All pull requests

Pull requests

• [Test] glm53/glm53_flash: Turn off crash dumps in the fault tests

  #55·jhinpan opened Oct 23, 2026Collaborator·3 of 5 tasks·

  ·

• [Feat] [MI355X] qwen38_flash_next: Run the MTP draft on the monokernel

  #53·amd-cgilli opened Oct 23, 2026·5 tasks done·

  ·

• [Feat] [MI355X] qwen38_flash_next: Add check mode

  #52·amd-cgilli opened Oct 23, 2026·5 tasks done·

  ·

• [Feat] [MI355X] qwen38_flash_next: Serve NEXTN verify steps on the monokernel

  #51·amd-cgilli opened Oct 23, 2026·5 tasks done·

  ·

• [Feat] [MI355X] qwen38_flash_next: Add the gfx950 kernel

  #50·amd-cgilli opened Oct 23, 2026·5 tasks done·

  ·

• [Fix] common: Import SGLang's monokernel interface from its contract-named paths

  #49·BBuf opened Oct 23, 2026Contributor·3 of 5 tasks·

  ·

• [Feat] [MI355X] dsv4: Add whole-layer DeepSeek-V4-Pro mono kernels (TP8)

  #48·1am9trash opened Oct 23, 2026Contributor·5 tasks·

  ·

• [Perf] [MI355X] glm53: Score the paged indexer keys per wave

  #42·jhinpan opened Oct 22, 2026Collaborator·4 of 5 tasks·

  ·

• [Test] deepseek_v41: Check capture and replay allocations

  #38·amd-pedghazi opened Oct 22, 2026Contributor·3 of 5 tasks·

  ·
  2

• [Feat] [MI355X] qwen38_a95b_mxfp4: Add serving checks for the Qwen3.8 MoE monokernel

  #35·yichiche opened Oct 22, 2026Contributor·

  ·

• [Feat] [MI355X] qwen38_a95b_mxfp4: Register the Qwen3.8 MoE decode monokernel spec

  #34·yichiche opened Oct 22, 2026Contributor·

  ·

• [Feat] [MI355X] qwen38_a95b_mxfp4: Add the gfx950 GDN and MXFP4 decode kernels

  #33·yichiche opened Oct 22, 2026Contributor·

  ·

• [Feat] [MI355X] qwen38_flash_next: Add the Qwen3.8-Flash-Next 180B monokernel

  #29·amd-cgilli opened Oct 22, 2026·5 tasks done·

  ·
  1

• [Feat] [MI355X] qwen3_8_dense: Add Qwen3.8-27B dense decode monokernels

  #28·vorapolsiloai opened Oct 22, 2026·4 tasks done·

  ·

• [Feat] [MI355X] muse_spark: Whole-layer monokernels for Muse Spark (TEP8)

  #26·zijiecode opened Oct 22, 2026Collaborator·5 tasks·

  ·

• [Perf] [MI355X] glm53: Select a long row's indexer top-k on up to 16 CTAs

  #24·jhinpan opened Oct 22, 2026Collaborator·4 of 5 tasks·

  ·

• [Feat] [MI355X] glm53_flash: Cover layers 0-2, 44 and the logits

  #23·heslami opened Oct 21, 2026Contributor·4 of 5 tasks·

  ·
  8

• [Perf] common: Make guarded mailbox waits as cheap as unguarded ones

  #19·amd-asalykov opened Oct 21, 2026·5 tasks done·

  ·
  3

• [Feat] [MI355X] qwen38_moe: Add Qwen3.8 MoE decode monokernels

  #16·omirosh opened Oct 20, 2026·5 tasks done·

  ·

• [Feat] [AMD] kimi_k3: Kimi-K3 layer monokernel on gfx950

  #13·billishyahao opened Oct 19, 2026Collaborator·4 of 5 tasks·

  ·
  2

• [Feat] [AMD] minimax_m3: MiniMax-M3 monokernels on gfx950 at TP2, TP4 and TP8

  #9·sahirema opened Oct 17, 2026·

  ·
  1