All pull requests

Pull requests

• perf(hip): band the prefill WMMA GEMM's blocks so row blocks share weights

  #1648·AbrahamCete opened Oct 30, 2026Contributor·8 of 9 tasks·

  ·

• deepseek: fail fast on a lost CUDA context and on stalled engine operations

  #1646·jayleaton opened Oct 30, 2026Contributor·7 of 9 tasks·

  ·

• server: initialize media TLS safely and bound socket waits

  #1645·jayleaton opened Oct 30, 2026Contributor·7 of 9 tasks·

  ·

• deepseek: read the pack's content identity once a boot

  #1644·jayleaton opened Oct 30, 2026Contributor·7 of 9 tasks·

  ·

• deepseek: bound CUDA weight and persistent-form upload stages

  #1643·jayleaton opened Oct 30, 2026Contributor·8 of 9 tasks·

  ·

• deepseek: prefer executable EXL3 siblings in CUDA rank plans

  #1642·jayleaton opened Oct 30, 2026Contributor·8 of 9 tasks·

  ·

• deepseek: admit supported dsq-1 checkpoint metadata

  #1641·jayleaton opened Oct 30, 2026Contributor·8 of 9 tasks·

  ·

• deepseek: build CPU asset generators before CUDA images

  #1640·jayleaton opened Oct 30, 2026Contributor·7 of 9 tasks·

  ·

• glm cuda: the load packs a layer's routed experts on 8 threads into 192 MiB of mapped pinned memory the GPU copies up, with 16 readers ahead

  #1636·Bizuayeu opened Oct 30, 2026Contributor·9 tasks done·

  ·

• lfm3_omni: LFM3-Omni-2B on Metal, speech in, text and speech out (part of #1627)

  #1635·drowzeys opened Oct 30, 2026Contributor·4 of 5 tasks·

  ·
  1

• lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508)

  #1634·jschmied opened Oct 30, 2026Contributor·

  ·

• Nemotron 3.5 Transcribe: audio input, the 65,536-token window, an hour of audio, MTP drafts

  #1632·BobClawblaw opened Oct 30, 2026Contributor·

  ·
  1

• Nemotron 3.5 Transcribe on CUDA: the language model and the audio encoder

  #1631·BobClawblaw opened Oct 30, 2026Contributor·

  ·
  2

• cuda experts: a bf16 format for a checkpoint's own bf16 experts

  #1630·BobClawblaw opened Oct 30, 2026Contributor·

  ·

• kernel contract: Nemotron's pocket kernels (nemotron_ops.cu)

  #1629·BobClawblaw opened Oct 30, 2026Contributor·

  ·

• relu^2 keeps a NaN: grouped experts, ck experts and Nemotron's pocket

  #1628·BobClawblaw opened Oct 30, 2026Contributor·

  ·

• glm cuda: DFlash2 reads complete prompt taps at TP>1

  #1623·BadAd84 opened Oct 29, 2026Contributor·

  ·

• hip flashnext: WMMA sparse attention for prompt chunks (sparse=wmma)

  #1622·gilby opened Oct 29, 2026Contributor·8 of 9 tasks·

  ·
  1

  1

• kolibri1 CUDA: the full layers' caches in pages of the core pool, every stream leasing; sliding layers keep their rings

  #1620·jschmied opened Oct 29, 2026Contributor·

  ·

• gemma4 CUDA: the global layers' caches in pages of the core pool, every stream leasing; sliding layers keep their rings

  #1619·jschmied opened Oct 29, 2026Contributor·

  ·

• hip flashnext: random-access advice on the n-gram table mappings

  #1615·gilby opened Oct 29, 2026Contributor·9 tasks done·

  ·
  1

  1

• hip: decode router over the bf16 rows the checkpoint holds

  #1614·gilby opened Oct 29, 2026Contributor·9 tasks done·

  ·
  1

• hip flashnext: take the MTP head's root from the index

  #1613·gilby opened Oct 29, 2026Contributor·8 of 9 tasks·

  ·
  1

  1

• hip: load a shared expert wider than the routed experts beside the stack

  #1612·gilby opened Oct 29, 2026Contributor·8 of 9 tasks·

  ·
  1

  1

• hip flashnext: parallel n-gram gather, release weight pages after upload

  #1611·gilby opened Oct 29, 2026Contributor·9 tasks done·

  ·
  1

  1