Method _load_pipeline

backend/python/diffusers/backend.py:218–492

Load a diffusers pipeline dynamically using the dynamic loader. This method uses load_diffusers_pipeline() for most pipelines, falling back to explicit handling only for pipelines requiring custom initialization (e.g., quantization, special VAE handling). A

(self, request, modelFile, fromSingleFile, torchType, variant, device_map=None)

class BackendServicer(backend_pb2_grpc.BackendServicer):

    def _load_pipeline(self, request, modelFile, fromSingleFile, torchType, variant, device_map=None):
        """
        Load a diffusers pipeline dynamically using the dynamic loader.

        This method uses load_diffusers_pipeline() for most pipelines, falling back
        to explicit handling only for pipelines requiring custom initialization
        (e.g., quantization, special VAE handling).

        Args:
            request: The gRPC request containing pipeline configuration
            modelFile: Path to the model file (for single file loading)
            fromSingleFile: Whether to use from_single_file() vs from_pretrained()
            torchType: The torch dtype to use
            variant: Model variant (e.g., "fp16")
            device_map: Device mapping strategy (e.g., "auto" for multi-GPU)

        Returns:
            The loaded pipeline instance
        """
        pipeline_type = request.PipelineType

        # Handle IMG2IMG request flag with default pipeline
        if request.IMG2IMG and pipeline_type == "":
            pipeline_type = "StableDiffusionImg2ImgPipeline"

        # ================================================================
        # Special cases requiring custom initialization logic
        # Only handle pipelines that truly need custom code (quantization,
        # special VAE handling, etc.). All other pipelines use dynamic loading.
        # ================================================================

        # FluxTransformer2DModel - requires quantization and custom transformer loading
        if pipeline_type == "FluxTransformer2DModel":
            dtype = torch.bfloat16
            bfl_repo = os.environ.get("BFL_REPO", "ChuckMcSneed/FLUX.1-dev")

            transformer = FluxTransformer2DModel.from_single_file(modelFile, torch_dtype=dtype, device_map=device_map)
            quantize(transformer, weights=qfloat8)
            freeze(transformer)
            text_encoder_2 = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype, device_map=device_map)
            quantize(text_encoder_2, weights=qfloat8)
            freeze(text_encoder_2)

            pipe = FluxPipeline.from_pretrained(bfl_repo, transformer=None, text_encoder_2=None, torch_dtype=dtype, device_map=device_map)
            pipe.transformer = transformer
            pipe.text_encoder_2 = text_encoder_2

            if request.LowVRAM:
                pipe.enable_model_cpu_offload()
            return pipe

        # WanPipeline - requires special VAE with float32 dtype
        if pipeline_type == "WanPipeline":
            vae = AutoencoderKLWan.from_pretrained(
                request.Model,
                subfolder="vae",
                torch_dtype=torch.float32,
                device_map=device_map