Added nvfp4 official support.

[BuffMcBigHuge]

Summary

Adds official support for NVFP4 quantization, enabling ~4x weight memory reduction on Blackwell GPUs (SM >= 10.0). NVFP4 uses NVIDIA's E2M1 format and hardware-accelerated Tensor Core kernels via comfy-kitchen.

What's new

• NVFP4 quantization option – Users with Blackwell GPUs (RTX 50xx, B100, etc.) can select nvfp4 (Blackwell) in the quantization dropdown when their hardware supports it.
• Shared quantization pipeline – FP8 and NVFP4 logic is centralized in quantization_utils.py, replacing duplicated FP8 logic across pipelines.
• Hardware detection – The server exposes supports_nvfp4 in the hardware info API based on CUDA device capability (SM >= 10.0). The UI only shows the NVFP4 option when supported.
• VACE compatibility – VACE components now support both FP8 and NVFP4 quantization when enabled.

Technical details

• NVFP4 uses comfy-kitchen's QuantizedTensor with TensorCoreNVFP4Layout for hardware-accelerated matmul on Blackwell.
• FP8 (fp8_e4m3fn) continues to use torchao for Ada+ GPUs (SM >= 8.9), with ~2x weight memory reduction.
• Dependencies: Adds comfy-kitchen[cublas]>=0.1.0 (Linux/Windows) and torchaudio==2.9.1 for future audio support.
• Fallback: If a user selects NVFP4 and later switches to a non-Blackwell GPU (e.g. from persisted state), the UI resets to fp8_e4m3fn.

Pipelines updated

All pipelines that support quantization now use the shared apply_quantization():

• krea_realtime_video
• longlive
• memflow
• reward_forcing
• streamdiffusionv2

Files changed

Area	Changes
Backend	New quantization_utils.py, enum update, 6 pipelines refactored, VACE mixin, hardware info API
Frontend	TypeScript types, supportsNvfp4 wiring, quantization dropdown with conditional NVFP4 option, persisted state reset
Deps	comfy-kitchen[cublas], torchaudio in pyproject.toml

[yondonfu]

I'll look into this more later, but on the first run with this branch with default LongLive settings:

2026-02-13 16:32:10,997 - scope.server.pipeline_manager - INFO - Initial load params: {'height': 320, 'width': 576, 'quantization': 'nvfp4', 'vace_enabled': False}
2026-02-13 16:32:10,998 - scope.server.pipeline_manager - INFO - VACE disabled by load_params, skipping VACE configuration
2026-02-13 16:32:11,804 - scope.core.pipelines.wan2_1.vace.mixin - INFO - _init_vace: No vace_path provided, VACE disabled
Loaded diffusion LoRA in 2.525s
2026-02-13 16:32:14,329 - scope.core.pipelines.wan2_1.lora.mixin - INFO - _init_loras: Found 0 LoRA configs to load
2026-02-13 16:32:15,419 - scope.core.pipelines.quantization_utils - INFO - Skipped 600 LoRA adapter layers
2026-02-13 16:32:15,419 - scope.core.pipelines.quantization_utils - INFO - Quantizing 301 Linear layers to NVFP4
2026-02-13 16:32:15,555 - scope.core.pipelines.quantization_utils - INFO - GPU memory before NVFP4 quantization: 3.30 GB
2026-02-13 16:32:15,562 - scope.server.pipeline_manager - ERROR - Failed to load pipeline longlive: CUDA kernel launch failed: CUDA driver version is insufficient for CUDA runtime version. If this error persists, consider removing the models directory 'C:\Users\yondo\.daydream-scope\models' and re-downloading models.
2026-02-13 16:32:15,568 - scope.server.pipeline_manager - ERROR - Failed to load pipeline: longlive
2026-02-13 16:32:15,568 - scope.server.pipeline_manager - ERROR - Some pipelines failed to load

Would be helpful to note the CUDA driver version required.

EDIT: I updated to the the latest driver version on my PC (see details below) and it runs now. Will share test results separately.

 NVIDIA-SMI 591.74                 Driver Version: 591.74         CUDA Version: 13.1

[varshith15]

why re-export?