File size: 2,177 Bytes
6571814
 
 
 
 
dd2bdb6
 
6571814
 
 
 
3174181
 
6571814
 
 
 
 
 
 
 
 
 
 
3174181
dd2bdb6
 
3174181
6571814
 
 
 
 
 
 
 
 
 
 
 
 
 
dd2bdb6
6571814
dd2bdb6
 
 
 
6571814
 
 
dd2bdb6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
from typing import Any
from typing import Callable
from typing import ParamSpec
import spaces
import torch
from spaces.zero.torch.aoti import ZeroGPUCompiledModel
from spaces.zero.torch.aoti import ZeroGPUWeights
from torch.utils._pytree import tree_map

P = ParamSpec('P')

TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length', min=1024, max=16384)
TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length', min=64, max=1024)

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},  # vid_freqs
        {0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM},   # txt_freqs
    ),
}

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):

    blocks = pipeline.transformer.transformer_blocks

    @spaces.GPU(duration=1200)
    def compile_block():
        block = blocks[0]
        with spaces.aoti_capture(block) as call:
            pipeline(*args, **kwargs)

        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
        # Only merge keys that exist in call.kwargs
        for key, value in TRANSFORMER_DYNAMIC_SHAPES.items():
            if key in call.kwargs:
                dynamic_shapes[key] = value

        with torch.no_grad():
            exported = torch.export.export(
                mod=block,
                args=call.args,
                kwargs=call.kwargs,
                dynamic_shapes=dynamic_shapes,
            )

        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS).archive_file

    archive_file = compile_block()
    for block in blocks:
        weights = ZeroGPUWeights(block.state_dict())
        block.forward = ZeroGPUCompiledModel(archive_file, weights)