All pull requests

Pull requests

• [DSL][Runtime] Add scoped ktrace usage and CUDA Graph replay

  #1263·coderfeli opened Oct 11, 2026Collaborator·

  ·

• [Kernel] Clean low-level IR across FlyDSL kernels

  #1262·coderfeli opened Oct 10, 2026Collaborator·

  ·

• [DSL] Expose scalar buffer load API

  #1261·coderfeli opened Oct 10, 2026Collaborator·

  ·

• [fix] Build the zipped_divide guide from the tiler shape

  #1258·xudoyuan opened Oct 10, 2026Collaborator·1 task·

  ·

• [Perf] Fuse GLM-5.2 TP4 indexer for MTP4 C1/C2

  #1256·coderfeli opened Oct 10, 2026Collaborator·

  ·

• [Perf][gfx1250] Preload the kernel arguments of the BF16 and A8W8 GEMMs

  #1254·jhinpan opened Oct 10, 2026Collaborator·3 of 4 tasks·

  ·

• [Bugfix][Test] Round MXFP8 E8M0 scales up in the test quantizers

  #1253·jhinpan opened Oct 9, 2026Collaborator·4 of 5 tasks·

  ·

• [Kernel][FA] Document the gfx950 attention kernels and the 8xD head-dim protocol (8/8)

  #1252·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Route flydsl_flash_attn_func to the new gfx950 bf16/f16 kernels (7/8)

  #1251·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Add gfx950 attention backward (dQ and dK/dV) ported from AOTriton 0.14 (6/8)

  #1250·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Port gfx950 attention ALiBi, sink, paged KV, split-K and XCD swizzle to the new forward (5/8)

  #1249·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Add the gfx950 attention forward kernel (dual-wave and wide bodies) (4/8)

  #1248·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Add gfx950 attention metadata, knobs and traits config (3/8)

  #1247·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Add arch-neutral attention common, abi, philox and dispatch modules (2/8)

  #1246·xinyazhang opened Oct 9, 2026Contributor·

  ·

• [Kernel][FA] Fix ds_read_tr waitcnt hazard and stagger SCC clobber (1/8)

  #1245·xinyazhang opened Oct 9, 2026Contributor·

  ·
  2

• [Doc] Add GLM and Kimi-K3 mono-kernel performance reports

  #1244·james-huang09 opened Oct 9, 2026Collaborator·1 task·

  ·

• [llvm] bump llvm to 1df311a6

  #1243·jli-melchior opened Oct 9, 2026Collaborator·1 task·

  ·

• [fix] Correct zipped_divide tiler hierarchy handling

  #1242·xudoyuan opened Oct 9, 2026Collaborator·1 task·

  ·

• [Bugfix] Shared numerics, streams, and tests

  #1240·dimitri91209 opened Oct 9, 2026Contributor·3 of 5 tasks·

  ·
  1

• [Windows] Add native ROCm JIT support

  #1239·Yasei-no-otoko opened Oct 8, 2026·

  ·
  7

• Migrate kernels tiled tdm

  #1238·xudoyuan opened Oct 8, 2026Collaborator·1 task·

  ·

• [Bugfix] Compute dynamic layout sizes in i64 and name oversized tensor dims

  #1236·Phil-amd opened Oct 8, 2026Member·

  ·
  

• [Bugfix] Sync waves before the RMSNorm/LayerNorm quant kernels reuse their reduction slot

  #1235·jhinpan opened Oct 7, 2026Collaborator·1 of 5 tasks·

  ·

• [Kernel]feat(dsv4): add the DeepSeek-V4 decode MonoKernel

  #1234·amd-weisun opened Oct 7, 2026Contributor·

  ·
  11

• [Perf][gfx1250] Add a prefill config selector to the BF16 GEMM

  #1232·jhinpan opened Oct 6, 2026Collaborator·3 of 5 tasks·

  ·