All pull requests

Pull requests

• check_env resolves the MLflow Databricks auth route the way the SDK does

  bugmedium
  #449·amazloumi opened Nov 2, 2026Member·9 tasks done·

  ·
  1

  1

• Add a FineVision reader and a parquet-to-Arrow converter to the VLM example

  #431·timothyngo opened Nov 1, 2026Collaborator·4 tasks done·

  ·
  5

• Add PixMo readers to the VLM example, with PixMo-Count points in percent

  #430·timothyngo opened Nov 1, 2026Collaborator·3 tasks done·

  ·
  4

• Add image-text documents, chat rendering and testbed readers to the VLM example

  #429·timothyngo opened Nov 1, 2026Collaborator·3 tasks done·

  ·
  5

• Every pipeline stage saves a checkpoint it can resume from

  bugcore
  #422·amazloumi opened Oct 30, 2026Member·9 tasks done·

  ·
  4

  1

• Reject or mask VLM prompts that leave no room for a caption, and keep an all-masked loss on the graph

  bugcore
  #421·amazloumi opened Oct 30, 2026Member·14 tasks done·

  ·
  2

  1

• Refuse resumes that lost train_state.pt and warn when a resume overrides load_path

  #414·Naeemkh opened Oct 29, 2026Member·4 of 6 tasks·

  ·
  2

  7

• Run the LR schedule, freeze stages and shutdown check on NaN-skipped steps

  core
  #413·Naeemkh opened Oct 29, 2026Member·4 of 6 tasks·

  ·
  3

  4

• Compute the HSDP data-parallel rank from mesh coordinates

  core
  #412·Naeemkh opened Oct 29, 2026Member·4 of 6 tasks·

  ·
  1

  4

• Build moe_shared_experts as N experts and fix the MoE MFU expert width

  core
  #411·Naeemkh opened Oct 29, 2026Member·4 of 6 tasks·

  ·
  1

  4

• Honor checkpoint.export_dtype in dcp-to-hf and record the written dtype

  #409·Naeemkh opened Oct 28, 2026Member·4 of 6 tasks·

  ·
  1

  1

• Reject tensor and expert parallelism for the VLM arches they cannot run

  core
  #408·Naeemkh opened Oct 28, 2026Member·4 of 6 tasks·

  ·
  4

  4

• Read .bin token shards with a declared dtype and each file by its own suffix

  core
  #407·Naeemkh opened Oct 28, 2026Member·4 of 6 tasks·

  ·
  2

  4

• Keep the checkpoint a save is writing out of retention

  core
  #404·Naeemkh opened Oct 28, 2026Member·4 of 6 tasks·

  ·
  1

  5

• Move the MLflow code into the kempnerforge.metrics.mlflow subpackage

  coreenhancement
  #403·amazloumi opened Oct 27, 2026Member·6 of 8 tasks·

  ·
  1

• Make MLflow a core dependency

  coredependenciesenhancement
  #402·amazloumi opened Oct 27, 2026Member·7 of 9 tasks·

  ·
  1

  3

• Add a typed contract and pinned manifest for the image-understanding recipe example

  #286·amazloumi opened Oct 16, 2026Member·5 tasks done·

  ·
  1

  1

• Restore each rank's own dataloader and RNG state on resume

  core
  #285·amazloumi opened Oct 16, 2026Member·8 tasks done·

  ·
  1

  2

• Add Transformer constructor knobs to omit the token embedding and output head

  core
  #284·amazloumi opened Oct 15, 2026Member·6 tasks done·

  ·
  1

  1

• Add caller-defined bidirectional visibility groups to packed causal attention

  core
  #283·amazloumi opened Oct 15, 2026Member·7 tasks done·

  ·
  1

  3

• Accept explicit per-token position ids in Transformer.forward

  core
  #282·amazloumi opened Oct 15, 2026Member·6 tasks done·

  ·
  1

  2

• Validate and benchmark the FlexAttention packed path

  low
  #207·timothyngo opened Sep 20, 2026Collaborator·8 tasks done·

  ·
  3