mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
8709fe8b53
test: add test cases for 0.19 release (#3608) * fix test name * add quickstart test for nemotron-ultra * add rcca multi-node test case for deepseek-v3 * add rcca info --------- squash (#3642) fix: nvbugs/5187237: fix deterministic mode crash (#3448) * nvbugs/5187237 nvbugs/5112075: fix deterministic mode error * remove waive * Revert "remove waive" This reverts commit 0bf5486d19906d692bfb7a6262333c296b0087ac. * revert ar fusion --------- update fp8 doc (#3647) tests: change qa perf test to trtllm-bench (#3619) fix: FP8 quantized lm_head (NvBug 5214229) (#3567) infra: Add PR approval protection for the release branch (#3634) fix: nvbugs/5231298: pytorch allreduce issue (#3673) Fix: nvbugs/5222698 variable not defined (#3630) * Fix: nvbugs/5222698 variable not defined * Tidy code --------- test:sync waives.txt from main branch by disabling test_perf/gpt_350m-cppmanager case (#3685) test:restore fp8 kv cache testing for L0 (#3671) doc: Update DeepSeek perf docs (#3693) * Update DeepSeek perf docs * update * Apply suggestions from code review --------- tests: waive test_llm_multi_node (#3664) fix: update test_user_buffers_mm_add_prologue atol (#3711) Fix: cherry-pick hmac encryption from main branch (#3635) * security fix cherry-pick changes from main * fix hmac in remote mpi session (#3649) --------- Un-waive DS-V3-Lite tests. (#3621) fix: FP8 kv accuracy (#3675) * fix FP8 kv accuracy * update doc --------- Fix script options for engines. (#3622) unwaive multi-node test (#3721) chore : Split more tests out of gpt tests (#3524) (#3674) doc:add torch examples link into torch backend documentation (#3749) test: Get Eagle tests working (#3593) (#3722) Waive L0 test (#3756) waive failed case in perf test, change default max_batch_size to 512 and write config.json to output log (#3656) Update ds v3 parameters in stress test. (#3676) waive gemma on L20 (#3766) https://nvbugs/5141291: Fix convert.py script for Qwen model. (#3758) Include Qwen2VLDecoderLayer in the smooth_qwen2_model function. fix: PP4 fixes and cleanup (#3688) remove benchmark test list (#3643) skip disagg deepseek test if sm!=90 (#3720) test: skip failed cases on B200 (#3710) * add skip condition to tests * fix error --------- test: [nvbug: 5234494] skip_pre_ada for fp8 cases (#3718) * skip_pre_ada for fp8 cases * update * update after rebase --------- add know issue to deepseek doc. (#3800) Fix ModelOpt Mixtral AWQ OOM (#3714) (#3761) Waive L0 tests (#3826) fix: Reduce memory usage in fused moe op associated with AutoTuning and fix moe fallback issue. (#3793) * Reduce memory usage in fused moe op associated with AutoTuning. * Replace pre-defined bucket size strategy with a generating function based on the tune_max_num_tokens. * Add free_memory logic of workspace in min_latency_mode fused moe path. * Fix fused_moe fallback issue. (#3652) min_latency_mode is only set to False during warmup phase. Thus when it becomes true during inference, all tactics fall back to the default one and thus cause perf regression. --------- [doc] Better document for Draft-Target-Model (DTM) speculative decoding (#3797) Fix pre-commit Fix again Address some review comments for the MI Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com> Co-authored-by: Zhanrui Sun <184402041+ZhanruiSunCh@users.noreply.github.com>
387 lines
17 KiB
Python
387 lines
17 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 pytest
|
|
from defs.common import (convert_weights, get_dummy_spec_decoding_heads,
|
|
venv_check_call)
|
|
from defs.conftest import skip_post_blackwell, skip_pre_ada
|
|
from defs.trt_test_alternative import check_call
|
|
|
|
|
|
@skip_post_blackwell
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("batch_size", [1, 8], ids=['bs1', 'bs8'])
|
|
@pytest.mark.parametrize("data_type", ['float16'])
|
|
@pytest.mark.parametrize("eagle_model_roots", ["EAGLE-Vicuna-7B-v1.3"],
|
|
indirect=True)
|
|
def test_llm_eagle_1gpu(batch_size, data_type, use_dynamic_tree,
|
|
eagle_model_roots, eagle_example_root,
|
|
llm_datasets_root, llm_rouge_root, llm_venv, cmodel_dir,
|
|
engine_dir):
|
|
print("Build engines...")
|
|
model_name = "eagle"
|
|
|
|
model_dir = convert_weights(llm_venv=llm_venv,
|
|
example_root=eagle_example_root,
|
|
cmodel_dir=cmodel_dir,
|
|
model=model_name,
|
|
model_path=eagle_model_roots,
|
|
data_type=data_type)
|
|
|
|
build_cmd = [
|
|
"trtllm-build",
|
|
f"--checkpoint_dir={model_dir}",
|
|
f"--output_dir={engine_dir}",
|
|
f"--gpt_attention_plugin={data_type}",
|
|
f"--gemm_plugin={data_type}",
|
|
f"--max_beam_width=1",
|
|
"--remove_input_padding=enable",
|
|
"--context_fmha=enable",
|
|
"--use_paged_context_fmha=enable",
|
|
"--max_input_len=1024",
|
|
"--max_seq_len=1536",
|
|
f"--max_batch_size={batch_size}",
|
|
"--paged_kv_cache=enable",
|
|
'--speculative_decoding_mode=eagle',
|
|
]
|
|
|
|
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
|
|
|
|
print("Run run...")
|
|
run_cmd = [
|
|
f"{eagle_example_root}/../run.py",
|
|
"--max_output_len=100",
|
|
f"--tokenizer_dir={eagle_model_roots[0]}",
|
|
"--log_level=verbose",
|
|
f"--engine_dir={engine_dir}",
|
|
]
|
|
if use_dynamic_tree:
|
|
run_cmd.extend(
|
|
[f"--eagle_dynamic_tree_max_top_k={3}", "--eagle_use_dynamic_tree"])
|
|
|
|
venv_check_call(llm_venv, run_cmd)
|
|
|
|
print("Run summarize...")
|
|
summary_cmd = [
|
|
f"{eagle_example_root}/../summarize.py", "--test_trt_llm",
|
|
"--hf_model_dir", f"{eagle_model_roots[0]}", "--tokenizer_dir",
|
|
f"{eagle_model_roots[0]}", f"--engine_dir={engine_dir}",
|
|
"--check_accuracy", "--tensorrt_llm_rouge1_threshold=24",
|
|
"--eagle_choices=[[0], [0, 0], [1], [0, 1], [2], [0, 0, 0], [1, 0], [0, 2], [3], [0, 3], [4], [0, 4], [2, 0], [0, 5], [0, 0, 1], [5], [0, 6], [6], [0, 7], [0, 1, 0], [1, 1], [7], [0, 8], [0, 0, 2], [3, 0], [0, 9], [8], [9], [1, 0, 0], [0, 2, 0], [1, 2], [0, 0, 3], [4, 0], [2, 1], [0, 0, 4], [0, 0, 5], [0, 0, 0, 0], [0, 1, 1], [0, 0, 6], [0, 3, 0], [5, 0], [1, 3], [0, 0, 7], [0, 0, 8], [0, 0, 9], [6, 0], [0, 4, 0], [1, 4], [7, 0], [0, 1, 2], [2, 0, 0], [3, 1], [2, 2], [8, 0], [0, 5, 0], [1, 5], [1, 0, 1], [0, 2, 1], [9, 0], [0, 6, 0], [0, 0, 0, 1], [1, 6], [0, 7, 0]]",
|
|
f"--max_ite=40", f"--batch_size={batch_size}",
|
|
f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}"
|
|
]
|
|
if use_dynamic_tree:
|
|
summary_cmd.extend(
|
|
[f"--eagle_dynamic_tree_max_top_k={3}", "--eagle_use_dynamic_tree"])
|
|
|
|
venv_check_call(llm_venv, summary_cmd)
|
|
|
|
|
|
# TODO: remove skip_post_blackwell after Speculative decoding is supported.
|
|
@skip_post_blackwell
|
|
@skip_pre_ada
|
|
@pytest.mark.parametrize("batch_size", [8], ids=['bs8'])
|
|
@pytest.mark.parametrize("data_type", ['float16'])
|
|
@pytest.mark.parametrize("eagle_model_roots", ["llama3.1-eagle-8b-hf_v0.5"],
|
|
indirect=True)
|
|
def test_llm_eagle_1gpu_modelopt_ckpt(batch_size, data_type, eagle_model_roots,
|
|
eagle_example_root, llm_datasets_root,
|
|
llm_rouge_root, llm_venv, cmodel_dir,
|
|
engine_dir):
|
|
print("Build engines...")
|
|
model_name = "eagle"
|
|
|
|
# Although the datatype is float16, the actual weights are FP8.
|
|
# The datatype in the convert stage is used for the input and output of the plugin.
|
|
|
|
model_dir = convert_weights(llm_venv=llm_venv,
|
|
example_root=eagle_example_root,
|
|
cmodel_dir=cmodel_dir,
|
|
model=model_name,
|
|
model_path=eagle_model_roots,
|
|
data_type=data_type)
|
|
|
|
build_cmd = [
|
|
"trtllm-build",
|
|
f"--checkpoint_dir={model_dir}",
|
|
f"--output_dir={engine_dir}",
|
|
f"--max_beam_width=1",
|
|
"--use_paged_context_fmha=enable",
|
|
f"--max_batch_size={batch_size}",
|
|
"--speculative_decoding_mode=eagle",
|
|
"--multiple_profiles=enable" # also test multiple_profiles
|
|
]
|
|
|
|
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
|
|
|
|
print("Run run...")
|
|
|
|
run_cmd = [
|
|
f"{eagle_example_root}/../run.py", f"--engine_dir={engine_dir}",
|
|
f"--tokenizer_dir={eagle_model_roots}",
|
|
"--eagle_choices=[[0], [0, 0], [1], [0, 1], [2], [0, 0, 0], [1, 0], [0, 2], [3], [0, 3], [4], [0, 4], [2, 0], [0, 5], [0, 0, 1], [5], [0, 6], [6], [0, 7], [0, 1, 0], [1, 1], [7], [0, 8], [0, 0, 2], [3, 0], [0, 9], [8], [9], [1, 0, 0], [0, 2, 0], [1, 2], [0, 0, 3], [4, 0], [2, 1], [0, 0, 4], [0, 0, 5], [0, 0, 0, 0], [0, 1, 1], [0, 0, 6], [0, 3, 0], [5, 0], [1, 3], [0, 0, 7], [0, 0, 8], [0, 0, 9], [6, 0], [0, 4, 0], [1, 4], [7, 0], [0, 1, 2], [2, 0, 0], [3, 1], [2, 2], [8, 0], [0, 5, 0], [1, 5], [1, 0, 1], [0, 2, 1], [9, 0], [0, 6, 0], [0, 0, 0, 1], [1, 6], [0, 7, 0]]",
|
|
"--max_output_len=100"
|
|
]
|
|
|
|
venv_check_call(llm_venv, run_cmd)
|
|
|
|
|
|
def test_with_dummy_eagle(hf_model_root,
|
|
use_dynamic_tree,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
batch_size=8,
|
|
data_type="bfloat16"):
|
|
print("Build engines...")
|
|
model_name = "eagle"
|
|
|
|
print("Creating dummy Eagle heads...")
|
|
get_dummy_spec_decoding_heads(hf_model_dir=hf_model_root,
|
|
save_dir=llm_venv.get_working_directory(),
|
|
mode='eagle')
|
|
|
|
eagle_model_root = os.path.join(llm_venv.get_working_directory(), 'fp8')
|
|
|
|
ckpt_dir = convert_weights(llm_venv=llm_venv,
|
|
example_root=eagle_example_root,
|
|
cmodel_dir=cmodel_dir,
|
|
model=model_name,
|
|
model_path=eagle_model_root,
|
|
data_type=data_type)
|
|
|
|
build_cmd = [
|
|
"trtllm-build",
|
|
f"--checkpoint_dir={ckpt_dir}",
|
|
f"--output_dir={engine_dir}",
|
|
f"--gpt_attention_plugin={data_type}",
|
|
f"--gemm_plugin={data_type}",
|
|
f"--max_beam_width=1",
|
|
"--remove_input_padding=enable",
|
|
"--context_fmha=enable",
|
|
"--use_paged_context_fmha=enable",
|
|
"--max_input_len=1024",
|
|
"--max_seq_len=1536",
|
|
f"--max_batch_size={batch_size}",
|
|
"--paged_kv_cache=enable",
|
|
'--speculative_decoding_mode=eagle',
|
|
]
|
|
check_call(" ".join(build_cmd), shell=True, env=llm_venv._new_env)
|
|
|
|
print("Run run...")
|
|
run_cmd = [
|
|
f"{eagle_example_root}/../run.py",
|
|
"--max_output_len=100",
|
|
f"--tokenizer_dir={hf_model_root}",
|
|
"--log_level=verbose",
|
|
f"--engine_dir={engine_dir}",
|
|
]
|
|
if use_dynamic_tree:
|
|
run_cmd.extend(
|
|
[f"--eagle_dynamic_tree_max_top_k={3}", "--eagle_use_dynamic_tree"])
|
|
|
|
venv_check_call(llm_venv, run_cmd)
|
|
|
|
print("Run summarize...")
|
|
summary_cmd = [
|
|
f"{eagle_example_root}/../summarize.py", "--test_trt_llm",
|
|
"--hf_model_dir", f"{hf_model_root}", "--tokenizer_dir",
|
|
f"{hf_model_root}", f"--engine_dir={engine_dir}",
|
|
"--eagle_choices=[[0], [0, 0], [1], [0, 1], [2], [0, 0, 0], [1, 0], [0, 2], [3], [0, 3], [4], [0, 4], [2, 0], [0, 5], [0, 0, 1], [5], [0, 6], [6], [0, 7], [0, 1, 0], [1, 1], [7], [0, 8], [0, 0, 2], [3, 0], [0, 9], [8], [9], [1, 0, 0], [0, 2, 0], [1, 2], [0, 0, 3], [4, 0], [2, 1], [0, 0, 4], [0, 0, 5], [0, 0, 0, 0], [0, 1, 1], [0, 0, 6], [0, 3, 0], [5, 0], [1, 3], [0, 0, 7], [0, 0, 8], [0, 0, 9], [6, 0], [0, 4, 0], [1, 4], [7, 0], [0, 1, 2], [2, 0, 0], [3, 1], [2, 2], [8, 0], [0, 5, 0], [1, 5], [1, 0, 1], [0, 2, 1], [9, 0], [0, 6, 0], [0, 0, 0, 1], [1, 6], [0, 7, 0]]",
|
|
f"--max_ite=40", f"--batch_size={batch_size}",
|
|
f"--dataset_dir={llm_datasets_root}", f"--rouge_dir={llm_rouge_root}"
|
|
]
|
|
if use_dynamic_tree:
|
|
summary_cmd.extend(
|
|
[f"--eagle_dynamic_tree_max_top_k={3}", "--eagle_use_dynamic_tree"])
|
|
|
|
venv_check_call(llm_venv, summary_cmd)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("llama_model_root",
|
|
['llama-v2-7b-hf', 'llama-3.1-8b', 'llama-3.2-1b'],
|
|
indirect=True)
|
|
def test_llama_eagle_1gpu(llama_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=llama_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("code_llama_model_root", ['CodeLlama-7b-Instruct'],
|
|
indirect=True)
|
|
def test_codellama_eagle_1gpu(code_llama_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=code_llama_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("llm_mistral_model_root", ['mistral-7b-v0.1'],
|
|
indirect=True)
|
|
def test_mistral_eagle_1gpu(llm_mistral_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=llm_mistral_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("mistral_nemo_model_root", ['Mistral-Nemo-12b-Base'],
|
|
indirect=True)
|
|
def test_mistral_nemo_eagle_1gpu(mistral_nemo_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=mistral_nemo_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("llm_qwen_model_root", [
|
|
"qwen_7b_chat", "qwen1.5_7b_chat", "qwen2_7b_instruct",
|
|
"qwen2_0.5b_instruct", "qwen2.5_1.5b_instruct"
|
|
],
|
|
indirect=True)
|
|
def test_qwen_eagle_1gpu(llm_qwen_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=llm_qwen_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|
|
|
|
|
|
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
|
|
ids=['eagle1', 'eagle2'])
|
|
@pytest.mark.parametrize("llm_phi_model_root", [
|
|
"phi-2", "Phi-3-mini-128k-instruct", "Phi-3-small-128k-instruct",
|
|
"Phi-3.5-mini-instruct"
|
|
],
|
|
indirect=True)
|
|
def test_phi_eagle_1gpu(llm_phi_model_root,
|
|
eagle_example_root,
|
|
llm_datasets_root,
|
|
llm_rouge_root,
|
|
llm_venv,
|
|
cmodel_dir,
|
|
engine_dir,
|
|
use_dynamic_tree,
|
|
batch_size=8,
|
|
data_type='bfloat16'):
|
|
|
|
test_with_dummy_eagle(hf_model_root=llm_phi_model_root,
|
|
eagle_example_root=eagle_example_root,
|
|
llm_venv=llm_venv,
|
|
cmodel_dir=cmodel_dir,
|
|
engine_dir=engine_dir,
|
|
batch_size=batch_size,
|
|
data_type=data_type,
|
|
use_dynamic_tree=use_dynamic_tree,
|
|
llm_datasets_root=llm_datasets_root,
|
|
llm_rouge_root=llm_rouge_root)
|