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  dfbfacc
- Fix optional SparsePagedAttention type resolution
  
  Copilot
  and
  kunal-vaishnavi
  authored
  6700329
- Support shared rotary caches in sparse indexer
  kunal-vaishnavi
  committed
  15d73ba
- Accept padding masks in sparse attention indexer
  kunal-vaishnavi
  committed
  3d6bf1a
- Fix remaining CUDA test compilation
  
  Copilot
  and
  kunal-vaishnavi
  committed
  ad5e934
- Support shared rotary caches in sparse indexer
  kunal-vaishnavi
  committed
  8ab1114
- Accept padding masks in sparse attention indexer
  kunal-vaishnavi
  committed
  445e7d3
- Address VarlenNGramHashMapping review feedback
  
  Copilot
  and
  kunal-vaishnavi
  committed
  a9e357a
- Fix optional SparsePagedAttention type resolution
  
  Copilot
  and
  kunal-vaishnavi
  committed
  ba39dc5
- Fix CUDA conflicts in Qwen operator updates
  kunal-vaishnavi
  committed
  3461e03
- Fuse query normalization into SparseAttentionIndexer
  kunal-vaishnavi
  committed
  c6efe24
- Merge branch 'stack-32526' into stack-32528
  kunal-vaishnavi
  committed
  6a13c37
- Fuse query normalization in WebGPU sparse indexer
  kunal-vaishnavi
  committed
  fe9d059
- Merge branch 'stack-32528' into stack-32618
  kunal-vaishnavi
  committed
  a73ff17
- Fuse query normalization into PackedSparseAttentionIndexer
  kunal-vaishnavi
  committed
  f32b9bf
- Merge branch 'stack-32618' into stack-32656
  kunal-vaishnavi
  committed
  851cc44
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  eb63f12
- Merge remote-tracking branch 'origin/main' into stack-32526
  kunal-vaishnavi
  committed
  efa7f56
- Merge branch 'stack-32526' into stack-32528
  kunal-vaishnavi
  committed
  72d02b1
- Merge branch 'stack-32528' into stack-32618
  kunal-vaishnavi
  committed
  96f00c6
- Merge branch 'stack-32618' into stack-32656
  kunal-vaishnavi
  committed
  61e9e7a
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  3949037
- Restrict SparseAttentionIndexer to integer masks
  kunal-vaishnavi
  committed
  9c8ed58
- Merge branch 'stack-32526' into stack-32528
  kunal-vaishnavi
  committed
  7d485c8
- Use integer masks in WebGPU sparse indexer
  kunal-vaishnavi
  committed
  bc922c2
- Merge branch 'stack-32528' into stack-32618
  kunal-vaishnavi
  committed
  324128f
- Merge branch 'stack-32618' into stack-32656
  kunal-vaishnavi
  committed
  c007083
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  b65729f
Commits on Sep 19, 2026
- Make DSA device validation opt-in
  
  Copilot
  and
  kunal-vaishnavi
  committed
  a2288d3
- Test default DSA validation mode
  
  Copilot
  and
  kunal-vaishnavi
  committed
  0c87280
- Fix segmented WebGPU test provider setup
  
  Copilot
  and
  kunal-vaishnavi
  committed
  c3a106b
- Exercise segmented aliased WebGPU state
  
  Copilot
  and
  kunal-vaishnavi
  committed
  608aa74
- Hoist QSA query preparation out of block scoring
  kunal-vaishnavi
  committed
  0e649ff
- Merge branch 'stack-32526' into stack-32528
  kunal-vaishnavi
  committed
  75d6b8a
- Merge branch 'stack-32528' into stack-32618
  kunal-vaishnavi
  committed
  b150814
- Merge branch 'stack-32618' into stack-32656
  kunal-vaishnavi
  committed
  240f37a
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  c3f101c
- Hoist packed QSA query preparation out of scoring
  kunal-vaishnavi
  committed
  04c7a92
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  2d8c42e
- Guard QSA decode performance in CUDA tests
  kunal-vaishnavi
  committed
  cdbf1fe
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  02a0fce
- Enforce QSA decode latency budgets
  kunal-vaishnavi
  committed
  7b9d8bc
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  842a598
- Optimize CUDA branchwise RMS normalization
  kunal-vaishnavi
  committed
  d944a2c
Commits on Sep 20, 2026
- Optimize small inner-axis CUDA Split
  kunal-vaishnavi
  committed
  712ec03
- Support packed QK in CUDA sparse attention indexer
  kunal-vaishnavi
  committed
  1aec0ff
- Merge branch 'stack-32526' into stack-32528
  kunal-vaishnavi
  committed
  dbf7d34
- Support packed QK in WebGPU sparse attention indexer
  kunal-vaishnavi
  committed
  01dc6e1
- Merge branch 'stack-32528' into stack-32618
  kunal-vaishnavi
  committed
  7f41361
- Merge branch 'stack-32618' into stack-32656
  kunal-vaishnavi
  committed
  31dd38b
- Merge branch 'stack-32656' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  f959e1c
- Fix packed GatedDeltaNet shape inference
  kunal-vaishnavi
  committed
  bb4c28a
- Support packed QKV in CUDA GatedDeltaNet
  kunal-vaishnavi
  committed
  56aabd0
- Support packed QKV in WebGPU GatedDeltaNet
  kunal-vaishnavi
  committed
  44f2a1d
Commits on Sep 21, 2026
- Optimize prefix-causal sparse attention indexing
  kunal-vaishnavi
  committed
  59253d6
- Optimize prefix-causal sparse attention indexing
  kunal-vaishnavi
  committed
  98c77d8
- Optimize QSA TopK and block scoring
  kunal-vaishnavi
  committed
  548ff66
- Optimize QSA TopK and block scoring
  kunal-vaishnavi
  committed
  a46ecd6
- Test multi-row QSA partial TopK
  kunal-vaishnavi
  committed
  5a01fad
- Fix multi-row QSA partial TopK stride
  kunal-vaishnavi
  committed
  726ee42
- Distribute QSA TopK and cache block representatives
  kunal-vaishnavi
  committed
  c095912
- Distribute QSA TopK and cache block representatives
  kunal-vaishnavi
  committed
  0d91053
Commits on Sep 22, 2026
- Implement hierarchical exact QSA selection
  kunal-vaishnavi
  committed
  cbc0463
- Implement hierarchical exact QSA selection
  kunal-vaishnavi
  committed
  1f1c1ea
- Use Env for sparse attention test variables
  kunal-vaishnavi
  committed
  b9699f1
- Use Env for sparse attention test variables
  kunal-vaishnavi
  committed
  9ba9d85
Commits on Sep 26, 2026
Commits on Sep 27, 2026
- Optimize packed QSA selection
  kunal-vaishnavi
  committed
  4da5794
- Support packed QK indexer inputs
  kunal-vaishnavi
  committed
  5f468d5
Commits on Sep 28, 2026
- Optimize CUDA SparsePagedAttention with split KV
  kunal-vaishnavi
  committed
  e93b8d2
- Merge branch 'optimize-sparse-paged-attention' into update-pr-32533
  kunal-vaishnavi
  committed
  1c35723
- Merge branch 'update-pr-32533' into update-pr-32613
  kunal-vaishnavi
  committed
  11a9d04
- Merge branch 'update-pr-32613' into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  5161035
- Optimize packed QSA prefill scoring
  kunal-vaishnavi
  committed
  5e186e8
- Optimize sparse paged attention prefill
  kunal-vaishnavi
  committed
  c58e358
- Capture n-gram state updates for partial acceptance
  kunal-vaishnavi
  committed
  838d0e3
- Gate variable-length convolution state capture
  kunal-vaishnavi
  committed
  30423f0
- Capture packed indexer state updates
  kunal-vaishnavi
  committed
  9f08e21
- Fix compact state capture validation regressions
  kunal-vaishnavi
  committed
  60c403f
Commits on Sep 29, 2026
- Honor deterministic compute for CUDA MatMulNBits tactics
  kunal-vaishnavi
  committed
  15f3a9a
Commits on Oct 2, 2026
- Fix FP4 GEMV expert-map fallback for unsupported fused prologues
  kunal-vaishnavi
  committed
  73d7c8e
- Add CUDA QMoE block-scaled FP8 support with routed-expert compaction
  
  Akshay Sonawane (apsonawane)
  authored and
  kunal-vaishnavi
  committed
  6d4142c
- Fix CUDA QMoE registration for independent FP8 scale types
  
  authored and
  kunal-vaishnavi
  committed
  8a30344
- Fix FP8 QMoE layout regression test and documentation
  
  authored and
  kunal-vaishnavi
  committed
  b8c3a21
- Fix QMoE CI expectations and older PyTorch test handling
  
  authored and
  kunal-vaishnavi
  committed
  0d419c8
- Extend routed-expert compaction to NVFP4 QMoE fallback
  kunal-vaishnavi
  committed
  8689050
Commits on Oct 3, 2026
- Keep packed sparse indexer state-update gate on CPU for CUDA graphs
  kunal-vaishnavi
  committed
  2fd134e
Commits on Oct 5, 2026
- Address QMoE block FP8 review
  
  authored and
  kunal-vaishnavi
  committed
  77294ac
- Fix QMoE documentation whitespace
  
  authored and
  kunal-vaishnavi
  committed
  94f10b3
- Fix QMoE dequant lambda return type
  
  authored and
  kunal-vaishnavi
  committed
  27ab528
- Fix QMoE packed-prefill tiling and CI review regressions
  kunal-vaishnavi
  committed
  01913e3
- [CUDA] Avoid repacking NVFP4 QMoE decode weights (#32325)
  
  Tianlei Wu (tianleiwu)
  authored and
  kunal-vaishnavi
  committed
  b1b8ab6
- Support offline NVFP4 QMoE weights through weights_prepacked
  kunal-vaishnavi
  committed
  0d765f9
- Fuse block-scaled FP8 QMoE GEMV and GEMM without weight scratch
  kunal-vaishnavi
  committed
  746b594
- Preserve FP8 fused layouts and validate CUDA graph replay
  kunal-vaishnavi
  committed
  bc02ae0
- Keep shared FP8 fused routing independent of PR-only snapshots
  kunal-vaishnavi
  committed
  3050a77
Commits on Oct 6, 2026
- Use native contiguous K-packed NVFP4 GEMV loads
  kunal-vaishnavi
  committed
  52d0a73
- Merge origin/main into kvaishnavi/qwen-38-flash
  kunal-vaishnavi
  committed
  d7177b6
- Support 80 expanded rows in native K-packed NVFP4 GEMV
  kunal-vaishnavi
  committed
  3e164d2
Commits on Oct 7, 2026
- Add stable sparse selection merge for CUDA and WebGPU
  kunal-vaishnavi
  committed
  41fe0ba
Commits on Oct 8, 2026
- Implement flat IndexShare packed sparse indexer reuse
  kunal-vaishnavi
  committed
  55b790a
Commits on Oct 9, 2026
Commits on Oct 10, 2026
Commits on Oct 15, 2026
- Optimize Qwen CUDA sparse attention, indexer, and quantized MoE kernels
  Akshay Sonawane (apsonawane)
  committed
  e483408
- Address CUDA QMoE review feedback
  
  Akshay Sonawane (apsonawane)
  and
  Copilot
  committed
  71ee3c9
- Restrict QMoE calibration capture permissions
  
  Akshay Sonawane (apsonawane)
  and
  Copilot
  committed
  19336e1
Commits on Oct 19, 2026
Commits on Oct 20, 2026

Commits

Commits on Sep 15, 2026

Commits on Sep 16, 2026

Commits on Sep 17, 2026

Commits on Sep 18, 2026