All pull requests

Pull requests

• Replace block pointers in the NVFP4 fake-quant Triton kernel

  #3194·0z5a opened Nov 3, 2026·

  ·
  1

  5

• [12/12] Expose the fixed DLA pipeline through the API and CLI

  #3192·mgohil-png opened Nov 3, 2026Contributor·12/12·

  ·
  7

• [11/12] Add optional DLA FP16 precision and terminal projection

  #3191·mgohil-png opened Nov 3, 2026Contributor·11/12·

  ·
  7

• [10/12] Add connected DLA reshape reduction and TopK patterns

  #3190·mgohil-png opened Nov 3, 2026Contributor·10/12·

  ·
  7

• [9/12] Add DLA logical rewrites and float consumer views

  #3189·mgohil-png opened Nov 3, 2026Contributor·9/12·

  ·
  5

• [8/12] Add fixed-sequence DLA LSTM decomposition

  #3188·mgohil-png opened Nov 3, 2026Contributor·8/12·

  ·
  5

• [7/12] Add DLA normalization and native GQA layout adapters

  #3187·mgohil-png opened Nov 3, 2026Contributor·7/12·

  ·
  6

• [6/12] Add DLA indexing reduction and axis layout rules

  #3186·mgohil-png opened Nov 3, 2026Contributor·6/12·

  ·
  6

• [5/12] Add DLA shape and rank layout rules

  #3185·mgohil-png opened Nov 3, 2026Contributor·5/12·

  ·
  10

• [4/12] Add DLA quantization and FP32 preprocessing

  #3184·mgohil-png opened Nov 3, 2026Contributor·4/12·

  ·
  17

• [3/12] Lower DLA indexing and fixed matrix operators

  #3183·mgohil-png opened Nov 3, 2026Contributor·3/12·

  ·
  6

• [2/12] Add DLA layout planning and boundary passes

  #3182·mgohil-png opened Nov 3, 2026Contributor·2/12·

  ·
  10

• [1/12] Replace legacy DLA helpers with an IR foundation

  #3181·mgohil-png opened Nov 3, 2026Contributor·1/12·

  ·
  18

• [NVBug 7029942] Fix Triton NVFP4 fake quantization ignoring the configured block size

  #3180·yingguo-trt opened Nov 3, 2026Contributor·

  ·
  13

• Write GGML layers as ModelOpt W8A16 in convert_ggml_to_int8.py, leaving other layers as exported

  #3176·cjluo-nv opened Nov 2, 2026Collaborator·

  ·
  5

• Add dflash_detach_target_hidden to sever the draft-to-base gradient path

  #3175·yeyu-nvidia opened Nov 2, 2026Contributor·5 tasks done·

  ·
  4

• [NVBug 6938145] Keep HF_TOKEN off argv and use a private temp dir in the MRCR gym recipe

  #3174·cjluo-nv opened Nov 2, 2026Collaborator·

  ·
  11

• [OMNIML-6340] Add NVFP4 weight-only local-shift GSQ

  #3173·ajrasane opened Nov 2, 2026Contributor·

  ·
  2

• Support image-text calibration in the FSDP2 PTQ flow

  #3172·sugunav14 opened Nov 2, 2026Contributor·

  ·
  16

• [OMNIML-6370] Fix AutoQuantize resume from a calibration-only checkpoint

  #3169·meenchen opened Nov 2, 2026Contributor·

  ·
  4

• [chore]: weekly bump of uv.lock on main (2026-11-02)

  #3168·github-actions[bot] opened Nov 2, 2026Bot·

  ·
  2

• fix static nvfp4 fake quant for unpromoted weight quantizers in TensorQuantizer

  #3167·saloni-garg opened Nov 2, 2026·

  ·
  1

  3

• Apply the legacy "default" quant_cfg key first wherever it appears

  #3166·wguo7 opened Nov 1, 2026·

  ·
  2

• Bump the github-actions group across 2 directories with 4 updates

  dependenciesgithub_actions
  #3165·dependabot[bot] opened Nov 1, 2026Bot·

  ·
  4

• Add experimental NVFP4 QAD adapter for NeMo AutoModel experts

  #3164·Edwardf0t1 opened Nov 1, 2026Contributor·

  ·
  7