All pull requests

Pull requests

• Remove Incorrect Low and High Per Kernel Impact Estimates from JSON

  #1103·tsrikris opened Oct 9, 2026Contributor·

  ·
  1

• Add busy ratio metric to trace splitter metadata

  #1102·kyle-hoffmeyer opened Oct 9, 2026Collaborator·

  ·
  3

• Run pull-request tests related to the changed files

  #1101·gabeweisz opened Oct 8, 2026Collaborator·3 tasks·

  ·
  3

• fix(perfmodel): sweep tile shapes in the Triton MXFP4/MXFP6 microbenchmark

  #1098·mehdi-saeedi opened Oct 7, 2026Contributor·

  ·
  6

• fix(perfmodel): use the 16-bit matrix peak for MX formats on GPUs without native FP4/FP6

  #1097·mehdi-saeedi opened Oct 7, 2026Contributor·

  ·
  2

• fix(perfmodel): roof Triton bf16 x mxfp4 MoE GEMMs at the activation width

  #1096·mehdi-saeedi opened Oct 7, 2026Contributor·

  ·
  2

• Generate perf reports from one YAML config

  #1094·ajassani opened Oct 7, 2026Collaborator·1 of 2 tasks·

  ·
  1

• Feat/trace health check

  #1088·kyle-hoffmeyer opened Oct 6, 2026Collaborator·

  ·
  1

• feat(ci): agent-driven PR self-review

  #1084·Ahmedhasssan-aig opened Oct 6, 2026Contributor·

  ·
  1

• Update xDiT patches

  #1083·kyle-hoffmeyer opened Oct 5, 2026Collaborator·

  ·

• Feat/perf model validation skill

  #1079·devalshahamd opened Oct 3, 2026Contributor·

  ·

• PerfModel: plug your own op models into perf reports with a (category, params, arch) API

  #1071·ajassani opened Oct 1, 2026Collaborator·

  ·
  1

• Support DSPARK draft/verify traces and map some different MLA kernels to MLA rooflines

  #1068·rkarhila-amd opened Sep 30, 2026·2 tasks·

  ·

• fix(kernel-source): require a name match for a single Triton def

  #1062·Ahmedhasssan-aig opened Sep 25, 2026Contributor·1 task·

  ·

• Point skills-checks at skillscope, which now owns the reference checks

  #1059·johnl-amd opened Sep 25, 2026·

  ·
  3

• fix: include Parent op category and Parent cpu_op in kernel_summary merge keys

  #1054·kyle-hoffmeyer opened Sep 23, 2026Collaborator·

  ·
  1

• fix: include parent_module in ops_summary merge keys to prevent cross-product

  #1053·kyle-hoffmeyer opened Sep 22, 2026Collaborator·

  ·
  1

  4

• fix: from_file with capture_trace_filepath no longer crashes

  #1047·kyle-hoffmeyer opened Sep 18, 2026Collaborator·

  ·
  1

  1

• Fix --sheets all when kernel_summary is missing

  #1046·kyle-hoffmeyer opened Sep 18, 2026Collaborator·

  ·
  1

  1

• added workflow dispatch

  #1037·kyle-hoffmeyer opened Sep 15, 2026Collaborator·

  ·
  1

• Inference report: add vLLM AsyncLLM request-lifecycle sheets

  #1034·ajassani opened Sep 15, 2026Collaborator·1 of 3 tasks·

  ·
  1

• TraceIndex: query roofline bound and utilization as typed columns

  #1025·ajassani opened Sep 11, 2026Collaborator·1 of 3 tasks·

  ·
  1

• Inference split: keep host ops between step annotations

  #1022·ajassani opened Sep 10, 2026Collaborator·1 of 3 tasks·

  ·
  1

  5

• Improve operation categorization for Megatron traces

  #994·poznano-amd opened Sep 1, 2026Contributor·

  ·
  1

• Add TraceLens model architecture visualizer

  #991·gabeweisz opened Sep 1, 2026Collaborator·4 tasks done·

  ·
  2