CookieApril/enginex-ascend-910-vllm
1
0
Fork 0
forked from EngineX-Ascend/enginex-ascend-910-vllm
Code Pull Requests Activity

init v0.11.0rc0

Browse Source
This commit is contained in:
luopingyi
2025-10-14 10:38:28 +08:00
parent 67afd0ea78
commit 66dc16f966
278 changed files with 28130 additions and 11708 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
tests/e2e/multicard/test_expert_parallel.py
View File

@@ -14,14 +14,24 @@ def test_e2e_ep_correctness(model_name):
]
max_tokens = 5
with VllmRunner(model_name, tensor_parallel_size=2,
enforce_eager=True) as vllm_model:
# FIXME: Really strange that chunked prefill might lead to different results, investigate further
with VllmRunner(
model_name,
tensor_parallel_size=2,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=True) as vllm_model:
tp_output = vllm_model.generate_greedy(example_prompts, max_tokens)
with VllmRunner(model_name,
tensor_parallel_size=2,
enable_expert_parallel=True,
enforce_eager=True) as vllm_model:
with VllmRunner(
model_name,
tensor_parallel_size=2,
enable_expert_parallel=True,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=True) as vllm_model:
ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(

60
tests/e2e/multicard/test_offline_inference_distributed.py
View File

@@ -23,6 +23,7 @@ Run `pytest tests/test_offline_inference.py`.
import os
from unittest.mock import patch
import pytest
from modelscope import snapshot_download # type: ignore
from vllm import SamplingParams
@@ -30,6 +31,15 @@ from tests.e2e.conftest import VllmRunner
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
QWEN_DENSE_MODELS = [
"vllm-ascend/Qwen3-8B-W8A8", "vllm-ascend/Qwen2.5-0.5B-Instruct-W8A8"
]
DEEPSEEK_W4A8_MODELS = [
"vllm-ascend/DeepSeek-V3-W4A8-Pruing",
"vllm-ascend/DeepSeek-V3.1-W4A8-puring"
]
def test_models_distributed_QwQ():
example_prompts = [
@@ -61,8 +71,8 @@ def test_models_distributed_DeepSeek_multistream_moe():
additional_config={
"torchair_graph_config": {
"enabled": True,
"enable_multistream_moe": True,
},
"enable_multistream_moe": True,
"ascend_scheduler_config": {
"enabled": True,
},
@@ -104,14 +114,15 @@ def test_models_distributed_Qwen3_W4A8DYNAMIC():
vllm_model.generate_greedy(example_prompts, max_tokens)
@pytest.mark.parametrize("model", DEEPSEEK_W4A8_MODELS)
@patch.dict(os.environ, {"VLLM_ASCEND_MLA_PA": "1"})
def test_models_distributed_DeepSeek_W4A8DYNAMIC():
def test_models_distributed_DeepSeek_W4A8DYNAMIC(model):
prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download("vllm-ascend/DeepSeek-V3-W4A8-Pruing"),
snapshot_download(model),
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
@@ -150,3 +161,46 @@ def test_sp_for_qwen3_moe() -> None:
enable_expert_parallel=True,
enforce_eager=True) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)
@pytest.mark.parametrize("enforce_eager", [True, False])
@pytest.mark.parametrize("model", QWEN_DENSE_MODELS)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE": "1"})
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_FLASHCOMM": "1"})
def test_models_distributed_Qwen_Dense_with_flashcomm_v1(model, enforce_eager):
example_prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download(model),
max_model_len=8192,
enforce_eager=enforce_eager,
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
@pytest.mark.parametrize("enforce_eager", [True, False])
@pytest.mark.parametrize("model", QWEN_DENSE_MODELS)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE": "1"})
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"})
def test_models_distributed_Qwen_Dense_with_prefetch_mlp_weight(
model, enforce_eager):
example_prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download(model),
max_model_len=8192,
enforce_eager=enforce_eager,
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)

42
tests/e2e/multicard/test_prefix_caching.py
View File

@@ -116,20 +116,22 @@ def test_prefix_cache_with_ascend_scheduler(model: str,
prefix_cache_output = vllm_model.generate_greedy(
INPUT_PROMPTS, max_tokens)
with VllmRunner(model,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
'enable_prefix_caching': True,
"enable_chunked_prefill": True,
},
},
enforce_eager=True,
max_model_len=2048,
tensor_parallel_size=2,
gpu_memory_utilization=0.7) as vllm_model:
chunk_prefill_prefix_cache_output = vllm_model.generate_greedy(
INPUT_PROMPTS, max_tokens)
# TODO: enable apc and chunked prefill with ascend scheduler will lead accuracy problem.
# Disable it now. Fix it or drop the ascend scheduler in the future.
# with VllmRunner(model,
# additional_config={
# 'ascend_scheduler_config': {
# 'enabled': True,
# 'enable_prefix_caching': True,
# "enable_chunked_prefill": True,
# },
# },
# enforce_eager=True,
# max_model_len=2048,
# tensor_parallel_size=2,
# gpu_memory_utilization=0.7) as vllm_model:
# chunk_prefill_prefix_cache_output = vllm_model.generate_greedy(
# INPUT_PROMPTS, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_output,
@@ -138,9 +140,9 @@ def test_prefix_cache_with_ascend_scheduler(model: str,
name_1="prefix_cache_output",
)
check_outputs_equal(
outputs_0_lst=chunk_prefill_prefix_cache_output,
outputs_1_lst=prefix_cache_output,
name_0="chunk_prefill_prefix_cache_output",
name_1="prefix_cache_output",
)
# check_outputs_equal(
# outputs_0_lst=chunk_prefill_prefix_cache_output,
# outputs_1_lst=prefix_cache_output,
# name_0="chunk_prefill_prefix_cache_output",
# name_1="prefix_cache_output",
# )

1
tests/e2e/multicard/test_qwen3_moe.py
View File

@@ -66,7 +66,6 @@ def test_models_distributed_Qwen3_MOE_W8A8():
max_model_len=8192,
tensor_parallel_size=2,
quantization="ascend",
enforce_eager=True,
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)

3
tests/e2e/multicard/test_torchair_graph_mode.py
View File

@@ -22,6 +22,8 @@ Run `pytest tests/multicard/test_torchair_graph_mode.py`.
import os
from typing import Dict
import pytest
from tests.e2e.conftest import VllmRunner
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
@@ -153,6 +155,7 @@ def _pangu_torchair_test_fixture(
print(f"Generated text: {vllm_output[i][1]!r}")
@pytest.mark.skip("skipping test_e2e_pangu_with_torchair")
def test_e2e_pangu_with_torchair():
additional_config = {
"torchair_graph_config": {

188
tests/e2e/multicard/test_weight_loader.py Normal file
View File

@@ -0,0 +1,188 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# 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.
#
"""
Compare the outputs of vLLM with and without aclgraph.
Run `pytest tests/multicard/test_external_launcher.py`.
"""
import os
import subprocess
import sys
import pytest
import torch_npu
MOE_MODELS = ["Qwen/Qwen3-30B-A3B"]
MODELS = ["Qwen/Qwen3-8B"]
DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10]
@pytest.mark.parametrize("model", MOE_MODELS)
def test_external_launcher_eager(model):
script = script = "/usr/local/python3.11.13/bin/python3.11/__w/vllm-ascend/tests/examples/test_weight_loader.py"
env = os.environ.copy()
# TODO: Change to 2 when ci machine has 4 cards
cmd = [
sys.executable,
str(script),
"--model",
model,
"--tp-size",
"2",
"--proc-per-node",
"2",
"--trust-remote-code",
"--enforce-eager",
"--enable-expert-parallel",
"--enable-sleep-mode",
"--model-weight-gib",
"20",
]
print(f"Running subprocess: {' '.join(cmd)}")
proc = subprocess.run(
cmd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
timeout=600,
)
output = proc.stdout.decode()
print(output)
assert "TP RANKS: [0]" in output
assert "TP RANKS: [1]" in output
assert "Generated text:" in output
assert proc.returncode == 0
@pytest.mark.parametrize("model", MOE_MODELS)
def test_external_launcher_aclgraph(model):
script = "/usr/local/python3.11.13/bin/python3.11/__w/vllm-ascend/tests/examples/test_weight_loader.py"
env = os.environ.copy()
# TODO: Change to 2 when ci machine has 4 cards
cmd = [
sys.executable,
str(script),
"--model",
model,
"--tp-size",
"2",
"--proc-per-node",
"2",
"--trust-remote-code",
"--enable-expert-parallel",
"--enable-sleep-mode",
"--model-weight-gib",
"20",
]
print(f"Running subprocess: {' '.join(cmd)}")
proc = subprocess.run(
cmd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
timeout=600,
)
output = proc.stdout.decode()
print(output)
assert "TP RANKS: [0]" in output
assert "TP RANKS: [1]" in output
assert "Generated text:" in output
assert proc.returncode == 0
@pytest.mark.parametrize("model", MODELS)
def test_external_launcher_dense(model):
script = "/usr/local/python3.11.13/bin/python3.11/__w/vllm-ascend/tests/examples/test_weight_loader.py"
env = os.environ.copy()
# TODO: Change to 2 when ci machine has 4 cards
cmd = [
sys.executable,
str(script),
"--model",
model,
"--tp-size",
"2",
"--proc-per-node",
"2",
"--trust-remote-code",
"--enable-sleep-mode",
"--model-weight-gib",
"20",
]
print(f"Running subprocess: {' '.join(cmd)}")
proc = subprocess.run(
cmd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
timeout=600,
)
output = proc.stdout.decode()
print(output)
assert "TP RANKS: [0]" in output
assert "TP RANKS: [1]" in output
assert "Generated text:" in output
assert proc.returncode == 0
@pytest.mark.parametrize("model", MODELS)
def test_external_launcher_dense_eager(model):
script = "/usr/local/python3.11.13/bin/python3.11/__w/vllm-ascend/tests/examples/test_weight_loader.py"
env = os.environ.copy()
# TODO: Change to 2 when ci machine has 4 cards
cmd = [
sys.executable,
str(script),
"--model",
model,
"--tp-size",
"2",
"--proc-per-node",
"2",
"--trust-remote-code",
"--enforce-eager",
"--enable-sleep-mode",
"--model-weight-gib",
"20",
]
print(f"Running subprocess: {' '.join(cmd)}")
proc = subprocess.run(
cmd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
timeout=600,
)
output = proc.stdout.decode()
print(output)
assert "TP RANKS: [0]" in output
assert "TP RANKS: [1]" in output
assert "Generated text:" in output
assert proc.returncode == 0