o ٷi@sUdZddlmZddlmZddlmZddlmZddl m Z er)ddl m Z ee Zee d Zeeeefed <eeZd ed Bd ed BfddZded Bd dfddZgdZd S)aQuantization support for diffusion models. This module provides a unified interface for quantizing diffusion transformers using various methods (FP8, etc.). It wraps vLLM's quantization infrastructure while allowing diffusion-model-specific defaults and optimizations. Example usage: from vllm_omni.diffusion.quantization import ( get_diffusion_quant_config, get_vllm_quant_config_for_layers, ) # Create FP8 config for diffusion model diff_config = get_diffusion_quant_config("fp8") # Get vLLM config to pass to linear layers vllm_config = get_vllm_quant_config_for_layers(diff_config) # Use in model initialization linear_layer = QKVParallelLinear(..., quant_config=vllm_config) ) TYPE_CHECKING) init_logger)DiffusionQuantizationConfig)DiffusionFp8Config)DiffusionGgufConfig)QuantizationConfig)fp8gguf_QUANT_CONFIG_REGISTRY quantizationNreturncKs^|dus |dkr dS|}|tvrtd|dtt|}td||di|S)aFactory function to create quantization config for diffusion models. Args: quantization: Quantization method name ("fp8", etc.) or None to disable **kwargs: Method-specific parameters passed to the config constructor Returns: DiffusionQuantizationConfig instance or None if quantization is disabled Raises: ValueError: If the quantization method is not supported Example: # Default FP8 with dynamic activation scaling config = get_diffusion_quant_config("fp8") # FP8 with custom parameters config = get_diffusion_quant_config( "fp8", activation_scheme="static", ignored_layers=["proj_out"], ) NnonezUnknown quantization method: z. Supported methods: z*Creating diffusion quantization config: %s)lowerr ValueErrorSUPPORTED_QUANTIZATION_METHODSloggerinfo)r kwargs config_clsrr]/home/ubuntu/.local/lib/python3.10/site-packages/vllm_omni/diffusion/quantization/__init__.pyget_diffusion_quant_config2s rdiffusion_quant_configzQuantizationConfig | NonecCs|durdS|S)aGet the vLLM QuantizationConfig to pass to linear layers. This extracts the underlying vLLM config from a DiffusionQuantizationConfig, which can then be passed to vLLM linear layers (QKVParallelLinear, etc.). Args: diffusion_quant_config: The diffusion quantization config, or None Returns: vLLM QuantizationConfig instance, or None if input is None N)get_vllm_quant_config)rrrr get_vllm_quant_config_for_layers[sr)rrrrrr)__doc__typingr vllm.loggerrbaserr rr r3vllm.model_executor.layers.quantization.base_configr__name__rr dictstrtype__annotations__listkeysrrr__all__rrrrs0        )