TensorRT-LLMs/tests/integration/defs/examples/test_mpt.py
Kaiyu Xie 2631f21089
Update (#2978)
Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>
2025-03-23 16:39:35 +08:00

416 lines
17 KiB
Python

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Module test_mpt test mpt examples."""
import os
import pytest
from defs.common import (convert_weights, generate_summary_cmd, venv_check_call,
venv_mpi_check_call)
from defs.conftest import skip_pre_ada
from defs.trt_test_alternative import check_call
@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize("data_type", ['bfloat16'])
@pytest.mark.parametrize("use_plugins", [True, False],
ids=['enable_plugins', 'disable_plugins'])
@pytest.mark.parametrize(
"context_fmha_type",
['enable_context_fmha', 'enable_context_fmha_fp32_acc', 'disable_fmha'])
def test_llm_mpt_7b_1node_4gpus(mpt_example_root, llm_venv,
llm_mpt_7b_model_root, llm_datasets_root,
llm_rouge_root, cmodel_dir, engine_dir,
data_type, use_plugins, context_fmha_type):
"mpt 7b test on 4gpus"
print("Converting MPT weights...")
model_name = os.path.basename(llm_mpt_7b_model_root)
ckpt_dir = convert_weights(llm_venv=llm_venv,
example_root=mpt_example_root,
cmodel_dir=cmodel_dir,
model=model_name,
model_path=llm_mpt_7b_model_root,
data_type=data_type,
gpus=4)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={4}",
f"--max_input_len={2048}",
f"--max_seq_len={2560}",
f"--workers={4}",
]
if use_plugins:
if context_fmha_type == "enable_fmha":
build_cmd.append("--context_fmha=enable")
elif context_fmha_type == "disable_fmha":
build_cmd.append("--context_fmha=disable")
build_cmd.extend([
f"--gpt_attention_plugin={data_type}", f"--gemm_plugin={data_type}"
])
else:
build_cmd.extend([
"--gpt_attention_plugin=disable",
"--gemm_plugin=disable",
"--context_fmha=disable",
"--paged_kv_cache=disable",
"--remove_input_padding=disable",
])
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running inference...")
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_7b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=18,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
if context_fmha_type == "enable_context_fmha_fp32_acc":
summary_cmd.append("--enable_context_fmha_fp32_acc")
venv_mpi_check_call(llm_venv, ["mpirun", "-n", "4", "--allow-run-as-root"],
summary_cmd)
@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize("data_type", ['bfloat16'])
@pytest.mark.parametrize("use_plugins", [True, False],
ids=['enable_plugins', 'disable_plugins'])
@pytest.mark.parametrize(
"context_fmha_type",
['enable_context_fmha', 'enable_context_fmha_fp32_acc', 'disable_fmha'])
def test_llm_mpt_30b_1node_4gpus(mpt_example_root, llm_venv,
llm_mpt_30b_model_root, llm_datasets_root,
llm_rouge_root, cmodel_dir, engine_dir,
data_type, use_plugins, context_fmha_type):
"mpt 30b test on 4gpus"
print("Converting MPT weights...")
model_name = os.path.basename(llm_mpt_30b_model_root)
ckpt_dir = convert_weights(llm_venv=llm_venv,
example_root=mpt_example_root,
cmodel_dir=cmodel_dir,
model=model_name,
model_path=llm_mpt_30b_model_root,
data_type=data_type,
gpus=4)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={4}",
f"--max_input_len={1024}",
f"--max_seq_len={1124}",
f"--workers={4}",
]
if use_plugins:
if context_fmha_type == "enable_fmha":
build_cmd.append("--context_fmha=enable")
elif context_fmha_type == "disable_fmha":
build_cmd.append("--context_fmha=disable")
build_cmd.extend([
f"--gpt_attention_plugin={data_type}", f"--gemm_plugin={data_type}"
])
else:
build_cmd.extend([
"--gpt_attention_plugin=disable",
"--gemm_plugin=disable",
"--context_fmha=disable",
"--paged_kv_cache=disable",
"--remove_input_padding=disable",
])
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running inference...")
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_30b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=17,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
if context_fmha_type == "enable_context_fmha_fp32_acc":
summary_cmd.append("--enable_context_fmha_fp32_acc")
venv_mpi_check_call(llm_venv, ["mpirun", "-n", "4", "--allow-run-as-root"],
summary_cmd)
@pytest.mark.parametrize(
"context_fmha_type",
['enable_context_fmha', 'enable_context_fmha_fp32_acc', 'disable_fmha'])
def test_llm_mpt_7b_1node_1gpu(mpt_example_root, llm_venv,
llm_mpt_7b_model_root, llm_datasets_root,
llm_rouge_root, cmodel_dir, engine_dir,
context_fmha_type):
"mpt-7b test on one gpu"
ckpt_dir = convert_weights(llm_venv, mpt_example_root, cmodel_dir, "mpt-7b",
llm_mpt_7b_model_root)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={2}",
f"--max_input_len={1024}",
f"--max_beam_width={5}",
"--gemm_plugin=float16",
]
if context_fmha_type == "enable_fmha":
build_cmd.append("--context_fmha=enable")
elif context_fmha_type == "disable_fmha":
build_cmd.append("--context_fmha=disable")
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running inference...")
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_7b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=20,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
if context_fmha_type == "enable_context_fmha_fp32_acc":
summary_cmd.append("--enable_context_fmha_fp32_acc")
venv_check_call(llm_venv, summary_cmd)
# transformers compatibility issues
# ImportError: cannot import name '_expand_mask' from 'transformers.models.bloom.modeling_bloom'
def test_llm_mpt_125m_summary(mpt_example_root, llm_venv,
llm_mpt_125m_model_root, llm_datasets_root,
llm_rouge_root, cmodel_dir, engine_dir,
update_transformers):
"mpt-125m summary test"
ckpt_dir = convert_weights(llm_venv=llm_venv,
example_root=mpt_example_root,
cmodel_dir=cmodel_dir,
model="mpt-125m",
model_path=llm_mpt_125m_model_root,
data_type="float32")
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={8}",
f"--max_input_len={924}",
f"--max_seq_len={1024}",
"--gpt_attention_plugin=float32",
"--gemm_plugin=float32",
]
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running summary...")
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_125m_model_root,
engine_dir=engine_dir,
batch_size=1,
data_type="fp32",
tensorrt_llm_rouge1_threshold=10,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
venv_check_call(llm_venv, summary_cmd)
@skip_pre_ada
def test_llm_mpt_7b_fp8_summary(mpt_example_root, llm_mpt_7b_model_root,
llm_datasets_root, llm_rouge_root, llm_venv,
engine_dir, qcache_dir):
"Build & Run mpt 7b with fp8."
# Quantize HF mpt 7b checkpoint into FP8 format
quantize_cmd = [
f"{mpt_example_root}/../quantization/quantize.py",
f"--model_dir={llm_mpt_7b_model_root}",
f"--calib_dataset={llm_datasets_root}/cnn_dailymail",
"--dtype=float16",
"--qformat=fp8",
"--kv_cache_dtype=fp8",
f"--output_dir={qcache_dir}/quantized_fp8",
]
venv_check_call(llm_venv, quantize_cmd)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={qcache_dir}/quantized_fp8/",
f"--output_dir={engine_dir}",
f"--max_input_len={1024}",
"--gpt_attention_plugin=float16",
"--gemm_plugin=float16",
"--remove_input_padding=enable",
]
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print('Run mpt-7b fp8...')
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_7b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=20,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
venv_check_call(llm_venv, summary_cmd)
def test_llm_mpt_7b_awq_int4_summary(mpt_example_root, llm_mpt_7b_model_root,
llm_datasets_root, llm_rouge_root,
llm_venv, engine_dir, qcache_dir):
"Build & Run mpt 7b with awq int4 gpus"
# Quantize HF mpt-7b checkpoint into int4 format
quantize_cmd = [
f"{mpt_example_root}/../quantization/quantize.py",
f"--model_dir={llm_mpt_7b_model_root}",
f"--calib_dataset={llm_datasets_root}/cnn_dailymail",
"--dtype=float16",
"--qformat=int4_awq",
"--calib_size=32",
f"--output_dir={qcache_dir}/quantized_int4",
]
venv_check_call(llm_venv, quantize_cmd)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={qcache_dir}/quantized_int4/",
f"--output_dir={engine_dir}",
f"--max_batch_size={64}",
f"--max_input_len={1024}",
"--gemm_plugin=float16",
"--gpt_attention_plugin=float16",
]
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print('Run mpt-7b awq int4...')
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_7b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=20,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
venv_check_call(llm_venv, summary_cmd)
@pytest.mark.parametrize("data_type", ['int8', 'int4'])
def test_llm_mpt_7b_weight_only(mpt_example_root, llm_venv,
llm_mpt_7b_model_root, llm_datasets_root,
llm_rouge_root, cmodel_dir, engine_dir,
data_type):
"mpt-7b test with weight only"
ckpt_dir = convert_weights(llm_venv=llm_venv,
example_root=mpt_example_root,
cmodel_dir=cmodel_dir,
model="mpt-7b",
model_path=llm_mpt_7b_model_root,
weight_only_precision=data_type)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={64}",
f"--max_input_len={1024}",
"--gemm_plugin=float16",
"--gpt_attention_plugin=float16",
]
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running inference...")
# For weight-only int4, mpt-7b has bad accuracy while mpt-125m and
# mpt-30b has comparable accuracy with FP16.
summary_cmd = generate_summary_cmd(mpt_example_root,
hf_model_dir=llm_mpt_7b_model_root,
engine_dir=engine_dir,
data_type="fp16",
tensorrt_llm_rouge1_threshold=20,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
venv_check_call(llm_venv, summary_cmd)
# transformers compatibility issues
# ImportError: cannot import name '_expand_mask' from 'transformers.models.bloom.modeling_bloom'
@pytest.mark.skip_less_device(2)
@pytest.mark.parametrize("num_beams", [1, 4],
ids=lambda num_beams: f'nb:{num_beams}')
def test_llm_replit_code_v1_5_3b_1node_2gpus(mpt_example_root, llm_venv,
llm_replit_code_v1_5_3b_model_root,
llm_datasets_root, llm_rouge_root,
cmodel_dir, engine_dir, num_beams,
update_transformers):
"replit code v1_5 3b test with 2gpus"
ckpt_dir = convert_weights(llm_venv=llm_venv,
example_root=mpt_example_root,
cmodel_dir=cmodel_dir,
model="mpt_replit_code",
model_path=llm_replit_code_v1_5_3b_model_root,
data_type="bfloat16",
gpus=2)
print("Building engines...")
build_cmd = [
"trtllm-build",
f"--checkpoint_dir={ckpt_dir}",
f"--output_dir={engine_dir}",
f"--max_batch_size={16}",
f"--max_input_len={1024}",
f"--max_beam_width={num_beams}",
"--gemm_plugin=bfloat16",
"--gpt_attention_plugin=bfloat16",
"--workers=2",
]
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
print("Running inference...")
summary_cmd = generate_summary_cmd(
mpt_example_root,
hf_model_dir=llm_replit_code_v1_5_3b_model_root,
engine_dir=engine_dir,
data_type="fp16",
num_beams=num_beams,
tensorrt_llm_rouge1_threshold=10,
dataset_dir=llm_datasets_root,
rouge_dir=llm_rouge_root)
venv_mpi_check_call(llm_venv, ["mpirun", "-n", "2", "--allow-run-as-root"],
summary_cmd)