transformers

83e7525a - Derive packed-sequence boundaries from position_ids in gated-DeltaNet models

Derive packed-sequence boundaries from position_ids in gated-DeltaNet models Linear-attention layers only separate packed sequences when given cu_seq_lens_q and seq_idx, so a padding-free batch described by position_ids only (e.g. DataCollatorWithFlattening defaults) mixed documents. Derive them once per forward when missing, so attention layers reuse them too.