All pull requests

Pull requests

• fix(server): name the lifecycle stage in every API's 503

  #616·YevheniiKotyrlo opened Oct 6, 2026Contributor·

  ·

• fix(engine): start a single rank without a rendezvous port

  #615·YevheniiKotyrlo opened Oct 6, 2026Contributor·

  ·

• fix(server): leak no semaphores when a signal stops the server

  #614·YevheniiKotyrlo opened Oct 6, 2026Contributor·

  ·

• fix(server)!: answer /health with 503 until the server accepts requests

  #613·YevheniiKotyrlo opened Oct 6, 2026Contributor·

  ·
  1

• fix(moe): route moe_align to the atomic-free staged path below sm_70

  #612·wac81 opened Oct 6, 2026·

  ·

• perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s)

  #610·SuperHotDogCat opened Oct 6, 2026·

  ·
  2

• fix(models): preserve glm-5.2 router correction bias in fp32

  #609·ischencheng opened Oct 5, 2026·

  ·

• perf(pre-sm70): pipeline the block-fp8 GEMMs and halve the extend-attention K/V re-reads on Pascal

  #605·wac81 opened Oct 5, 2026·

  ·

• fix(server): soft-accept OpenAI response_format json_object/json_schema

  #604·berdachuk opened Oct 5, 2026·5 tasks·

  ·

• feat(deepseek-v41): support text and vision serving

  new-model
  #603·jason-fxz opened Oct 5, 2026Collaborator·

  ·

• fix(moe): fit Qwen3.8 Flash Next FP8 banks across host and GPU memory

  #602·yunwei37 opened Oct 4, 2026Contributor·

  ·
  1

  1

• feat(engine): --embed-device cpu keeps the input embedding in RAM

  #596·KarrAcaRn opened Oct 4, 2026·

  ·

• feat(server): add ft info and a pre-load memory preflight

  #595·KarrAcaRn opened Oct 3, 2026·

  ·
  1

• feat(qwen4_exp): support ModelOpt QAD checkpoints (NVFP4 PLE tables, MXFP8 dense)

  #594·skavans opened Oct 3, 2026·

  ·

• fix(checkpoint): refuse converting into a directory that already holds an FTW

  #593·Avicennasis opened Oct 3, 2026·

  ·

• fix(mm): classify image fetch errors and re-check redirect targets

  #592·Avicennasis opened Oct 3, 2026·

  ·

• perf(moe): fold the shared-expert gate epilogue into hc_combine_norm

  #591·skavans opened Oct 2, 2026·

  ·

• perf(model): fuse hc combine with the next block's RMSNorm

  #590·skavans opened Oct 2, 2026·

  ·

• fix(checkpoint): resolve HF repo Ids before conversion

  #589·rbrisita opened Oct 2, 2026·

  ·

• feat(rocm): run on RDNA3/RDNA4 with torch 2.14/2.15 and ROCm 7.14/10

  #587·zhangnju opened Oct 1, 2026·3 tasks done·

  ·
  1

• refactor(hf): merge an override into the section's own dict, with no fallback

  #586·YevheniiKotyrlo opened Oct 1, 2026Contributor·

  ·

• perf(linear): run a single-row bf16 projection as one GEMV

  #585·YevheniiKotyrlo opened Sep 30, 2026Contributor·

  ·

• fix(engine): size memory pools against the VRAM other processes leave free on Windows

  #583·YevheniiKotyrlo opened Sep 30, 2026Contributor·

  ·
  1

• test(quant): skip declined local checkpoints and label cache paths portably

  #582·YevheniiKotyrlo opened Sep 30, 2026Contributor·

  ·

• fix(windows): read checkpoints without O_DIRECT, posix_fadvise or PROT_READ

  #581·YevheniiKotyrlo opened Sep 30, 2026Contributor·

  ·