EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Drop torchair (#4814)

Browse Source 
aclgraph is stable and fast now. Let's drop torchair graph mode now.

TODO: some logic to adapt torchair should be cleaned up as well. We'll
do it in the following PR.

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
This commit is contained in:
wangxiyuan
2025-12-10 09:20:40 +08:00
committed by GitHub
parent ba9cda9dfd
commit 835b4c8f1d
84 changed files with 77 additions and 16881 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tests/e2e/multicard/test_offline_inference_distributed.py
View File

@@ -78,9 +78,6 @@ def test_models_distributed_DeepSeek_multistream_moe():
tensor_parallel_size=2,
distributed_executor_backend="mp",
additional_config={
"torchair_graph_config": {
"enabled": True,
},
"enable_multistream_moe": True,
"refresh": True,
},
@@ -144,17 +141,12 @@ def test_models_distributed_DeepSeek_W4A8DYNAMIC(model):
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download(model),
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
enforce_eager=True,
enable_expert_parallel=True,
additional_config={"torchair_graph_config": {
"enabled": False,
}},
) as vllm_model:
with VllmRunner(snapshot_download(model),
dtype="auto",
tensor_parallel_size=2,
quantization="ascend",
enforce_eager=True,
enable_expert_parallel=True) as vllm_model:
vllm_model.generate_greedy(prompts, max_tokens)

290
tests/e2e/multicard/test_torchair_graph_mode.py
View File

@@ -1,290 +0,0 @@
#
# 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.
# This file is a part of the vllm-ascend project.
#
"""Compare the short outputs of HF and vLLM when using greedy sampling.
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"
def _deepseek_torchair_test_fixture(
additional_config: Dict,
*,
tensor_parallel_size=2,
use_v1_schduler=False,
):
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
kwargs = {}
if not use_v1_schduler:
kwargs = {
"refresh": True,
}
additional_config.update(**kwargs)
with VllmRunner(
"vllm-ascend/DeepSeek-V3-Pruning",
dtype="half",
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend="mp",
additional_config=additional_config,
) as vllm_model:
# use greedy sampler to make sure the generated results are fix
vllm_output = vllm_model.generate_greedy(example_prompts, 5)
# NOTE: vllm-ascend/DeepSeek-V3-Pruning is a random weight of
# DeepSeek-V3 with 2 hidden layers, thus the golden results seems
# inaccurate. This will only change if accuracy improves with the
# official weights of DeepSeek-V3.
golden_results = [
'Hello, my name is下载早点向前很有่อง',
'The president of the United States isSender)## physiological Albany',
'The capital of France is Rocky转角 hospitalizedinterval sparked',
'The future of AI is её asegο BIOS一扫',
]
assert len(golden_results) == len(vllm_output)
for i in range(len(vllm_output)):
assert golden_results[i] == vllm_output[i][1]
print(f"Generated text: {vllm_output[i][1]!r}")
def test_e2e_deepseekv3_with_torchair():
additional_config = {
"torchair_graph_config": {
"enabled": True,
},
}
_deepseek_torchair_test_fixture(additional_config)
def test_e2e_deepseekv3_with_torchair_ms_mla():
additional_config = {
"torchair_graph_config": {
"enabled": True,
"enable_multistream_mla": True,
},
}
_deepseek_torchair_test_fixture(additional_config)
@pytest.mark.skip("accuracy test failed. Fix me")
def test_e2e_deepseekv3_with_torchair_v1scheduler():
additional_config = {
"torchair_graph_config": {
"enabled": True,
},
}
_deepseek_torchair_test_fixture(additional_config, use_v1_schduler=True)
def _pangu_torchair_test_fixture(
additional_config: Dict,
*,
tensor_parallel_size=2,
):
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# torchair is only work without chunked-prefill now
kwargs = {
"refresh": True,
}
additional_config.update(**kwargs)
with VllmRunner(
"vllm-ascend/pangu-pro-moe-pruing",
dtype="half",
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend="mp",
additional_config=additional_config,
enable_expert_parallel=True,
) as vllm_model:
# use greedy sampler to make sure the generated results are fix
vllm_output = vllm_model.generate_greedy(example_prompts, 5)
# NOTE: vllm-ascend/pangu-pro-moe-pruing is only part of PanguProMoE
# with 2 hidden layers, thus the golden results seems inaccurate.
# This will only change if accuracy changes with the official weights
# of PanguProMoE.
golden_results = [
'Hello, my name is Remempondeprecatedmiot忱',
'The president of the United States is Remem下的一个 rever ceremoni Segnali',
'The capital of France is Rememvoud administrativ Remem投',
'The future of AI isotope Segnali Zoeken精细化 supus',
]
assert len(golden_results) == len(vllm_output)
for i in range(len(vllm_output)):
assert golden_results[i] == vllm_output[i][1]
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": {
"enabled": True,
},
}
_pangu_torchair_test_fixture(additional_config)
def _qwen_torchair_test_fixture(
model,
tp,
enable_expert_parallel,
):
# The current access control does not support 16 cards,
# so the MC2 operator in Qwen's graph mode cannot run.
# Once 16-card support is available,
# this e2e can be switched to graph mode.
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
additional_config = {
"torchair_graph_config": {
"enabled": False,
},
"refresh": True,
}
with VllmRunner(
model,
dtype="half",
tensor_parallel_size=tp,
distributed_executor_backend="mp",
enforce_eager=True,
additional_config=additional_config,
enable_expert_parallel=enable_expert_parallel,
) as vllm_model:
# use greedy sampler to make sure the generated results are fix
vllm_output = vllm_model.generate_greedy(example_prompts, 5)
# NOTE: vllm-ascend/pangu-pro-moe-pruing is only part of PanguProMoE
# with 2 hidden layers, thus the golden results seems inaccurate.
# This will only change if accuracy changes with the official weights
# of PanguProMoE.
golden_results = [
'Hello, my name is Remempondeprecatedmiot忱',
'The president of the United States is Remem下的一个 rever ceremoni Segnali',
'The capital of France is Rememvoud administrativ Remem投',
'The future of AI isotope Segnali Zoeken精细化 supus',
]
assert len(golden_results) == len(vllm_output)
for i in range(len(vllm_output)):
print(f"Generated text: {vllm_output[i][1]!r}")
def test_e2e_qwen2_with_torchair():
_qwen_torchair_test_fixture("Qwen/Qwen2.5-0.5B-Instruct", 2, False)
def test_e2e_qwen3_moe_with_torchair():
_qwen_torchair_test_fixture("Qwen/Qwen3-30B-A3B", 2, True)
# test deepseek-v2-lite
def _deepseek_v2_lite_torchair_test_fixure(
additional_config: Dict,
*,
tensor_parallel_size=2,
use_v1_schduler=False,
):
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
kwargs = {}
if not use_v1_schduler:
kwargs = {
"refresh": True,
}
additional_config.update(**kwargs)
with VllmRunner(
"deepseek-ai/DeepSeek-V2-Lite",
dtype="half",
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend="mp",
additional_config=additional_config,
) as vllm_model:
vllm_output = vllm_model.generate_greedy(example_prompts, 5)
# NOTE: deepseek-ai/DeepSeek-V2-Lite is a random weight of
# DeepSeek-V2-Lite with 2 hidden layers, thus the golden results seems
# inaccurate. This will only change if accuracy improves with the
# official weights of DeepSeek-V2-Lite.
for i in range(len(vllm_output)):
generated_text = vllm_output[i][1]
assert len(
generated_text.strip()) > 0, f"The {i}-th output is null, failed"
def test_e2e_deepseekv2lite_with_torchair():
additional_config = {
"torchair_graph_config": {
"enabled": True,
},
}
_deepseek_v2_lite_torchair_test_fixure(additional_config)
def test_e2e_deepseekv2lite_with_torchair_v1scheduler():
additional_config = {
"torchair_graph_config": {
"enabled": True,
},
}
_deepseek_v2_lite_torchair_test_fixure(additional_config,
use_v1_schduler=True)
# kv_cache enable e2e test
def test_e2e_deepseekv2lite_with_nz():
additional_config = {
"torchair_graph_config": {
"enabled": True,
"enable_kv_nz": True,
},
}
_deepseek_v2_lite_torchair_test_fixure(additional_config)