使用 dynamo 后端编译 GPT2

此脚本说明了在流行的 GPT2 模型上使用 dynamo 后端的 Torch-TensorRT 工作流程。

导入和模型定义

import torch
import torch_tensorrt
from transformers import AutoModelForCausalLM, AutoTokenizer
from utils import export_llm, generate

# Define the parameters and initialize the model
MAX_TOKENS = 32
DEVICE = torch.device("cuda:0")

# Define the GPT2 model from hugging face
# kv_cache is not supported in Torch-TRT currently.
# CPU is used here so that GPU memory is reserved for TRT compilation.
with torch.no_grad():
    tokenizer = AutoTokenizer.from_pretrained("gpt2")
    model = (
        AutoModelForCausalLM.from_pretrained(
            "gpt2",
            pad_token_id=tokenizer.eos_token_id,
            use_cache=False,
            attn_implementation="eager",
        )
        .eval()
        .half()
    )

对示例输入提示进行标记并获取 pytorch 模型输出

prompt = "I enjoy walking with my cute dog"
model_inputs = tokenizer(prompt, return_tensors="pt")
input_ids = model_inputs["input_ids"]

# Auto-regressive generation loop for greedy decoding using PyTorch model
# We use a custom generate function which is very similar to the huggingface one.
pyt_gen_tokens = generate(model, input_ids, MAX_TOKENS, tokenizer.eos_token_id)

使用 dynamo 后端使用 Torch-TensorRT 进行编译并生成 TensorRT 输出

# Export the GPT2 model into an ExportedProgram which is input of TRT compilation
# To compile the model in FP16, we do the following
# 1) Cast the model to FP16 via model.half()
# 2) Enable use_explicit_typing=True. Certain layers are explicitly casted to FP32 within the pytorch model and this flag respects this behavior during TRT compilation
# 3) Enable use_fp32_acc=True. This ensures all the matmuls are accumulated in FP32 precision (similar to PyTorch)
gpt2_ep = export_llm(model, input_ids, max_seq_len=1024)
trt_model = torch_tensorrt.dynamo.compile(
    gpt2_ep,
    inputs=[input_ids],
    enabled_precisions={torch.float32},
    truncate_double=True,
    device=DEVICE,
    disable_tf32=True,
    use_explicit_typing=True,
    use_fp32_acc=True,
)

# Auto-regressive generation loop for greedy decoding using TensorRT model
# We use a custom generate function which is very similar to the huggingface one.
# Move inputs to GPU
input_ids = input_ids.to(DEVICE)
trt_gen_tokens = generate(trt_model, input_ids, MAX_TOKENS, tokenizer.eos_token_id)

解码 PyTorch 和 TensorRT 的输出句子

print("=============================")
print(
    "Pytorch model generated text: ",
    tokenizer.decode(pyt_gen_tokens[0], skip_special_tokens=True),
)
print("=============================")
print(
    "TensorRT model generated text: ",
    tokenizer.decode(trt_gen_tokens[0], skip_special_tokens=True),
)

# Prompt : What is parallel programming ?

# =============================
# Pytorch model generated text: The parallel programming paradigm is a set of programming languages that are designed to be used in parallel. The main difference between parallel programming and parallel programming is that

# =============================
# TensorRT model generated text: The parallel programming paradigm is a set of programming languages that are designed to be used in parallel. The main difference between parallel programming and parallel programming is that