All pull requests

Pull requests

• fix: preserve Qwen4Exp MTP decode numerics

  CUDAdocumentationggmlmodel
  #234·pwilkin opened Oct 29, 2026Collaborator·

  ·

• cuda : Gated DeltaNet decode fusion with the state read in place

  CUDAggml
  #231·Nathanw1014 opened Oct 27, 2026·

  ·

• rpc, sched: batch cross-backend split input reads; dual Strix Halo notes

  documentationggml
  #230·dzannotti opened Oct 27, 2026Collaborator·

  ·

• vulkan : fuse the qwen4exp QSA indexer scorer into its GEMM epilogue (QSA_SCORE)

  ggmltestingVulkan
  #229·Nathanw1014 opened Oct 27, 2026·

  ·

• vulkan : qwen4exp complete-block selection after TOP_K in one dispatch (QSA_EXPAND)

  ggmltestingVulkan
  #228·Nathanw1014 opened Oct 27, 2026·

  ·

• cuda : mm_ids_helper_512_10 skips expert ids outside [0, 512)

  CUDAggml
  #227·gaetan-puleo opened Oct 26, 2026Collaborator·

  ·

• cuda : MMB BF16 shadows only for weights, dropped when the weight memory changes

  CUDAggml
  #226·gaetan-puleo opened Oct 26, 2026Collaborator·

  ·

• model : fix qwen4exp layer inputs for DFlash drafters

  model
  #225·Nathanw1014 opened Oct 26, 2026·

  ·

• vulkan : stream the q5_1 MoE down weights in MUL_MAT_ID(+MUL)+CPY(f16)

  ggmltestingVulkan
  #224·Nathanw1014 opened Oct 26, 2026·

  ·

• vulkan : MoE gate + up + SwiGLU in one weight-streaming MUL_MAT_ID launch (q4_K / q5_K, f16 and f32 chains)

  ggmltestingVulkan
  #223·Nathanw1014 opened Oct 26, 2026·

  ·
  8

• qwen4exp : q8_0 K/V for the QSA selected-key kernels, share score-strip input views

  CUDAggmlmodel
  #219·iggy opened Oct 25, 2026·

  ·

• llama : extract embeddings at the tensor row width (qwen4exp / deepseek4 abort)

  #218·gaetan-puleo opened Oct 25, 2026Collaborator·

  ·
  2

• llama : synchronize before writing the inputs of a reused graph

  #217·gaetan-puleo opened Oct 25, 2026Collaborator·

  ·

• qwen4exp : multi-slot compute memory (split-stream strips, bias upload, compact visibility)

  model
  #216·gaetan-puleo opened Oct 25, 2026Collaborator·3/3·

  ·

• qwen4exp : 1M context (uint32 QSA block ids, depth-bounded indexer scratch, RoPE-scaled slots)

  CUDAggmlmodelserver
  #215·gaetan-puleo opened Oct 25, 2026Collaborator·2/3·

  ·
  1

• llama : stage mmap -> integrated GPU weight copies through a host buffer

  #214·gaetan-puleo opened Oct 24, 2026Collaborator·

  ·

• vulkan : exact int8 views for IQ2/IQ3 and ternary types on the int8 coopmat MMQ

  ggmltestingVulkan
  #208·Nathanw1014 opened Oct 21, 2026·

  ·

• vulkan : EXL3 matmul ops (EXL3_MUL_MAT and EXL3_MUL_MAT_ID)

  conversionCUDAdocumentationggmlmodeltestingVulkan
  #204·Nathanw1014 opened Oct 20, 2026·

  ·

• feat: exl3 support

  conversionCUDAdocumentationggmlmodeltesting
  #201·pwilkin opened Oct 19, 2026Collaborator·

  ·
  1

• hip: share IQ3 decode across verify widths

  CUDAggmltesting
  #196·BonkusCheemus opened Oct 17, 2026·

  ·
  2

• server : keep context checkpoints in unlinked files (--ctx-checkpoints-dir)

  server
  #180·rcfa opened Oct 15, 2026·

  ·
  1

• server : purge the least recently used idle slot and keep it in the prompt cache

  server
  #176·rcfa opened Oct 14, 2026·

  ·
  1

• server : save the draft context state with the slot

  server
  #171·rcfa opened Oct 12, 2026·

  ·
  1

• qsa : skip foreign sequences when filling scan tails

  #165·iTechMedic opened Oct 11, 2026·

  ·

• qsa : keep the incremental QSA prefix across mrope images (target and MTP draft)

  #163·iTechMedic opened Oct 9, 2026·

  ·
  5