All pull requests

Pull requests

• CompositeSampler: avoid full-vocabulary sort for top-p/min-p without topK

  #524·BhargavKumarNath opened Nov 13, 2026·3 of 4 tasks·

  ·
  1

• Grammar sessions are reused only for the same stop tokens, and padded logits are masked

  #522·Lewis300 opened Nov 12, 2026Contributor·3 of 8 tasks·

  ·
  1

• fix(vlm): keep Muse Glimmer prefill causal

  #519·Coding-Professional opened Nov 12, 2026·3 tasks done·

  ·
  1

• docs: list SAM 3 Video in the model catalog

  #518·Coding-Professional opened Nov 12, 2026·3 tasks done·

  ·

• Gemma 4: pack the draft head for a CPU runtime

  #514·carinapeng opened Nov 11, 2026Contributor·

  ·

• Speculative decoding: run the Gemma 4 draft head on the CPU

  #513·carinapeng opened Nov 11, 2026Contributor·

  ·
  2

• Dynamic Flux2

  #510·kevchengcodes opened Nov 10, 2026Contributor·4 of 8 tasks·

  ·
  1

  

• Rewind to the shared prefix on mid-sequence divergence instead of resetting the KV cache

  #506·PratikDhanave opened Nov 9, 2026Contributor·3 of 8 tasks·

  ·
  1

• Serve image input over the chat API

  #501·stikves opened Nov 7, 2026Contributor·8 tasks·

  ·
  

• Add SmolLM3 support and export configurations (#493)

  #496·sirohikartik opened Nov 7, 2026Contributor·5 of 8 tasks·

  ·
  1

  10

• Fix check_divisibility IndexError on 1-D-weight modules, add unit tests

  #495·PratikDhanave opened Nov 7, 2026Contributor·

  ·
  3

• Document SDPA width abort and .aimodelc load, size and cache behavior in common_issues

  #487·john-rocky opened Nov 7, 2026Contributor·3 of 8 tasks·

  ·

• Add Qwen3-Embedding-0.6B to coreai.embedding.export

  #486·john-rocky opened Nov 6, 2026Contributor·4 of 8 tasks·

  ·
  1

  13

• Run one-token prefill chunks on main

  #432·john-rocky opened Nov 5, 2026Contributor·3 tasks done·

  ·
  1

  2

• Mutable weights: tensors held on device that change after export

  #405·carinapeng opened Nov 3, 2026Contributor·3 of 8 tasks·

  ·
  1

• Add ModelScope as optional model download backend

  #364·yrk111222 opened Oct 22, 2026·5 of 8 tasks·

  ·
  30