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

59
tests/e2e/310p/test_offline_inference_parallel_310p.py
View File

@@ -1,59 +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.
import pytest
import vllm # noqa: F401
import vllm_ascend # noqa: F401
from tests.e2e.conftest import VllmRunner
# Pangu local model path
MODELS = [
"IntervitensInc/pangu-pro-moe-model",
]
# set additional config for ascend scheduler and torchair graph
ADDITIONAL_CONFIG = [{
"additional_config": {
"torchair_graph_config": {
"enabled": True
}
}
}]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float16"])
@pytest.mark.parametrize("max_tokens", [5])
@pytest.mark.parametrize("enfore_eager", [True, False])
@pytest.mark.parametrize("additional_config", ADDITIONAL_CONFIG)
def test_pangu_model(model: str, dtype: str, max_tokens: int,
enfore_eager: bool, additional_config: dict) -> None:
if enfore_eager:
additional_config = {}
example_prompts = [
"Hello, my name is",
"The future of AI is",
]
with VllmRunner(model,
tensor_parallel_size=4,
dtype=dtype,
max_model_len=1024,
enforce_eager=True,
enable_expert_parallel=True,
additional_config=additional_config,
distributed_executor_backend="mp") as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)

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)

4
tests/e2e/nightly/features/test_mtpx_deepseek_r1_0528_w8a8.py
View File

@@ -73,7 +73,6 @@ async def test_models(model: str, mode: str) -> None:
"VLLM_RPC_TIMEOUT": "3600000",
"VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000"
}
additional_config: dict[str, Any] = {}
speculative_config = {"num_speculative_tokens": 2, "method": "mtp"}
compilation_config = {
"cudagraph_capture_sizes": [56],
@@ -104,7 +103,6 @@ async def test_models(model: str, mode: str) -> None:
["--speculative-config",
json.dumps(speculative_config)])
server_args.extend(["--gpu-memory-utilization", "0.92"])
additional_config["torchair_graph_config"] = {"enabled": True}
aisbench_cases = aisbench_gsm8k
if mode == "mtp3":
env_dict["HCCL_OP_EXPANSION_MODE"] = "AIV"
@@ -117,9 +115,7 @@ async def test_models(model: str, mode: str) -> None:
server_args.extend(
["--compilation-config",
json.dumps(compilation_config)])
additional_config["torchair_graph_config"] = {"enabled": False}
aisbench_cases = aisbench_aime
server_args.extend(["--additional-config", json.dumps(additional_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,
}

7
tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py
View File

@@ -74,13 +74,6 @@ async def test_models(model: str) -> None:
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
}
additional_config = {
"torchair_graph_config": {
"enabled": True,
"enable_multistream_moe": False,
"enable_multistream_mla": True,
"graph_batch_size": [16],
"use_cached_graph": True
},
"chunked_prefill_for_mla": True,
"enable_weight_nz_layout": True
}

12
tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py
View File

@@ -29,7 +29,6 @@ MODELS = [
]
MODES = [
"torchair",
"single",
"aclgraph",
"aclgraph_mlapo",
@@ -78,13 +77,6 @@ async def test_models(model: str, mode: str) -> None:
}
speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
additional_config = {
"torchair_graph_config": {
"enabled": True,
"enable_multistream_moe": False,
"enable_multistream_mla": True,
"graph_batch_sizes": [16],
"use_cached_graph": True
},
"chunked_prefill_for_mla": True,
"enable_weight_nz_layout": True
}
@@ -99,12 +91,8 @@ async def test_models(model: str, mode: str) -> None:
]
if mode == "single":
server_args.append("--enforce-eager")
additional_config["torchair_graph_config"] = {"enabled": False}
if mode == "aclgraph":
additional_config["torchair_graph_config"] = {"enabled": False}
if mode == "aclgraph_mlapo":
env_dict["VLLM_ASCEND_ENABLE_MLAPO"] = "1"
additional_config["torchair_graph_config"] = {"enabled": False}
server_args.extend(["--additional-config", json.dumps(additional_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,

3
tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
View File

@@ -68,9 +68,6 @@ async def test_models(model: str) -> None:
"cudagraph_mode": "FULL_DECODE_ONLY"
}
additional_config: dict[str, Any] = {
"torchair_graph_config": {
"enabled": True
},
"enable_shared_expert_dp": False,
"multistream_overlap_shared_expert": False,
"dynamic_eplb": True,

26
tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py
View File

@@ -72,27 +72,13 @@ async def test_models(model: str, tp_size: int, dp_size: int,
port = get_open_port()
env_dict = {"HCCL_BUFFSIZE": "1024", "VLLM_ASCEND_ENABLE_MLAPO": "0"}
server_args = [
"--no-enable-prefix-caching",
"--enable-expert-parallel",
"--no-enable-prefix-caching", "--enable-expert-parallel",
"--tensor-parallel-size",
str(tp_size),
"--data-parallel-size",
str(dp_size),
"--port",
str(port),
"--max-model-len",
"16384",
"--max-num-batched-tokens",
"16384",
"--block-size",
"16",
"--trust-remote-code",
"--quantization",
"ascend",
"--gpu-memory-utilization",
"0.9",
"--additional-config",
'{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}',
str(tp_size), "--data-parallel-size",
str(dp_size), "--port",
str(port), "--max-model-len", "16384", "--max-num-batched-tokens",
"16384", "--block-size", "16", "--trust-remote-code", "--quantization",
"ascend", "--gpu-memory-utilization", "0.9"
]
if full_graph:
server_args += [

64
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-A2-torchair.yaml
View File

@@ -1,64 +0,0 @@
test_name: "test DeepSeek-R1-W8A8 torchair on A2"
model: "vllm-ascend/DeepSeek-R1-0528-W8A8"
num_nodes: 2
npu_per_node: 8
env_common:
VLLM_USE_MODELSCOPE: true
HCCL_BUFFSIZE: 1024
SERVER_PORT: 8080
OMP_PROC_BIND: false
OMP_NUM_THREADS: 10
deployment:
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
--host 0.0.0.0
--port $SERVER_PORT
--data-parallel-size 4
--data-parallel-size-local 2
--data-parallel-address $LOCAL_IP
--data-parallel-rpc-port 13399
--no-enable-prefix-caching
--max-num-seqs 16
--tensor-parallel-size 4
--max-model-len 36864
--max-num-batched-tokens 6000
--enable-expert-parallel
--trust-remote-code
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
--headless
--data-parallel-size 4
--data-parallel-rpc-port 13399
--data-parallel-size-local 2
--data-parallel-start-rank 2
--data-parallel-address $MASTER_IP
--no-enable-prefix-caching
--max-num-seqs 16
--tensor-parallel-size 4
--max-model-len 36864
--max-num-batched-tokens 6000
--enable-expert-parallel
--trust-remote-code
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
benchmarks:
acc:
case_type: accuracy
dataset_path: vllm-ascend/gsm8k
request_conf: vllm_api_general_chat
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
max_out_len: 32768
batch_size: 512
baseline: 95
threshold: 5

8
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-EPLB.yaml
View File

@@ -58,7 +58,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
'{"enable_prefill_optimizations":true,"enable_weight_nz_layout":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
-
server_cmd: >
@@ -96,7 +96,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
'{"enable_prefill_optimizations":true,"enable_weight_nz_layout":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -135,7 +135,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":true,"enable_multistream_mla":true,"graph_batch_sizes":[28],"use_cached_graph":true,"enable_super_kernel":false},"multistream_overlap_shared_expert":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
'{"multistream_overlap_shared_expert":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -173,7 +173,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":true,"enable_multistream_mla":true,"graph_batch_sizes":[28],"use_cached_graph":true,"enable_super_kernel":false},"multistream_overlap_shared_expert":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
'{"multistream_overlap_shared_expert":true,"dynamic_eplb":true,"num_iterations_eplb_update":2048,"num_wait_worker_iterations":200}'
benchmarks:
perf:
case_type: performance

8
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8.yaml
View File

@@ -57,7 +57,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
'{"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
@@ -95,7 +95,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
'{"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -134,7 +134,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":true,"enable_multistream_mla":true,"graph_batch_sizes":[28],"use_cached_graph":true,"enable_super_kernel":false},"multistream_overlap_shared_expert":true}'
'{"multistream_overlap_shared_expert":true}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -172,7 +172,7 @@ deployment:
}
}'
--additional-config
'{"torchair_graph_config":{"enabled":true,"enable_multistream_mla":true,"graph_batch_sizes":[28],"use_cached_graph":true,"enable_super_kernel":false},"multistream_overlap_shared_expert":true}'
'{"multistream_overlap_shared_expert":true}'
benchmarks:
perf:
case_type: performance

1
tests/e2e/nightly/multi_node/config/models/DeepSeek-V3.yaml
View File

@@ -82,7 +82,6 @@ deployment:
--trust-remote-code
--no-enable-prefix-caching
--gpu-memory-utilization 0.9
--additional-config '{"torchair_graph_config":{"enabled":true}}'
--kv-transfer-config
'{"kv_connector": "MooncakeConnector",
"kv_role": "kv_consumer",

2
tests/e2e/nightly/multi_node/config/models/DeepSeek-V3_2-Exp-bf16.yaml
View File

@@ -29,7 +29,6 @@ deployment:
--trust-remote-code
--no-enable-prefix-caching
--gpu-memory-utilization 0.9
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
-
server_cmd: >
@@ -49,5 +48,4 @@ deployment:
--trust-remote-code
--no-enable-prefix-caching
--gpu-memory-utilization 0.92
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
benchmarks:

106
tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py
View File

@@ -1,106 +0,0 @@
from __future__ import annotations
import pytest
from vllm import SamplingParams
from vllm.config import CompilationConfig, CUDAGraphMode
from tests.e2e.conftest import VllmRunner
@pytest.fixture
def sampling_config():
return SamplingParams(temperature=0, max_tokens=256, ignore_eos=False)
@pytest.fixture
def model_name():
return "wemaster/deepseek_mtp_main_random_bf16"
def mtp_torchair_correctness(
sampling_config: SamplingParams,
model_name: str,
graph_mode: CUDAGraphMode = CUDAGraphMode.PIECEWISE,
):
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
'''
Compare the outputs of a original LLM and a speculative LLM
should be the same when using mtp speculative decoding.
'''
with VllmRunner(model_name,
tensor_parallel_size=1,
gpu_memory_utilization=0.7,
max_model_len=256,
enforce_eager=False,
additional_config={
"torchair_graph_config": {
"enabled": True,
"use_cached_graph": False,
"graph_batch_sizes": [1, 2, 4],
},
"multistream_overlap_shared_expert": "True"
}) as ref_llm:
ref_outputs = ref_llm.generate(example_prompts, sampling_config)
graph_mode_str = "PIECEWISE"
if graph_mode == CUDAGraphMode.FULL:
graph_mode_str = "FULL"
with VllmRunner(model_name,
tensor_parallel_size=1,
max_num_seqs=256,
gpu_memory_utilization=0.7,
distributed_executor_backend="mp",
enable_expert_parallel=True,
speculative_config={
"method": "mtp",
"num_speculative_tokens": 1,
},
enforce_eager=False,
max_model_len=2000,
compilation_config=CompilationConfig(
cudagraph_mode=graph_mode_str),
additional_config={
"torchair_graph_config": {
"enabled": True,
"use_cached_graph": False,
"graph_batch_sizes": [1, 2, 4],
},
"multistream_overlap_shared_expert": "True"
}) as spec_llm:
spec_outputs = spec_llm.generate(example_prompts, sampling_config)
matches = 0
misses = 0
for ref_output, spec_output in zip(ref_outputs, spec_outputs):
ref_token_ids = ref_output[0][0]
spec_token_ids = spec_output[0][0]
if ref_token_ids == spec_token_ids[:len(ref_token_ids)]:
matches += 1
else:
misses += 1
print(f"ref_output: {ref_output[1][0]}")
print(f"spec_output: {spec_output[1][0]}")
# Heuristic: expect at least 66% of the prompts to match exactly
# Upon failure, inspect the outputs to check for inaccuracy.
assert matches > int(0.66 * len(ref_outputs))
def test_mtp_torchair_correctness_piecewise(
sampling_config: SamplingParams,
model_name: str,
):
mtp_torchair_correctness(sampling_config, model_name)
def test_mtp_torchair_correctness_full(
sampling_config: SamplingParams,
model_name: str,
):
mtp_torchair_correctness(sampling_config, model_name, CUDAGraphMode.FULL)