All pull requests

Pull requests

• Fake-quantize all HF fused NVFP4 experts in one kernel launch

  #3011·sychen52 opened Oct 25, 2026Contributor·

  ·
  2

• perf(speculative): fuse the DFlash-family loss with the lm_head projection

  #3010·h-guo18 opened Oct 25, 2026Contributor·

  ·
  4

• [Puzzletron] Use the model's head_dim for the KV-cache memory estimate

  #3009·soksof opened Oct 25, 2026·

  ·
  1

  14

• Skip Qwen3.8's QSA indexer loop during calibration when it keeps every token

  #3008·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  5

• [OMNIML-5802] Drop Python 3.10 support and adopt Python 3.11+ features

  #3007·kevalmorabia97 opened Oct 25, 2026Collaborator·

  ·
  9

• Require accelerate>=1.15.0 for llm_qat

  cherry-pick-0.48.0
  #3006·kevalmorabia97 opened Oct 25, 2026Collaborator·

  ·
  11

• Document installing flash-linear-attention for linear-attention models

  #3003·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  5

• Run fused-MoE expert GPTQ updates on all of this process's GPUs

  #3002·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  3

• perf(speculative): deliver training batches out of order; compile + FlexAttention in the DFlash2 recipe

  #3001·h-guo18 opened Oct 25, 2026Contributor·

  ·
  4

• Restore each module's training mode after calibration

  #3000·shoumikhin opened Oct 25, 2026·

  ·
  4

• feat(launcher): prepare required assets before submission

  #2999·ChenhanYu opened Oct 25, 2026Collaborator·

  ·
  6

• Fix _get_generator is_manual flag and seed=0 handling

  #2998·udsy19 opened Oct 24, 2026·

  ·
  2

• [6938144] Stop making /workspace world-writable in the vLLM serve example image

  #2997·shengliangxu opened Oct 24, 2026Collaborator·

  ·
  8

• [OMNIML-6284] Make awq_lite independent of how the calibration tokens are batched

  #2996·shengliangxu opened Oct 24, 2026Collaborator·

  ·
  16

• [OMNIML-6284] Add a PTQ scale-comparison tool and hf_ptq --deterministic

  #2995·shengliangxu opened Oct 24, 2026Collaborator·

  ·
  37

• refactor(speculative): draft DSpark and Domino through one AR-validation loop

  #2994·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• feat(speculative): DFlash-family loss terms in config; one teacher with an opt-in softcap

  #2993·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• feat(speculative): log the DFlash family's draft_metrics from the example trainer

  #2992·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• feat(speculative): D-PACE position weighting for DSpark and Domino

  #2991·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• [2/2] refactor(speculative): train the DFlash family through one forward with shared loss terms

  #2990·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• [1/2] refactor(speculative): hand the DFlash-family loss one context per training step

  #2989·h-guo18 opened Oct 24, 2026Contributor·

  ·
  2

• Fix tensor-parallel gradient scaling in Top-K KL distillation

  #2988·wiederse opened Oct 24, 2026·

  ·
  11

• Drop unverifiable claims from the external-draft docs

  #2987·phicksy opened Oct 24, 2026Contributor·

  ·
  7

• fix(speculative): weigh DFlash-family losses per anchor position, not per sequence

  #2986·h-guo18 opened Oct 24, 2026Contributor·

  ·
  4

• Add GGML IQ1/IQ2 x NVFP4-activation (W1A4/W2A4) calibration and an NVFP4 converter

  #2985·cjluo-nv opened Oct 24, 2026Collaborator·

  ·
  4