o ۷i@s0dZddlmZddlmZGdddeZdS)z3FP8 quantization config for diffusion transformers.) Fp8Config)DiffusionQuantizationConfigc@sBeZdZdZeZ   d dedeedBdeedBfddZ dS) DiffusionFp8Configa`FP8 quantization config optimized for diffusion transformers. Uses dynamic activation scaling (no calibration dataset needed) and online weight quantization from BF16/FP16 checkpoints. Device Compatibility: - Turing (SM 75+): Weight-only FP8 via Marlin kernel - Ada/Hopper (SM 89+): Full W8A8 FP8 with native hardware support The kernel selection is automatic based on GPU capability. Args: activation_scheme: Activation quantization scheme. - "dynamic": Per-token dynamic scaling (default, no calibration) - "static": Single per-tensor scale (requires calibration) weight_block_size: Block size for block-wise weight quantization. Format: [block_n, block_k]. If None, uses per-tensor scaling. ignored_layers: List of layer name patterns to skip quantization. dynamicNactivation_schemeweight_block_sizeignored_layerscCs,||_||_|p g|_td|||d|_dS)NF)is_checkpoint_fp8_serializedrrr )rrr r _vllm_config)selfrrr r Z/home/ubuntu/vllm_env/lib/python3.10/site-packages/vllm_omni/diffusion/quantization/fp8.py__init__"s  zDiffusionFp8Config.__init__)rNN) __name__ __module__ __qualname____doc__rquant_config_clsstrlistintrr r r rr s  rN)r+vllm.model_executor.layers.quantization.fp8rbaserrr r r rs