All pull requests

Pull requests

• Feed decoder projections their weight untransposed on CUDA

  #807·y-chan opened Oct 29, 2026·

  ·
  6

  

• Isolate A10 CI Python dependencies to fix PEP 668 failures

  #806·eserscor opened Oct 29, 2026·

  ·
  5

• Offer the static KV cache to GroupQueryAttention

  #805·y-chan opened Oct 29, 2026·

  ·
  1

• Give speech-language decoders the static KV cache

  #804·y-chan opened Oct 29, 2026·

  ·
  1

• Add fixed-capacity Gemma4 vision exports

  #803·ananyaa-ms opened Oct 29, 2026·

  ·
  4

• Consider offline static weight-layout folding separately

  #801·titaiwangms opened Oct 28, 2026Contributor·

  ·
  2

• Add opt-in Drama 3 CUDA 13.0 consumer qualification adapter

  #800·justinchuby opened Oct 28, 2026Member·

  ·
  2

• Add isolated Drama 3 runtime audit and benchmark consumer

  #799·justinchuby opened Oct 28, 2026Member·

  ·
  1

• Add portable source-ordered RoPE for Drama 3

  #798·justinchuby opened Oct 28, 2026Member·

  ·
  2

• Integrate portable Drama 3 GDN and explicit continuation context

  #797·justinchuby opened Oct 28, 2026Member·

  ·
  2

• Scope Drama 3 source-ordered RMS normalization

  #796·justinchuby opened Oct 28, 2026Member·

  ·
  2

• Scope Drama 3 source-layout projections with reusable output-major GEMM

  #795·justinchuby opened Oct 28, 2026Member·

  ·
  2

• Support per-channel quantized export and configurable MatMulNBits acc…

  #794·neilmsft opened Oct 27, 2026·

  ·
  17

• Add Drama 3 ONNX speech graphs and sampling

  #793·justinchuby opened Oct 27, 2026Member·

  ·
  12

• [Flint] Add experimental CUDA BF16 RecursiveSI Flint onboarding

  #792·rui-ren opened Oct 25, 2026Contributor·

  ·
  5

• Fix Gemma4 GGUF shared-KV import and tied quantized output head

  #791·rui-ren opened Oct 25, 2026Contributor·

  ·
  6

• Bound GGUF affine requantization scratch allocations

  #790·rui-ren opened Oct 25, 2026Contributor·

  ·
  3

• docs: investigate external Bolt Mamba precision boundaries

  #786·rui-ren opened Oct 22, 2026Contributor·

  ·
  1

• Add opt-in sparse IQ2 MoE lowering for Nemotron-H Omni

  #785·rui-ren opened Oct 22, 2026Contributor·

  ·
  3

• Allow canonical Olive INT3 expert overrides in native QMoE export

  #783·jiafatom opened Oct 21, 2026·

  ·
  1

• Implement opt-in Qwen4-Exp ngram/main split

  #780·titaiwangms opened Oct 19, 2026Contributor·

  ·
  4

• Add CUDA Nemotron 3.5 Bolt Omni export with native mixed quantization

  #779·rui-ren opened Oct 15, 2026Contributor·

  ·
  24

• Document Qwen4-Exp MTP forward contract

  #777·titaiwangms with CopilotAI opened Oct 14, 2026·

  ·
  1

  6

  

• Propose opt-in Qwen4-Exp ngram/main split contract

  #776·titaiwangms with CopilotAI opened Oct 14, 2026·

  ·
  1

  8

  

• Add K2 Horizon model (llama.cpp dense + MoVA support)

  #763·gramalingam opened Oct 8, 2026Collaborator·

  ·
  18