FLA prepare_inputs_for_generation sliced input_ids to the last token when transformers passed its own non-empty DynamicCache, so generate() ignored the prompt. Now only an FLA Cache is trusted and input_ids are sliced by tokens actually seen. Verified on GPU: generate logits match teacher-forced forward (worst cos 0.9996, was 0.68). Weights unchanged.
Support fp32/fp16 loading (BF16 autocast) and padded batches (unpad via cu_seqlens)
FLA chunk_delta_rule rejects fp32, so non-BF16 weights now run the blocks under BF16 autocast (training numerics). attention_mask with padding is now handled by unpacking rows into cu_seqlens varlen sequences instead of raising. Weights unchanged.
