Repository navigation

[TRTLLM-10929][feat] add fp8 combine in moe_a2a - #11844

Merged

dc3671 merged 14 commits into

NVIDIA:mainfrom

dc3671:add-fp8-combine

Mar 16, 2026

Merged

[TRTLLM-10929][feat] add fp8 combine in moe_a2a#11844
dc3671 merged 14 commits into
NVIDIA:mainfrom
dc3671:add-fp8-combine

Commits on Mar 12, 2026
- feat: add fp8 combine for moe all2all
  dc3671
  committed
  f84bd43
- use vectorized_quant
  dc3671
  committed
  3abeadd
- add in-place quant for payload_in_workspace
  dc3671
  committed
  eaeb467
- use model_config.use_low_precision_combine to enable fp8_combine
  dc3671
  committed
  4a3ca49
- remove redundant configs
  dc3671
  committed
  f0ea523
- add e2e test for low_precision_moe_combine
  dc3671
  committed
  e522b8a
- update accoarding to review comments
  dc3671
  committed
  ff66c67
- loose test atol for fp8
  dc3671
  committed
  08471f5
- add multi-gpu e2e accuracy test for low_precision_moe_combine
  dc3671
  committed
  8b90aeb
- update moe all2all microbenchmark
  dc3671
  committed
  3727f06
- fix build
  dc3671
  committed
  b9adf78
- fix test
  dc3671
  committed
  bb6422e
- improve vectorized_quant_impl
  dc3671
  committed
  49e60d7
Commits on Mar 13, 2026
- fix test
  dc3671
  committed
  fb19919