File size: 5,558 Bytes
809f5b5 |
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 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: MIT
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.
import types
from pathlib import Path
import tensorrt as trt
import torch
from cache_diffusion.cachify import CACHED_PIPE, get_model
from cuda import cudart
from diffusers.models.transformers.transformer_sd3 import SD3Transformer2DModel
from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
from trt_pipeline.config import ONNX_CONFIG
from trt_pipeline.models.sd3 import sd3_forward
from trt_pipeline.models.sdxl import (
cachecrossattnupblock2d_forward,
cacheunet_forward,
cacheupblock2d_forward,
)
from polygraphy.backend.trt import (
CreateConfig,
Profile,
engine_from_network,
network_from_onnx_path,
save_engine,
)
from torch.onnx import export as onnx_export
from .utils import Engine
def replace_new_forward(backbone):
if backbone.__class__ == UNet2DConditionModel:
backbone.forward = types.MethodType(cacheunet_forward, backbone)
for upsample_block in backbone.up_blocks:
if (
hasattr(upsample_block, "has_cross_attention")
and upsample_block.has_cross_attention
):
upsample_block.forward = types.MethodType(
cachecrossattnupblock2d_forward, upsample_block
)
else:
upsample_block.forward = types.MethodType(cacheupblock2d_forward, upsample_block)
elif backbone.__class__ == SD3Transformer2DModel:
backbone.forward = types.MethodType(sd3_forward, backbone)
def get_input_info(dummy_dict, info: str = None, batch_size: int = 1):
return_val = [] if info == "profile_shapes" or info == "input_names" else {}
def collect_leaf_keys(d):
for key, value in d.items():
if isinstance(value, dict):
collect_leaf_keys(value)
else:
value = (value[0] * batch_size,) + value[1:]
if info == "profile_shapes":
return_val.append((key, value)) # type: ignore
elif info == "profile_shapes_dict":
return_val[key] = value # type: ignore
elif info == "dummy_input":
return_val[key] = torch.ones(value).half().cuda() # type: ignore
elif info == "input_names":
return_val.append(key) # type: ignore
collect_leaf_keys(dummy_dict)
return return_val
def get_total_device_memory(backbone):
max_device_memory = 0
for _, engine in backbone.engines.items():
max_device_memory = max(max_device_memory, engine.engine.device_memory_size)
return max_device_memory
def load_engines(backbone, engine_path: Path, batch_size: int = 1):
backbone.engines = {}
for f in engine_path.iterdir():
if f.is_file():
eng = Engine()
eng.load(str(f))
backbone.engines[f"{f.stem}"] = eng
_, shared_device_memory = cudart.cudaMalloc(get_total_device_memory(backbone))
for engine in backbone.engines.values():
engine.activate(shared_device_memory)
backbone.cuda_stream = cudart.cudaStreamCreate()[1]
for block_name in backbone.engines.keys():
backbone.engines[block_name].allocate_buffers(
shape_dict=get_input_info(
ONNX_CONFIG[backbone.__class__][block_name]["dummy_input"],
"profile_shapes_dict",
batch_size,
),
device=backbone.device,
batch_size=batch_size,
)
# TODO: Free and clean up the origin pytorch cuda memory
def warm_up(backbone, batch_size: int = 1):
print("Warming-up TensorRT engines...")
for name, engine in backbone.engines.items():
dummy_input = get_input_info(
ONNX_CONFIG[backbone.__class__][name]["dummy_input"], "dummy_input", batch_size
)
_ = engine(dummy_input, backbone.cuda_stream)
def teardown(pipe):
backbone = get_model(pipe)
for engine in backbone.engines.values():
del engine
cudart.cudaStreamDestroy(backbone.cuda_stream)
del backbone.cuda_stream
def load_unet_trt(unet, engine_path: Path, batch_size: int = 1):
backbone = unet
engine_path.mkdir(parents=True, exist_ok=True)
replace_new_forward(backbone)
load_engines(backbone, engine_path, batch_size)
warm_up(backbone, batch_size)
backbone.use_trt_infer = True
|