-
Notifications
You must be signed in to change notification settings - Fork 50
Pull requests: goodfire-ai/param-decomp
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
Bump actions/cache from 5 to 6
dependencies
Pull requests that update a dependency file
github_actions
Pull requests that update GitHub Actions code
#904
opened Jun 29, 2026 by
dependabot
Bot
Loading…
perf(hsdp): donate train-step buffers — 2→4 seq/GPU + ~3.6× throughput (+ V/U hoist)
#903
opened Jun 28, 2026 by
ocg-goodfire
Collaborator
Loading…
fix(ci): make stacked-parity forward pins portable (CI flake)
#901
opened Jun 25, 2026 by
danbraunai-goodfire
Collaborator
Loading…
fix(jax): correctness + multi-host robustness fixes from feature/jax review
#900
opened Jun 25, 2026 by
danbraunai-goodfire
Collaborator
Loading…
feat(losses): split batch-invariant frequency-minimality out of imp-min
#899
opened Jun 25, 2026 by
ocg-goodfire
Collaborator
Loading…
fix(load_run): replicate HF prefix without cross-host allgather (dp>=64 OOM)
#898
opened Jun 25, 2026 by
danbraunai-goodfire
Collaborator
Loading…
Compact per-step train log with elapsed<eta
#897
opened Jun 25, 2026 by
danbraunai-goodfire
Collaborator
Loading…
Add experiments/mnist VPD family (memorization-vs-generalization study)
#890
opened Jun 24, 2026 by
lee-goodfire
Collaborator
•
Draft
fix(slow_eval): chunkwise CI fn slow-eval on multi-host GPU (bf16 + gather)
#888
opened Jun 23, 2026 by
danbraunai-goodfire
Collaborator
Loading…
fix(eval): make the slow-eval tier work on multi-host GPU
#885
opened Jun 22, 2026 by
danbraunai-goodfire
Collaborator
Loading…
Bump actions/checkout from 6 to 7
dependencies
Pull requests that update a dependency file
github_actions
Pull requests that update GitHub Actions code
#881
opened Jun 22, 2026 by
dependabot
Bot
Loading…
L18-23 6-layer MLP VPD ablation sweep: configs + generator
#878
opened Jun 19, 2026 by
ocg-goodfire
Collaborator
Loading…
docs: imp-min parity + 1→9-layer / batch scaling analysis
#869
opened Jun 17, 2026 by
ocg-goodfire
Collaborator
Loading…
Smooth-L0 (Geman-McClure) importance-minimality loss
#852
opened Jun 16, 2026 by
danbraunai-goodfire
Collaborator
Loading…
deps: upgrade to transformers v5 (load HF targets in native dtype)
#844
opened Jun 16, 2026 by
danbraunai-goodfire
Collaborator
Loading…
papers: add the VPD post (Interpreting Language Model Parameters)
#562
opened Jun 12, 2026 by
ocg-goodfire
Collaborator
Loading…
fix(lm): retry HTTP 408 from the HF CDN in the hub retry backend
#561
opened Jun 12, 2026 by
danbraunai-goodfire
Collaborator
Loading…
Migrate VPD to JAX: train + analyze in one framework; retire torch to oracle
#560
opened Jun 11, 2026 by
ocg-goodfire
Collaborator
Loading…
Load HF targets in their native dtype, not fp32
#559
opened Jun 11, 2026 by
Antovigo
Collaborator
Loading…
perf(multipool): fuse PPGD source-grad into the main backward
#556
opened Jun 9, 2026 by
danbraunai-goodfire
Collaborator
Loading…
clean up orphaned .snapshot_scratch partials
#555
opened Jun 9, 2026 by
danbraunai-goodfire
Collaborator
Loading…
refactor(3-pool): SUM-grad convention (supersedes #545; fixes 2nd stoch→V/U instance)
#546
opened Jun 2, 2026 by
ocg-goodfire
Collaborator
•
Draft
fix(3-pool): scale PPGD's CI grad by n_ci to survive the CI-pool AVG
#545
opened Jun 2, 2026 by
danbraunai-goodfire
Collaborator
Loading…
metrics: split frequency-minimality out of importance-minimality (batch-invariant)
#543
opened May 30, 2026 by
ocg-goodfire
Collaborator
•
Draft
7.5-pool distributed strategy
#530
opened May 27, 2026 by
danbraunai-goodfire
Collaborator
Loading…
7 tasks
Previous Next
ProTip!
Updated in the last three days: updated:>2026-06-25.