SageAttention was only compiled for A100 (sm80) and H100 (sm90).
RTX 5090 (Blackwell sm120) has no compatible kernel, causing ComfyUI
to crash during generation with "Connection reset by peer".
Switch to PyTorch's native SDPA which works on all architectures.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Empty backend string still triggered inductor. Removing compile_args
connection from WanVideoModelLoader nodes (131, 132) ensures no
torch.compile is applied to the transformer blocks.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The inductor backend was causing cudaErrorInvalidValue during Triton
kernel autotuning on Blackwell architecture. Setting backend to empty
string disables torch.compile entirely.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Disable torch.compile (inductor -> disabled) to reduce cold start time
- Fix handler to detect video type from file extension, not output key
- Fix HTML to check filename extension for video display
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The frontend-to-API conversion was using outdated widget names that
don't match the current WanVideo node API. Using the exported API
format workflow directly bypasses this issue.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
