fix(dflash): honor config.is_causal in DFlash training masks [1/3]

[maocheng23]

Split of #905, part 1 of 3 (the actual bug fix). Stacked on upstream main 53398a8.

Motivation

SGLang's DFlash model (v0.5.18+, _get_dflash_attention_type) decides causality inside the draft block from config.is_causal:

• unset: full layers bidirectional, sliding layers causal
• false: all layers bidirectional
• true: all layers causal

SpecForge ignores the field. It always trains sliding layers causal and full layers bidirectional, so any draft with is_causal: false and sliding layers trains causal but is served bidirectional.

Changes

• create_dflash_sdpa_mask / create_dflash_block_mask take is_causal and resolve it with SGLang's rule via resolve_dflash_is_causal(is_causal, layer_type): an explicit value applies to every layer, unset is the per-layer-type default.
• The online wrapper passes the draft config's is_causal.
• Sliding layers also get the window's left bound inside the draft block (kv_off >= q_off - (W-1)). There is no right bound, because SGLang backends disagree on one for non-causal windows (FA3 symmetric (W-1, W-1), FlashInfer/Triton left-only). This only differs from before when sliding_window < block_size, which no shipped config has (W = 2048/4096, block 8/16).
• benchmark_dflash2_conv.py passes the config's is_causal too.
• Function names and signatures are otherwise unchanged, so no callers need edits.

⚠️ This changes training for both shipped DFlash2 recipes

configs/qwen3.6-27b-dflash2.json and configs/qwen3.8-27b-dflash2.json set is_causal: false and use only sliding layers. After this PR they train bidirectional inside the block (which is what SGLang v0.5.18+ already serves). This needs a decision before merging:

• (a) keep as-is, so training follows the config and matches serving;
• (b) remove is_causal from both configs, which keeps today's causal training and makes serving causal too.

Causal vs bidirectional inside the block has never been ablated cleanly. Existing checkpoints can be served to match how they were trained, without retraining, by deleting is_causal from the served config.json.

Tests

tests/test_utils/test_dflash_mask.py: the reference mask is extended with is_causal. New tests cover dense vs reference, flex mask_mod vs dense, and flex block occupancy for window ∈ {None, 1, 3, 9} × is_causal ∈ {None, False, True}, plus an in-block spot check and a check that the training wrapper reads draft_model.config.is_causal.

Run on CPU (torch 2.14, transformers 5.12.1): every tests/test_utils/test_dflash* and tests/test_modeling/test_dflash* file passes when run alone.

Co-authored-by @cih9088 (original #905); every commit carries a Co-authored-by: Inhyuk Cho trailer.

🤖 Generated with Claude Code