[hiroki-abe-58]

On Windows, the RTX 5090 EdenLive preset (configs/edenlive/realtime/wan22_5b_ar_int8_convrot_5090.json) stops before the first chunk. There are two separate failures.

Problem

1. SM120 SageAttention. sage_attn2_sm120_cached calls sageattention._fused and _qattn_sm89 functions as module attributes. The Windows SageAttention wheels that the Windows guide links to (woct0rdho, stable-ABI builds) register these kernels as torch ops instead (torch.ops.sageattention_fused, torch.ops.sageattention_qattn_sm89), so the modules have no such attributes:
   AttributeError: module 'sageattention._fused' has no attribute 'quant_per_block_int8_fuse_sub_mean_cuda'
2. Fused GELU → ConvRot INT8 quantizer. It is built with cpp_extension.load(is_python_module=False) and opened with ctypes. On Windows a DLL only exports functions that are marked for export, so the entry point is missing:
   AttributeError: function 'lightx2v_prepare_convrot_gelu' not found
prepare_convrot_gelu_fusion only catches OSError and RuntimeError, so the model load fails, instead of falling back to the unfused path as its docstring describes.

Change

• sage_sm120_cached.py: one helper returns the module functions when the module has them, and otherwise the torch-ops namespaces (stable-ABI builds). Builds with module functions behave as before.
• convrot_gelu_quant.cu: export the two extern "C" entry points under _WIN32. The macro is empty on other platforms.
• fused_gelu_linear.py: a missing entry point (AttributeError) now counts as "fusion unavailable", like a failed build.

Validation

Setup: Windows 11, RTX 5090 (driver 610.88), torch 2.14.1+cu130, SageAttention 2.2.0+cu130torch2.10.0andhigher.post6 (abi3), comfy-kitchen 0.2.39, CUDA 13.1 with MSVC 2022 (Build Tools 17.14).

Kernel checks (EdenLive 5B attention shapes: 24 heads, head dim 128, 2640 query and 15840 key/value tokens, BF16):

check	main (d2d7424)	this PR
SageAttn2Sm120CachedWeight.apply vs SDPA	AttributeError	cosine 0.99924
apply_prepared(prepare_sage_sm120_kv(...)) vs SDPA	AttributeError	cosine 0.99924
prepare_convrot_gelu_fusion()	AttributeError	True; dumpbin /exports lists both functions
fused_gelu_convrot_int8_quant(x) vs comfy-kitchen's quantize_int8_rowwise_convrot64(gelu(x), 256), x: [2640, 14336] BF16	-	INT8 values and scales bit-identical

End to end. I ran python -m lightx2v.infer --model_cls wan22_edenlive_ar --task i2v with the 5090 preset's settings: INT8 ConvRot DiT and proxy encoder, sage_attn2_sm120_cached, use_compile, warmup, use_cuda_graph, and LightTAE streaming decode. Three kernels have no Windows build, so I substituted:

• torch_real_rope for FlashInfer RoPE;
• torch RMSNorm for sgl-kernel;
• BF16 T5 on the CPU for FP8-SGL.

Input: the first 120 frames of the cryo-crossing sample at 1280×704.

• main: stops at model load (ConvRot) or at the first attention call (Sage), depending on the config.
• This PR: 117 frames at 22.3 FPS generation. That is the sum of the request's chunk times; the one-time compile warmup of about 108 s is excluded. Peak reserved memory was 24.0 GiB, and the frames follow the proxy video.

For comparison, on the same machine without sage_attn2_sm120_cached, compile and CUDA graphs:

• BF16 with torch kernels: 7.9 FPS;
• with SageAttention 2 and Triton norms: 11.0 FPS;
• with INT8 ConvRot added: 16.8 FPS.

Not changed here: building the extension needs the MSVC environment, and vcvars64.bat sets Platform=x64. Windows environment variable names are case-insensitive, so lightx2v_platform reads it as PLATFORM and stops with Unsupported PLATFORM: x64 unless PLATFORM=cuda is set. I can add a line about this to the Windows guide if that would help.

🤖 Generated with Claude Code