TensorRT-LLMs/tensorrt_llm/tools/multimodal_builder.py
Yechan Kim 8ba8699f66
[TRTLLM-8310][feat] Add Qwen3-VL-MoE (#9689)
Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com>
2025-12-15 20:05:20 -08:00

1840 lines
73 KiB
Python

import math
import os
import shutil
import sys
import tarfile
from time import time
import yaml
# isort: off
import torch
import tensorrt as trt
from pathlib import Path
from tensorrt_llm._utils import torch_dtype_to_str, to_json_file
from tensorrt_llm.builder import Builder
from tensorrt_llm.logger import logger
from transformers import (AutoConfig, AutoModel, AutoModelForCausalLM,
AutoModelForVision2Seq, AutoProcessor,
Blip2ForConditionalGeneration, Blip2Processor,
FuyuForCausalLM, FuyuProcessor,
LlavaForConditionalGeneration, NougatProcessor,
Pix2StructForConditionalGeneration,
VisionEncoderDecoderModel, CLIPVisionModel)
# isort: on
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from safetensors.torch import load_model, save_file
from transformers import CLIPImageProcessor
from ..runtime.session import Session
def add_multimodal_arguments(parser):
parser.add_argument(
'--model_type',
type=str,
default=None,
choices=[
'blip2', 'llava', 'llava_next', 'llava_onevision',
'llava_onevision_lmms', 'vila', 'nougat', 'cogvlm', 'fuyu',
'pix2struct', 'neva', 'kosmos-2', 'video-neva', 'phi-3-vision',
'phi-4-multimodal', 'mllama', 'internvl', 'qwen2_vl',
'internlm-xcomposer2', 'qwen2_audio', 'pixtral', 'eclair'
],
help="Model type")
parser.add_argument(
'--model_path',
type=str,
default=None,
help=
"Huggingface repo, local directory with weights or path to checkpoint file"
)
parser.add_argument('--vila_path',
type=str,
default=None,
help="Path to VILA source code directory")
parser.add_argument('--output_dir',
type=str,
default=None,
help="Directory where visual TRT engines are saved")
parser.add_argument('--max_batch_size',
type=int,
default=4,
help="Maximum batch size for input images")
parser.add_argument(
'--max_hw_dims',
type=int,
default=5184,
help=
"Maximum multiply of h and w after patching for input images for qwen2_vl"
)
parser.add_argument(
'--min_hw_dims',
type=int,
default=128,
help=
"Minimum multiply of h and w after patching for input images for qwen2_vl"
)
parser.add_argument(
'--num_mul_bins',
type=int,
default=128,
help="Number of Mel frequency bins of input audios for qwen2_audio")
parser.add_argument(
'--max_mel_seq_len',
type=int,
default=3000,
help=
"Maximum Mel frequency feature lengths of input audios for qwen2_audio")
return parser
class MultimodalEngineBuilder:
def __init__(self, args):
args.device = torch.device(
"cuda") if torch.cuda.is_available() else "cpu"
if args.output_dir is None:
# default path to save the engines
model_name = args.model_path.split('/')[-1]
args.output_dir = f'tmp/trt_engines/{model_name}/multimodal_encoder'
os.makedirs(args.output_dir, exist_ok=True)
self.args = args
def build(self):
args = self.args
if args.model_type == 'blip2':
build_blip2_engine(args)
elif args.model_type == 'internlm-xcomposer2':
build_interlm_xcomposer2_engine(args)
elif args.model_type == 'pix2struct':
build_pix2struct_engine(args)
elif 'llava' in args.model_type:
build_llava_engine(args)
elif args.model_type == 'vila':
assert args.vila_path is not None, "Please clone and provide VILA source code path"
build_vila_engine(args)
elif args.model_type == 'nougat':
build_nougat_engine(args)
elif args.model_type == 'cogvlm':
build_cogvlm_engine(args)
elif args.model_type == 'fuyu':
build_fuyu_engine(args)
elif args.model_type == 'neva':
build_neva_engine(args)
elif args.model_type == 'video-neva':
build_video_neva_engine(args)
elif args.model_type == 'kosmos-2':
build_kosmos_engine(args)
elif args.model_type == 'phi-3-vision':
build_phi_engine(args)
elif args.model_type == 'phi-4-multimodal':
build_phi4mm_engine(args)
elif args.model_type == 'mllama':
build_mllama_engine(args)
elif args.model_type == 'internvl':
build_internvl_engine(args)
elif args.model_type == 'qwen2_vl':
build_qwen2_vl_engine(args)
elif args.model_type == 'qwen2_audio':
build_qwen2_audio_engine(args)
elif args.model_type == "pixtral":
build_pixtral_engine(args)
elif args.model_type == "eclair":
build_eclair_engine(args)
else:
raise RuntimeError(f"Invalid model type {args.model_type}")
def export_onnx(model,
input,
onnx_dir,
onnx_name='model.onnx',
input_names=['input'],
output_names=['encoder_output'],
dynamic_axes={'input': {
0: 'batch'
}},
logger=trt.Logger(trt.Logger.INFO)):
logger.log(trt.Logger.INFO, f"Exporting onnx to {onnx_dir}/{onnx_name}")
os.makedirs(onnx_dir, exist_ok=True)
torch.onnx.export(
model,
input,
f'{onnx_dir}/{onnx_name}',
opset_version=17,
input_names=input_names,
output_names=output_names,
dynamic_axes=dynamic_axes,
# Required for pytorch>=2.9.0 as dynamo becomes the default and introduces bugs as it does not support opset_version=17 natively
dynamo=False)
def build_trt_engine(model_type,
input_sizes,
onnx_dir,
engine_dir,
max_batch_size,
dtype=torch.float16,
model_params=None,
onnx_name='model.onnx',
engine_name='model.engine',
delete_onnx=True,
logger=trt.Logger(trt.Logger.INFO)):
"""Build TensorRT engine from ONNX model.
Args:
model_params (dict): Optional model specific parameters, e.g.:
- qwen2_vl_dim (int): Dimension for Qwen2-VL model
- min_hw_dims (int): Minimum HW dimensions
- max_hw_dims (int): Maximum HW dimensions
- num_frames (int): Number of frames for video models
"""
model_params = model_params or {}
onnx_file = f'{onnx_dir}/{onnx_name}'
engine_file = f'{engine_dir}/{engine_name}'
config_file = f'{engine_dir}/config.json'
logger.log(trt.Logger.INFO, f"Building TRT engine to {engine_file}")
builder = trt.Builder(logger)
network = builder.create_network(
1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
profile = builder.create_optimization_profile()
config_args = {
"precision": torch_dtype_to_str(dtype),
"model_type": model_type,
"strongly_typed": False,
"max_batch_size": max_batch_size,
"model_name": "multiModal"
}
if "num_frames" in model_params:
config_args["num_frames"] = model_params["num_frames"]
config_wrapper = Builder().create_builder_config(**config_args)
config = config_wrapper.trt_builder_config
parser = trt.OnnxParser(network, logger)
with open(onnx_file, 'rb') as model:
if not parser.parse(model.read(), os.path.abspath(onnx_file)):
logger.log(trt.Logger.ERROR, "Failed parsing %s" % onnx_file)
for error in range(parser.num_errors):
logger.log(trt.Logger.ERROR, parser.get_error(error))
logger.log(trt.Logger.INFO, "Succeeded parsing %s" % onnx_file)
nBS = -1
nMinBS = 1
nOptBS = max(nMinBS, int(max_batch_size / 2))
nMaxBS = max_batch_size
# input sizes can be:
# - integer list, when inputs are constant size images. e.g. [3, H, W]
# - list of integer lists, when inputs are dynamic size images. e.g. [[1, 1, 2700], [1, 500, 2700], [1, 4096, 2700]]
# - list of list of integer lists, when there are many inputs and each input have dynamic size. e.g.
# [[[1, 1, 2700], [1, 500, 2700], [1, 4096, 2700]], [[1, 1], [1, 1], [1,1]]]
assert isinstance(input_sizes, list), "input_sizes must be a list"
if model_type == "qwen2_vl":
input_images = network.get_input(0)
inputT = network.get_input(1)
attenstion_mask = network.get_input(2)
qwen2_vl_dim = model_params.get('qwen2_vl_dim', 0)
min_hw_dims = model_params.get('min_hw_dims', 0)
max_hw_dims = model_params.get('max_hw_dims', 0)
assert min_hw_dims > 0, "min_hw_dims must be positive for qwen2_vl"
assert max_hw_dims > 0, "max_hw_dims must be positive for qwen2_vl"
multi_size_min = min_hw_dims
multi_size_max = max_hw_dims * max_batch_size
multi_size_opt = max(multi_size_min, int(multi_size_max / 2))
inputT.shape = [-1, *input_sizes]
profile.set_shape(inputT.name, [multi_size_min, *input_sizes],
[multi_size_opt, *input_sizes],
[multi_size_max, *input_sizes])
input_images.shape = [-1, qwen2_vl_dim]
profile.set_shape(input_images.name, [multi_size_min, qwen2_vl_dim],
[multi_size_opt, qwen2_vl_dim],
[multi_size_max, qwen2_vl_dim])
attenstion_mask.shape = [1, -1, -1]
profile.set_shape(attenstion_mask.name,
[1, multi_size_min, multi_size_min],
[1, multi_size_opt, multi_size_opt],
[1, multi_size_max, multi_size_max])
elif model_type == "qwen2_audio":
inputT = network.get_input(0)
mask = network.get_input(1)
num_mul_bins = model_params.get('num_mul_bins', 0)
max_mel_seq_len = model_params.get('max_mel_seq_len', 0)
assert num_mul_bins > 0, "num_mul_bins must be positive for qwen2_audio"
assert max_mel_seq_len > 0, "max_mel_seq_len must be positive for qwen2_audio"
inputT.shape = [nBS, num_mul_bins, max_mel_seq_len]
max_seq_len = (max_mel_seq_len - 2) // 2 + 1
mask.shape = [nBS, 1, max_seq_len, max_seq_len]
profile.set_shape(
inputT.name,
[nMinBS, num_mul_bins, max_mel_seq_len],
[nOptBS, num_mul_bins, max_mel_seq_len],
[nMaxBS, num_mul_bins, max_mel_seq_len],
)
profile.set_shape(
mask.name,
[nMinBS, 1, max_seq_len, max_seq_len],
[nOptBS, 1, max_seq_len, max_seq_len],
[nMaxBS, 1, max_seq_len, max_seq_len],
)
else:
if isinstance(input_sizes[0], int):
logger.log(trt.Logger.INFO, f"Processed input sizes {input_sizes}")
inputT = network.get_input(0)
inputT.shape = [nBS, *input_sizes]
min_size = opt_size = max_size = input_sizes
profile.set_shape(inputT.name, [nMinBS, *min_size],
[nOptBS, *opt_size], [nMaxBS, *max_size])
elif isinstance(input_sizes[0], list) and isinstance(
input_sizes[0][0], list):
for idx, input_size in enumerate(input_sizes):
assert len(input_size) == 3
inputT = network.get_input(idx)
min_size, opt_size, max_size = input_size
profile.set_shape(inputT.name, [nMinBS, *min_size],
[nOptBS, *opt_size], [nMaxBS, *max_size])
elif len(input_sizes) == 3 and isinstance(input_sizes[0], list):
inputT = network.get_input(0)
min_size, opt_size, max_size = input_sizes
profile.set_shape(inputT.name, [nMinBS, *min_size],
[nOptBS, *opt_size], [nMaxBS, *max_size])
logger.log(
trt.Logger.INFO,
f"Processed min/opt/max input sizes {min_size}/{opt_size}/{max_size}"
)
else:
raise ValueError(f"invalid input sizes: {input_sizes}")
config.add_optimization_profile(profile)
t0 = time()
engine_string = builder.build_serialized_network(network, config)
t1 = time()
if engine_string is None:
raise RuntimeError("Failed building %s" % (engine_file))
else:
logger.log(trt.Logger.INFO,
"Succeeded building %s in %d s" % (engine_file, t1 - t0))
logger.log(trt.Logger.INFO, 'Recording engine output shape in config')
engine_session = Session.from_serialized_engine(engine_string)
output_tensor_name = network.get_output(0).name
output_shape = engine_session.engine.get_tensor_shape(
output_tensor_name)
output_shape = list(output_shape)
config_wrapper.output_shape = output_shape
os.makedirs(engine_dir, exist_ok=True)
with open(engine_file, 'wb') as f:
f.write(engine_string)
# Clear onnx files since we no longer need them after a successful engine build
if delete_onnx:
shutil.rmtree(onnx_dir)
Builder.save_config(config_wrapper, config_file)
def build_blip2_engine(args):
processor = Blip2Processor.from_pretrained(args.model_path)
raw_image = Image.new('RGB', [10, 10]) # dummy image
prompt = "Question: what is this? Answer:"
inputs = processor(raw_image, prompt,
return_tensors="pt").to(args.device, torch.float16)
image = inputs['pixel_values']
class Blip2VisionWrapper(torch.nn.Module):
def __init__(self, vision_model, qformer, projector, query_tokens):
super().__init__()
self.vision_model = vision_model
self.qformer = qformer
self.projector = projector
self.query_tokens = query_tokens
def forward(self, image):
features = self.vision_model(image)[0]
qformer_output = self.qformer(query_embeds=self.query_tokens,
encoder_hidden_states=features,
return_dict=True)
return self.projector(qformer_output.last_hidden_state)
model = Blip2ForConditionalGeneration.from_pretrained(args.model_path,
dtype=torch.float16)
blip2_llm = ""
if model.language_model.config.architectures[
0] == 'T5ForConditionalGeneration':
blip2_llm = "t5"
elif model.language_model.config.architectures[0] == 'OPTForCausalLM':
blip2_llm = "opt"
wrapper = Blip2VisionWrapper(model.vision_model, model.qformer,
model.language_projection, model.query_tokens)
wrapper.to(args.device)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type + "-" + blip2_llm, # blip2-t5 or blip2-opt
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_interlm_xcomposer2_engine(args):
model = AutoModel.from_pretrained(args.model_path,
trust_remote_code=True).to(torch.float16)
raw_image = Image.new('RGB', [10, 10])
image = model.vis_processor(raw_image).unsqueeze(0).to(
args.device, torch.float16)
class InternLMXComposer2VisionWrapper(torch.nn.Module):
def __init__(self, vision_model, vision_proj):
super().__init__()
self.vision_model = vision_model
self.vision_proj = vision_proj
def forward(self, image):
return self.vision_proj(self.vision_model(image))
wrapper = InternLMXComposer2VisionWrapper(model.vit, model.vision_proj)
wrapper.to(args.device)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_pix2struct_engine(args):
processor = AutoProcessor.from_pretrained(args.model_path)
raw_image = Image.new('RGB', [10, 10]) # dummy image
dtype = torch.float16
inputs = processor(text="dummy", images=raw_image, return_tensors="pt")
image = inputs['flattened_patches'].to(args.device, dtype)
class pix2structVisionWrapper(torch.nn.Module):
def __init__(self, encoder):
super().__init__()
self.encoder = encoder
def forward(self, image):
attention_mask = (image.abs().sum(dim=-1) != 0)
vision_x = self.encoder.embeddings(image)
img_features = self.encoder.encoder(vision_x,
attention_mask=attention_mask)
img_features = self.encoder.layernorm(img_features[0])
return img_features
model = Pix2StructForConditionalGeneration.from_pretrained(args.model_path,
dtype=dtype)
wrapper = pix2structVisionWrapper(model.encoder.to(args.device))
# input shape: batch size, number of patches, hidden dimension
# attention mask shape: batch size, number of patches
# The number of image patches can vary depending on the image size, but it typically
# falls within a relatively narrow range. To improve performance, we can avoid using
# dynamic axis for the input patches and instead use a fixed number of patches along
# with an attention mask.
export_onnx(wrapper, (image, ),
f'{args.output_dir}/onnx',
input_names=['input'],
dynamic_axes={'input': {
0: 'batch'
}})
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2]], # Number of Patches, Hidden Dimension
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
dtype=torch.bfloat16)
def build_llava_engine(args):
processor = AutoProcessor.from_pretrained(args.model_path)
if args.model_type == "llava":
raw_image = Image.new('RGB', [10, 10]) # dummy image
image = processor(text="dummy", images=raw_image,
return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)
class LlavaVisionWrapper(torch.nn.Module):
def __init__(self, tower, projector, feature_layer):
super().__init__()
self.tower = tower
self.projector = projector
self.feature_layer = feature_layer
def forward(self, image):
all_hidden_states = self.tower(
image, output_hidden_states=True).hidden_states
features = all_hidden_states[self.feature_layer][:, 1:]
return self.projector(features)
hf_config = AutoConfig.from_pretrained(args.model_path)
hf_config.vision_config._attn_implementation = "eager"
# Need to setup at hf_config._attn_implementation after transformers >= 4.46
hf_config._attn_implementation = "eager"
model = LlavaForConditionalGeneration.from_pretrained(
args.model_path, dtype=torch.float16, config=hf_config)
wrapper = LlavaVisionWrapper(
model.vision_tower.to(args.device),
model.multi_modal_projector.to(args.device),
model.config.vision_feature_layer)
elif args.model_type == "llava_next":
from transformers import LlavaNextForConditionalGeneration
raw_image = Image.new('RGB', [512, 512])
image = processor(text="dummy", images=raw_image,
return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)[0]
class LlavaNextVisionWrapper(torch.nn.Module):
def __init__(self, vision_tower, projector):
super().__init__()
self.vision_tower = vision_tower
self.projector = projector
def forward(self, pixel_values):
image_features = self.vision_tower(pixel_values,
output_hidden_states=True)
selected_image_feature = image_features.hidden_states[-2][:, 1:]
image_features = self.projector(selected_image_feature)
return image_features # (bs, 576, c)
hf_config = AutoConfig.from_pretrained(args.model_path)
hf_config.vision_config._attn_implementation = "eager"
model = LlavaNextForConditionalGeneration.from_pretrained(
args.model_path, dtype=torch.float16, config=hf_config)
wrapper = LlavaNextVisionWrapper(
model.vision_tower.vision_model.to(args.device),
model.multi_modal_projector.to(args.device),
)
elif args.model_type == "llava_onevision_lmms":
from llava.mm_utils import process_images
from llava.model.builder import load_pretrained_model
_, model, processor, _ = load_pretrained_model(args.model_path,
None,
args.model_type,
torch_dtype="float16")
raw_image = Image.new('RGB', [512, 512])
image = process_images([raw_image], processor,
model.config).squeeze(0).to(
args.device, torch.float16)
class LlavaQwenVisionWrapper(torch.nn.Module):
def __init__(self, vision_tower, projector):
super().__init__()
self.vision_tower = vision_tower
self.projector = projector
def forward(self, pixel_values):
image_features = self.vision_tower(pixel_values)
image_features = self.projector(image_features)
return image_features # (sigma(bs, patches_i), 729, c)
wrapper = LlavaQwenVisionWrapper(model.get_model().get_vision_tower(),
model.get_model().mm_projector)
elif args.model_type == "llava_onevision":
from transformers import LlavaOnevisionForConditionalGeneration
raw_image = Image.new('RGB', [512, 512])
image = processor(text="dummy", images=raw_image,
return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)[0]
class LlavaOnevisionVisionWrapper(torch.nn.Module):
def __init__(self, vision_tower, projector, config):
super().__init__()
self.vision_tower = vision_tower
self.projector = projector
self.config = config
def forward(self, pixel_values):
image_features = self.vision_tower(pixel_values,
output_hidden_states=True)
selected_image_feature = image_features.hidden_states[
self.config.vision_feature_layer]
image_features = self.projector(selected_image_feature)
return image_features # (sigma(bs, patches_i), 729, c)
model = LlavaOnevisionForConditionalGeneration.from_pretrained(
args.model_path, dtype=torch.float16)
wrapper = LlavaOnevisionVisionWrapper(
model.vision_tower.to(args.device),
model.multi_modal_projector.to(args.device), model.config)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
if args.model_type == "llava_next":
image_newline = model.model.image_newline.data
tensor_img_newline = {"image_newline": image_newline}
save_file(tensor_img_newline,
os.path.join(args.output_dir, "image_newlines.safetensors"))
if args.model_type == "llava_onevision":
image_newline = model.model.image_newline.data
tensor_img_newline = {"image_newline": image_newline}
save_file(tensor_img_newline,
os.path.join(args.output_dir, "image_newlines.safetensors"))
if args.model_type == "llava_onevision_lmms":
image_newline = model.model.image_newline.data
tensor_img_newline = {"image_newline": image_newline}
save_file(tensor_img_newline,
os.path.join(args.output_dir, "image_newlines.safetensors"))
def build_vila_engine(args):
# Note: VILA model is not in public HF model zoo yet. We need to explicitly import from the git repo
sys.path.append(args.vila_path)
from llava.model import LlavaLlamaConfig, LlavaLlamaModel # noqa
from transformers import AutoModel
model = AutoModel.from_pretrained(
args.model_path,
device_map='auto',
)
vision_tower = model.get_vision_tower()
image_processor = vision_tower.image_processor
raw_image = Image.new('RGB', [10, 10]) # dummy image
image = image_processor(images=raw_image,
return_tensors="pt")['pixel_values']
if isinstance(image, list):
image = image[0].unsqueeze(0)
image = image.to(args.device, torch.float16)
class VilaVisionWrapper(torch.nn.Module):
def __init__(self, tower, projector):
super().__init__()
self.tower = tower
self.projector = projector
def forward(self, image):
features = self.tower(image)
return self.projector(features)
model = AutoModel.from_pretrained(
args.model_path,
device_map='auto',
)
wrapper = VilaVisionWrapper(model.get_vision_tower().to(args.device),
model.mm_projector.to(args.device))
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_nougat_engine(args):
processor = NougatProcessor.from_pretrained(args.model_path)
raw_image = Image.new('RGB', [10, 10]) # dummy image
image = processor(raw_image, return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)
class SwinEncoderWrapper(torch.nn.Module):
def __init__(self, encoder):
super().__init__()
self.encoder = encoder
def forward(self, image):
return self.encoder(image).last_hidden_state
model = VisionEncoderDecoderModel.from_pretrained(args.model_path,
dtype=torch.float16)
swin_encoder = model.get_encoder().to(args.device)
wrapper = SwinEncoderWrapper(swin_encoder)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_cogvlm_engine(args):
hf_config = AutoConfig.from_pretrained(args.model_path,
trust_remote_code=True)
image_size = hf_config.vision_config['image_size']
dtype = hf_config.torch_dtype
image = torch.empty(1,
3,
image_size,
image_size,
dtype=dtype,
device=args.device) # dummy image
class CogVlmVisionWrapper(torch.nn.Module):
def __init__(self, encoder):
super().__init__()
self.encoder = encoder
def forward(self, image):
return self.encoder(image)
cogvlm = AutoModelForCausalLM.from_pretrained(args.model_path,
dtype=dtype,
trust_remote_code=True)
vit_encoder = cogvlm.model.vision.to(args.device).eval()
wrapper = CogVlmVisionWrapper(vit_encoder)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
dtype=dtype)
def build_fuyu_engine(args):
processor = FuyuProcessor.from_pretrained(args.model_path)
raw_image = Image.new('RGB', [10, 10])
image = processor(text="dummy", images=raw_image,
return_tensors="pt")['image_patches'][0].to(
args.device, torch.float16).unsqueeze(0)
class FuyuEncoderWrapper(torch.nn.Module):
def __init__(self, linear):
super().__init__()
self.linear = linear.to(torch.float16)
def forward(self, patches):
return self.linear(patches).flatten(0, 1)
model = FuyuForCausalLM.from_pretrained(args.model_path,
dtype=torch.float16)
vision_encoder = model.vision_embed_tokens
wrapper = FuyuEncoderWrapper(vision_encoder).to(args.device)
export_onnx(wrapper,
image,
f'{args.output_dir}/onnx',
dynamic_axes={'input': {
0: 'batch',
2: 'patch'
}})
build_trt_engine(
args.model_type,
# [nImgs, nImgPatches, nDims]
# nImgs is always one since each query has exactly one image
# nImgPatches depends on image size (patch size: 30x30)
# nDims is 30x30x3=2700 (patch size x color channels)
[[1, 1, 2700], [1, 500, 2700], [1, 4096, 2700]],
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_neva_engine(args):
# extract NeMo checkpoint
with tarfile.open(args.model_path) as tar:
nemo_config = yaml.safe_load(tar.extractfile("./model_config.yaml"))
try:
# trained without TP
mp0_weights = torch.load(tar.extractfile("./model_weights.ckpt"),
map_location=args.device)
except KeyError:
# trained with TP
mp0_weights = torch.load(
tar.extractfile("./mp_rank_00/model_weights.ckpt"),
map_location=args.device)
vision_config = nemo_config["mm_cfg"]["vision_encoder"]
class VisionEncoderWrapper(torch.nn.Module):
def __init__(self, encoder, connector):
super().__init__()
self.encoder = encoder
self.connector = connector
def forward(self, images):
vision_x = self.encoder(pixel_values=images,
output_hidden_states=True)
vision_x = vision_x.hidden_states[-2]
vision_x = vision_x[:, 1:]
vision_x = self.connector(vision_x)
return vision_x
vision_path = vision_config["from_pretrained"]
joined_path = os.path.join(os.path.dirname(args.model_path),
os.path.basename(vision_path))
if os.path.isdir(joined_path):
vision_path = joined_path
encoder = AutoModel.from_pretrained(vision_path,
dtype=torch.bfloat16,
trust_remote_code=True)
vision_encoder = encoder.vision_model
hf_config = encoder.config
dtype = hf_config.torch_dtype
# connector
assert nemo_config["mm_cfg"]["mm_mlp_adapter_type"] == "mlp2x_gelu"
vision_connector = torch.nn.Sequential(
torch.nn.Linear(vision_config["hidden_size"],
nemo_config["hidden_size"],
bias=True), torch.nn.GELU(),
torch.nn.Linear(nemo_config["hidden_size"],
nemo_config["hidden_size"],
bias=True)).to(dtype=dtype)
key_prefix = "model.embedding.word_embeddings.adapter_layer.mm_projector_adapter.mm_projector"
for layer in range(0, 3, 2):
vision_connector[layer].load_state_dict({
'weight':
mp0_weights[f"{key_prefix}.{layer}.weight"].to(dtype),
'bias':
mp0_weights[f"{key_prefix}.{layer}.bias"].to(dtype),
})
# export the whole wrapper
wrapper = VisionEncoderWrapper(vision_encoder,
vision_connector).to(args.device, dtype)
image_size = hf_config.vision_config.image_size
dummy_image = torch.empty(
1, 3, image_size, image_size, dtype=dtype,
device=args.device) # dummy image shape [B, C, H, W]
export_onnx(wrapper, dummy_image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[3, image_size, image_size], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
dtype=dtype)
def build_video_neva_engine(args):
# extract NeMo checkpoint
with tarfile.open(args.model_path) as tar:
nemo_config = yaml.safe_load(tar.extractfile("./model_config.yaml"))
try:
# trained without TP
mp0_weights = torch.load(tar.extractfile("./model_weights.ckpt"),
map_location=args.device)
except KeyError:
# trained with TP
mp0_weights = torch.load(
tar.extractfile("./mp_rank_00/model_weights.ckpt"),
map_location=args.device)
vision_config = nemo_config["mm_cfg"]["vision_encoder"]
class VisionEncoderWrapper(torch.nn.Module):
def __init__(self, encoder, connector):
super().__init__()
self.encoder = encoder
self.connector = connector
def forward(self, images):
b, num_frames, c, h, w = images.shape
images = images.view(b * num_frames, c, h, w)
vision_x = self.encoder(
pixel_values=images, #[(B num_frames), C, H, W]
output_hidden_states=True)
vision_x = vision_x.hidden_states[-2]
vision_x = vision_x[:, 1:]
# reshape back to [B, num_frames, img_size, hidden_size]
vision_x = vision_x.view(b, num_frames, -1, vision_x.shape[-1])
vision_x = self.connector(vision_x)
return vision_x
encoder = AutoModel.from_pretrained(vision_config["from_pretrained"],
dtype=torch.bfloat16,
trust_remote_code=True,
attn_implementation="eager")
vision_encoder = encoder.vision_model
hf_config = encoder.config
dtype = hf_config.torch_dtype
# connector
assert nemo_config["mm_cfg"]["mm_mlp_adapter_type"] == "linear"
vision_connector = torch.nn.Linear(vision_config["hidden_size"],
nemo_config["hidden_size"],
bias=True)
key_prefix = "model.embedding.word_embeddings.adapter_layer.mm_projector_adapter.mm_projector"
vision_connector.load_state_dict({
'weight':
mp0_weights[f"{key_prefix}.weight"].to(dtype),
'bias':
mp0_weights[f"{key_prefix}.bias"].to(dtype),
})
# export the whole wrapper
wrapper = VisionEncoderWrapper(vision_encoder,
vision_connector).to(args.device, dtype)
image_size = hf_config.vision_config.image_size
num_frames = nemo_config['data']['num_frames']
dummy_video = torch.empty(1,
num_frames,
3,
image_size,
image_size,
dtype=dtype,
device=args.device) # dummy image
export_onnx(wrapper, dummy_video, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[num_frames, 3, image_size, image_size], # [num_frames, 3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
dtype=dtype,
model_params={'num_frames': num_frames})
def build_kosmos_engine(args):
processor = AutoProcessor.from_pretrained(args.model_path)
raw_image = Image.new('RGB', [10, 10]) # dummy image
image = processor(text="dummy", images=raw_image,
return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)
class VisionEncoderWrapper(torch.nn.Module):
def __init__(self, encoder, connector):
super().__init__()
self.encoder = encoder
self.connector = connector
def forward(self, images):
vision_x = self.encoder(images, output_hidden_states=True)
img_features = self.encoder.model.post_layernorm(
vision_x.last_hidden_state)
img_features = F.normalize(img_features, dim=-1)
img_features, _ = self.connector(img_features)
return img_features
model = AutoModelForVision2Seq.from_pretrained(args.model_path,
dtype=torch.float16)
wrapper = VisionEncoderWrapper(
model.vision_model.to(args.device),
model.image_to_text_projection.to(args.device))
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size)
def build_phi_engine(args):
logger.warning(
"Skipping TRT engine build for Phi-3 vision encoder. MultimodalModelRunner will use PyTorch vision encoder. Flash/SDPA attention in CLIP encoder is not compatible with torch.onnx.export and eager attention is unstable in PyTorch."
)
# Dump config.json needed by model runner
config_args = {
"builder_config": {
"precision": torch_dtype_to_str(torch.float16),
"model_type": "phi-3-vision",
}
}
to_json_file(config_args, args.output_dir + "/config.json")
return
processor = AutoProcessor.from_pretrained(args.model_path,
trust_remote_code=True,
num_crops=16)
raw_image = Image.new('RGB', [10, 10]) # dummy image
image = processor(text="<|image_1|>\ndummy",
images=raw_image,
return_tensors="pt")['pixel_values'].to(
args.device, torch.float16)
image = image.flatten(0, 1)
class Phi3VisionWrapper(torch.nn.Module):
def __init__(self, vision_model):
super().__init__()
self.vision_model = vision_model
def forward(self, pixel_values):
return self.vision_model.get_img_features(pixel_values).reshape(
1, pixel_values.shape[0], -1, self.vision_model.image_dim_out)
model = AutoModelForCausalLM.from_pretrained(args.model_path,
dtype=torch.float16,
trust_remote_code=True)
vision_model = model.model.vision_embed_tokens
# Replace img_processor that uses flash attention with eager attention
clip_config = vision_model.img_processor.config
clip_config._attn_implementation = 'eager'
del vision_model.img_processor
vision_model.img_processor = CLIPVisionModel(clip_config).to(torch.float16)
vision_model = vision_model.to(args.device)
wrapper = Phi3VisionWrapper(vision_model)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
num_crops = processor.image_processor.num_crops
build_trt_engine(args.model_type,
[image.shape[1], image.shape[2], image.shape[3]],
f'{args.output_dir}/onnx', args.output_dir,
args.max_batch_size * (num_crops + 1))
def build_phi4mm_engine(args):
logger.warning(
"Skipping TRT engine build for Phi-4-multimodal encoder. MultimodalModelRunner will use PyTorch vision & audio encoder. Flash/SDPA attention in CLIP encoder is not compatible with torch.onnx.export and eager attention is unstable in PyTorch."
)
# Dump config.json needed by model runner
config_args = {
"builder_config": {
"precision": torch_dtype_to_str(torch.float16),
"model_type": "phi-4-multimodal",
}
}
os.makedirs(os.path.join(args.output_dir, "vision"), exist_ok=True)
os.makedirs(os.path.join(args.output_dir, "audio"), exist_ok=True)
to_json_file(config_args,
os.path.join(args.output_dir, "vision", "config.json"))
to_json_file(config_args,
os.path.join(args.output_dir, "audio", "config.json"))
return
# Following code works ok with eager mode attention. Leaving it here so that it could
# be used once issues in torch / onnx mentioned above resolved.
processor = AutoProcessor.from_pretrained(args.model_path,
trust_remote_code=True)
raw_image = Image.new('RGB', [10, 10]) # dummy image
import numpy as np
audio_feature_size = 500
audio_compression_rate = 8
audio_sampling_rate = 16000
audio_len = int((audio_feature_size * audio_compression_rate + 2) *
audio_sampling_rate / 100)
raw_audio = (np.zeros(audio_len), audio_sampling_rate) # dummy audio
inputs = processor(text="<|image_1|><|audio_1|>\ndummy",
images=[raw_image],
audios=[raw_audio],
return_tensors="pt")
img_embeds = inputs['input_image_embeds'].to(args.device, torch.float16)
img_attention_mask = inputs['image_attention_mask'].to(
args.device, torch.bool)
img_embeds = img_embeds.flatten(0, 1) # (2, 3, 448, 448)
img_attention_mask = img_attention_mask.flatten(0, 1) # (2, 32, 32)
aud_embeds = inputs['input_audio_embeds'].to(args.device,
torch.float16) # (1, 4000, 80)
aud_len, aud_dim = aud_embeds.shape[1:]
aud_embeds = torch.cat(
[aud_embeds,
aud_embeds.new_zeros(1, 4000 - aud_len, aud_dim)], dim=1)
aud_attention_mask = torch.ones(1, aud_embeds.shape[1]).to(
args.device, torch.bool)
aud_attention_mask[0, aud_len:] = 0
class Phi4VisionWrapper(torch.nn.Module):
def __init__(self, vision_model):
super().__init__()
self.vision_model = vision_model
@torch.no_grad
def forward(self, img_embeds, attention_mask):
features = self.vision_model.get_img_features(
img_embeds, attention_mask)
return self.vision_model.img_projection(features)
class Phi4AudioWrapper(torch.nn.Module):
def __init__(self, audio_model):
super().__init__()
self.audio_model = audio_model
@torch.no_grad
def forward(self, aud_embeds, attention_mask):
features, _ = self.audio_model.encoder(aud_embeds, attention_mask)
speech_out = self.audio_model.audio_projection['speech'](features)
vision_out = self.audio_model.audio_projection['vision'](features)
return torch.cat((speech_out, vision_out), dim=-1)
model = AutoModelForCausalLM.from_pretrained(args.model_path,
dtype='auto',
trust_remote_code=True)
vision_model = model.model.embed_tokens_extend.image_embed
vision_model = vision_model.to(args.device, torch.float16)
vision_model.eval()
vision_wrapper = Phi4VisionWrapper(vision_model)
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
part_name = 'vision'
onnx_dir = f"{args.output_dir}/{part_name}/onnx"
export_onnx(vision_wrapper,
input=(img_embeds, img_attention_mask),
onnx_dir=onnx_dir,
input_names=['input', 'attention_mask'],
dynamic_axes={
'input': {
0: "batch"
},
'attention_mask': {
0: "batch"
}
})
build_trt_engine(
args.model_type,
input_sizes=[[list(img_embeds.shape[1:]) for _ in range(3)],
[list(img_attention_mask.shape[1:]) for _ in range(3)]],
onnx_dir=onnx_dir,
engine_dir=f"{args.output_dir}/{part_name}",
max_batch_size=args.max_batch_size,
engine_name=f"visual_encoder.engine",
dtype=torch.float16)
audio_model = model.model.embed_tokens_extend.audio_embed
audio_model = audio_model.to(args.device, torch.float16)
audio_model.eval()
audio_wrapper = Phi4AudioWrapper(audio_model)
part_name = 'audio'
onnx_dir = f"{args.output_dir}/{part_name}/onnx"
export_onnx(audio_wrapper,
input=(aud_embeds, aud_attention_mask),
onnx_dir=onnx_dir,
input_names=['input', 'attention_mask'],
dynamic_axes={
'input': {
0: "batch"
},
'attention_mask': {
0: 'batch'
}
})
build_trt_engine(
args.model_type,
input_sizes=[[list(aud_embeds.shape[1:]) for _ in range(3)],
[list(aud_attention_mask.shape[1:]) for _ in range(3)]],
onnx_dir=onnx_dir,
engine_dir=f"{args.output_dir}/{part_name}",
max_batch_size=args.max_batch_size,
engine_name=f"audio_encoder.engine",
dtype=torch.float16)
def build_mllama_engine(args):
class MLLaMAVisionWrapper(torch.nn.Module):
def __init__(self, vision_model, output_proj):
super().__init__()
self.vision_model = vision_model
self.output_proj = output_proj
def forward(self, pixel_values, aspect_ratio_ids, aspect_ratio_mask):
out = self.vision_model(pixel_values, aspect_ratio_ids,
aspect_ratio_mask).last_hidden_state
out = self.output_proj(out)
return out
processor = AutoProcessor.from_pretrained(args.model_path)
# MllamaForConditionalGeneration requires transformers >= 4.45, which is
# conflict with limitation of other multimodal models.
from transformers import MllamaForConditionalGeneration
model = MllamaForConditionalGeneration.from_pretrained(args.model_path,
dtype='auto',
device_map='auto')
# Check if the model structure is updated to transformers >= 4.52.0
if hasattr(model, 'model') and hasattr(model.model, 'vision_model'):
vision_model = model.model.vision_model
multi_modal_projector = model.model.multi_modal_projector
else:
# transformers < 4.52.0
vision_model = model.vision_model
multi_modal_projector = model.multi_modal_projector
wrapper = MLLaMAVisionWrapper(vision_model, multi_modal_projector)
model_dtype = model.dtype
image = Image.new('RGB', [2048, 2688]) # dummy image
inputs = processor(images=image,
return_tensors="pt").to(model_dtype).to(model.device)
# inputs["pixel_values"]: torch.Size([1, 1, 4, 3, 448, 448])
# inputs["aspect_ratio_ids"]: torch.Size([1, 1])
# inputs["aspect_ratio_mask"]: torch.Size([1, 1, 4])
export_onnx(
wrapper,
input=tuple([value for key, value in inputs.items()]),
onnx_dir=f'{args.output_dir}/onnx',
input_names=[key for key in inputs],
output_names=['encoder_output'],
dynamic_axes={key: {
0: "batch"
}
for key in inputs},
)
build_trt_engine(
args.model_type,
[[list(inputs[key].shape[1:]) for _ in range(3)] for key in inputs],
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
model_dtype,
)
def build_internvl_engine(args):
raw_image = Image.new('RGB', [10, 10]) # Dummy image
if 'InternVL2-26B' in args.model_path:
image_processor = AutoProcessor.from_pretrained(
'OpenGVLab/InternViT-6B-448px-V1-5')
else:
image_processor = CLIPImageProcessor.from_pretrained(
'OpenGVLab/InternViT-300M-448px')
image = image_processor(images=raw_image, return_tensors='pt').pixel_values
image = image.to(args.device, torch.float16)
class InternvlVisionWrapper(torch.nn.Module):
def __init__(self, model, downsample_ratio=0.5, layer_idx=-1):
super().__init__()
self.vision_model = model.vision_model
self.mlp1 = model.mlp1
self.downsample_ratio = downsample_ratio
self.layer_idx = layer_idx
def pixel_shuffle(self, x, scale_factor=0.5):
n, w, h, c = x.size()
# N, W, H, C --> N, W, H * scale, C // scale
x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
# N, W, H * scale, C // scale --> N, H * scale, W, C // scale
x = x.permute(0, 2, 1, 3).contiguous()
# N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
x = x.view(n, int(h * scale_factor), int(w * scale_factor),
int(c / (scale_factor * scale_factor)))
x = x.permute(0, 2, 1, 3).contiguous()
return x
def forward(self, image):
immde_res = self.vision_model(image, output_hidden_states=True)
vit_embeds = immde_res.hidden_states[self.layer_idx]
vit_embeds = vit_embeds[:, 1:, :]
h = w = int(vit_embeds.shape[1]**0.5)
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
vit_embeds_px = self.pixel_shuffle(
vit_embeds, scale_factor=self.downsample_ratio)
vit_embeds_px = vit_embeds_px.reshape(vit_embeds_px.shape[0], -1,
vit_embeds_px.shape[-1])
vit_embeds_mlp = self.mlp1(vit_embeds_px)
return vit_embeds_mlp
model = AutoModelForCausalLM.from_pretrained(args.model_path,
dtype=torch.float16,
trust_remote_code=True,
use_flash_attn=False).to(
args.device)
max_num_crops = model.config.max_dynamic_patch
wrapper = InternvlVisionWrapper(model, model.config.downsample_ratio,
model.config.select_layer)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(args.model_type,
[image.shape[1], image.shape[2], image.shape[3]],
f'{args.output_dir}/onnx', args.output_dir,
args.max_batch_size * max_num_crops)
def compute_rotary_pos_emb(grid_thw, hf_config, VisionRotaryEmbedding):
head_dim = hf_config.vision_config.embed_dim // hf_config.vision_config.num_heads
rotary_pos_emb_func = VisionRotaryEmbedding(head_dim // 2)
hf_config.vision_config.spatial_merge_size
def rot_pos_emb(grid_thw, rotary_pos_emb_func):
pos_ids = []
for t, h, w in grid_thw:
hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
hpos_ids = hpos_ids.reshape(
h // hf_config.vision_config.spatial_merge_size,
hf_config.vision_config.spatial_merge_size,
w // hf_config.vision_config.spatial_merge_size,
hf_config.vision_config.spatial_merge_size,
)
hpos_ids = hpos_ids.permute(0, 2, 1, 3)
hpos_ids = hpos_ids.flatten()
wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
wpos_ids = wpos_ids.reshape(
h // hf_config.vision_config.spatial_merge_size,
hf_config.vision_config.spatial_merge_size,
w // hf_config.vision_config.spatial_merge_size,
hf_config.vision_config.spatial_merge_size,
)
wpos_ids = wpos_ids.permute(0, 2, 1, 3)
wpos_ids = wpos_ids.flatten()
pos_ids.append(
torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
pos_ids = torch.cat(pos_ids, dim=0)
max_grid_size = grid_thw[:, 1:].max()
rotary_pos_emb_full = rotary_pos_emb_func(max_grid_size)
rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
return rotary_pos_emb
rotary_pos_emb = rot_pos_emb(grid_thw, rotary_pos_emb_func)
return rotary_pos_emb
def build_qwen2_vl_engine(args):
import transformers
from qwen_vl_utils import process_vision_info
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from transformers.models.qwen2_vl.configuration_qwen2_vl import \
Qwen2VLVisionConfig
from transformers.models.qwen2_vl.modeling_qwen2_vl import (
Qwen2VisionTransformerPretrainedModel, Qwen2VLVisionBlock,
VisionAttention, VisionRotaryEmbedding)
model = Qwen2VLForConditionalGeneration.from_pretrained(
args.model_path,
dtype=torch.float32,
device_map="cpu",
attn_implementation="eager")
hf_config = AutoConfig.from_pretrained(args.model_path)
qwen2_vl_dim = hf_config.vision_config.in_chans * hf_config.vision_config.patch_size * hf_config.vision_config.patch_size * hf_config.vision_config.temporal_patch_size
processor = AutoProcessor.from_pretrained(args.model_path)
messages = [{
"role":
"user",
"content": [
{
"type":
"image",
"image":
"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
},
{
"type": "text",
"text": "Describe this picture?"
},
],
}]
text = processor.apply_chat_template(messages,
tokenize=False,
add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
for i in range(len(image_inputs)):
image_inputs[i] = image_inputs[i].resize(
(image_inputs[i].size[0] // 2, image_inputs[i].size[1] // 2))
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs
image = inputs['pixel_values'].to(torch.float16)
image_grid_thw = inputs['image_grid_thw']
cu_seqlens = torch.repeat_interleave(
image_grid_thw[:, 1] * image_grid_thw[:, 2],
image_grid_thw[:, 0]).cumsum(dim=0, dtype=torch.int32)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
seq_length = image.shape[0]
attention_mask = torch.full([1, seq_length, seq_length],
torch.finfo(image.dtype).min,
device=image.device,
dtype=image.dtype)
for i in range(1, len(cu_seqlens)):
attention_mask[..., cu_seqlens[i - 1]:cu_seqlens[i],
cu_seqlens[i - 1]:cu_seqlens[i]] = 0
rotary_pos_emb = compute_rotary_pos_emb(image_grid_thw, hf_config,
VisionRotaryEmbedding)
class VisionAttentionOpt(VisionAttention):
def __init__(self, config: Qwen2VLVisionConfig):
# Fallback for compatibility with older transformers versions (for certain nvbugs/tests)
if transformers.__version__ >= '4.53.0':
super().__init__(config)
self.head_dim = config.embed_dim // config.num_heads
else:
num_heads = config.num_heads
dim = config.embed_dim
super().__init__(dim, num_heads)
self.head_dim = dim // num_heads
def forward(self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
rotary_pos_emb: torch.Tensor = None) -> torch.Tensor:
seq_length = hidden_states.shape[0]
q, k, v = self.qkv(hidden_states).reshape(seq_length, 3,
self.num_heads,
-1).permute(1, 0, 2,
3).unbind(0)
# Copied from transformers.models.llama.modeling_qwen2_vl in v4.48
def rotate_half(x):
x1 = x[..., :x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2:]
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb_vision(
tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
orig_dtype = tensor.dtype
tensor = tensor.float()
cos = freqs.cos()
sin = freqs.sin()
cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
output = (tensor * cos) + (rotate_half(tensor) * sin)
output = output.to(orig_dtype)
return output
q = apply_rotary_pos_emb_vision(q.unsqueeze(0),
rotary_pos_emb).squeeze(0)
k = apply_rotary_pos_emb_vision(k.unsqueeze(0),
rotary_pos_emb).squeeze(0)
q = q.transpose(0, 1)
k = k.transpose(0, 1)
v = v.transpose(0, 1)
attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(
self.head_dim)
attn_weights = attn_weights + attention_mask
attn_weights = nn.functional.softmax(attn_weights,
dim=-1,
dtype=torch.float32).to(
q.dtype)
attn_output = torch.matmul(attn_weights, v)
attn_output = attn_output.transpose(0, 1)
attn_output = attn_output.reshape(seq_length, -1)
attn_output = self.proj(attn_output)
return attn_output
class Qwen2VLVisionBlockOpt(Qwen2VLVisionBlock):
def __init__(self, config, attn_implementation: str = "eager") -> None:
super().__init__(config)
self.attn = VisionAttentionOpt(config)
def forward(self, hidden_states, attention_mask,
rotary_pos_emb) -> torch.Tensor:
hidden_states = hidden_states + self.attn(
self.norm1(hidden_states),
attention_mask=attention_mask,
rotary_pos_emb=rotary_pos_emb)
hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
return hidden_states
class Qwen2VisionTransformerPretrainedModelOpt(
Qwen2VisionTransformerPretrainedModel):
def __init__(self, config) -> None:
super().__init__(config)
self.blocks = nn.ModuleList([
Qwen2VLVisionBlockOpt(config, config._attn_implementation)
for _ in range(config.depth)
])
def forward(self, hidden_states: torch.Tensor,
rotary_pos_emb: torch.Tensor,
attention_mask: torch.Tensor) -> torch.Tensor:
hidden_states = self.patch_embed(hidden_states)
for blk in self.blocks:
hidden_states = blk(hidden_states,
attention_mask=attention_mask,
rotary_pos_emb=rotary_pos_emb)
res = self.merger(hidden_states)
return res
class VisionEncoderWrapper(torch.nn.Module):
def __init__(self, model):
super().__init__()
self.visual = Qwen2VisionTransformerPretrainedModelOpt._from_config(
model.config.vision_config,
dtype=torch.float32,
)
self.visual.load_state_dict(model.visual.state_dict())
def forward(self, images, rotary_pos_emb, attention_mask):
img_features = self.visual(images, rotary_pos_emb, attention_mask)
return img_features
wrapper = VisionEncoderWrapper(model)
dynamic_axes = {
'input': {
0: 'hw'
},
'rotary_pos_emb': {
0: 'hw'
},
'attention_mask': {
1: 'hw',
2: 'hw'
}
}
export_onnx(wrapper, (image, rotary_pos_emb, attention_mask),
f'{args.output_dir}/onnx',
input_names=['input', 'rotary_pos_emb', 'attention_mask'],
output_names=['encoder_output'],
dynamic_axes=dynamic_axes)
rotary_pos_emb_dim = hf_config.vision_config.embed_dim // hf_config.vision_config.num_heads // 2
build_trt_engine(args.model_type, [rotary_pos_emb_dim],
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
model_params={
'qwen2_vl_dim': qwen2_vl_dim,
'min_hw_dims': args.min_hw_dims,
'max_hw_dims': args.max_hw_dims
})
def build_qwen2_audio_engine(args):
from transformers import Qwen2AudioForConditionalGeneration
model = Qwen2AudioForConditionalGeneration.from_pretrained(
args.model_path, dtype=torch.float16)
# dummy audio features, dtype is float32
audio = torch.randn(1,
args.num_mul_bins,
args.max_mel_seq_len,
device=args.device)
max_seq_len = (args.max_mel_seq_len - 2) // 2 + 1
mask = torch.zeros((audio.size(0), 1, max_seq_len, max_seq_len),
device=args.device,
dtype=torch.float16)
class AudioEncoderWrapper(torch.nn.Module):
def __init__(self, audio_tower, multi_modal_projector):
super(AudioEncoderWrapper, self).__init__()
self.audio_tower = audio_tower
self.multi_modal_projector = multi_modal_projector
def forward(self, x, mask):
audio_outputs = self.audio_tower(x, attention_mask=mask)
selected_audio_feature = audio_outputs.last_hidden_state
audio_features = self.multi_modal_projector(selected_audio_feature)
return audio_features
wrapper = AudioEncoderWrapper(model.audio_tower,
model.multi_modal_projector)
wrapper.eval().to(args.device)
del model # To save memory
dynamic_axes = {
"input": {
0: "batch"
},
"mask": {
0: "batch"
},
"output": {
0: "batch"
},
}
export_onnx(wrapper, (audio, mask),
f'{args.output_dir}/onnx',
input_names=["input", "mask"],
output_names=["output"],
dynamic_axes=dynamic_axes)
build_trt_engine(args.model_type, [],
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
model_params={
'num_mul_bins': args.num_mul_bins,
'max_mel_seq_len': args.max_mel_seq_len
})
def build_pixtral_engine(args):
processor = AutoProcessor.from_pretrained(args.model_path)
hf_config = AutoConfig.from_pretrained(args.model_path)
vision_config = hf_config.vision_config
raw_image = Image.new(
'RGB',
[vision_config.image_size, vision_config.image_size]) # dummy image
inputs = processor(text="dummy", images=[raw_image], return_tensors="pt")
pixel_values = inputs["pixel_values"].to(args.device, torch.bfloat16)
attention_mask = torch.zeros(
1, vision_config.image_size // vision_config.patch_size,
vision_config.image_size // vision_config.patch_size).to(
args.device, torch.bfloat16)
# isort: off
from transformers.models.pixtral.modeling_pixtral import \
apply_rotary_pos_emb
from transformers import Mistral3ForConditionalGeneration
from transformers.models.pixtral.modeling_pixtral import (PixtralAttention,
PixtralVisionModel
)
from transformers.models.mistral3.modeling_mistral3 import (
Mistral3MultiModalProjector, Mistral3PatchMerger)
# isort: on
@torch.no_grad
def attn_forward(self,
hidden_states,
attention_mask,
position_embeddings,
output_attentions=False):
batch, patches, _ = hidden_states.size()
q = self.q_proj(hidden_states)
k = self.k_proj(hidden_states)
v = self.v_proj(hidden_states)
q = q.view(batch, patches, self.num_heads,
self.head_dim).transpose(1, 2)
k = k.view(batch, patches, self.num_heads,
self.head_dim).transpose(1, 2)
v = v.view(batch, patches, self.num_heads,
self.head_dim).transpose(1, 2)
cos, sin = position_embeddings
q, k = apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=0)
# attention_mask is of shape [batch, patches].
mask = attention_mask[:, None, None, :]
attn_output = torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=mask).transpose(1, 2).contiguous()
attn_output = attn_output.reshape(batch, patches, -1)
attn_output = self.o_proj(attn_output)
return attn_output, None
@torch.no_grad
def vision_tower_forward(self, pixel_values, attention_mask):
patch_embeds = self.patch_conv(pixel_values) # (bs, c, h, w)
patch_embeds = patch_embeds.flatten(2).transpose(1, 2) # (bs, h*w, c)
attention_mask = attention_mask.flatten(1) # (bs, h*w)
patch_embeds = self.ln_pre(patch_embeds)
position_ids = self.position_ids.flatten() # (h*w, )
position_embeddings = self.patch_positional_embedding(
patch_embeds, position_ids)
out = self.transformer(patch_embeds,
attention_mask=attention_mask,
position_embeddings=position_embeddings,
output_hidden_states=False,
output_attentions=False,
return_dict=False)[0]
return out
@torch.no_grad
def patch_merger_forward(self, image_features, attention_mask):
h, w = attention_mask.shape[-2:]
bs, n, d = image_features.shape
image_grid = image_features.view(bs, h, w, d).permute(0, 3, 1, 2)
image_features = torch.nn.functional.unfold(image_grid, 2,
stride=2).transpose(1, 2)
image_features = self.merging_layer(image_features)
return image_features
@torch.no_grad
def mm_projector_forward(self, image_features, attention_mask):
image_features = self.norm(image_features)
image_features = self.patch_merger(image_features, attention_mask)
hidden_states = self.linear_2(self.act(self.linear_1(image_features)))
return hidden_states
class PixtralVisionWrapper(torch.nn.Module):
def __init__(self, vision_tower, mm_projector):
super().__init__()
self.vision_tower = vision_tower
self.mm_projector = mm_projector
@torch.no_grad
def forward(self, pixel_values, attention_mask):
features = self.vision_tower(pixel_values, attention_mask)
out = self.mm_projector(features, attention_mask)
return out
model = Mistral3ForConditionalGeneration.from_pretrained(args.model_path,
dtype="auto")
vision_tower = model.vision_tower
mm_projector = model.multi_modal_projector
height = width = vision_config.image_size // vision_config.patch_size
mesh = torch.meshgrid(torch.arange(height),
torch.arange(width),
indexing="ij")
h_grid, v_grid = torch.stack(mesh, dim=-1).chunk(2, -1)
ids = h_grid[..., 0] * width + v_grid[..., 0]
vision_tower.register_buffer("position_ids", ids)
PixtralAttention.forward = attn_forward
PixtralVisionModel.forward = vision_tower_forward
Mistral3PatchMerger.forward = patch_merger_forward
Mistral3MultiModalProjector.forward = mm_projector_forward
vision_tower = vision_tower.to(args.device, torch.bfloat16)
mm_projector = mm_projector.to(args.device, torch.bfloat16)
vision_tower.eval()
mm_projector.eval()
wrapper = PixtralVisionWrapper(vision_tower, mm_projector)
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
part_name = 'vision'
onnx_dir = f"{args.output_dir}/{part_name}/onnx"
export_onnx(wrapper,
input=(pixel_values, attention_mask),
onnx_dir=onnx_dir,
input_names=['input', 'attention_mask'],
dynamic_axes={
'input': {
0: "batch"
},
'attention_mask': {
0: "batch"
}
})
build_trt_engine(
args.model_type,
input_sizes=[[list(pixel_values.shape[1:]) for _ in range(3)],
[list(attention_mask.shape[1:]) for _ in range(3)]],
onnx_dir=onnx_dir,
engine_dir=args.output_dir,
max_batch_size=args.max_batch_size,
engine_name=f"model.engine",
dtype=torch.bfloat16)
def build_eclair_engine(args):
class RadioWithNeck(torch.nn.Module):
def __init__(self):
super().__init__()
try:
self.model_encoder = torch.hub.load("NVlabs/RADIO",
"radio_model",
version="radio_v2.5-h")
except Exception as e:
raise RuntimeError(
f"Failed to load RADIO model from torch.hub: {e}")
self.model_encoder.summary_idxs = torch.tensor(4)
self.conv1 = torch.nn.Conv1d(1280, 1024, 1)
self.layer_norm1 = torch.nn.LayerNorm(1024,
eps=1e-6,
elementwise_affine=True)
self.conv2 = torch.nn.Conv2d(1024,
1024,
kernel_size=(1, 4),
stride=(1, 4),
padding=0,
bias=False)
self.layer_norm2 = torch.nn.LayerNorm(1024,
eps=1e-6,
elementwise_affine=True)
@torch.no_grad
def forward(self, pixel_values):
_, feature = self.model_encoder(pixel_values)
output = self.conv1(feature.permute(0, 2, 1)).permute(0, 2, 1)
output = self.layer_norm1(output).permute(0, 2, 1)
b, d, _ = output.shape
h = pixel_values.shape[-2] // 16
w = pixel_values.shape[-1] // 16
output = self.conv2(output.reshape(b, d, h, w))
output = output.flatten(-2, -1).permute(0, 2, 1)
output = self.layer_norm2(output)
return output
processor = NougatProcessor.from_pretrained(args.model_path)
model = VisionEncoderDecoderModel.from_pretrained("facebook/nougat-base")
model.encoder = RadioWithNeck()
model.decoder.resize_token_embeddings(len(processor.tokenizer))
model.config.decoder_start_token_id = processor.tokenizer.eos_token_id # 2
model.config.pad_token_id = processor.tokenizer.pad_token_id # 1
checkpoint_path = os.path.join(args.model_path, "model.safetensors")
if not os.path.exists(checkpoint_path):
raise FileNotFoundError(
f"Model checkpoint not found at {checkpoint_path}")
load_model(model, checkpoint_path)
wrapper = model.encoder.to(args.device)
# temporary fix due to TRT onnx export bug
for block in wrapper.model_encoder.model.blocks:
block.attn.fused_attn = False
image = torch.randn((1, 3, 2048, 1648),
device=args.device,
dtype=torch.bfloat16)
export_onnx(wrapper, image, f'{args.output_dir}/onnx')
build_trt_engine(
args.model_type,
[image.shape[1], image.shape[2], image.shape[3]], # [3, H, W]
f'{args.output_dir}/onnx',
args.output_dir,
args.max_batch_size,
dtype=torch.bfloat16,
engine_name='visual_encoder.engine')