All pull requests

Pull requests

• properly pass --clear-cache-threshold to LoRA trainer

  #2000·Lumonopolis opened Oct 22, 2026·

  ·

• fix(gemma3n): apply KV sharing without cache (#1674)

  #1999·Yu-Haikuo opened Oct 22, 2026·

  ·
  1

• Recursive Flint model support

  new_model
  #1998·pcuenca opened Oct 21, 2026Contributor·

  ·
  5

• fix: handle nested parameter tags in Qwen3 Coder parser

  #1996·rafiimam opened Oct 21, 2026·

  ·
  1

• Fix nested parameter tags in qwen3_coder tool parser

  #1993·axiom-of-choice opened Oct 19, 2026Contributor·

  ·
  1

  1

• Add NVIDIA Nemotron-Labs Dory model

  new_model
  #1988·pmichalas opened Oct 17, 2026·1 task done·

  ·
  1

• feat(qwen3_5): honour explicit layer_types over full_attention_interval

  #1987·hermitdave opened Oct 17, 2026·

  ·

• Add --loss-chunk-size to reduce LoRA training memory with large vocabularies

  #1983·spicysharkco opened Oct 16, 2026·

  ·

• nemotron_h: support Nemotron 3.5 Bolt (heterogeneous Puzzle expert counts, omni key layout)

  new_model
  #1982·vichzp opened Oct 16, 2026·

  ·

• Fix pipeline load when a shard holds another stage's scales

  #1980·ashokgelal opened Oct 15, 2026·

  ·

• Fix multi-token attention mask in RotatingKVCache

  #1978·robertcprice opened Oct 15, 2026·

  ·

• Fix silent drop of unparseable tool calls in server

  #1972·santanu20 opened Oct 14, 2026·

  ·
  1

• qwen3_moe honors rope_scaling; handles transformers >= 5

  #1968·hierophantos opened Oct 13, 2026·

  ·

• Qwen3-Coder parser loses precision when parsing large integer tool arguments

  #1967·anidotdev opened Oct 12, 2026·

  ·

• Fix /v1/models sending 200 before the cache scan can fail

  #1965·HamedShams opened Oct 12, 2026·

  ·
  1

  1

• Read RoPE settings from transformers 5 rope_parameters

  #1964·JayeshSuryavanshi opened Oct 11, 2026·

  ·

• Support qwen3_5_text and qwen3_5_moe_text model types

  #1963·JayeshSuryavanshi opened Oct 11, 2026·

  ·

• Use mx.fast.gated_delta_update on GPUs without Metal, including Apple Silicon under Mesa/Vulkan.

  #1961·joshuaswarren opened Oct 9, 2026·

  ·

• Fix Gemma4 loading with per-layer attention configs

  #1954·dogukanveziroglu opened Oct 8, 2026Contributor·

  ·
  1

  1

• fix(server): match the xtc_threshold range to the sampler's

  #1952·NishchayMahor opened Oct 6, 2026Contributor·

  ·

• Validate XTC arguments outside the compiled function, so one bad argument stops breaking all sampling

  #1951·NishchayMahor opened Oct 6, 2026Contributor·

  ·

• Skip blocking logprob reads when logprobs are not requested

  #1950·lawrencehui opened Oct 6, 2026·

  ·
  1

• Add K2-Horizon (IFM) model support

  #1948·Avicennasis opened Oct 6, 2026·1 task done·

  ·

• add support for the (German language) Kolibri 1 model by Aleph Alpha

  #1945·velaia opened Oct 4, 2026·

  ·
  1

• Enforce the prompt cache byte budget continuously and make the /health endpoint stall-aware

  #1944·anubhavagr opened Oct 4, 2026·

  ·
  1

  1