# 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. """Module test_qwen test qwen examples.""" import csv import os import pytest from defs.common import (convert_weights, test_multi_lora_support, venv_check_call, venv_mpi_check_call) from defs.conftest import (get_device_count, get_device_memory, get_sm_version, skip_post_blackwell, skip_pre_ada) from defs.trt_test_alternative import check_call # skip trt flow cases on post-Blackwell-Ultra if get_sm_version() >= 103: pytest.skip( "TRT workflow tests are not supported on post Blackwell-Ultra architecture", allow_module_level=True) @pytest.mark.parametrize( "context_fmha_type", ['enable_fmha', 'disable_fmha', 'enable_fmha_fp32_acc']) @pytest.mark.parametrize("use_weight_only", [pytest.param(True, marks=skip_post_blackwell), False], ids=["enable_weight_only", "disable_weight_only"]) @pytest.mark.parametrize( "remove_input_padding", [True, False], ids=["enable_remove_input_padding", "disable_remove_input_padding"]) @pytest.mark.parametrize( "paged_kv_cache", [True, False], ids=["enable_paged_kv_cache", "disable_paged_kv_cache"]) @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen2.5_7b_instruct", "qwen_7b_chat", "qwen1.5_0.5b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct", "qwen2_vl_7b_instruct", "qwen2_0.5b_instruct", "qwen2.5_0.5b_instruct", "qwen2.5_1.5b_instruct" ], indirect=True) def test_llm_qwen_single_gpu_summary( qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, context_fmha_type, use_weight_only, remove_input_padding, paged_kv_cache, ): "Build & run Qwen-7B-Chat on single gpu." print("Converting checkpoint...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype, use_weight_only=use_weight_only) print("Building engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", f"--max_seq_len=2048", ] if context_fmha_type == "enable_fmha": build_cmd.append("--context_fmha=enable") elif context_fmha_type == "disable_fmha": build_cmd.append("--context_fmha=disable") if remove_input_padding: build_cmd.append("--remove_input_padding=enable") else: build_cmd.append("--remove_input_padding=disable") if paged_kv_cache: build_cmd.append("--paged_kv_cache=enable") else: build_cmd.append("--paged_kv_cache=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--tensorrt_llm_rouge1_threshold=22", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] if context_fmha_type == "enable_fmha_fp32_acc": summary_cmd.append("--enable_context_fmha_fp32_acc") venv_check_call(llm_venv, summary_cmd) @pytest.mark.skip_less_device_memory(40000) @pytest.mark.parametrize( "context_fmha_type", ['enable_fmha', 'disable_fmha', 'enable_fmha_fp32_acc']) @pytest.mark.parametrize("use_weight_only", [True, False], ids=["enable_weight_only", "disable_weight_only"]) @pytest.mark.parametrize( "remove_input_padding", [True, False], ids=["enable_remove_input_padding", "disable_remove_input_padding"]) @pytest.mark.parametrize( "paged_kv_cache", [True, False], ids=["enable_paged_kv_cache", "disable_paged_kv_cache"]) @pytest.mark.parametrize("llm_qwen_model_root", ["qwen1.5_moe_a2.7b_chat"], indirect=True) def test_llm_qwen_moe_single_gpu_summary( qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, context_fmha_type, use_weight_only, remove_input_padding, paged_kv_cache, ): "Build & run Qwen1.5-MoE-A2.7B-Chat on single gpu." print("Converting checkpoint...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype, use_weight_only=use_weight_only) print("Building engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", f"--max_seq_len=2048", ] if context_fmha_type == "enable_fmha": build_cmd.append("--context_fmha=enable") elif context_fmha_type == "disable_fmha": build_cmd.append("--context_fmha=disable") if remove_input_padding: build_cmd.append("--remove_input_padding=enable") else: build_cmd.append("--remove_input_padding=disable") if paged_kv_cache: build_cmd.append("--paged_kv_cache=enable") else: build_cmd.append("--paged_kv_cache=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--tensorrt_llm_rouge1_threshold=22", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] if context_fmha_type == "enable_fmha_fp32_acc": summary_cmd.append("--enable_context_fmha_fp32_acc") venv_check_call(llm_venv, summary_cmd) # Delete this case refer to https://nvbugs/5072417 # @pytest.mark.parametrize("llm_lora_model_root", ["Ko-QWEN-7B-Chat-LoRA"], # indirect=True) # @pytest.mark.parametrize("llm_qwen_model_root", ["qwen_7b_chat"], indirect=True) # def test_llm_qwen_7b_single_gpu_lora( # qwen_example_root, # llm_qwen_model_root, # llm_venv, # cmodel_dir, # engine_dir, # llm_lora_model_root, # ): # "run Qwen lora test on single gpu." # print("Build engines...") # dtype = 'float16' # model_name = os.path.basename(llm_qwen_model_root) # ckpt_dir = convert_weights(llm_venv=llm_venv, # example_root=qwen_example_root, # cmodel_dir=cmodel_dir, # model=model_name, # model_path=llm_qwen_model_root, # data_type=dtype) # print("Build engines...") # build_cmd = [ # "trtllm-build", # f"--checkpoint_dir={ckpt_dir}", # f"--output_dir={engine_dir}", # "--lora_plugin=auto", # "--gemm_plugin=auto", # f"--lora_dir={llm_lora_model_root}", # ] # check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) # ref_1 = [ # 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, # 151644, 872, 198, 126246, 144370, 91145, 11, 137601, 29326, 86034, # 12802, 5140, 98734, 19391, 35711, 30, 151645, 198, 151644, 77091, 198, # 126246, 144370, 91145, 0, 134561, 58677, 78125, 21329, 66019, 124685, # 134619, 94152, 28626, 17380, 11, 134637, 20401, 138520, 19391, 143603 # ] # ref_2 = [ # 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, # 151644, 872, 198, 126246, 144370, 91145, 11, 137601, 29326, 86034, # 12802, 5140, 98734, 19391, 35711, 30, 151645, 198, 151644, 77091, 198, # 126246, 144370, 91145, 0, 134561, 330, 48, 1103, 54, 268, 1, 78952, 13, # 151645, 198, 151643, 151643, 151643, 151643, 151643 # ] # input_text = "안녕하세요, 혹시 이름이 뭐에요?" # print("Run inference with lora id 0...") # venv_check_call(llm_venv, [ # f"{qwen_example_root}/../run.py", # "--max_output_len=20", # f"--input_text={input_text}", # "--lora_task_uids=0", # f"--tokenizer_dir={llm_qwen_model_root}", # f"--engine_dir={engine_dir}", # f"--output_csv={llm_venv.get_working_directory()}/use_lora.csv", # "--use_py_session", # ]) # with open(f"{llm_venv.get_working_directory()}/use_lora.csv") as f: # predict = csv.reader(f) # predict = next(predict) # predict = [int(p) for p in predict] # assert ref_1 == predict # print("Run inference with lora id -1...") # venv_check_call(llm_venv, [ # f"{qwen_example_root}/../run.py", # "--max_output_len=20", # f"--input_text={input_text}", # "--lora_task_uids=-1", # f"--tokenizer_dir={llm_qwen_model_root}", # f"--engine_dir={engine_dir}", # f"--output_csv={llm_venv.get_working_directory()}/no_lora.csv", # "--use_py_session", # ]) # with open(f"{llm_venv.get_working_directory()}/no_lora.csv") as f: # predict = csv.reader(f) # predict = next(predict) # predict = [int(p) for p in predict] # assert ref_2 == predict @pytest.mark.parametrize("llm_lora_model_root", ["Qwen1.5-7B-Chat-750Mb-lora"], indirect=True) @pytest.mark.parametrize("llm_qwen_model_root", ["qwen1.5_7b_chat"], indirect=True) def test_llm_qwen1_5_7b_single_gpu_lora( qwen_example_root, llm_qwen_model_root, llm_venv, cmodel_dir, engine_dir, llm_lora_model_root, ): "run Qwen1.5 lora test on single gpu." print("Build engines...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", "--lora_plugin=auto", "--gemm_plugin=auto", f"--lora_dir={llm_lora_model_root}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) ref_1 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 40, 2776, 458, 15235, 7881, 553, 5264, 15469, 11, 773, 358, 1513, 944, 614, 264, 829, 304, 279, 8606, 5530 ] ref_2 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 40, 1079, 1207, 16948, 11, 264, 3460, 4128, 1614, 3465, 553, 54364, 14817, 13, 151645, 151645, 151645, 151645, 151645, 151645 ] input_text = "What is your name?" print("Run inference with lora id 0...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=0", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/use_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/use_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_1 == predict print("Run inference with lora id -1...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=-1", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/no_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/no_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_2 == predict @pytest.mark.skip_less_device_memory(60000) @pytest.mark.parametrize("llm_lora_model_root", ["Upcycled-Qwen1.5-MoE2.7B-LoRA"], indirect=True) @pytest.mark.parametrize("llm_qwen_model_root", ["qwen1.5_moe_a2.7b_chat"], indirect=True) def test_llm_qwen1_5_moe_single_gpu_lora( qwen_example_root, llm_qwen_model_root, llm_venv, cmodel_dir, engine_dir, llm_lora_model_root, ): "run Qwen1.5 MoE lora test on single gpu." print("Build engines...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", "--lora_plugin=auto", "--moe_plugin=disable", f"--lora_dir={llm_lora_model_root}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) ref_1 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935 ] ref_2 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 40, 1079, 1207, 16948, 11, 264, 3460, 4128, 1614, 3465, 553, 54364, 14817, 13, 151645, 151645, 151645, 151645, 151645, 151645 ] input_text = "What is your name?" print("Run inference with lora id 0...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=0", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/use_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/use_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_1 == predict print("Run inference with lora id -1...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=-1", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/no_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/no_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_2 == predict @pytest.mark.skip_less_device_memory(60000) @pytest.mark.parametrize("llm_lora_model_root", ["Upcycled-Qwen1.5-MoE2.7B-LoRA"], indirect=True) @pytest.mark.parametrize("llm_qwen_model_root", ["qwen1.5_moe_a2.7b_chat"], indirect=True) def test_llm_qwen1_5_moe_plugin_single_gpu_lora( qwen_example_root, llm_qwen_model_root, llm_venv, cmodel_dir, engine_dir, llm_lora_model_root, ): "run Qwen1.5 MoE lora test on single gpu." print("Build engines...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", "--lora_plugin=auto", "--moe_plugin=auto", f"--lora_dir={llm_lora_model_root}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) ref_1 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935, 151935 ] ref_2 = [ 151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 3838, 374, 697, 829, 30, 151645, 198, 151644, 77091, 198, 40, 1079, 1207, 16948, 11, 264, 3460, 4128, 1614, 3465, 553, 54364, 14817, 13, 151645, 151645, 151645, 151645, 151645, 151645 ] input_text = "What is your name?" print("Run inference with lora id 0...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=0", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/use_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/use_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_1 == predict print("Run inference with lora id -1...") venv_check_call(llm_venv, [ f"{qwen_example_root}/../../../run.py", "--max_output_len=20", f"--input_text={input_text}", "--lora_task_uids=-1", f"--tokenizer_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", f"--output_csv={llm_venv.get_working_directory()}/no_lora.csv", "--use_py_session", ]) with open(f"{llm_venv.get_working_directory()}/no_lora.csv") as f: predict = csv.reader(f) predict = next(predict) predict = [int(p) for p in predict] assert ref_2 == predict @skip_post_blackwell @pytest.mark.parametrize("use_weight_only", [True, False], ids=['enable_weight_only', 'disable_weight_only']) @pytest.mark.parametrize("use_gemm_plugin", [True, False], ids=["enable_gemm_plugin", "disable_gemm_plugin"]) @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen2.5_7b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct", "qwen2_vl_7b_instruct" ], indirect=True) def test_llm_qwen_7b_int8_kv_1node_1gpus(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, use_weight_only, use_gemm_plugin): "Build & Run Qwen-7B-Chat int8 kv cache" print("Converting checkpoint...") dtype = 'float16' model_name = "Qwen-chat-7b-int8-kv" ckpt_dir = convert_weights( llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype, int8_kv_cache=True, use_weight_only=use_weight_only, weight_only_precision='int8' if use_weight_only else None, calib_dataset=f"{llm_datasets_root}/ccdv/cnn_dailymail") print("Building engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--gpt_attention_plugin={dtype}" ] if use_gemm_plugin: build_cmd.append(f"--gemm_plugin={dtype}") else: build_cmd.append("--gemm_plugin=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", "--tensorrt_llm_rouge1_threshold=22", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] venv_check_call(llm_venv, summary_cmd) @pytest.mark.skip_less_device_memory(40000) @pytest.mark.parametrize("num_beams", [1, 4], ids=lambda num_beams: f'nb:{num_beams}') @pytest.mark.parametrize( "tp_pp_size", [(2, 1), (4, 1), (2, 2)], ids=lambda tp_pp_size: f'tp{tp_pp_size[0]}pp{tp_pp_size[1]}') @pytest.mark.parametrize("use_plugin", ["enable_plugin", "disable_plugin"]) @pytest.mark.parametrize( "context_fmha_type", ['enable_fmha', 'disable_fmha', 'enable_fmha_fp32_acc']) @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen2.5_7b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct", "qwen2_vl_7b_instruct" ], indirect=True) def test_llm_qwen_7b_multi_gpus_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, num_beams, tp_pp_size, use_plugin, context_fmha_type): "Run qwen on multi gpus" tp_size, pp_size = tp_pp_size world_size = tp_size * pp_size if get_device_count() < world_size: pytest.skip(f"devices are less than {world_size}.") print("Converting checkpoint...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype, gpus=world_size, tp_size=tp_size, pp_size=pp_size, workers=world_size) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", "--max_seq_len=8192", f"--max_beam_width={num_beams}", f"--workers={world_size}" ] if use_plugin: build_cmd.append(f"--gemm_plugin={dtype}") build_cmd.append(f"--gpt_attention_plugin={dtype}") else: build_cmd.append("--gemm_plugin=disable") build_cmd.append("--gpt_attention_plugin=disable") if context_fmha_type == "enable_fmha": build_cmd.append("--context_fmha=enable") elif context_fmha_type == "disable_fmha": build_cmd.append("--context_fmha=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--num_beams={num_beams}", "--tensorrt_llm_rouge1_threshold=24", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] if context_fmha_type == "enable_fmha_fp32_acc": summary_cmd.append("--enable_context_fmha_fp32_acc") venv_mpi_check_call( llm_venv, ["mpirun", "-n", f"{world_size}", "--allow-run-as-root"], summary_cmd) @skip_post_blackwell @pytest.mark.parametrize("num_beams", [1, 2, 4], ids=lambda num_beams: f'nb:{num_beams}') @pytest.mark.parametrize("per_token_channel", [True, False], ids=["enable_ptpc", "disable_ptpc"]) @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen_7b_chat", "qwen_14b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct", "qwen2_vl_7b_instruct", "qwen2.5_7b_instruct" ], indirect=True) def test_llm_qwen_smooth_quant_single_gpu_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, engine_dir, cmodel_dir, per_token_channel, num_beams): "Run qwen with smooth quant on single gpu" if "14B" in llm_qwen_model_root and get_device_memory() < 80000: pytest.skip("GPU memory is insufficient.") print("Converting checkpoint...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights( llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype, smoothquant=0.5, per_token=per_token_channel, per_channel=per_token_channel, calib_dataset=f"{llm_datasets_root}/ccdv/cnn_dailymail") print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", f"--max_beam_width={num_beams}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--max_seq_len={8192}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--num_beams={num_beams}", "--tensorrt_llm_rouge1_threshold=24", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] venv_check_call(llm_venv, summary_cmd) @skip_post_blackwell @pytest.mark.parametrize("num_beams", [1, 4], ids=lambda num_beams: f'nb:{num_beams}') @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen_7b_chat_int4", "qwen1.5_14b_chat_int4", "qwen1.5_7b_chat_awq", "qwen2_7b_awq", "qwen2.5_14b_instruct_int4" ], indirect=True) def test_llm_qwen_int4_single_gpu_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, num_beams): "Run qwen with gptq on single gpu" print("Converting checkpoint...") dtype = 'float16' model_name = os.path.basename(llm_qwen_model_root) ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model=model_name, model_path=llm_qwen_model_root, data_type=dtype) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", f"--max_beam_width={num_beams}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--max_seq_len={8192}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--num_beams={num_beams}", "--tensorrt_llm_rouge1_threshold=24", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] venv_check_call(llm_venv, summary_cmd) @skip_post_blackwell @pytest.mark.parametrize("num_beams", [1, 4], ids=lambda num_beams: f'nb:{num_beams}') @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen_7b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct", "qwen2_vl_7b_instruct", "qwen2.5_7b_instruct" ], indirect=True) def test_llm_qwen_awq_single_gpu_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, engine_dir, num_beams, qcache_dir): "Build & run int4 awq on 1 gpus." print("Quantizing model...") dtype = 'float16' output_dir = f"{qcache_dir}/quantized_int4-awq" quantize_cmd = [ f"{qwen_example_root}/../../../quantization/quantize.py", f"--model_dir={llm_qwen_model_root}", f"--calib_dataset={llm_datasets_root}/cnn_dailymail", f"--output_dir={output_dir}", "--awq_block_size=128", "--qformat=int4_awq", "--dtype=float16", "--calib_size=32", ] venv_check_call(llm_venv, quantize_cmd) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={output_dir}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", f"--max_beam_width={num_beams}", f"--output_dir={engine_dir}", "--max_seq_len=8192", f"--max_batch_size={8}", ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", "--hf_model_dir", f"{llm_qwen_model_root}", "--data_type", "fp16", "--check_accuracy", f"--engine_dir={engine_dir}", f"--num_beams={num_beams}", "--tensorrt_llm_rouge1_threshold=22.8", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] venv_check_call(llm_venv, summary_cmd) @pytest.mark.skip_less_device(8) @pytest.mark.skip_less_device_memory(80000) @pytest.mark.parametrize( "context_fmha", ["context_fmha", "context_fmha_fp32_acc", "disable_fmha"]) @pytest.mark.parametrize( "tp_pp_size", [(8, 1), (4, 2)], ids=lambda tp_pp_size: f'tp{tp_pp_size[0]}pp{tp_pp_size[1]}') @pytest.mark.parametrize( "llm_qwen_model_root", ["qwen2.5_72b_chat", "qwen1.5_72b_chat", "qwen2_72b_instruct"], indirect=True) def test_llm_qwen_1node_8gpus_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, context_fmha, tp_pp_size): "Run qwen with smooth quant on single gpu" print("Converting checkpoint...") dtype = 'float16' tp_size, pp_size = tp_pp_size world_size = tp_size * pp_size ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model="qwen-72b", model_path=llm_qwen_model_root, data_type=dtype, tp_size=4, pp_size=2) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--gemm_plugin={dtype}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--workers={world_size}", "--max_beam_width=4", "--max_seq_len=8192", "--max_input_len=2048", ] if context_fmha == "context_fmha": build_cmd.append("--context_fmha=enable") elif context_fmha == "disable_fmha": build_cmd.append("--context_fmha=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", f"--hf_model_dir={llm_qwen_model_root}", "--data_type=fp16", "--check_accuracy", f"--engine_dir={engine_dir}", "--num_beams=4", "--tensorrt_llm_rouge1_threshold=22", "--max_input_length=2048", "--output_len=2048", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] if context_fmha == "context_fmha_fp32_acc": summary_cmd.append("--enable_context_fmha_fp32_acc") venv_mpi_check_call( llm_venv, ["mpirun", "-n", f"{world_size}", "--allow-run-as-root"], summary_cmd) @pytest.mark.skip_less_device(4) @pytest.mark.skip_less_device_memory(40000) @pytest.mark.parametrize( "context_fmha", ["context_fmha", "context_fmha_fp32_acc", "disable_fmha"]) @pytest.mark.parametrize( "tp_pp_size", [(4, 1), (2, 2)], ids=lambda tp_pp_size: f'tp{tp_pp_size[0]}pp{tp_pp_size[1]}') @pytest.mark.parametrize("llm_qwen_model_root", ["qwen2_57b_a14b"], indirect=True) def test_llm_qwen_moe_multi_gpu_summary(qwen_example_root, llm_qwen_model_root, llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir, engine_dir, context_fmha, tp_pp_size): "Run qwen with smooth quant on single gpu" print("Converting checkpoint...") dtype = 'float16' tp_size, pp_size = tp_pp_size world_size = tp_size * pp_size ckpt_dir = convert_weights(llm_venv=llm_venv, example_root=qwen_example_root, cmodel_dir=cmodel_dir, model="qwen-57b", model_path=llm_qwen_model_root, data_type=dtype, tp_size=tp_size, pp_size=pp_size) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={ckpt_dir}", f"--gemm_plugin={dtype}", f"--output_dir={engine_dir}", f"--max_batch_size={8}", f"--workers={world_size}", "--max_input_len=2048", "--max_seq_len=8192", ] if context_fmha == "context_fmha": build_cmd.append("--context_fmha=enable") elif context_fmha == "disable_fmha": build_cmd.append("--context_fmha=disable") check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", f"--hf_model_dir={llm_qwen_model_root}", "--data_type=fp16", "--check_accuracy", f"--engine_dir={engine_dir}", "--tensorrt_llm_rouge1_threshold=20", "--max_input_length=2048", "--output_len=2048", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}" ] if context_fmha == "context_fmha_fp32_acc": summary_cmd.append("--enable_context_fmha_fp32_acc") venv_mpi_check_call( llm_venv, ["mpirun", "-n", f"{world_size}", "--allow-run-as-root"], summary_cmd) @skip_pre_ada @pytest.mark.parametrize("dtype", ['float16', 'bfloat16']) @pytest.mark.parametrize("qformat", ['fp8']) @pytest.mark.parametrize("llm_qwen_model_root", [ "qwen2.5_7b_instruct", "qwen2_7b_instruct", "qwen2_vl_7b_instruct", "qwen2_0.5b_instruct", "qwen2.5_0.5b_instruct", "qwen2.5_1.5b_instruct" ], indirect=True) def test_llm_hf_qwen_quantization_1gpu(dtype, llm_qwen_model_root, llm_venv, cmodel_dir, engine_dir, qwen_example_root, llm_datasets_root, llm_rouge_root, qformat): "Run qwen quantization tests" print("Convert checkpoint by modelopt...") convert_cmd = [ f"{qwen_example_root}/../../../quantization/quantize.py", f"--model_dir={llm_qwen_model_root}", f"--calib_dataset={llm_datasets_root}/cnn_dailymail", f"--dtype={dtype}", f"--qformat={qformat}", f"--kv_cache_dtype={qformat}", f"--output_dir={cmodel_dir}", ] venv_check_call(llm_venv, convert_cmd) print("Build engines...") build_cmd = [ "trtllm-build", f"--checkpoint_dir={cmodel_dir}", f"--output_dir={engine_dir}", f"--gemm_plugin={dtype}", f"--gpt_attention_plugin={dtype}", "--max_seq_len=4096", f"--max_batch_size={8}", "--use_fp8_context_fmha=disable", "--use_paged_context_fmha=disable" ] check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env) print("Run evaluation...") threshold_score = 21 mmlu_score = None if "0.5B" in llm_qwen_model_root: mmlu_score = 46.0 elif "1.5B" in llm_qwen_model_root: mmlu_score = 58.5 elif "7B" in llm_qwen_model_root: mmlu_score = 71.5 # Install custom jinja to overcome ImportError from transformers library. # ImportError: apply_chat_template requires jinja2>=3.1.0 to be installed. llm_venv.run_cmd(['-m', 'pip', 'install', 'jinja2==3.1.0']) # Run MMLU for Qwen 2.5 models. if '2.5' in llm_qwen_model_root: mmlu_cmd = [ "trtllm-eval", f"--model={engine_dir}", f"--tokenizer={llm_qwen_model_root}", "--backend=tensorrt", "mmlu", f"--dataset_path={llm_datasets_root}/mmlu", "--check_accuracy", f"--accuracy_threshold={mmlu_score}" ] check_call(" ".join(mmlu_cmd), shell=True, env=llm_venv._new_env) else: summary_cmd = [ f"{qwen_example_root}/../../../summarize.py", "--test_trt_llm", f"--hf_model_dir={llm_qwen_model_root}", f"--engine_dir={engine_dir}", "--check_accuracy", f"--tensorrt_llm_rouge1_threshold={threshold_score}", f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}", ] venv_check_call(llm_venv, summary_cmd) @skip_pre_ada @pytest.mark.parametrize( "llm_qwen_model_root", ["qwen2_0.5b_instruct", "qwen2.5_0.5b_instruct", "qwen2.5_1.5b_instruct"], indirect=True) def test_llm_hf_qwen_multi_lora_1gpu(llm_qwen_model_root, llm_venv, cmodel_dir, engine_dir, qwen_example_root, llm_datasets_root, qformat='fp8', dtype='bfloat16'): "Run Qwen models with multiple dummy LoRAs." print("Convert checkpoint by modelopt...") convert_cmd = [ f"{qwen_example_root}/../../../quantization/quantize.py", f"--model_dir={llm_qwen_model_root}", f"--calib_dataset={llm_datasets_root}/cnn_dailymail", f"--dtype={dtype}", f"--qformat={qformat}", f"--kv_cache_dtype={qformat}", f"--output_dir={cmodel_dir}", ] venv_check_call(llm_venv, convert_cmd) test_multi_lora_support( hf_model_dir=llm_qwen_model_root, tllm_ckpt_dir=cmodel_dir, engine_dir=engine_dir, llm_venv=llm_venv, example_root=qwen_example_root, num_loras=2, lora_rank=8, target_hf_modules=["q_proj", "k_proj", "v_proj"], target_trtllm_modules=["attn_q", "attn_k", "attn_v"], zero_lora_weights=True, )