| Age | Commit message (Collapse) | Author |
|---|
|
Layer 0 has no incoming edges structurally (no prior layers), but
receives the embedding as input. The cascading gate was killing its
outgoing edges (hard: g=0, soft: g=0.5), causing nll_hard to be
~2x worse than baseline. Fix: set g=1 for layer 0 nodes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
|
- NLL loss was shifting labels twice (olmo_labels already shifted,
then code did logits[:,:-1] vs labels[:,1:]). Fixed in 9 locations:
trainer, pipeline, olmo_graph, sanity_check, eval.
- Head U/V weights init with std=0.01 (was Kaiming ~5.7 std) so
UV^T≈0 at init, ensuring Z≈logit_bias=15 and A≈0.953.
- Updated SVD rank test to subtract logit_bias before checking.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
|
- Add learnable logit_bias=15.0 to PredictorMLP, so σ(15/τ_init) ≈ 0.95
at init, reproducing dense connectivity instead of random A≈0.25
- Fix dtype mismatch: cast A to model dtype (bfloat16) in DAGFormerOLMo.forward
- Fix YAML lr parsing: add type coercion in TrainConfig.from_yaml
- Fix device mismatch: call self.to(device) in StructurePredictor.__init__
- Add python -u for unbuffered SLURM output, TOKENIZERS_PARALLELISM=false
- Delete stale eval_cache.pt (built with buggy MLP input code)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
|
- olmo_graph.py: Modified OLMo2-1B forward with per-head routing via 256x256 adjacency matrix A
- Proportional attribution for post-norm decomposition
- All 6 GPU sanity checks pass (baseline diff = 0.000001)
- predictor.py: Qwen3-Embedding encoder + MLP decoder + Gumbel-Sigmoid + cascading gate
- pipeline.py: End-to-end glue (predictor -> A -> OLMo -> NLL)
- trainer.py: Full training loop with DDP, gradient accumulation, eval, checkpointing
- dolma.py: Streaming Dolma v1.7 with sequence packing
- 43/43 unit tests pass
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
