optimization.py

"""
AOT compilation optimization for Qwen-Image-Edit pipeline.
"""

import gc
from typing import Any
from typing import Callable
from typing import ParamSpec
from torchao.quantization import quantize_
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
import spaces
import torch
from torch.utils._pytree import tree_map


P = ParamSpec('P')


TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length')
TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length')

TRANSFORMER_DYNAMIC_SHAPES = {
    'hidden_states': {
        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
    },
    'encoder_hidden_states': {
        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
    },
    'encoder_hidden_states_mask': {
        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
    },
    'image_rotary_emb': ({
        0: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
    }, {
        0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
    }),
}


INDUCTOR_CONFIGS = {
    'conv_1x1_as_mm': True,
    'epilogue_fusion': False,
    'coordinate_descent_tuning': True,
    'coordinate_descent_check_all_directions': True,
    'max_autotune': True,
    'triton.cudagraphs': True,
}


def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):

    @spaces.GPU(duration=1000)
    def compile_transformer():

        # Run warmup pass to capture transformer inputs (needs all components on GPU)
        with spaces.aoti_capture(pipeline.transformer) as call:
            pipeline(*args, **kwargs)

        # Offload text encoder to CPU to free ~16GB during memory-intensive torch.export
        # This happens AFTER warmup but BEFORE export to avoid device mismatch
        text_encoder_device = next(pipeline.text_encoder.parameters()).device
        pipeline.text_encoder.to('cpu')
        gc.collect()
        torch.cuda.empty_cache()

        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES

        # quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())

        exported = torch.export.export(
            mod=pipeline.transformer,
            args=call.args,
            kwargs=call.kwargs,
            dynamic_shapes=dynamic_shapes,
        )

        compiled = spaces.aoti_compile(exported, INDUCTOR_CONFIGS)

        # Move text encoder back to original device
        pipeline.text_encoder.to(text_encoder_device)

        return compiled

    spaces.aoti_apply(compile_transformer(), pipeline.transformer)