# 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 _utils import numpy as np import tensorrt as trt import torch from polygraphy.backend.trt import CreateConfig, EngineFromNetwork, TrtRunner from polygraphy.logger import G_LOGGER from tensorrt_llm.quantization.quantize import (qserve_pack_reorder_per_channel, qserve_pack_reorder_per_group) G_LOGGER.severity = 0 import tensorrt_llm from tensorrt_llm import Tensor from tensorrt_llm.quantization.functional import (qserve_gemm_per_channel, qserve_gemm_per_group) sys.path.append(os.path.join(os.path.dirname(__file__), '..')) from utils.util import skip_pre_ampere class TestQServeGemm(unittest.TestCase): def setUp(self): tensorrt_llm.logger.set_level('error') def _case_qserve_gemm_per_group(self, m, n, k, group_size=128, seed=123456, dtype="float16"): torch.manual_seed(seed) # Initialize qact (int8 range: -128 to 127) qact = torch.randint(-128, 128, (m, k), dtype=torch.int8) # Initialize act_scales (float16) act_scales = torch.rand( m, 1, dtype=torch.float16) * 0.1 # small positive values # Initialize qweight (int4 range: 0 to 15, as per qserve_quantize_weight) qweight = torch.randint(0, 16, (n, k), dtype=torch.int8) # Initialize s1_scales (float16) s1_scales = torch.rand(n, 1, dtype=torch.float16) * 0.1 # Initialize s2_scales (int) s2_scales = torch.randint(1, 16, (n, k // group_size), dtype=torch.int8) # Initialize s2_zeros (int) s2_zeros = torch.randint(0, 16, (n, k // group_size), dtype=torch.int8) # Ground truth matmul ref = _utils.gt_qserve_gemm_per_group(qact, act_scales, qweight, s1_scales, s2_scales, s2_zeros).cpu().numpy() # Prepare data for QServe gemm kernel qweight, s1_scales, s2_scales, s2_zeros = qserve_pack_reorder_per_group( qweight, s1_scales, s2_scales, s2_zeros, group_size) # Create builder builder = tensorrt_llm.Builder() builder.strongly_typed = False # Test need to run in weekly typed mode # Create empty network network = builder.create_network() # Allow SQ plugin of dtype type network.plugin_config.set_qserve_plugins("float16") with tensorrt_llm.net_guard(network): qact_trt = Tensor( name='qact', shape=qact.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) act_scales_trt = Tensor( name='act_scales', shape=act_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) qweight_trt = Tensor( name='qweight', shape=qweight.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) s1_scales_trt = Tensor( name='s1_scales', shape=s1_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) s2_scales_trt = Tensor( name='s2_scales', shape=s2_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) s2_zeros_trt = Tensor( name='s2_zeros', shape=s2_zeros.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) output = qserve_gemm_per_group(qact_trt, act_scales_trt, qweight_trt, s1_scales_trt, s2_scales_trt, s2_zeros_trt) output.mark_output('output', dtype) engine = EngineFromNetwork( (builder.trt_builder, network.trt_network), config=CreateConfig(int8=True, fp16=True, memory_pool_limits={ trt.MemoryPoolType.WORKSPACE: 32 * 1024 * 1024 })) # Infer engine with TrtRunner(engine) as runner: outputs = runner.infer( feed_dict={ 'qact': qact.numpy(), 'act_scales': act_scales.numpy(), 'qweight': qweight.numpy(), 's1_scales': s1_scales.numpy(), 's2_scales': s2_scales.numpy(), 's2_zeros': s2_zeros.numpy() }) output = outputs['output'] # Allow difference by one code point. self.assertTrue(np.allclose(output, ref, rtol=0, atol=np.spacing(ref))) def _case_qserve_gemm_per_channel(self, m, n, k, seed=123456, dtype="float16"): torch.manual_seed(seed) # Initialize qact (int8 range: -128 to 127) qact = torch.randint(-128, 128, (m, k), dtype=torch.int8) # Initialize act_scales (float16) act_scales = torch.rand( m, 1, dtype=torch.float16) * 0.1 # small positive values # Compute act_sums act_sums = torch.sum(qact.float() * act_scales.float(), dim=1, keepdim=True).half() # Initialize qweight (int4 range: 0 to 15, as per qserve_quantize_weight) qweight = torch.randint(0, 16, (n, k), dtype=torch.int8) # Initialize s1_scales (float16) s1_scales = torch.rand(n, 1, dtype=torch.float16) * 0.1 # Initialize s1_szeros (float16) s1_zeros = torch.randint(1, 16, (n, 1)).half() # Ground truth matmul ref = _utils.gt_qserve_gemm_per_channel(qact, act_scales, act_sums, qweight, s1_scales, s1_zeros).cpu().numpy() # Prepare data for QServe gemm kernel qweight, s1_scales, s1_szeros = qserve_pack_reorder_per_channel( qweight, s1_scales, s1_zeros) # Create builder builder = tensorrt_llm.Builder() builder.strongly_typed = False # Test need to run in weekly typed mode # Create empty network network = builder.create_network() # Allow SQ plugin of dtype type network.plugin_config.set_qserve_plugins("float16") with tensorrt_llm.net_guard(network): qact_trt = Tensor( name='qact', shape=qact.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) act_scales_trt = Tensor( name='act_scales', shape=act_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) act_sums_trt = Tensor( name='act_sums', shape=act_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) qweight_trt = Tensor( name='qweight', shape=qweight.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("int8")) s1_scales_trt = Tensor( name='s1_scales', shape=s1_scales.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) s1_szeros_trt = Tensor( name='s1_szeros', shape=s1_szeros.shape, dtype=tensorrt_llm._utils.str_dtype_to_trt("float16")) output = qserve_gemm_per_channel(qact_trt, act_scales_trt, act_sums_trt, qweight_trt, s1_scales_trt, s1_szeros_trt) output.mark_output('output', dtype) engine = EngineFromNetwork( (builder.trt_builder, network.trt_network), config=CreateConfig(int8=True, fp16=True, memory_pool_limits={ trt.MemoryPoolType.WORKSPACE: 32 * 1024 * 1024 })) # Infer engine with TrtRunner(engine) as runner: outputs = runner.infer( feed_dict={ 'qact': qact.numpy(), 'act_scales': act_scales.numpy(), 'act_sums': act_sums.numpy(), 'qweight': qweight.numpy(), 's1_scales': s1_scales.numpy(), 's1_szeros': s1_szeros.numpy(), }) output = outputs['output'] # Allow some difference. self.assertTrue(np.allclose(output, ref, rtol=1e-2, atol=0.25)) @skip_pre_ampere def test_qserve_gemm_per_group(self, dtype='float16'): bs = 2 inseq = 16 hidden_size = 768 # qkv_gemm self._case_qserve_gemm_per_group(bs * inseq, 3 * hidden_size, hidden_size, dtype=dtype) # mlp_gemm_1 self._case_qserve_gemm_per_group(bs * inseq, 4 * hidden_size, hidden_size, dtype=dtype) @skip_pre_ampere def test_qserve_gemm_per_channel(self, dtype='float16'): bs = 2 inseq = 16 hidden_size = 768 # qkv_gemm self._case_qserve_gemm_per_channel(bs * inseq, 3 * hidden_size, hidden_size, dtype=dtype) # mlp_gemm_1 self._case_qserve_gemm_per_channel(bs * inseq, 4 * hidden_size, hidden_size, dtype=dtype) def test_qserve_gemm_per_group_no_plugin(self): # Create builder builder = tensorrt_llm.Builder() # Create empty network network = builder.create_network() with tensorrt_llm.net_guard(network): # Gemm ootb should fail with self.assertRaisesRegex( TypeError, "QServe Quant GEMM is only supported with plugin"): qserve_gemm_per_group(None, None, None, None, None, None) def test_qserve_gemm_per_channel_no_plugin(self): # Create builder builder = tensorrt_llm.Builder() # Create empty network network = builder.create_network() with tensorrt_llm.net_guard(network): # Gemm ootb should fail with self.assertRaisesRegex( TypeError, "QServe Quant GEMM is only supported with plugin"): qserve_gemm_per_channel(None, None, None, None, None, None) if __name__ == '__main__': unittest.main()