Repository navigation

[Perf] beam: Add the SM103 single-token BF16 MoE GEMV follow-up - #46401

Draft

Jiminator wants to merge 47 commits into

mainfrom

beam-sm103-moe-gemv

Draft

[Perf] beam: Add the SM103 single-token BF16 MoE GEMV follow-up#46401
Jiminator wants to merge 47 commits into
mainfrom
beam-sm103-moe-gemv

Commits on Oct 19, 2026
- [Chore] docker: Add Beam CUDA image (1/5)
  Jiminator
  committed
  46c7fed
- [Feat] kernels: Add Beam fused ops, router GEMMs and all-reduce boundary (2/5)
  
  committed
  65f3a8b
Commits on Oct 20, 2026
- Move the Dynamo and grpcio-tools pins out of the Beam image
  Jiminator
  committed
  3cd22c7
- Merge branch 'beam-stack-01-docker' into beam-stack-02-kernels
  Jiminator
  committed
  0edec36
- Move the nvidia-cuda-tileiras pin to its own PR
  Jiminator
  committed
  1a5baa7
- Move the cuTile router GEMM to its own PR
  Jiminator
  committed
  e5f9360
- Drop the Beam image from this PR
  Jiminator
  committed
  cb50998
Commits on Oct 21, 2026
- Drop the #42930 pick in zero_pages.py
  Jiminator
  committed
  82fa70c
- Drop the kernels layer's tests and benchmarks
  Jiminator
  committed
  b80a244
- Merge branch 'main' into beam-stack-02-kernels
  Jiminator
  committed
  44f81b0
- Drop the KernelSpec fields main removed
  Jiminator
  committed
  a25899c
- Remove the fused-op instances only the dropped tests used
  Jiminator
  committed
  7899e54
- Register the BF16x3 router helpers in the kernel inventory
  Jiminator
  committed
  2fbb520
- Merge branch 'main' into beam-stack-02-kernels
  Jiminator
  committed
  26385e0
- Allocate the BF16x3 router weight split per call
  Jiminator
  committed
  7565cdb
- Validate the BF16x3 split's operands and drop its unused torch path
  Jiminator
  committed
  faddbb3
- Move the BF16x3 weight split to the quantization group
  Jiminator
  committed
  7a33c19
- Correct the fast-math SiLU, BF16x3 and Beam all-reduce comments
  Jiminator
  committed
  6a4c90a
- Keep the BF16 tiny GEMM split independent of the token count
  Jiminator
  committed
  86f16b6
- Drop the Beam fused ops' torch_compile overrides and native fallbacks
  Jiminator
  committed
  77e25c4
- Require the post-norm weight in the Beam sandwich norm
  Jiminator
  committed
  d347ff8
- Use SPDX license headers in the Beam kernel files
  Jiminator
  committed
  a8e836f
- Drop unused helper defaults and type the Beam all-reduce entry point
  Jiminator
  committed
  3b5abfb
- Keep eager rounding in the Beam fused ops' torch_compile backend
  Jiminator
  committed
  14e419b
- Reject unsupported dtypes in the Beam Triton fused ops
  Jiminator
  committed
  2434661
- Check the BF16x3 split weight's rank before unpacking it
  Jiminator
  committed
  4b1354f
- Prefetch FP32 tiny GEMM weights before the PDL wait
  Jiminator
  committed
  5ba437b
- Drop the FP32 router tiny GEMM split override
  Jiminator
  committed
  d104fa3
- Drop the unverified FP32 router GEMV crossover claim
  Jiminator
  committed
  01a68eb
- Launch the Beam all-reduce boundary with one CTA per row
  Jiminator
  committed
  0f27929
- Keep eager sigmoid in the Beam gate's compiled native path
  Jiminator
  committed
  c906c12
- State the measured margin behind the FP32 router GEMV row cap
  Jiminator
  committed
  7a4d7e3
- Register the Beam sandwich norm through the existing fused-norm loop
  Jiminator
  committed
  f0a51da
Commits on Oct 22, 2026
Commits on Oct 23, 2026
- Merge branch 'main' into beam-stack-02-kernels
  Jiminator
  committed
  b51ec78
Commits on Oct 24, 2026
- Merge branch 'main' into beam-stack-02-kernels
  Jiminator
  authored
  ceb0d8e
- Merge branch 'main' into beam-stack-02-kernels
  Jiminator
  authored
  1b16a41
Commits on Oct 25, 2026

Commits

Commits on Oct 19, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 22, 2026

Commits on Oct 23, 2026

Commits on Oct 24, 2026

Commits on Oct 25, 2026