Repository navigation

[WS1][CUDA][MiniMax-H3] Add FP32 video input projection - #562

Merged

maxiaosong1124 merged 28 commits into

RL-Align:test-h3from

walker-ai:feat/h3-video-input-projection-fp32

Oct 28, 2026

Merged

[WS1][CUDA][MiniMax-H3] Add FP32 video input projection#562
maxiaosong1124 merged 28 commits into
RL-Align:test-h3from
walker-ai:feat/h3-video-input-projection-fp32

Commits on Oct 20, 2026
- Add H3 FP32 video input projection with Triton and CUTLASS
  walker-ai
  committed
  9a991e3
- Reuse shared GPU reduction tree for H3 FP32 input projection
  walker-ai
  committed
  ed7f0db
- Optimize canonical H3 FP32 projection reductions and packing
  walker-ai
  committed
  dbfb1a2
- Optimize H3 FP32 projection leaf loops and bias paths
  walker-ai
  committed
  ac95108
- Optimize H3 FP32 projection chunks and CUTLASS layouts
  walker-ai
  committed
  51d7821
- Optimize H3 FP32 projection chunks and Triton tiles
  walker-ai
  committed
  ca13589
- Fuse H3 CUTLASS bias and widen FP32 chunk merges
  walker-ai
  committed
  541c737
- Optimize H3 Triton FP32 GEMM and canonical row reuse
  walker-ai
  committed
  96b39cf
- Remove H3 FP32 projection CUTLASS candidate
  walker-ai
  committed
  e9ebcee
- Separate canonical linear backend from projection matmul
  walker-ai
  committed
  0a152fb
- Remove caller-controlled canonical ordering option
  walker-ai
  committed
  d38dff5
- Fix H3 FP32 chunk accuracy and extend regression coverage
  walker-ai
  committed
  b6be602
- Simplify H3 FP32 bias tree scheduling
  walker-ai
  committed
  3fc97a0
- Align H3 projection naming and documentation style
  walker-ai
  committed
  fc0b332
- Clarify H3 projection block configuration names
  walker-ai
  committed
  19df90d
- Restore shared FP32 GEMM compiler fusion defaults
  walker-ai
  committed
  682ece6
- Refactor compact tree workspace planning
  walker-ai
  committed
  88adb4c
- Remove compact tree storage after H20 breakdown
  walker-ai
  committed
  5d60da6
- Document H3 projection GEMM dispatch and scheduling
  walker-ai
  committed
  6d621f3
- Document H3 projection interface and canonical linear gradients
  walker-ai
  committed
  d3fb659
- Strengthen H3 projection validation and stabilize backend identity
  walker-ai
  committed
  f517a37
- Make H3 projection reproduction use seeded synthetic inputs
  walker-ai
  committed
  c2e2067
- Record incremental peak memory for H3 projection benchmarks
  walker-ai
  committed
  9fd297b
Commits on Oct 21, 2026
- Fix FP32 GEMM condition for Triton 3.0 CI
  walker-ai
  committed
  520065f
- Remove projection weight fixtures and simplify reproduction setup
  walker-ai
  committed
  38193cf
- Cover H3 video and conditioned task layouts
  walker-ai
  committed
  8bd28a9
Commits on Oct 26, 2026
- Merge branch 'test-h3' into feat/h3-video-input-projection-fp32
  maxiaosong1124
  authored
  2678985
Commits on Oct 28, 2026
- Merge branch 'test-h3' into feat/h3-video-input-projection-fp32
  maxiaosong1124
  authored
  6f42d88

Commits

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 26, 2026

Commits on Oct 28, 2026