All pull requests

Pull requests

• ci(rocm): add ROCm 10.0.0 Dockerfile and validation workflow

  #535·LZ-QWQ opened Sep 22, 2026·

  ·

• fix(server): clamp /v1/models context_length to the allocated KV pool

  #531·oscar-investmatic opened Sep 20, 2026·

  ·
  1

• fix(sampling): apply presence and frequency penalties

  #527·taking-lying-flat opened Sep 19, 2026Contributor·

  ·

• feat(kvcache): host-resident prefix tier for hybrid GDN models (snapshots + KV pages)

  #525·alvarorsouza-arch opened Sep 19, 2026·

  ·

• feat(glm5_next): serve GLM-5.3-Flash with tensor parallelism

  #507·calvarado2004 opened Sep 17, 2026·

  ·

• fix(scheduler): preserve pending hybrid checkpoint across prefill chunks

  #505·eamars opened Sep 17, 2026·

  ·
  1

• feat(server): serve per-request inference metrics under --enable-metr…

  #504·FantasticCode2019 opened Sep 17, 2026·

  ·
  1

• feat(deepseek_v4): serve image input on the V4-Flash-Vision release

  #502·void-mckenzie opened Sep 17, 2026·

  ·
  1

• perf(moe): single-launch prefill buffer invalidation without hidden sync

  #500·alvarorsouza-arch opened Sep 17, 2026·

  ·
  3

• feat(kvcache): host-RAM KV tier for sparse-attention models (--kv-host-pages)

  #499·alvarorsouza-arch opened Sep 17, 2026·

  ·
  1

• feat(qwen4_exp): support channel-wise fp8 dense weights and per-row PLE loaders

  #498·alvarorsouza-arch opened Sep 17, 2026·

  ·

• fix(scheduler): wait for rank subscribers before first broadcast

  #495·Cyber-Marty opened Sep 17, 2026·

  ·
  1

• fix(gemma4): load mixed-quant GGUF checkpoints

  #494·Poyo-Yon opened Sep 16, 2026·

  ·
  1

• fix(gemma4): load mixed compressed-tensors nvfp4

  #493·Poyo-Yon opened Sep 16, 2026·

  ·
  1

• refactor(moe): extract hybrid decode orchestration

  #491·zihaomu opened Sep 16, 2026·

  ·
  1

• fix(swa): keep the prompt head's windowed KV so fan-out can reuse a shared prefix

  #488·huhoo opened Sep 15, 2026·

  ·
  1

• fix(server): hoist system messages to front for templates requiring system-first

  #487·rafaeldrincon opened Sep 15, 2026·

  ·

• feat(scheduler): interleave decode steps into long prefill runs

  #484·huhoo opened Sep 15, 2026·

  ·
  1

  2

• feat(server): expose POST /v1/tokenize endpoint for client-side token counting

  feature
  #477·rafaeldrincon opened Sep 14, 2026·

  ·
  1

  1

• feat(server): add --default-thinking-mode flag for server-wide thinking default

  #476·rafaeldrincon opened Sep 14, 2026·

  ·
  1

• fix(tokenizer): respect explicit enable_thinking=false even when tools present

  #475·rafaeldrincon opened Sep 14, 2026·

  ·
  1

• fix(engine): reject fa attention backend below sm_90 at config time

  #468·SanjayS45 opened Sep 14, 2026·

  ·
  1

• fix(utils): tolerate coalesced msgpack frames in zmq pull queues

  #466·Cyber-Marty opened Sep 14, 2026·

  ·
  1

• fix(scheduler): match stop strings with incremental decoding

  #464·taking-lying-flat opened Sep 14, 2026Contributor·

  ·

• fix(server): loopback-bind the TP rendezvous store, add --dist-port

  #461·shikhakath opened Sep 13, 2026·

  ·
  1