All pull requests

Pull requests

• Control vectors: a native GLP steering vector for Qwen3.8-27B on Metal

  #53·BlivionIaG opened Oct 27, 2026Owner·9 tasks done·

  ·

• [Draft] EXL3 Flash Next: DeltaNet heads split across the two ranks for decode rows, opt-in (slice 7, #1419)

  #52·BlivionIaG opened Oct 26, 2026Owner·

  ·

• [Draft] EXL3 Flash Next: drafts on by default for one user, guarded by measured tokens a ms (slice 6, #1419)

  #51·BlivionIaG opened Oct 26, 2026Owner·

  ·

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

  #49·BlivionIaG opened Oct 19, 2026Owner·9 tasks done·

  ·

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

  #48·BlivionIaG opened Oct 19, 2026Owner·9 tasks done·

  ·

• EXL3 Qwen3.8-Flash-Next on the native CUDA engine (WIP, to be split)

  #47·BlivionIaG opened Oct 18, 2026Owner·

  ·

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

  #46·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #45·BlivionIaG opened Oct 18, 2026Owner·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

  #44·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #43·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #42·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #41·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #40·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #39·BlivionIaG opened Oct 18, 2026Owner·9 tasks done·

  ·

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

  #38·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #37·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #36·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #35·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #34·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #33·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #32·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #31·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #30·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #29·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·

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

  #28·BlivionIaG opened Oct 17, 2026Owner·9 tasks done·

  ·