mombeeru/FLUX.2-dev

1"""2"""3 4from typing import Any5from typing import Callable6from typing import ParamSpec7import spaces8import torch9from spaces.zero.torch.aoti import ZeroGPUCompiledModel10from spaces.zero.torch.aoti import ZeroGPUWeights11from torch.utils._pytree import tree_map12 13P = ParamSpec('P')14 15TRANSFORMER_IMAGE_DIM = torch.export.Dim('image_seq_length', min=4096, max=16384) # min: 0 images, max: 3 (1024x1024) images16 17TRANSFORMER_DYNAMIC_SHAPES = {18    'double': {19        'hidden_states': {20            1: TRANSFORMER_IMAGE_DIM,21        },22        'image_rotary_emb': (23            {0: TRANSFORMER_IMAGE_DIM + 512},24            {0: TRANSFORMER_IMAGE_DIM + 512},25        ),26    },27    'single': {28        'hidden_states': {29            1: TRANSFORMER_IMAGE_DIM + 512,30        },31        'image_rotary_emb': (32            {0: TRANSFORMER_IMAGE_DIM + 512},33            {0: TRANSFORMER_IMAGE_DIM + 512},34        ),35    },36}37 38INDUCTOR_CONFIGS = {39    'conv_1x1_as_mm': True,40    'epilogue_fusion': False,41    'coordinate_descent_tuning': True,42    'coordinate_descent_check_all_directions': True,43    'max_autotune': True,44    'triton.cudagraphs': True,45}46 47def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):48 49    blocks = {50        'double': pipeline.transformer.transformer_blocks,51        'single': pipeline.transformer.single_transformer_blocks,52    }53 54    @spaces.GPU(duration=1200)55    def compile_block(blocks_kind: str):56        block = blocks[blocks_kind][0]57        with spaces.aoti_capture(block) as call:58            pipeline(*args, **kwargs)59 60        dynamic_shapes = tree_map(lambda t: None, call.kwargs)61        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES[blocks_kind]62 63        with torch.no_grad():      64            exported = torch.export.export(65                mod=block,66                args=call.args,67                kwargs=call.kwargs,68                dynamic_shapes=dynamic_shapes,69            )70 71        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS).archive_file72 73    for blocks_kind in ('double', 'single'):74        archive_file = compile_block(blocks_kind)75        for block in blocks[blocks_kind]:76            weights = ZeroGPUWeights(block.state_dict())77            block.forward = ZeroGPUCompiledModel(archive_file, weights)78