All pull requests

Pull requests

• [catalog] Deprecate the CalibForge cohorts

  agent-generatedagentic-lint
  #11197·penfever opened Nov 11, 2026Member·

  ·
  3

• [sft] Unquote OpenCode prompts and drop empty system turns

  #11193·Min-Li opened Nov 11, 2026Contributor·

  ·
  3

• [iris] Grant submitter access to the shared H100 clusters

  #11192·GarlGuo opened Nov 11, 2026·

  ·
  1

• [grug] Hand off hero SFT to another rack count

  agent-generatedagentic-lint
  #11189·ravwojdyla opened Nov 11, 2026Member·

  ·
  12

• [long-loop] Start RL from a pinned HF export

  agent-generatedagentic-lint
  #11185·karansag opened Nov 10, 2026Contributor·

  ·

• [WIP][finelog] Prune segments by catalog min/max of flagged columns

  agent-generatedagentic-lint
  #11184·HRL9000 opened Nov 10, 2026Collaborator·

  ·
  6

• [iris] Admit Snowball parity CI to cw-us-east-02a

  agent-generatedagentic-lintweaver
  #11182·loom-oa-dev[bot] opened Nov 10, 2026Bot·

  ·
  3

• [rl] Let the Snowball quick loop use InfiniBand for NCCL

  agent-generatedagentic-lint
  #11178·mcwitt opened Nov 10, 2026Member·

  ·
  1

• [finelog] Forward rows that already carry a cluster name

  agent-generatedagentic-lint
  #11176·HRL9000 opened Nov 10, 2026Collaborator·

  ·
  3

• [rl] Mirror Hub models on CPU while retaining node fallback

  agent-generatedagentic-lint
  #11161·karansag opened Nov 10, 2026Contributor·1/2·

  ·
  19

• [inference] Prepare Hugging Face weights for direct marin-serve iris

  agent-generatedagentic-lint
  #11160·karansag opened Nov 10, 2026Contributor·2/2·

  ·
  12

• [grug] Pipeline June step logging and check the loss in the step

  #11159·mminutoli opened Nov 10, 2026Contributor·

  ·
  5

• [grug] Checkpoint June stacked blocks without prevent_cse

  #11158·mminutoli opened Nov 10, 2026Contributor·

  ·
  6

• [grug] Gather June ZeRO-1 expert weights in the first microbatch

  #11157·mminutoli opened Nov 10, 2026Contributor·

  ·
  5

• [grug] Shard June expert optimizer state over the replicas (ZeRO-1)

  #11156·mminutoli opened Nov 10, 2026Contributor·

  ·
  7

• [grug] Overlap June expert-gradient reductions with the last microbatch

  #11155·mminutoli opened Nov 10, 2026Contributor·

  ·
  5

• [grug] Reduce June accumulated gradients once per step

  #11154·mminutoli opened Nov 10, 2026Contributor·

  ·
  5

• [grug] Accumulate June gradients over microbatches

  #11153·mminutoli opened Nov 10, 2026Contributor·

  ·
  1

• [grug] Carry a compute-dtype parameter copy in the June train state

  #11152·mminutoli opened Nov 10, 2026Contributor·

  ·
  4

• [grug] Reduce June replicated gradients and routing stats after the scan

  #11151·mminutoli opened Nov 10, 2026Contributor·

  ·
  4

• [grug] Add June layer-scan options and a half-RoPE without concat

  #11150·mminutoli opened Nov 10, 2026Contributor·

  ·
  16

• [grug] Add Pallas-Triton routing metadata kernels for ring_filtered

  #11148·mminutoli opened Nov 10, 2026Contributor·

  ·
  15

• [grug] Add fused add + RMS norm + GatedNorm options to the June model

  #11147·mminutoli opened Nov 10, 2026Contributor·

  ·
  29

• [grug] Add ring_filtered send-gather and token-sum options

  #11146·mminutoli opened Nov 10, 2026Contributor·

  ·
  3

• [levanter] Write the batched_xla CE weight gradient in place

  #11145·mminutoli opened Nov 10, 2026Contributor·

  ·
  23