All pull requests

Pull requests

• perf(hip): one row's routed fc1 and shared up in one launch, both with the relu^2 epilogue

  #1103·BlivionIaG opened Oct 19, 2026Contributor·

  ·

• perf(hip): the planned decode walk in one launch a query head for spans up to 4096 keys

  #1102·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(nemotron): fewer launches a HIP decode token: q, k and v grouped, the residual and next norm merged, Mamba windows in three launches

  #1101·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): the argmax of a wide logits row in parts across the card

  #1100·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): the planned decode walk reads q and writes its rows in the activation type

  #1099·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): Nemotron's relu^2 as the stream tile's epilogue for one-row MoE layers

  #1098·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): a Mamba-2 window in three launches, each row's state kept in its snapshot slot

  #1097·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): the residual add and the RMS norm in one launch, the Nemotron MoE sum among its inputs

  #1096·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(nemotron): HIP prompts in 2048-row chunks, so the routed experts get enough rows a tile

  #1095·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(nemotron): lane rounds, lane costs and graph replay for the Nemotron HIP engine, and the decode-path test

  #1094·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(nemotron): tensor parallelism on HIP: followers run rank 0's steps, the cache and drafts agreed across ranks

  #1093·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(nemotron): Nemotron-H on the native HIP engine, one GPU: admission, loader, forward and lane backend

  #1092·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• perf(hip): the planned values walk splits a KV head's query heads while its grid is small

  #1091·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(hip): Nemotron-H MoE layer on the shared expert plan and products, its pick ranked on 512 threads

  #1090·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• refactor(hip): the routed MoE's tile choice is public, its rule tested

  #1089·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(hip): Nemotron Mamba-2 mixer kernels, launches, ops and tests

  #1088·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• feat(hip): a HIP family can refuse --slide and --mtp-confidence by name before it loads

  #1087·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• fix(hip): refuse --kv-cache fp8 by name until HIP has fp8 caches

  #1086·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• refactor(hip): shared HIP serving pieces and lane cost timing in hip_serve; Qwen3.5 on them

  #1085·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• fix(qwen3_5): no prompt cache step goes to the other ranks once the group stopped

  #1084·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• fix(hip): preserve the MLX stream nonfinite classes

  #1083·BlivionIaG opened Oct 19, 2026Contributor·

  ·

• feat(hip): the kernel contract runs on HIP

  #1082·BlivionIaG opened Oct 19, 2026Contributor·9 tasks done·

  ·

• qwen CUDA: long-prompt prefill 5.4x faster (4096-row chunks, tensor-core attention, the delta-rule chain)

  #1080·BobClawblaw opened Oct 19, 2026Contributor·

  ·

• qwen CUDA: Qwen3.8-27B decode 9-15% faster with DFlash2 (code +25%), replies unchanged

  #1079·BobClawblaw opened Oct 19, 2026Contributor·

  ·

• qwen CUDA: TF_DRAFT_NODES above 15 serves; a stop or cancel lands between prompt chunks

  #1078·BobClawblaw opened Oct 19, 2026Contributor·

  ·