QVAC-21981 fix: compose the walk attention mask explicitly (old-ggml CUDA compat)

[DmitryMalishev]

What

Removes the ABot walk's dependency on ggml's fused masked soft-max, whose CUDA kernel in the registry's sanctioned 2026-07-03 ggml line (50cf5630) rejects the walk's 2D-mask head broadcast — the first block aborted with ggml_cuda_compute_forward: SOFT_MAX failed, "invalid argument". That incompatibility is the only reason the monorepo addon carries a ggml overlay repin and why qvac-registry-vcpkg#284 initially proposed a ggml REF change (rejected in review — the ltx-lineage REF is not the diffusion line, and the branch tip carries an embargoed commit).

The change (one call-site, ABot-only): WanSelfAttention::forward routes the masked case through forward_kv with no cached context — the KV-cache formulation that composes scale → add(mask) → soft_max from primitive ops supported by every ggml revision in play. Only the ABot walk ever passes a mask; every other Wan model uses the nullptr default and is untouched. forward_kv was built during the KV-cache work to mirror the fused path op-for-op, so this is a switch between two already-gated formulations, not new math.

Validation (all against ggml 50cf5630, the registry REF)

• CPU golden replay (F16 DiT, 4 blocks, recompute path — the changed one) vs the PyTorch reference: WALK PASS at cosine 0.999951 / 0.998994 / 0.997401 / 0.992991 — identical to the fused path's gated results, i.e. the two formulations are numerically equivalent.
• CUDA (RTX 4050): recompute and KV walks both complete with coherent frames on the exact ggml revision that previously aborted in SOFT_MAX on block 0.
• Engine builds cleanly against 50cf5630 on CPU (MinGW) and CUDA (MSVC/nvcc) — first builds of this engine against the registry ggml line.

Regression safety: provably ABot-only

⚠️ TL;DR for review: the diff is +12 lines / 0 deletions in one function, inside if (mask != nullptr) — and only the ABot walk ever passes a mask. Every other workflow executes a byte-identical instruction path.

Verified call-graph containment:

1. WanSelfAttention::forward(…, mask = nullptr) receives a mask argument in exactly one place: WanAttentionBlock::forward (wan.hpp:1744), whose attn_mask is also defaulted nullptr.
2. WanAttentionBlock::forward receives that argument in exactly one place in the tree: abot_world.hpp:389 (the walk). The standard Wan generation path (wan.hpp:2223 — Wan 2.1 T2V/I2V, Wan 2.2 TI2V/A14B, VACE) omits it → nullptr → unchanged path.
3. SD/SDXL/SD3/FLUX/LTX/Ideogram/ESRGAN never instantiate Wan attention; other self_attn->forward matches in the repo (clip.hpp, llm.hpp, ernie_image.hpp, anima.hpp) are unrelated classes.

Even for ABot this is not a behavior change: the golden replay on the routed formulation reproduced the fused path's gates at identical cosines to six decimals. The monorepo's full 73-test GPU suite (incl. real Wan 2.1/2.2 and LTX generations) re-runs against the merged tip automatically when the registry port (qvac-registry-vcpkg#284) is adopted by tetherto/qvac#3352, before anything reaches consumers.

Effect downstream

• qvac-registry-vcpkg can publish stable-diffusion-cpp 2026-07-03#6 pointing at this branch's merge tip while keeping ggml at 50cf5630 (PR LYCORIS Model leejet/stable-diffusion.cpp#284 will repoint its sd REF here once merged).
• The monorepo addon can then delete both of its overlay ports (stable-diffusion-cpp and the ggml repin) and run entirely on registry ports.

🤖 Generated with Claude Code