fix: AMD gfx1201 (RDNA4/ROCm) — INT8 Triton f32 MFMA, LTX Video device fix, validate_settings KeyError

[athornbro]

Summary

Five fixes to make Wan2GP and LTX Video run correctly on AMD RDNA4 GPUs (RX 9070 / 9070 XT, gfx1201) under ROCm/HIP on Windows. All changes are no-ops on CUDA.

shared/kernels/quanto_int8_triton.py

HIP libdevice — Use triton.language.extra.hip.libdevice on ROCm instead of the CUDA libdevice. The CUDA libdevice triggers __nv_rintf has been dropped errors under Triton 3.6+ on AMD, silently killing the fused INT8 kernel.

f32 MFMA for int8 dot on gfx1201 — All three Triton INT8 GEMM kernels now cast int8 operands to float32 before tl.dot on HIP. gfx1201 bf16 MFMA only provides K=8 and K=16 intrinsics; Wan 2.2 uses K=256 (works) but LTX Video uses K=64 — no matching bf16 intrinsic → Triton compile crash. Float32 MFMA (K=2/4) tiles freely at any block_k. Accumulators updated from int32 to float32 accordingly.

models/ltx_video/models/transformers/attention.py

skip_layer_mask device mismatch — skip_layer_mask is constructed on CPU but hidden_states_a is on GPU, causing RuntimeError: Expected all tensors to be on the same device at the first LTX Video denoising step. Fixed with .to(hidden_states_a.device) on both the AttentionSkip and AttentionValues branches.

wgp.py

validate_settings KeyError fix — Media input keys (image_start, image_end, image_refs, video_guide, etc.) are absent from the payload when inputs are cleared or a different model tab is active. Changed from inputs["key"] to inputs.get("key") throughout, plus a friendly error for I2V without an input image.

asyncio socket warning — Suppress spurious socket.send() raised exception log noise on Windows.

plugins/wan2gp-video-mask-creator/plugin.py

ROCm CPU LAPACK fix — ROCm Windows PyTorch ships without CPU LAPACK, so nn.init.orthogonal_ fails during matanyone model construction. Monkey-patches orthogonal_ to retry the QR decomposition on the GPU when the CPU path raises a LAPACK error. Zero overhead on systems with working CPU LAPACK.

Test platform

• AMD Radeon RX 9070 XT (gfx1201 / RDNA4 / Wave32), ROCm 7.2, Windows 11
• Triton 3.7.0 (triton-windows)
• Tested: Wan 2.2 14B I2V int8 (4-step lightning) and LTX Video 0.9.8 13B int8

[deepbeepmeep]

thx, but I cant merge this PR as is :

• not all changes are no ops for nvidia: for instance there are int32 conversions transformed into float32 conversions. please provide either monkey patch centralized in one module dedicated to AMD or explicit AMD kernels
• wgp change is unrelated to AMD and should be in a separate PR