feat(dflash): wire caller-provided SWA mask through draft graph

[javierpazo]

feat(dflash): wire caller-provided SWA mask through draft graph

When the Qwen3.6 draft graph is built for a context that exceeds
the SWA window, the caller-provided sliding-window attention mask
must reach ggml_flash_attn_ext on the SWA layers.

Previously the mask was constructed and then nullified
post-construction in qwen3_dflash_graph, so SWA layers ran without
the intended visibility constraint at long contexts.

This change makes the wiring explicit and pinable as a contract:

• dflash_graph.h — DraftGraphInputs gains an optional
  attn_mask field. Documented as caller-owned, type F16, with
  shape [kv_len, q_len] (or padded [kv_pad, q_pad]), values
  0 for visible positions and -inf for masked positions.
  Two helpers added so callers do not reimplement the same logic:

  bool draft_graph_needs_swa_mask(const DraftWeights & w,
  int ctx_len);
  void build_draft_swa_mask(std::vector<uint16_t> & out,
  int ctx_len, int q_len,
  int swa_window);

  lm_head is normalised to a default-null member at the same
  time (small consistency fix; layout unchanged).

• qwen3_dflash_graph.cpp — when total_k > swa_window on a SWA
  layer, the graph wires the caller-provided mask through to
  ggml_flash_attn_ext and stops nullifying it. Layers that are
  not SWA still ignore the mask, as before.

• smoke_draft_graph.cpp and test_vs_oracle.cpp — small
  alignment so the existing tests can build / fill the draft SWA
  mask when ctx_len + q_len > swa_window. No new test scaffolding
  is added in this commit; the focused regression test lives in a
  separate PR (test(dflash): contract test for draft SWA mask wiring) so each PR keeps one concern.

Validation:

• Built and ran smoke_draft_graph and test_vs_oracle on
  RTX 6000 Ada (sm_89), Heretic Q4_K_M target, FP16 safetensors
  drafter, FA_WINDOW=0. Both tests pass before and after; the
  behaviour at ctx_len <= swa_window is unchanged (mask not
  needed and not consumed).
• At long context the SWA layers now respect the caller mask.

Verification vs existing community PRs:

COMP-COMPL with PR #94 (open, "support Qwen3.6-27B-DFlash draft
(SWA layers)", Quitetall) and PR #129 (open Draft, "sliding
window attention for Qwen3.6 draft model", howard0su).

• PR feat(dflash): support Qwen3.6-27B-DFlash draft (SWA layers) — 106 t/s on RTX 4090 #94 wires SWA via masks (same family as this PR).
• PR draft: sliding window attention for Qwen3.6 draft model #129 wires SWA via per-layer K/V truncation instead.

The interface added here (caller-mask field + helpers) is small
enough that it can survive either approach landing first. If
PR #94 lands first, this commit should rebase cleanly because it
formalises the mask path #94 already needs internally; if
PR #129 lands first, the mask path here remains useful for
callers that prefer mask semantics. Maintainers, happy to
coordinate ordering.

Author: Javier Pazo xabicasa@gmail.com

[cubic-dev-ai]

No issues found across 4 files