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
186 lines
7.6 KiB
Python
186 lines
7.6 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 tensorrt_llm.quantization import QuantMode
|
|
|
|
|
|
class TestQuantMode(unittest.TestCase):
|
|
|
|
def test_all(self):
|
|
# Set activations and weights flags.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
|
|
# Make sure _all returns True when asked for both ACTIVATIONS and INT8_WEIGHTS.
|
|
self.assertTrue(qm._all(QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS))
|
|
# Make sure _all returns False when asked only for ACTIVATIONS.
|
|
self.assertFalse(qm._all(QuantMode.ACTIVATIONS))
|
|
# Make sure _all returns True when asked only for ACTIVATIONS if limited to ACTIVATIONS flag.
|
|
self.assertTrue(
|
|
qm._all(QuantMode.ACTIVATIONS, mask=QuantMode.ACTIVATIONS))
|
|
|
|
def test_any(self):
|
|
# Set activations and weights flags.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
|
|
# Make sure _any returns True when asked for both ACTIVATIONS and INT8_WEIGHTS.
|
|
self.assertTrue(qm._any(QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS))
|
|
# Make sure _any returns True when asked only for ACTIVATIONS.
|
|
self.assertTrue(qm._any(QuantMode.ACTIVATIONS))
|
|
# Make sure _any returns False when asked for PER_TOKEN.
|
|
self.assertFalse(qm._any(QuantMode.PER_TOKEN))
|
|
|
|
def test_count(self):
|
|
# Make sure the COUNT value is as expected - change that test if you add a new flag.
|
|
self.assertEqual(QuantMode.COUNT.value, 1 << 12)
|
|
|
|
def test_from_description(self):
|
|
# Test weight only.
|
|
qm = QuantMode.from_description(True, False, False, False)
|
|
# Make sure only the INT8_WEIGHTS flag is set.
|
|
self.assertEqual(qm, QuantMode.INT8_WEIGHTS)
|
|
|
|
# Test weight only.
|
|
qm = QuantMode.use_weight_only()
|
|
# Make sure only the INT8_WEIGHTS flag is set.
|
|
self.assertEqual(qm, QuantMode.INT8_WEIGHTS)
|
|
|
|
# Test weight only (int4).
|
|
qm = QuantMode.from_description(True, False, False, False, False, True)
|
|
# Make sure only the INT4_WEIGHTS flag is set.
|
|
self.assertEqual(qm, QuantMode.INT4_WEIGHTS)
|
|
|
|
# Test weight only.
|
|
qm = QuantMode.use_weight_only(use_int4_weights=True)
|
|
# Make sure only the INT4_WEIGHTS flag is set.
|
|
self.assertEqual(qm, QuantMode.INT4_WEIGHTS)
|
|
|
|
# Test activation/weight per-tensor.
|
|
qm = QuantMode.from_description(True, True, False, False)
|
|
# The reference.
|
|
expected_qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
# Make sure ACTIVATIONS and INT8_WEIGHTS flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
# Test activation/weight per-tensor.
|
|
qm = QuantMode.use_smooth_quant()
|
|
# Make sure ACTIVATIONS and INT8_WEIGHTS flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
# Test activation/weight per-tensor & per-channel.
|
|
qm = QuantMode.from_description(True, True, False, True)
|
|
# The reference.
|
|
expected_qm = expected_qm | QuantMode.PER_CHANNEL
|
|
# Make sure ACTIVATIONS, INT8_WEIGHTS and PER_CHANNEL flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
# Test activation/weight per-tensor & per-channel.
|
|
qm = QuantMode.use_smooth_quant(per_channel=True)
|
|
# Make sure ACTIVATIONS, INT8_WEIGHTS and PER_CHANNEL flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
# Test activation/weight per-token & per-channel.
|
|
qm = QuantMode.from_description(True, True, True, True)
|
|
# The expected result.
|
|
expected_qm = expected_qm | QuantMode.PER_TOKEN
|
|
# Make sure all flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
# Test activation/weight per-token & per-channel.
|
|
qm = QuantMode.use_smooth_quant(True, True)
|
|
# Make sure all flags are set.
|
|
self.assertEqual(qm, expected_qm)
|
|
|
|
def test_per_channel(self):
|
|
# Set per-channel flag.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS | QuantMode.PER_CHANNEL
|
|
# Make sure it returns True for per-channel.
|
|
self.assertTrue(qm.has_per_channel_scaling())
|
|
# Do not set per-channel flag.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
# Make sure it returns False for per-channel.
|
|
self.assertFalse(qm.has_per_channel_scaling())
|
|
|
|
def test_per_token(self):
|
|
# Set per-token flag.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS | QuantMode.PER_TOKEN
|
|
# Make sure it returns True for per-token.
|
|
self.assertTrue(qm.has_per_token_dynamic_scaling())
|
|
# Make sure it returns False for per-tensor.
|
|
self.assertFalse(qm.has_act_static_scaling())
|
|
|
|
# Do not set per-token flag.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
# Make sure it returns False for per-token.
|
|
self.assertFalse(qm.has_per_token_dynamic_scaling())
|
|
# Make sure it returns True for per-tensor.
|
|
self.assertTrue(qm.has_act_static_scaling())
|
|
|
|
def test_weights_only(self):
|
|
# Set weights flags.
|
|
qm = QuantMode.INT8_WEIGHTS
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_weight_only())
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_int8_weight_only())
|
|
|
|
# Set weights flags.
|
|
qm = QuantMode.INT4_WEIGHTS
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_weight_only())
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_int4_weight_only())
|
|
|
|
# Set activations and weights flags.
|
|
qm = QuantMode.ACTIVATIONS | QuantMode.INT8_WEIGHTS
|
|
# Make sure it returns False for weight-only.
|
|
self.assertFalse(qm.is_weight_only())
|
|
|
|
def test_int8_kv_cache(self):
|
|
# Set int8 kv cache flags.
|
|
qm = QuantMode.INT8_KV_CACHE
|
|
# Make sure it returns True for kv_cache.
|
|
self.assertTrue(qm.has_int8_kv_cache())
|
|
# Make sure it returns True for any quantization.
|
|
self.assertTrue(qm.has_any_quant())
|
|
|
|
# Set weights flags.
|
|
qm = QuantMode.INT8_WEIGHTS
|
|
# Make sure it returns True for any quantization.
|
|
self.assertTrue(qm.has_any_quant())
|
|
# Set int8 KV cache flag.
|
|
qm = qm.set_int8_kv_cache()
|
|
# Make sure it returns True for kv_cache.
|
|
self.assertTrue(qm.has_int8_kv_cache())
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_weight_only())
|
|
# Make sure it returns True for weight-only.
|
|
self.assertTrue(qm.is_int8_weight_only())
|
|
|
|
def test_failure_quant(self):
|
|
# Expect failure if weights are not quantized, but activations are.
|
|
self.assertRaises(
|
|
ValueError,
|
|
lambda: QuantMode.from_description(False, True, False, False))
|
|
|
|
# Expect failure if per token and per channel quantization, but weights and activations are not quantized.
|
|
self.assertRaises(
|
|
ValueError,
|
|
lambda: QuantMode.from_description(False, False, True, True))
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|