897 B
897 B
Test Report
Version: 0.2.0
Executed synthetic tests
- Package source compiles as Python.
- Krea 2 architecture key validation.
- BF16 + FP16 global-ratio merge.
- BF16 output.
- FP16 output.
- FP32 output.
- Legacy FP8 scaled input using
_quantization_metadata + weight_scale. - Current FP8 scaled input using
comfy_quant + weight_scale. - FP8 input dequantization before merge.
- FP8 E4M3FN requantization after merge.
- Generation of
weight_scale,comfy_quantand_quantization_metadata. - Streaming safetensors write and safe_open validation.
- Mismatched FP8 quantization masks are rejected.
- Existing output names receive a numeric suffix.
Not yet executed
- Full 13.1GB Krea 2 FP8 checkpoint.
- Full 26.3GB Krea 2 BF16 checkpoint.
- Windows long-running save under the user's exact ComfyUI build.
- INT8 ConvRot, MXFP8 and NVFP4. These are intentionally unsupported.