All pull requests

Pull requests

• fix: ASTRA under FSDP

  #588·luciaquirke opened Oct 30, 2026Collaborator·

  ·

• feat: Multi-GPU output-token influence

  #582·brendanlong opened Oct 28, 2026Collaborator·

  ·

• Add Gemma Scope 2 SAE feature similarity baseline

  #579·luciaquirke opened Oct 26, 2026Collaborator·

  ·

• Add AttributionUnit for gradient and score rows

  #571·luciaquirke opened Oct 24, 2026Collaborator·

  ·

• fix: Weight TKFAC covariances by per-position traces and apply delta

  #566·flyingdoubleG opened Oct 24, 2026Contributor·

  ·
  8

• Default per-token scoring to output token influence

  #557·luciaquirke opened Oct 19, 2026Collaborator·

  ·
  1

• feat: allow retrain bank to be built without precomputed scores (#446)

  #553·vnayakde opened Oct 15, 2026·

  ·
  1

  2

• feat: AdamW-influence curvature and all-parameter unrolling

  #549·luciaquirke opened Oct 13, 2026Collaborator·

  ·
  1

• feat: Unrolled Gauss-Newton attribution

  #546·luciaquirke opened Oct 8, 2026Collaborator·

  ·
  1

• feat: invert the Kronecker factors in the projected space

  #543·luciaquirke opened Oct 7, 2026Collaborator·

  ·

• [WIP] Tensor-parallel metasmoothness and extra evals for filter retrains

  #538·luciaquirke opened Oct 5, 2026Collaborator·

  ·

• [WIP] Tensor-parallel MAGIC: sharded activations and narrow frozen weights

  #537·luciaquirke opened Oct 2, 2026Collaborator·

  ·

• [WIP] Let metasmoothness run under fsdp

  #535·luciaquirke opened Oct 1, 2026Collaborator·

  ·
  1

• Add input token influence scoring

  #526·brendanlong opened Sep 30, 2026Collaborator·

  ·
  2

• Span-level attribution

  #524·luciaquirke opened Sep 30, 2026Collaborator·

  ·

• Add per-chunk QLD rows to the leaderboard

  #522·luciaquirke opened Sep 30, 2026Collaborator·

  ·

• Somewhat optimize MoE gradient collection

  #493·luciaquirke opened Sep 25, 2026Collaborator·

  ·

• Checkpoint the KFAC fit for resumable Hessian fitting

  #470·prajak002 opened Sep 16, 2026Contributor·

  ·
  6

• Add sharded data-parallel runs sharing one run_path

  #295·luciaquirke opened Jun 5, 2026Collaborator·4 tasks done·

  ·

• Make identity/FOOF write covariance dirs, fix target_modules

  #274·smarter opened May 5, 2026Collaborator·

  ·

• [WIP] Test modular norm for MAGIC attribution

  #236·luciaquirke opened Apr 12, 2026Collaborator·

  ·