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

Revert "drop ascend scheduler" (#4580)

Browse Source 
Reverts vllm-project/vllm-ascend#4498
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
This commit is contained in:
Mengqing Cao
2025-11-29 22:20:48 +08:00
committed by GitHub
parent 4dbe4fd123
commit 517fd9272d
52 changed files with 2948 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/e2e/310p/test_offline_inference_parallel_310p.py
View File

@@ -24,12 +24,15 @@ from tests.e2e.conftest import VllmRunner
MODELS = [
"IntervitensInc/pangu-pro-moe-model",
]
# set additional config for torchair graph
# set additional config for ascend scheduler and torchair graph
ADDITIONAL_CONFIG = [{
"additional_config": {
"torchair_graph_config": {
"enabled": True
},
"ascend_scheduler_config": {
"enabled": True,
}
}
}]

21
tests/e2e/multicard/test_expert_parallel.py
View File

@@ -15,14 +15,23 @@ def test_e2e_ep_correctness(model_name):
max_tokens = 5
# FIXME: Really strange that chunked prefill might lead to different results, investigate further
with VllmRunner(model_name, tensor_parallel_size=2,
enforce_eager=False) as vllm_model:
with VllmRunner(
model_name,
tensor_parallel_size=2,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=False) 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=False) as vllm_model:
with VllmRunner(
model_name,
tensor_parallel_size=2,
enable_expert_parallel=True,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=False) as vllm_model:
ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(

16
tests/e2e/multicard/test_fused_moe_allgather_ep.py
View File

@@ -49,7 +49,13 @@ def test_generate_with_allgather():
tensor_parallel_size=2,
max_model_len=1024,
dtype="auto",
enable_expert_parallel=True) as vllm_model:
enable_expert_parallel=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"chunked_prefill_enabled": False,
},
}) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)
@@ -70,5 +76,11 @@ def test_generate_with_alltoall():
tensor_parallel_size=2,
max_model_len=1024,
dtype="auto",
enable_expert_parallel=True) as vllm_model:
enable_expert_parallel=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"chunked_prefill_enabled": False,
},
}) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)

14
tests/e2e/multicard/test_offline_inference_distributed.py
View File

@@ -82,6 +82,9 @@ def test_models_distributed_DeepSeek_multistream_moe():
"enabled": True,
},
"enable_multistream_moe": True,
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
},
) as vllm_model:
@@ -151,9 +154,14 @@ def test_models_distributed_DeepSeek_W4A8DYNAMIC(model):
quantization="ascend",
enforce_eager=True,
enable_expert_parallel=True,
additional_config={"torchair_graph_config": {
"enabled": False,
}},
additional_config={
"torchair_graph_config": {
"enabled": False,
},
"ascend_scheduler_config": {
"enabled": True,
}
},
) as vllm_model:
vllm_model.generate_greedy(prompts, max_tokens)

66
tests/e2e/multicard/test_prefix_caching.py
View File

@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Compare the with and without prefix caching on V1 scheduler."""
"""Compare the with and without prefix caching on V1 scheduler or AscendScheduler."""
import pytest
@@ -84,3 +84,67 @@ def test_prefix_cache_with_v1_scheduler(model: str, max_tokens: int) -> None:
name_0="vllm_output",
name_1="prefix_cache_output",
)
@pytest.mark.skip(reason="Fix me, the accuracy is not correct")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_tokens", [50])
def test_prefix_cache_with_ascend_scheduler(model: str,
max_tokens: int) -> None:
with VllmRunner(model,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
enforce_eager=False,
max_model_len=2048,
tensor_parallel_size=2,
gpu_memory_utilization=0.7) as vllm_model:
vllm_output = vllm_model.generate_greedy(INPUT_PROMPTS, max_tokens)
with VllmRunner(model,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
'enable_prefix_caching': True,
},
},
enforce_eager=False,
max_model_len=2048,
tensor_parallel_size=2,
gpu_memory_utilization=0.7) as vllm_model:
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,
outputs_1_lst=prefix_cache_output,
name_0="vllm_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",
# )

9
tests/e2e/multicard/test_qwen3_next.py
View File

@@ -24,7 +24,6 @@ Run `pytest tests/e2e/multicard/test_qwen3_next.py`.
import os
from unittest.mock import patch
import pytest
from modelscope import snapshot_download # type: ignore
from tests.e2e.conftest import VllmRunner
@@ -64,8 +63,6 @@ def test_models_distributed_Qwen3_NEXT_TP4_FULL_DECODE_ONLY():
del vllm_model
@pytest.mark.skip(
reason="Qwen3-Next + MTP doesn't work with chunked prefill. Fix Me")
def test_models_distributed_Qwen3_NEXT_MTP_TP4_SIMILARITY():
example_prompts = [
"Hello, my name is",
@@ -92,6 +89,12 @@ def test_models_distributed_Qwen3_NEXT_MTP_TP4_SIMILARITY():
gpu_memory_utilization=0.8,
distributed_executor_backend="mp",
enforce_eager=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"enable_chunked_prefill": False
}
},
speculative_config={
"method": "qwen3_next_mtp",
"num_speculative_tokens": 1

12
tests/e2e/multicard/test_torchair_graph_mode.py
View File

@@ -44,6 +44,9 @@ def _deepseek_torchair_test_fixture(
kwargs = {}
if not use_v1_schduler:
kwargs = {
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
additional_config.update(**kwargs)
@@ -117,6 +120,9 @@ def _pangu_torchair_test_fixture(
# torchair is only work without chunked-prefill now
kwargs = {
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
additional_config.update(**kwargs)
@@ -179,6 +185,9 @@ def _qwen_torchair_test_fixture(
"torchair_graph_config": {
"enabled": False,
},
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
@@ -235,6 +244,9 @@ def _deepseek_v2_lite_torchair_test_fixure(
kwargs = {}
if not use_v1_schduler:
kwargs = {
"ascend_scheduler_config": {
"enable": True,
},
"refresh": True,
}
additional_config.update(**kwargs)

6
tests/e2e/nightly/features/test_mtpx_deepseek_r1_0528_w8a8.py
View File

@@ -73,7 +73,11 @@ async def test_models(model: str, mode: str) -> None:
"VLLM_RPC_TIMEOUT": "3600000",
"VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000"
}
additional_config: dict[str, Any] = {}
additional_config: dict[str, Any] = {
"ascend_scheduler_config": {
"enabled": False
},
}
speculative_config = {
"num_speculative_tokens": 2,
"method": "deepseek_mtp"

3
tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py
View File

@@ -74,6 +74,9 @@ async def test_models(model: str) -> None:
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
}
additional_config = {
"ascend_scheduler_config": {
"enabled": False
},
"torchair_graph_config": {
"enabled": True,
"enable_multistream_moe": False,

7
tests/e2e/nightly/features/test_prefix_cache_qwen3_32b_int8.py
View File

@@ -68,7 +68,12 @@ aisbench_cases75 = [{
async def test_models(model: str) -> None:
port = get_open_port()
env_dict = {"TASK_QUEUE_ENABLE": "1", "HCCL_OP_EXPANSION_MODE": "AIV"}
additional_config = {"enable_weight_nz_layout": True}
additional_config = {
"ascend_scheduler_config": {
"enabled": False
},
"enable_weight_nz_layout": True
}
server_args = [
"--quantization", "ascend", "--reasoning-parser", "qwen3",
"--tensor-parallel-size", "4", "--port",

3
tests/e2e/nightly/features/test_qwen3_32b_int8_a3_feature_stack3.py
View File

@@ -83,7 +83,8 @@ async def test_models(model: str, tp_size: int) -> None:
"0.9", "--block-size", "128", "--max-num-seqs", "256",
"--enforce-eager", "--max-model-len", "35840",
"--max-num-batched-tokens", "35840", "--additional-config",
'{"enable_weight_nz_layout":true}', "--compilation-config",
'{"ascend_scheduler_config":{"enabled":true},"enable_weight_nz_layout":true}',
"--compilation-config",
'{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[1,8,24,48,60]}'
]
with RemoteOpenAIServer(model,

8
tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py
View File

@@ -33,6 +33,7 @@ MODES = [
"single",
"aclgraph",
"aclgraph_mlapo",
"no_chunkprefill",
]
prompts = [
@@ -81,6 +82,9 @@ async def test_models(model: str, mode: str) -> None:
"method": "deepseek_mtp"
}
additional_config = {
"ascend_scheduler_config": {
"enabled": False
},
"torchair_graph_config": {
"enabled": True,
"enable_multistream_moe": False,
@@ -108,6 +112,10 @@ async def test_models(model: str, mode: str) -> None:
if mode == "aclgraph_mlapo":
env_dict["VLLM_ASCEND_ENABLE_MLAPO"] = "1"
additional_config["torchair_graph_config"] = {"enabled": False}
if mode == "no_chunkprefill":
additional_config["ascend_scheduler_config"] = {"enabled": True}
i = server_args.index("--max-num-batched-tokens") + 1
server_args[i] = "36864"
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

@@ -71,6 +71,9 @@ async def test_models(model: str) -> None:
"cudagraph_mode": "FULL_DECODE_ONLY"
}
additional_config: dict[str, Any] = {
"ascend_scheduler_config": {
"enabled": False
},
"torchair_graph_config": {
"enabled": True
},

1
tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py
View File

@@ -92,6 +92,7 @@ async def test_models(model: str, tp_size: int, dp_size: int,
"--gpu-memory-utilization",
"0.9",
"--additional-config",
'{"ascend_scheduler_config":{"enabled":true},'
'"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}',
]
if full_graph:

5
tests/e2e/nightly/models/test_qwen2_5_vl_32b.py
View File

@@ -85,8 +85,9 @@ async def test_models(model: str, tp_size: int) -> None:
str(tp_size), "--port",
str(port), "--max-model-len", "30000", "--max-num-batched-tokens",
"40000", "--max-num-seqs", "400", "--trust-remote-code",
"--gpu-memory-utilization", "0.8", "--compilation_config",
'{"cudagraph_mode": "FULL_DECODE_ONLY"}'
"--gpu-memory-utilization", "0.8", "--additional-config",
'{"ascend_scheduler_config":{"enabled":false}}',
"--compilation_config", '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
]
request_keyword_args: dict[str, Any] = {
**api_keyword_args,

6
tests/e2e/nightly/models/test_qwen3_235b_a22b_w8a8_eplb.py
View File

@@ -60,7 +60,11 @@ async def test_models(model: str) -> None:
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"
}
additional_config: dict[str, Any] = {}
additional_config: dict[str, Any] = {
"ascend_scheduler_config": {
"enabled": False
},
}
compilation_config = {"cudagraph_mode": "FULL_DECODE_ONLY"}
server_args = [
"--quantization", "ascend", "--async-scheduling",

6
tests/e2e/nightly/models/test_qwen3_235b_w8a8.py
View File

@@ -63,6 +63,11 @@ async def test_models(model: str, mode: str) -> None:
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"
}
additional_config: dict[str, Any] = {
"ascend_scheduler_config": {
"enabled": False
},
}
compilation_config = {"cudagraph_mode": "FULL_DECODE_ONLY"}
server_args = [
"--quantization", "ascend", "--async-scheduling",
@@ -77,6 +82,7 @@ async def test_models(model: str, mode: str) -> None:
server_args.extend(
["--compilation-config",
json.dumps(compilation_config)])
server_args.extend(["--additional-config", json.dumps(additional_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,
}

2
tests/e2e/nightly/models/test_qwq_32b.py
View File

@@ -93,6 +93,8 @@ async def test_models(model: str, mode: str, tp_size: int) -> None:
server_args.remove(
'{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1, 8, 24, 48, 60]}'
)
server_args.append("--additional-config")
server_args.append('{"ascend_scheduler_config":{"enabled":true}}')
server_args.append("--enforce-eager")
request_keyword_args: dict[str, Any] = {
**api_keyword_args,

4
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-A2-torchair.yaml
View File

@@ -30,7 +30,7 @@ deployment:
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
--additional-config '{"ascend_scheduler_config":{"enabled":false},"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
@@ -51,7 +51,7 @@ deployment:
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
--additional-config '{"ascend_scheduler_config":{"enabled":false},"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
benchmarks:
acc:
case_type: accuracy

4
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-A2.yaml
View File

@@ -31,7 +31,7 @@ deployment:
--gpu-memory-utilization 0.9
--enforce-eager
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--additional-config '{"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
--additional-config '{"ascend_scheduler_config":{"enabled":false},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
@@ -53,5 +53,5 @@ deployment:
--gpu-memory-utilization 0.9
--enforce-eager
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--additional-config '{"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
--additional-config '{"ascend_scheduler_config":{"enabled":false},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
benchmarks:

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

@@ -50,7 +50,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
-
server_cmd: >
@@ -80,7 +80,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -111,7 +111,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -141,7 +141,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
benchmarks:
perf:
case_type: performance

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

@@ -49,7 +49,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
'{"ascend_scheduler_config":{"enabled":false},"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
@@ -79,7 +79,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--additional-config
'{"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
'{"ascend_scheduler_config":{"enabled":false},"torchair_graph_config":{"enabled":false,"enable_multistream_shared_expert":false},"enable_prefill_optimizations":true,"enable_weight_nz_layout":true}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -110,7 +110,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
-
server_cmd: >
vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
@@ -140,7 +140,7 @@ deployment:
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}'
--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}'
'{"ascend_scheduler_config":{"enabled":false},"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}'
benchmarks:
perf:
case_type: performance

4
tests/e2e/nightly/multi_node/config/models/DeepSeek-V3_2-Exp-bf16.yaml
View File

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

41
tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
View File

@@ -48,26 +48,27 @@ def mtp_correctness(sampling_config: SamplingParams,
if graph_mode == CUDAGraphMode.FULL:
graph_mode_str = "FULL_DECODE_ONLY"
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":
"deepseek_mtp",
"num_speculative_tokens":
num_speculative_tokens,
"disable_padded_drafter_batch":
disable_padded_drafter_batch,
},
enforce_eager=enforce_eager,
max_model_len=2000,
compilation_config=CompilationConfig(
cudagraph_mode=graph_mode_str,
cudagraph_capture_sizes=[12],
)) as spec_llm:
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": "deepseek_mtp",
"num_speculative_tokens": num_speculative_tokens,
"disable_padded_drafter_batch": disable_padded_drafter_batch,
},
enforce_eager=enforce_eager,
max_model_len=2000,
compilation_config=CompilationConfig(
cudagraph_mode=graph_mode_str,
cudagraph_capture_sizes=[12],
),
additional_config={"ascend_scheduler_config": {
"enabled": False
}}) as spec_llm:
spec_outputs = spec_llm.generate(example_prompts, sampling_config)
matches = 0

170
tests/e2e/singlecard/test_ascend_scheduler.py Normal file
View File

@@ -0,0 +1,170 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm import SamplingParams
from tests.e2e.conftest import VllmRunner
from tests.e2e.model_utils import check_outputs_equal
MODEL = "Qwen/Qwen3-0.6B"
@pytest.mark.parametrize("enforce_eager", [True, False])
def test_concurrent_partial_prefill(enforce_eager):
with VllmRunner(MODEL,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
max_num_seqs=3,
max_num_batched_tokens=8192,
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7) as vllm_model:
outputs = vllm_model.model.generate(["Hello my name is Robert and I"] *
3)
assert len(outputs) == 3
for output in outputs:
assert len(output.outputs) == 1
@pytest.mark.parametrize("enforce_eager", [True, False])
def test_prefix_cache_stats_is_recorded(enforce_eager):
with VllmRunner(MODEL,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
max_num_seqs=3,
max_num_batched_tokens=8192,
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7) as vllm_model:
# 17 tokens will make sure first 16 tokens are cached in a block
input_tokens = {"prompt_token_ids": [101] * 129}
_ = vllm_model.model.generate([input_tokens])
outputs = vllm_model.model.generate([input_tokens])
assert outputs[0].num_cached_tokens == 128
@pytest.mark.parametrize("max_tokens",
[4]) # cannot align results when max_tokens > 4
@pytest.mark.parametrize("chunked_prefill_token_size", [2048])
def test_chunked_prefill_with_ascend_scheduler(
max_tokens: int, chunked_prefill_token_size: int) -> None:
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
max_num_seqs = chunked_prefill_token_size
max_num_batched_tokens = chunked_prefill_token_size
with VllmRunner(MODEL,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
'enable_chunked_prefill': True,
},
},
max_num_seqs=max_num_seqs,
max_num_batched_tokens=max_num_batched_tokens,
max_model_len=2048,
gpu_memory_utilization=0.7) as vllm_model:
chunked_prefill_output = vllm_model.generate_greedy(
example_prompts, max_tokens)
with VllmRunner(MODEL,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
max_model_len=2048,
gpu_memory_utilization=0.7) as vllm_model:
vllm_output = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_output,
outputs_1_lst=chunked_prefill_output,
name_0="vllm_output",
name_1="chunked_prefill_output",
)
@pytest.mark.parametrize("max_tokens",
[4]) # cannot align results when max_tokens > 4
@pytest.mark.parametrize("chunked_prefill_token_size", [2048])
def test_chunked_prefill_with_scheduler_dynamic_batch(
max_tokens: int, chunked_prefill_token_size: int) -> None:
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
max_num_seqs = chunked_prefill_token_size
max_num_batched_tokens = chunked_prefill_token_size
with VllmRunner(MODEL,
additional_config={
'SLO_limits_for_dynamic_batch': 0,
},
max_num_seqs=max_num_seqs,
max_num_batched_tokens=max_num_batched_tokens,
max_model_len=2048,
gpu_memory_utilization=0.7) as vllm_model:
dynamic_batch_output = vllm_model.generate_greedy(
example_prompts, max_tokens)
with VllmRunner(MODEL,
additional_config={
'SLO_limits_for_dynamic_batch': -1,
},
max_model_len=2048,
gpu_memory_utilization=0.7) as vllm_model:
vllm_output = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_output,
outputs_1_lst=dynamic_batch_output,
name_0="vllm_output",
name_1="chunked_prefill_output",
)
def test_async_scheduling_eager() -> None:
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
] * 10
sampling_params = SamplingParams(temperature=0.2,
max_tokens=10,
stop_token_ids=None)
with VllmRunner(
"Qwen/Qwen2.5-0.5B-Instruct",
max_model_len=4096,
max_num_seqs=50,
dtype="bfloat16",
gpu_memory_utilization=0.9,
async_scheduling=True,
) as vllm_model:
vllm_model.generate(prompts, sampling_params=sampling_params)
def test_async_scheduling_with_full_graph() -> None:
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
] * 10
sampling_params = SamplingParams(temperature=0.2,
max_tokens=10,
stop_token_ids=None)
with VllmRunner("Qwen/Qwen3-8B",
max_model_len=4096,
max_num_seqs=50,
dtype="bfloat16",
gpu_memory_utilization=0.9,
async_scheduling=True,
compilation_config={"cudagraph_mode":
"FULL"}) as vllm_model:
vllm_model.generate(prompts, sampling_params=sampling_params)

82
tests/e2e/singlecard/test_chunked.py Normal file
View File

@@ -0,0 +1,82 @@
#
# 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/compile/test_aclgraph.py`.
"""
import gc
import pytest
import torch
from vllm import SamplingParams
from tests.e2e.conftest import VllmRunner
MODELS = ["Qwen/Qwen2.5-0.5B-Instruct"]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_tokens", [1])
def test_models(
model: str,
max_tokens: int,
) -> None:
prompts = ["The president of the United States is"]
sampling_params = SamplingParams(
max_tokens=max_tokens,
temperature=0.0,
)
with VllmRunner(model,
long_prefill_token_threshold=20,
enforce_eager=False) as vllm_model:
output1 = vllm_model.generate(prompts, sampling_params)
with VllmRunner(model,
enforce_eager=False,
additional_config={
'ascend_scheduler_config': {
'enabled': True
},
}) as vllm_model:
output2 = vllm_model.generate(prompts, sampling_params)
# Extract the generated token IDs for comparison
token_ids1 = output1[0][0][0]
token_ids2 = output2[0][0][0]
print(f"Token IDs 1: {token_ids1}")
print(f"Token IDs 2: {token_ids2}")
# Convert token IDs to tensors and calculate cosine similarity
# Take the length of a shorter sequence to ensure consistent dimensions
min_len = min(len(token_ids1), len(token_ids2))
tensor1 = torch.tensor(token_ids1[:min_len], dtype=torch.float32)
tensor2 = torch.tensor(token_ids2[:min_len], dtype=torch.float32)
# Calculate similarity using torch.cosine_similarity
similarity = torch.cosine_similarity(tensor1, tensor2, dim=0)
print(f"Token IDs cosine similarity: {similarity.item()}")
assert similarity > 0.95
gc.collect()
torch.npu.empty_cache()
torch.npu.reset_peak_memory_stats()

35
tests/e2e/singlecard/test_vlm.py
View File

@@ -20,6 +20,7 @@
Run `pytest tests/test_offline_inference.py`.
"""
import pytest
from vllm import SamplingParams
from vllm.assets.audio import AudioAsset
from vllm.assets.image import ImageAsset
@@ -54,6 +55,40 @@ def test_multimodal_vl(prompt_template):
assert output_str, "Generated output should not be empty."
@pytest.mark.skip(reason="This e2e test will stuck in multi-batch scenario. "
"Add this back after fixing the issue.")
def test_multimodal_ascend_scheduler(prompt_template):
image = ImageAsset("cherry_blossom") \
.pil_image.convert("RGB")
img_questions = [
"What is the content of this image?",
"Describe the content of this image in detail.",
"What's in the image?",
"Where is this image taken?",
]
images = [image] * len(img_questions)
prompts = prompt_template(img_questions)
with VllmRunner("Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=4096,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
mm_processor_kwargs={
"min_pixels": 28 * 28,
"max_pixels": 1280 * 28 * 28,
"fps": 1,
},
enforce_eager=True) as vllm_model:
outputs = vllm_model.generate_greedy(prompts=prompts,
images=images,
max_tokens=64)
assert len(outputs) == len(prompts)
for _, output_str in outputs:
assert output_str, "Generated output should not be empty."
def test_multimodal_audio():
audio_prompt = "".join([
f"Audio {idx+1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n"

134
tests/ut/core/test_schedule_config.py Normal file
View File

@@ -0,0 +1,134 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
#
# 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.
from vllm.config import SchedulerConfig
from tests.ut.base import TestBase
from vllm_ascend.core.schedule_config import AscendSchedulerConfig
class TestAscendSchedulerConfig(TestBase):
def setUp(self):
self.basic_scheduler_config = SchedulerConfig(
max_num_batched_tokens=8192,
max_model_len=8192,
is_multimodal_model=False,
send_delta_data=False,
)
def test_initialize_from_config_with_default(self):
# No additional config given, check the default value here.
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config, {})
self.assertEqual(ascend_config.enable_chunked_prefill, False)
self.assertEqual(ascend_config.policy, "fcfs")
self.assertEqual(ascend_config.scheduler_cls,
"vllm_ascend.core.scheduler.AscendScheduler")
self.assertEqual(ascend_config.max_num_encoder_input_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
def test_initialize_from_config_with_override(self):
# test override
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=False,
policy="fcfs",
scheduler_cls="vllm_ascend.core.scheduler.AscendScheduler",
max_num_batched_tokens=8192,
max_model_len=2048,
max_long_partial_prefills=1,
long_prefill_token_threshold=512,
),
)
self.assertEqual(ascend_config.enable_chunked_prefill, False)
self.assertEqual(ascend_config.policy, "fcfs")
self.assertEqual(ascend_config.scheduler_cls,
"vllm_ascend.core.scheduler.AscendScheduler")
self.assertEqual(ascend_config.max_num_batched_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
self.assertEqual(ascend_config.max_long_partial_prefills, 1)
self.assertEqual(ascend_config.long_prefill_token_threshold, 512)
def test_not_implemented_policy(self):
with self.assertRaises(NotImplementedError) as context:
AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
policy="custom_policy",
max_num_batched_tokens=8192,
max_model_len=2048,
),
)
self.assertIn(
"currently AscendScheduler only supports fcfs policy",
str(context.exception),
)
def test_no_override(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config, {})
self.assertEqual(ascend_config.max_num_encoder_input_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
def test_valid_config_with_multimodal(self):
config = AscendSchedulerConfig.initialize_from_config(
SchedulerConfig(is_multimodal_model=True,
max_num_batched_tokens=8192), {})
self.assertTrue(config.is_multimodal_model)
def test_valid_config_with_chunked_prefill(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=True,
max_num_batched_tokens=8192,
max_model_len=8192,
),
)
self.assertEqual(ascend_config.max_num_batched_tokens, 8192)
self.assertEqual(ascend_config.max_model_len, 8192)
self.assertTrue(ascend_config.enable_chunked_prefill)
def test_invalid_config_without_chunked_prefill(self):
with self.assertRaises(ValueError) as context:
AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=False,
max_num_batched_tokens=2048,
max_model_len=8192,
),
)
self.assertIn(
"Ascend scheduler is enabled without chunked prefill feature",
str(context.exception),
)
self.assertIn("max_num_batched_tokens (2048)", str(context.exception))
self.assertIn("max_model_len (8192)", str(context.exception))
def test_initialize_from_config_with_pd_transfer(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_pd_transfer=True,
decode_max_num_seqs=48,
max_num_batched_tokens=8192,
max_model_len=4096,
),
)
self.assertEqual(ascend_config.enable_pd_transfer, True)
self.assertEqual(ascend_config.decode_max_num_seqs, 48)

1473
tests/ut/core/test_scheduler.py Normal file
View File

File diff suppressed because it is too large Load Diff

1
tests/ut/ops/test_linear.py
View File

@@ -99,6 +99,7 @@ class TestAscendRowParallelLinear(BaseLinearTest):
ascend_config._ASCEND_CONFIG = MagicMock()
ascend_config._ASCEND_CONFIG.oproj_tensor_parallel_size = 2
ascend_config._ASCEND_CONFIG.ascend_scheduler_config.enabled = False
linear = AscendRowParallelLinear(
input_size=16,

7
tests/ut/ops/test_vocab_parallel_embedding.py
View File

@@ -209,7 +209,12 @@ class TestAscendLogitsProcessor(unittest.TestCase):
return_value=torch.randn(1, self.vocab_size)),
patch(
"vllm_ascend.ops.vocab_parallel_embedding.get_lmhead_tp_group.all_gather",
return_value=torch.randn(1, self.vocab_size))
return_value=torch.randn(1, self.vocab_size)),
patch(
"vllm_ascend.core.schedule_config.AscendSchedulerConfig.initialize_from_config",
return_value=MagicMock(max_num_batched_tokens=1000,
max_model_len=512,
enable_chunked_prefill=False))
]
for p in self.patches:

7
tests/ut/quantization/test_w8a8_dynamic.py
View File

@@ -33,6 +33,13 @@ class TestAscendW8A8FusedMoEMethod(TestBase):
mock_get_ep_group.return_value = mock_ep_group
mock_ascend_config = Mock()
# 创建一个具有具体属性的 Mock 对象来表示 ascend_scheduler_config
mock_ascend_scheduler_config = Mock()
mock_ascend_scheduler_config.enabled = False
mock_ascend_scheduler_config.max_num_batched_tokens = 1024
mock_ascend_scheduler_config.max_model_len = 2048
mock_ascend_config.ascend_scheduler_config = mock_ascend_scheduler_config
mock_ascend_config.torchair_graph_config = Mock(enabled=False)
mock_ascend_config.enable_chunked_prefill = False
mock_get_ascend_config.return_value = mock_ascend_config

9
tests/ut/test_ascend_config.py
View File

@@ -56,6 +56,9 @@ class TestAscendConfig(TestBase):
self.assertTrue(torchair_graph_config.enable_frozen_parameter)
self.assertFalse(torchair_graph_config.enable_kv_nz)
ascend_scheduler_config = ascend_config.ascend_scheduler_config
self.assertFalse(ascend_scheduler_config.enabled)
@_clean_up_ascend_config
def test_init_ascend_config_with_additional_config(self):
test_vllm_config = VllmConfig()
@@ -71,6 +74,9 @@ class TestAscendConfig(TestBase):
"enable_kv_nz": True
},
"multistream_overlap_shared_expert": True,
"ascend_scheduler_config": {
"enabled": True
},
"expert_map_path": "test_expert_map_path",
"refresh": True,
}
@@ -88,6 +94,9 @@ class TestAscendConfig(TestBase):
self.assertTrue(torchair_graph_config.enable_frozen_parameter)
self.assertTrue(torchair_graph_config.enable_kv_nz)
ascend_scheduler_config = ascend_config.ascend_scheduler_config
self.assertTrue(ascend_scheduler_config.enabled)
@_clean_up_ascend_config
def test_init_ascend_config_with_refresh(self):
test_vllm_config = VllmConfig()

26
tests/ut/test_platform.py
View File

@@ -32,6 +32,7 @@ class TestNPUPlatform(TestBase):
def mock_vllm_ascend_config():
mock_ascend_config = MagicMock()
mock_ascend_config.torchair_graph_config.enabled = False
mock_ascend_config.ascend_scheduler_config.enabled = False
mock_ascend_config.enable_shared_expert_dp = False
return mock_ascend_config
@@ -521,6 +522,31 @@ class TestNPUPlatform(TestBase):
self.platform.check_and_update_config(vllm_config)
self.assertEqual(vllm_config.compilation_config.custom_ops, [])
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._910_93)
@patch("vllm_ascend.ascend_config.check_ascend_config")
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch(
"vllm_ascend.core.recompute_schedule_config.RecomputeSchedulerConfig.initialize_from_config"
)
def test_check_and_update_config_ascend_scheduler_config(
self, mock_init_recompute, mock_init_ascend, mock_check_ascend,
mock_soc_version):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.ascend_scheduler_config.enabled = True
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.parallel_config.tensor_parallel_size = 1
mock_init_recompute.return_value = MagicMock()
with patch("vllm_ascend.core.schedule_config.AscendSchedulerConfig"
) as mock_scheduler:
from vllm_ascend import platform
importlib.reload(platform)
self.platform.check_and_update_config(vllm_config)
mock_scheduler.initialize_from_config.assert_called_once()
@patch('vllm_ascend.platform.get_ascend_config')
def test_get_attn_backend_cls_use_v1_and_mla(self, mock_get_ascend_config):
mock_config = MagicMock()

4
tests/ut/test_utils.py
View File

@@ -253,10 +253,12 @@ class TestUtils(TestBase):
model_path = os.path.join(os.path.dirname(__file__), "fake_weight")
test_model_config = ModelConfig(model=model_path, enforce_eager=True)
test_parallel_config = ParallelConfig()
ascend_config = {"ascend_scheduler_config": {"enabled": False}}
test_vllm_config = VllmConfig(
model_config=test_model_config,
compilation_config=test_compilation_config,
parallel_config=test_parallel_config)
parallel_config=test_parallel_config,
additional_config=ascend_config)
utils.update_aclgraph_sizes(test_vllm_config)
os.environ['HCCL_OP_EXPANSION_MODE'] = 'AIV'
utils.update_aclgraph_sizes(test_vllm_config)

2
tests/ut/torchair/models/test_torchair_deepseek_v2.py
View File

@@ -235,6 +235,8 @@ def test_torchair_deepseek_v2_mlp(mock_distributed, base_config):
hidden_act="silu",
quant_config=None)
assert isinstance(mlp.act_fn, TorchairDeepseekV2SiluAndMul)
ascend_config = MagicMock()
ascend_config._ASCEND_CONFIG.ascend_scheduler_config.enabled = False
with patch(
"vllm_ascend.torchair.models.torchair_deepseek_v2.QuantizationConfig"
) as mock_quant_config: