Yi-6B sets model_max_length=4096. encode() would clip long Wikipedia
pages before we pack seq_len chunks. Raise the cap so only our
chunker limits context.
CLI default attnres=off was overwriting block checkpoints on resume
so later loads hit Unexpected key(s). Only apply flags the user
passed. Track next_chunk so a budget-exit save does not skip the
untrained yield. 0.5b now uses 01-ai/Yi-6B (64k); refuse resume
when the ckpt tokenizer does not match.
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.
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.
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