mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
201 lines
8.3 KiB
Python
201 lines
8.3 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 os
|
|
import sys
|
|
import unittest
|
|
|
|
from parameterized import parameterized
|
|
|
|
from tensorrt_llm.layers import (Attention, ColumnLinear, GatedMLP, RmsNorm,
|
|
RowLinear)
|
|
from tensorrt_llm.mapping import Mapping
|
|
from tensorrt_llm.models import (GPTForCausalLM, LLaMAConfig, LLaMAForCausalLM,
|
|
PretrainedConfig)
|
|
from tensorrt_llm.models.modeling_utils import QuantConfig
|
|
from tensorrt_llm.quantization import QuantAlgo
|
|
|
|
# isort: off
|
|
from tensorrt_llm.quantization.layers import (
|
|
Fp8RowwiseAttention, Fp8RowwiseGatedMLP, Fp8RowwiseRmsNorm,
|
|
SmoothQuantAttention, SmoothQuantLayerNorm, SmoothQuantMLP,
|
|
WeightOnlyQuantColumnLinear, WeightOnlyQuantRowLinear)
|
|
# isort: on
|
|
from tensorrt_llm.quantization.quantize import quantize
|
|
|
|
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
|
|
from utils.util import unittest_name_func
|
|
|
|
|
|
class TestQuant(unittest.TestCase):
|
|
|
|
def test_weight_only_quant(self):
|
|
quant_algo = QuantAlgo.W8A16
|
|
config = {
|
|
'architecture': 'GPTForCausalLM',
|
|
'dtype': 'float16',
|
|
'num_hidden_layers': 2,
|
|
'num_attention_heads': 12,
|
|
'hidden_size': 768,
|
|
'vocab_size': 51200,
|
|
'max_position_embeddings': 1024,
|
|
'hidden_act': 'relu',
|
|
}
|
|
config = PretrainedConfig.from_dict(config)
|
|
model = GPTForCausalLM(config)
|
|
|
|
quant_model = quantize(model, QuantConfig(quant_algo))
|
|
|
|
self.assertTrue(hasattr(quant_model, 'quant_mode'))
|
|
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].attention.qkv,
|
|
WeightOnlyQuantColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].attention.dense,
|
|
WeightOnlyQuantRowLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].mlp.fc,
|
|
WeightOnlyQuantColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].mlp.proj,
|
|
WeightOnlyQuantRowLinear))
|
|
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[1].attention.qkv,
|
|
WeightOnlyQuantColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[1].attention.dense,
|
|
WeightOnlyQuantRowLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[1].mlp.fc,
|
|
WeightOnlyQuantColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[1].mlp.proj,
|
|
WeightOnlyQuantRowLinear))
|
|
|
|
self.assertTrue(isinstance(quant_model.lm_head, ColumnLinear))
|
|
|
|
def test_weight_only_quant_exclude_modules(self):
|
|
quant_algo = QuantAlgo.W8A16
|
|
config = {
|
|
'architecture': 'GPTForCausalLM',
|
|
'dtype': 'float16',
|
|
'num_hidden_layers': 1,
|
|
'num_attention_heads': 12,
|
|
'hidden_size': 768,
|
|
'vocab_size': 51200,
|
|
'max_position_embeddings': 1024,
|
|
'hidden_act': 'relu',
|
|
}
|
|
config = PretrainedConfig.from_dict(config)
|
|
model = GPTForCausalLM(config)
|
|
|
|
quant_model = quantize(
|
|
model,
|
|
QuantConfig(quant_algo,
|
|
exclude_modules=[
|
|
'fc', 'dense', 'vocab_embedding',
|
|
'position_embedding', 'block_embedding'
|
|
]))
|
|
|
|
self.assertTrue(hasattr(quant_model, 'quant_mode'))
|
|
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].attention.qkv,
|
|
WeightOnlyQuantColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].attention.dense,
|
|
RowLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].mlp.fc, ColumnLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.transformer.layers[0].mlp.proj,
|
|
WeightOnlyQuantRowLinear))
|
|
self.assertTrue(
|
|
isinstance(quant_model.lm_head, WeightOnlyQuantColumnLinear))
|
|
|
|
def test_convert_GPT_to_smooth_quant(self):
|
|
config = {
|
|
'architecture': 'GPTForCausalLM',
|
|
'dtype': 'float16',
|
|
'num_hidden_layers': 1,
|
|
'num_attention_heads': 1,
|
|
'hidden_size': 128,
|
|
'vocab_size': 1024,
|
|
'max_position_embeddings': 256,
|
|
'hidden_act': 'gelu',
|
|
}
|
|
config = PretrainedConfig.from_dict(config)
|
|
model = GPTForCausalLM(config)
|
|
|
|
quant_algo = QuantAlgo.W8A8_SQ_PER_TENSOR_PLUGIN
|
|
quant_config = QuantConfig(quant_algo)
|
|
quant_model = quantize(model, quant_config)
|
|
for layer in quant_model.transformer.layers:
|
|
assert isinstance(layer.input_layernorm, SmoothQuantLayerNorm)
|
|
assert isinstance(layer.post_layernorm, SmoothQuantLayerNorm)
|
|
assert isinstance(layer.mlp, SmoothQuantMLP)
|
|
assert isinstance(layer.attention, SmoothQuantAttention)
|
|
|
|
assert quant_model.quant_mode == quant_config.quant_mode
|
|
|
|
@parameterized.expand([(False, 1, 0), (True, 1, 0), (False, 2, 0),
|
|
(False, 2, 1), (True, 2, 0), (True, 2, 1)],
|
|
name_func=unittest_name_func)
|
|
def test_fp8_rowwise_quant(self, use_meta_recipe: bool, world_size: int,
|
|
rank: int):
|
|
mapping = Mapping(rank=rank, pp_size=world_size, world_size=world_size)
|
|
config = LLaMAConfig(architecture='LlamaForCausalLM',
|
|
dtype='float16',
|
|
hidden_size=128,
|
|
num_hidden_layers=16,
|
|
num_attention_heads=16,
|
|
vocab_size=1024,
|
|
hidden_act='silu',
|
|
position_embedding_type='rope_gpt_neox',
|
|
max_position_embeddings=256,
|
|
mapping=mapping)
|
|
model = LLaMAForCausalLM(config)
|
|
|
|
quant_algo = QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN
|
|
quant_config = QuantConfig(quant_algo, use_meta_recipe=use_meta_recipe)
|
|
|
|
quant_model = quantize(model, quant_config)
|
|
local_num_hidden_layers = len(quant_model.transformer.layers)
|
|
for local_layer_idx, layer in enumerate(quant_model.transformer.layers):
|
|
assert layer.layer_idx == local_layer_idx + rank * local_num_hidden_layers
|
|
if use_meta_recipe and (layer.layer_idx == 0
|
|
or layer.layer_idx == 15):
|
|
assert isinstance(layer.input_layernorm, RmsNorm)
|
|
assert isinstance(layer.attention, Attention)
|
|
assert isinstance(layer.post_layernorm, RmsNorm)
|
|
assert isinstance(layer.mlp, GatedMLP)
|
|
elif use_meta_recipe:
|
|
assert isinstance(layer.input_layernorm, RmsNorm)
|
|
assert isinstance(layer.attention, Attention)
|
|
assert isinstance(layer.post_layernorm, Fp8RowwiseRmsNorm)
|
|
assert isinstance(layer.mlp, Fp8RowwiseGatedMLP)
|
|
else:
|
|
assert isinstance(layer.input_layernorm, Fp8RowwiseRmsNorm)
|
|
assert isinstance(layer.attention, Fp8RowwiseAttention)
|
|
assert isinstance(layer.post_layernorm, Fp8RowwiseRmsNorm)
|
|
assert isinstance(layer.mlp, Fp8RowwiseGatedMLP)
|
|
|
|
assert quant_model.quant_mode == quant_config.quant_mode
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|