TensorRT-LLMs/tests/attention/test_gpt_attention_no_cache.py
石晓伟 8f91cff22e
TensorRT-LLM Release 0.15.0 (#2529)
Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>
2024-12-04 13:44:56 +08:00

309 lines
13 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.
import unittest
from collections import OrderedDict
# isort: off
import torch
# isort: on
import os
import sys
from parameterized import parameterized
import tensorrt_llm
from tensorrt_llm import Tensor, str_dtype_to_trt
from tensorrt_llm._utils import str_dtype_to_torch, torch_dtype_to_trt
from tensorrt_llm.functional import gpt_attention
from tensorrt_llm.models.generation_mixin import GenerationMixin
from tensorrt_llm.models.modeling_utils import get_kv_cache_type_from_legacy
from tensorrt_llm.plugin.plugin import ContextFMHAType
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from utils.util import unittest_name_func
class TestPluginNoCache(unittest.TestCase):
def setUp(self):
tensorrt_llm.logger.set_level('info')
@staticmethod
def build_engine(qkv_shape,
max_batch_size,
max_beam_width,
max_input_len,
max_seq_len,
num_kv_heads,
head_size,
dtype,
num_layers,
remove_input_padding,
context_fmha_type,
use_cache=True):
kv_dtype = str_dtype_to_trt(dtype)
hidden_size = num_kv_heads * head_size
num_tokens = max_batch_size * max_input_len
builder = tensorrt_llm.Builder()
builder_config = builder.create_builder_config(
name="attention",
precision=dtype,
)
net = builder.create_network()
net.plugin_config.to_legacy_setting()
net.plugin_config.gpt_attention_plugin = dtype
net.plugin_config.set_context_fmha(context_fmha_type)
net.plugin_config.remove_input_padding = remove_input_padding
kv_cache_type = get_kv_cache_type_from_legacy(
use_cache, net.plugin_config.paged_kv_cache)
with tensorrt_llm.net_guard(net):
inputs = GenerationMixin().prepare_attention_inputs(
max_batch_size=max_batch_size,
max_beam_width=max_beam_width,
max_input_len=max_input_len,
max_seq_len=max_seq_len,
num_kv_heads=num_kv_heads,
head_size=head_size,
num_layers=num_layers,
kv_dtype=kv_dtype,
remove_input_padding=remove_input_padding,
use_gpt_attention_plugin=True,
enable_ctx_gen_opt_profiles=False,
kv_cache_type=kv_cache_type,
)
if remove_input_padding:
qkv = Tensor(name="qkv",
shape=(-1, hidden_size * 3),
dtype=str_dtype_to_trt(dtype),
dim_range=OrderedDict([
('tokens', [(1, num_tokens // 2, num_tokens)]),
('hidden_size', [hidden_size * 3]),
]))
else:
qkv = Tensor(name="qkv",
shape=(-1, -1, hidden_size * 3),
dtype=str_dtype_to_trt(dtype),
dim_range=OrderedDict([
('batch_size', [(1, max_batch_size // 2,
max_batch_size)]),
('tokens', [(1, max_input_len // 2,
max_input_len)]),
('hidden_size', [hidden_size * 3]),
]))
sequence_length = inputs['sequence_length']
host_context_lengths = inputs['host_context_lengths']
host_max_attention_window_sizes = inputs[
'host_max_attention_window_sizes']
host_sink_token_length = inputs['host_sink_token_length']
context_lengths = inputs['context_lengths']
host_request_types = inputs['host_request_types']
host_past_key_value_lengths = inputs['host_past_key_value_lengths']
past_key_value = inputs['past_key_value']
if past_key_value:
past_key_value = past_key_value[0]
cache_indirection = inputs['cache_indirection']
host_runtime_perf_knobs_tensor = inputs['host_runtime_perf_knobs']
host_context_progress = inputs['host_context_progress']
outputs = gpt_attention(
qkv=qkv,
past_key_value=past_key_value,
sequence_length=sequence_length,
host_past_key_value_lengths=host_past_key_value_lengths,
host_max_attention_window_sizes=host_max_attention_window_sizes,
host_sink_token_length=host_sink_token_length,
context_lengths=context_lengths,
cache_indirection=cache_indirection,
host_request_types=host_request_types,
layer_idx=0,
num_heads=num_kv_heads,
num_kv_heads=num_kv_heads,
hidden_size_per_head=head_size,
q_scaling=1.0,
rotary_embedding_dim=0,
max_context_length=max_input_len,
host_context_lengths=host_context_lengths,
use_cache=use_cache,
host_runtime_perf_knobs=host_runtime_perf_knobs_tensor,
host_context_progress=host_context_progress)
net._mark_output(outputs[0],
'output',
dtype=str_dtype_to_trt(dtype))
if use_cache:
net._mark_output(outputs[1],
'present_key_value',
dtype=str_dtype_to_trt(dtype))
return builder.build_engine(net, builder_config)
@parameterized.expand([("float16", True, ContextFMHAType.disabled),
("float16", False, ContextFMHAType.enabled)],
name_func=unittest_name_func)
def test_plugin_no_cache(self, dtype: str, remove_input_padding: bool,
fmha_type: ContextFMHAType):
max_batch_size = 8
max_beam_width = 1
max_input_len = 128
max_seq_len = max_input_len
num_kv_heads = 16
head_size = 32
num_layers = 1
hidden_size = num_kv_heads * head_size
str_dtype_to_trt(dtype)
if remove_input_padding:
qkv_shape = (max_batch_size * max_input_len, hidden_size * 3)
out_shape = (max_batch_size * max_input_len, hidden_size)
else:
qkv_shape = (max_batch_size, max_input_len, hidden_size * 3)
out_shape = (max_batch_size, max_input_len, hidden_size)
qkv = torch.randn(
qkv_shape, dtype=str_dtype_to_torch(dtype), device="cuda") * 1e-3
sequence_length = torch.full([max_batch_size],
max_input_len,
dtype=torch.int32).cuda()
host_past_key_value_lengths = torch.zeros([max_batch_size],
dtype=torch.int32).cpu()
host_max_attention_window_sizes = torch.tensor([max_input_len],
dtype=torch.int32).cpu()
host_sink_token_length = torch.tensor([0], dtype=torch.int32).cpu()
context_lengths = torch.full([max_batch_size],
max_input_len,
dtype=torch.int32).cuda()
cache_indirection = torch.zeros(
[max_batch_size, max_beam_width, max_input_len],
dtype=torch.int32,
device='cuda')
host_request_types = torch.zeros([max_batch_size],
dtype=torch.int32).cpu()
host_context_lengths = torch.full([max_batch_size],
max_input_len,
dtype=torch.int32).cpu()
present_key_value = torch.zeros(
[max_batch_size, 2, num_kv_heads, max_input_len, head_size],
dtype=str_dtype_to_torch(dtype),
device='cuda')
output = torch.zeros(out_shape,
dtype=str_dtype_to_torch(dtype),
device="cuda")
output_nocache = torch.zeros(out_shape,
dtype=str_dtype_to_torch(dtype),
device="cuda")
perf_knob_tensor_size = 16
host_runtime_perf_knobs = torch.tensor([-1] * perf_knob_tensor_size,
dtype=torch.int64,
device='cpu')
host_context_progress = torch.tensor([0],
dtype=torch.int64,
device='cpu')
engine = TestPluginNoCache.build_engine(
qkv_shape=qkv_shape,
max_batch_size=max_batch_size,
max_beam_width=max_beam_width,
max_input_len=max_input_len,
max_seq_len=max_seq_len,
num_kv_heads=num_kv_heads,
head_size=head_size,
dtype=dtype,
num_layers=num_layers,
remove_input_padding=remove_input_padding,
context_fmha_type=fmha_type,
use_cache=False)
session = tensorrt_llm.runtime.Session.from_serialized_engine(engine)
inputs = {
'qkv': qkv,
'host_max_attention_window_sizes': host_max_attention_window_sizes,
'host_sink_token_length': host_sink_token_length,
'context_lengths': context_lengths,
'host_request_types': host_request_types,
'host_runtime_perf_knobs': host_runtime_perf_knobs,
'host_context_progress': host_context_progress
}
if remove_input_padding:
inputs['host_context_lengths'] = host_context_lengths
outputs = {
'output': output_nocache,
}
inputs_info = [
tensorrt_llm.runtime.TensorInfo(name,
torch_dtype_to_trt(tensor.dtype),
tensor.shape)
for name, tensor in inputs.items()
]
session.infer_shapes(inputs_info)
stream = torch.cuda.current_stream()
session.run(inputs=inputs, outputs=outputs, stream=stream.cuda_stream)
engine = TestPluginNoCache.build_engine(
qkv_shape=qkv_shape,
max_batch_size=max_batch_size,
max_beam_width=max_beam_width,
max_input_len=max_input_len,
max_seq_len=max_seq_len,
num_kv_heads=num_kv_heads,
head_size=head_size,
dtype=dtype,
num_layers=num_layers,
remove_input_padding=remove_input_padding,
context_fmha_type=fmha_type)
session = tensorrt_llm.runtime.Session.from_serialized_engine(engine)
inputs = {
'qkv': qkv,
'sequence_length': sequence_length,
'host_past_key_value_lengths': host_past_key_value_lengths,
'host_max_attention_window_sizes': host_max_attention_window_sizes,
'host_sink_token_length': host_sink_token_length,
'context_lengths': context_lengths,
'cache_indirection': cache_indirection,
'host_request_types': host_request_types,
'past_key_value_0': present_key_value,
'host_runtime_perf_knobs': host_runtime_perf_knobs,
'host_context_progress': host_context_progress
}
if remove_input_padding:
inputs['host_context_lengths'] = host_context_lengths
outputs = {
'output': output,
'present_key_value': present_key_value,
}
inputs_info = [
tensorrt_llm.runtime.TensorInfo(name,
torch_dtype_to_trt(tensor.dtype),
tensor.shape)
for name, tensor in inputs.items()
]
session.infer_shapes(inputs_info)
stream = torch.cuda.current_stream()
session.run(inputs=inputs, outputs=outputs, stream=stream.cuda_stream)
assert torch.equal(output, output_nocache)
if __name__ == "__main__":
unittest.main()