Skip to content

issues Search Results · language:Dune language:JavaScript language:Java language:Java language:JavaScript language:Python

Filter by

37.5M results  (1 s)

37.5M results

goodfire-ai/param-decomp
D4/R3 counter-based RNG layout invariance confirmed (JAX stronger than torch) (done)

Bookkeeping. SPEC_AUDIT D4/R3 CONFIRMED: JAX s fold_in(run_key, step) is identical across processes (param_decomp_jax/jax_single_pool/run.py:294), a STRONGER invariance than torch s deliberate per-rank ...
kind:done
parity
subsystem:spec-tests-crosscutting
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #769

goodfire-ai/param-decomp
S14′ fused persistent ascent reuses un-coeffed main backward grad — confirmed (done)

Bookkeeping. SPEC_AUDIT S14′ CONFIRMED: torch takes a second autograd.grad on the un-coeffed per-term loss (param_decomp/metrics/persistent_pgd_recon.py:217) before the main backward; JAX reuses the fused ...
kind:done
parity
subsystem:spec-tests-crosscutting
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #768

goodfire-ai/param-decomp
S8/D2 imp-min global per-component sum inside log2 confirmed semantically identical (done)

Bookkeeping. SPEC_AUDIT S8/D2 CONFIRMED: torch SUM-reduces per-component sums across ranks with an autograd-aware all_reduce inside the log2 (param_decomp/metrics/importance_minimality.py:161-167); JAX ...
kind:done
parity
subsystem:spec-tests-crosscutting
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #767

goodfire-ai/param-decomp
S17 faithfulness Σ‖Δ‖²/Σnumel confirmed at rtol 2e-4 (done)

Bookkeeping. SPEC_AUDIT S17 CONFIRMED: global mean sq-delta over all params, fp32 deltas outside autocast, equivalence fixture faith=0.10001952 matches at rtol 2e-4. torch param_decomp/metrics/faithfulness.py:14-25; ...
kind:done
parity
subsystem:spec-tests-crosscutting
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #766

goodfire-ai/param-decomp
S6: lower_leaky_hard custom VJP confirmed bit-identical to torch (done)

Bookkeeping. SPEC_AUDIT spot-checked the torch custom backward (param_decomp/ci_sigmoids.py:24-53) line-for-line against JAX _lhs_b (param_decomp_jax/jax_single_pool/ci_fn.py:34-48): nested where including ...
kind:done
parity
subsystem:spec-tests-crosscutting
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #765

microsoft/JARVIS
Tonnam

  • phawatrsaekwy666-cmyk
  • 5
  • Opened 
    1 minute ago
  • #4385

goodfire-ai/param-decomp
Adversary sources + optimizer state deliberately omitted from export

CONFIRMED (I9, S13/S14/S15): the export carries V/U + CI-fn params + frozen target ONLY. Persistent adversary sources and optimizer/Adam state are training-only and intentionally excluded — no torch consumer ...
kind:done
parity
subsystem:interop-export
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #759

goodfire-ai/param-decomp
Exported tensors are fp32 (V/U masters lossless, bf16 target upcasts exactly)

CONFIRMED (I8/N1): tensors emitted fp32 (_f32); V/U are already fp32 masters (lossless); the frozen bf16 target upcasts exactly, matching torch s fp32 HF load path. jax export.py:62-63,144-149.
kind:done
parity
subsystem:interop-export
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #758

goodfire-ai/param-decomp
Site-order permutation JAX KIND_ORDER → torch sorted() implemented and verified

CONFIRMED (I7, S10/S11): the one real trap . The export reorders in_proj ROW blocks (by d_in) and out-head COLUMN blocks + bias (by C) from JAX computation order to sorted(site_names); in_proj_b and per-block ...
kind:done
parity
subsystem:interop-export
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #757

goodfire-ai/param-decomp
CI-fn key mapping + orientation (in/out proj, blocks, inv_freq buffer, no transpose)

CONFIRMED (I6): in_proj_{w,b}→_input_projector.{W,b}, out_{w,b}→_output_head.{W,b} (NO transpose); blocks[i].w{q,k,v,o}→_blocks.{i}.attn.{q,k,v,out}_proj.weight; blocks[i].{w1,b1,w2,b2}→_blocks.{i}.mlp.{0,2}.{W,b}; ...
kind:done
parity
subsystem:interop-export
  • ocg-goodfire
  • Opened 
    1 minute ago
  • #756
Issue origami icon

Learn how you can use GitHub Issues to plan and track your work.

Save views for sprints, backlogs, teams, or releases. Rank, sort, and filter issues to suit the occasion. The possibilities are endless.Learn more about GitHub Issues
ProTip! Restrict your search to the title by using the in:title qualifier.
Issue origami icon

Learn how you can use GitHub Issues to plan and track your work.

Save views for sprints, backlogs, teams, or releases. Rank, sort, and filter issues to suit the occasion. The possibilities are endless.Learn more about GitHub Issues
ProTip! Restrict your search to the title by using the in:title qualifier.