Sacrebleu errors used to fall back to set-overlap unigrams, so any
English hyp scored ~70–80 against English refs and SFT reported
success 1.0. Use count-based char n-grams, keep BLEU failures from
clobbering chrF, and print a few hyps during eval.
Step 0 generate was writing a 5GB success=0 snapshot and leaving the
32GB card fragmented, so the next Adam step OOM'd after batch-32 eval.
Only promote _best after step 0 and empty_cache when eval returns.
train_sft used to torch.save only after the full epoch budget, so Ctrl+C
dropped all translation weights. Write _last every --ckpt-every steps
and on KeyboardInterrupt; write _best when frozen eval (success, chrF)
improves; --resume continues from _last.
OpenBayes sets SWANLAB_PROJECT as a string; swanlab>=0.9 parses that as
ProjectSettings and raises QuoteAwareEnvSettingsSource. Drop it, keep
SWANLAB_PROJ_NAME, and share run-id extraction with train_k3.
swanlab.init always opened a new experiment on --resume. Save the run
id in the checkpoint and pass resume=True, id=... on the next start.
--swanlab-id overrides; --swanlab-new forces a fresh experiment.
Every micro-step was hitting SwanLab, and every 100 steps wrote a 5GB
ckpt plus greedy decode. 0.5b now logs every 20, eval/held-out every 500,
saves _last every 1000, samples every 2000.
Python float scales and fp32 expert weights promoted SiTU outputs to
fp32, so index_add mixed BFloat16 dest with Float source and crashed
the 0.5b run. Cast packed weights and gate scalars to z.dtype.
Whole-mixer checkpoint plus T×T MLA scores OOM'd a 31GB GPU on backward.
Checkpoint each AttnRes block, run absorbed MLA through SDPA, and compute
CE in vocab chunks so [B,T,V] logits are never materialized.
--max-tokens is now the training budget; default --steps 2000 no longer
caps a 1B-token run at 250 optimizer steps.
Route with σ(W_r x), Top-k(s+b), then L1-normalize over the selected set.
Add Switch/GShard aux and router z-loss into train_k3 and train_sft.
Wiki parquet URLs honor HF_ENDPOINT for mirrored downloads.
Replace dense all-expert forward (16 experts × all tokens) with
permute-dispatch: sort token-expert pairs by expert id, pad to
[R, C, ℓ] (C = max tokens per expert), run 3 bmm calls for the
batched SiTU-GLU activation, then scatter-add weighted results back.
Routed expert FLOPs drop from R·N to R·C (C ≈ N·k/R under uniform
routing). SiTU parameter structure unchanged; checkpoint compatible.
Tests: sparse-vs-dense fwd/bwd equivalence, unselected expert zero
grad, last_capacity tracking.
dela