mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
258c7540c0
Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> open source f8c0381a2bc50ee2739c3d8c2be481b31e5f00bd (#2736) Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> Add note for blackwell (#2742) Update the docs to workaround the extra-index-url issue (#2744) update README.md (#2751) Fix github io pages (#2761) Update
218 lines
7.5 KiB
Python
218 lines
7.5 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
|
|
|
|
import torch
|
|
from parameterized import parameterized
|
|
|
|
import tensorrt_llm
|
|
|
|
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
|
|
from utils.util import skip_bf16_pre_ampere, unittest_name_func
|
|
|
|
FP8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
|
|
|
|
|
|
class TestDynamicFP8QuantDequant(unittest.TestCase):
|
|
|
|
def setUp(self):
|
|
torch.manual_seed(42)
|
|
tensorrt_llm.logger.set_level('error')
|
|
|
|
def _ref_quant(self, x_, x_scale_):
|
|
x_ = x_.float()
|
|
finfo = torch.finfo(torch.float8_e4m3fn)
|
|
inv_scale = x_scale_.float().reciprocal()
|
|
x_fp8_ = (x_ * inv_scale).clamp(min=finfo.min, max=finfo.max)
|
|
return x_fp8_.to(torch.float8_e4m3fn)
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_activation_scales(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
m = 11
|
|
n = 11
|
|
A = torch.randn((m, n), dtype=dtype).cuda()
|
|
B, s = torch.ops.tensorrt_llm.quantize_e4m3_activation(A)
|
|
s_ref = (torch.max(A.float().abs(), -1)[0].view(m, 1) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
B_ref = self._ref_quant(A, s_ref)
|
|
|
|
torch.testing.assert_close(s_ref, s)
|
|
torch.testing.assert_close(B.float(), B_ref.float())
|
|
|
|
B_s, _ = torch.ops.tensorrt_llm.static_quantize_e4m3_activation(
|
|
A, s.float())
|
|
|
|
torch.testing.assert_close(B_s.float(), B_ref.float())
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_weight_scales(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
m = 11
|
|
n = 11
|
|
A = torch.randn((m, n), dtype=dtype).cuda()
|
|
B, s = torch.ops.tensorrt_llm.quantize_e4m3_weight(A)
|
|
s_ref = (torch.max(A.float().abs(), 0)[0].view(1, n) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
B_ref = self._ref_quant(A, s_ref)
|
|
|
|
torch.testing.assert_close(s_ref, s)
|
|
torch.testing.assert_close(B.float(), B_ref.float())
|
|
|
|
B_s, _ = torch.ops.tensorrt_llm.static_quantize_e4m3_weight(
|
|
A, s.float())
|
|
|
|
torch.testing.assert_close(B_s.float(), B_ref.float())
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_per_tensor_scales(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
m = 11
|
|
n = 11
|
|
A = torch.randn((m, n), dtype=dtype).cuda()
|
|
B, s = torch.ops.tensorrt_llm.quantize_e4m3_per_tensor(A)
|
|
s_ref = (A.flatten().float().abs().max().view(1, 1) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
B_ref = self._ref_quant(A, s_ref)
|
|
|
|
torch.testing.assert_close(s_ref, s)
|
|
torch.testing.assert_close(B.float(), B_ref.float())
|
|
|
|
B_s, _ = torch.ops.tensorrt_llm.static_quantize_e4m3_per_tensor(
|
|
A, s.float())
|
|
|
|
torch.testing.assert_close(B_s.float(), B_ref.float())
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_dequantization_activation(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
n = 512
|
|
m = 1024
|
|
A = torch.randn((n, m), dtype=dtype).cuda()
|
|
|
|
assert A.stride() == (m, 1)
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_activation(A)
|
|
|
|
assert qA.shape == A.shape
|
|
assert qA.shape[:-1] == s.shape[:-1]
|
|
assert s.shape[-1] == 1
|
|
assert s.dtype == A.dtype
|
|
assert qA.dtype == torch.float8_e4m3fn
|
|
|
|
s_ref = (torch.max(A.float().abs(), -1)[0].view(n, 1) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
torch.testing.assert_close(s_ref, s)
|
|
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_activation(qA, s)
|
|
|
|
assert B.shape == A.shape
|
|
assert B.dtype == A.dtype
|
|
|
|
torch.testing.assert_close(A, B, atol=0.2, rtol=0.01)
|
|
|
|
# testing exact match
|
|
A = torch.randint(0, 8, (n, m), dtype=dtype).cuda()
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_activation(A)
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_activation(qA, s)
|
|
|
|
torch.testing.assert_close(A, B)
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_dequantization_weight(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
n = 512
|
|
m = 1024
|
|
A = torch.randn((n, m), dtype=dtype).cuda()
|
|
|
|
assert A.stride() == (m, 1)
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_weight(A)
|
|
|
|
assert qA.shape == A.shape
|
|
assert qA.shape[1:] == s.shape[1:]
|
|
assert s.shape[0] == 1
|
|
|
|
s_ref = (torch.max(A.float().abs(), 0)[0].view(1, m) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
torch.testing.assert_close(s_ref, s)
|
|
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_weight(qA, s)
|
|
|
|
torch.testing.assert_close(A, B, atol=0.2, rtol=0)
|
|
|
|
# testing exact match
|
|
A = torch.randint(0, 8, (n, m), dtype=dtype).cuda()
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_weight(A)
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_weight(qA, s)
|
|
|
|
torch.testing.assert_close(A, B)
|
|
|
|
@parameterized.expand([(torch.float32), (torch.float16), (torch.bfloat16)],
|
|
name_func=unittest_name_func)
|
|
def test_quantization_dequantization_per_tensor(self, dtype):
|
|
# Skip tests that are not supported in pre-ampere architecture
|
|
skip_bf16_pre_ampere(dtype)
|
|
|
|
n = 512
|
|
m = 1024
|
|
A = torch.randn((n, m), dtype=dtype).cuda()
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_per_tensor(A)
|
|
|
|
assert qA.shape == A.shape
|
|
assert qA.dim() == s.dim()
|
|
assert s.numel() == 1
|
|
|
|
s_ref = (A.flatten().float().abs().max().view(1, 1) /
|
|
FP8_E4M3_MAX).to(dtype)
|
|
torch.testing.assert_close(s_ref, s)
|
|
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_per_tensor(qA, s)
|
|
|
|
# per tensor is less accurate than others, so larger atol is used.
|
|
torch.testing.assert_close(A, B, atol=0.25, rtol=0)
|
|
|
|
# testing exact match
|
|
A = torch.randint(0, 8, (n, m), dtype=dtype).cuda()
|
|
|
|
qA, s = torch.ops.tensorrt_llm.quantize_e4m3_per_tensor(A)
|
|
B = torch.ops.tensorrt_llm.dequantize_e4m3_per_tensor(qA, s)
|
|
|
|
torch.testing.assert_close(A, B)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|