NaN fix v2 (root cause: exactly-zero embedding rows trigger gemma-backward NaN on torch 2.13, localised to layer-12 mlp grad_out via hooks; A-probe with zero token reproduced it): no zero rows anywhere - B uses 1e-4 no-info epsilon, resampler opener 1e-4
agent
884986c
NaN-grad fix: rsqrt-normalise resampler k/v (torch-2.13 sdpa backward NaNs with 4x-magnitude latent inputs); assert now NaN-aware with grad table
agent
f473997
grad-debug diagnostics on the zero-grad assert (D-condition crash is not locally reproducible)
agent
494b63d
run_pilot: condition A = learned soft prompt (zero-input resampler had no grad path - caught by the assert on real gemma); p_halt None guards for halting-free configs (fast_wide)
agent
dd20003
run_pilot: held-out split (20% never trained on), heldout_ce measurement, generation now on an UNSEEN position
agent
2e7e49c
allowlist gemma pilot scripts on the Space; --hf-checkpoint hub pull so the container doesn't need local ckpt files
agent
f7cd607
run_pilot: list-eos support (gemma-it has eos [1,106]); teacher-forcing slices target up to first eos (learn to stop)
agent
6341f4d
run_pilot: mock-gemma dry run green end-to-end; fixes: StepDescriptor t=0 branch, frozen-grad assert guards trainable-only, B zero-prefix is floor-only (no backward), generation shape
agent
33dbd59
gemma_decoder pilot: --checkpoint flag (trained-model source), textiser fixes (router-less guard, share clamp), requirements + transformers/huggingface_hub for Space