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

[CI] fix ci (#2464)

Browse Source 
### What this PR does / why we need it?
1. use action/checkout@v5 instead of v4
2. remove dbo test case because there is issue with it and will be
refactored later
3. make vllm-ascend compatible with vllm v0.10.1.1 and add CI for it
4. fix sampler api changes introduced by
https://github.com/vllm-project/vllm/pull/22387
6. fix qwen3 moe config changes intruoduced by
https://github.com/vllm-project/vllm/pull/20562
7. fix kvcache block changes introduced by
https://github.com/vllm-project/vllm/pull/23262

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
CI passed with existing test.


- vLLM version: v0.10.0
- vLLM main:
0c6e40bbaa

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-08-22 07:30:48 +08:00
committed by GitHub
parent 0ca3f48c90
commit b0403f8d8a
27 changed files with 389 additions and 199 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -78,26 +78,6 @@ def test_models_distributed_DeepSeek_multistream_moe():
vllm_model.generate_greedy(example_prompts, max_tokens)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DBO": "1"})
def test_models_distributed_DeepSeek_dbo():
example_prompts = ["The president of the United States is"] * 41
dtype = "half"
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
with VllmRunner(
"deepseek-ai/DeepSeek-V2-Lite",
dtype=dtype,
tensor_parallel_size=2,
distributed_executor_backend="mp",
) as vllm_model:
model_arch = 'DeepseekV2ForCausalLM'
registed_models = ModelRegistry.models
assert registed_models[
model_arch].module_name == "vllm_ascend.models.deepseek_dbo"
assert registed_models[
model_arch].class_name == "CustomDeepseekDBOForCausalLM"
vllm_model.generate(example_prompts, sampling_params)
@pytest.mark.skip(
reason=
"deepseek dbo dose not consider the support on half precision float, will enable this ut after we actually support it"

294
tests/ut/core/test_scheduler.py
View File

@@ -13,7 +13,7 @@ from vllm.v1.core.kv_cache_utils import (get_request_block_hasher,
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
KVCacheGroupSpec)
from vllm.v1.outputs import DraftTokenIds, ModelRunnerOutput
from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.request import Request, RequestStatus
from vllm.v1.structured_output import StructuredOutputManager
@@ -21,6 +21,11 @@ from tests.ut.base import TestBase
from vllm_ascend.core.scheduler import AscendScheduler
from vllm_ascend.utils import vllm_version_is
if not vllm_version_is("0.10.1.1"):
from vllm.v1.outputs import DraftTokenIds
else:
DraftTokenIds = None
EOS_TOKEN_ID = 50256
MODEL = "Qwen3-0.6B"
ENABLE_PREFIX_CACHING = None
@@ -66,16 +71,33 @@ def create_requests(
def make_output(scheduler):
return ModelRunnerOutput(
req_ids=[req.request_id for req in scheduler.running],
req_id_to_index={
req.request_id: i
for i, req in enumerate(scheduler.running)
},
sampled_token_ids=[[1000]] * len(scheduler.running),
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
req_ids = [req.request_id for req in scheduler.running]
req_id_to_index = {
req.request_id: i
for i, req in enumerate(scheduler.running)
}
sampled_token_ids = [[1000]] * len(scheduler.running)
logprobs = None
if vllm_version_is("0.10.1.1"):
modelrunner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_id_to_index,
sampled_token_ids=sampled_token_ids,
spec_token_ids=None,
logprobs=logprobs,
prompt_logprobs_dict={},
pooler_output=[],
)
else:
modelrunner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_id_to_index,
sampled_token_ids=sampled_token_ids,
logprobs=logprobs,
prompt_logprobs_dict={},
pooler_output=[],
)
return modelrunner_output
class TestAscendScheduler(TestBase):
@@ -271,8 +293,7 @@ class TestAscendScheduler(TestBase):
req.num_computed_tokens = req.num_tokens
scheduler.requests[req.request_id] = req
scheduler.running.append(req)
if not vllm_version_is("0.9.2"):
req.status = RequestStatus.RUNNING
req.status = RequestStatus.RUNNING
scheduler_output = SchedulerOutput(scheduled_new_reqs=[],
scheduled_cached_reqs=[],
@@ -291,18 +312,33 @@ class TestAscendScheduler(TestBase):
free_encoder_input_ids=[],
structured_output_request_ids={},
grammar_bitmask=None)
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[EOS_TOKEN_ID], [10, 11]
], # First request hits EOS, second continues
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[EOS_TOKEN_ID], [
10, 11
]], # First request hits EOS, second continues
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[EOS_TOKEN_ID], [
10, 11
]], # First request hits EOS, second continues
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output, model_output)
@@ -325,8 +361,7 @@ class TestAscendScheduler(TestBase):
req.num_computed_tokens = req.num_tokens
scheduler.requests[req.request_id] = req
scheduler.running.append(req)
if not vllm_version_is("0.9.2"):
req.status = RequestStatus.RUNNING
req.status = RequestStatus.RUNNING
scheduler_output = SchedulerOutput(scheduled_new_reqs=[],
scheduled_cached_reqs=[],
@@ -346,18 +381,31 @@ class TestAscendScheduler(TestBase):
free_encoder_input_ids=[],
structured_output_request_ids={},
grammar_bitmask=None)
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 42, 12],
[13, 14]], # First request hits stop token
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 42, 12],
[13, 14]], # First request hits stop token
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 42, 12],
[13, 14]], # First request hits stop token
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output, model_output)
@@ -379,8 +427,7 @@ class TestAscendScheduler(TestBase):
req.num_computed_tokens = req.num_tokens
scheduler.requests[req.request_id] = req
scheduler.running.append(req)
if not vllm_version_is("0.9.2"):
req.status = RequestStatus.RUNNING
req.status = RequestStatus.RUNNING
scheduler_output = SchedulerOutput(scheduled_new_reqs=[],
scheduled_cached_reqs=[],
@@ -401,18 +448,31 @@ class TestAscendScheduler(TestBase):
structured_output_request_ids={},
grammar_bitmask=None)
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 11, 12],
[13]], # First request exceeds max_tokens
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 11, 12],
[13]], # First request exceeds max_tokens
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_output = ModelRunnerOutput(
req_ids=[req.request_id for req in requests],
req_id_to_index={
req.request_id: i
for i, req in enumerate(requests)
},
sampled_token_ids=[[10, 11, 12],
[13]], # First request exceeds max_tokens
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output, model_output)
# Verify first request stopped due to length
@@ -448,13 +508,24 @@ class TestAscendScheduler(TestBase):
structured_output_request_ids={},
grammar_bitmask=None)
model_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[EOS_TOKEN_ID, 10, 11]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[EOS_TOKEN_ID, 10, 11]],
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[EOS_TOKEN_ID, 10, 11]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output, model_output)
@@ -505,13 +576,23 @@ class TestAscendScheduler(TestBase):
512)
# Model output of the first request.
model_runner_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[0]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[0]],
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_runner_output = ModelRunnerOutput(
req_ids=[requests[0].request_id],
req_id_to_index={requests[0].request_id: 0},
sampled_token_ids=[[0]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output0,
model_runner_output)
@@ -521,13 +602,23 @@ class TestAscendScheduler(TestBase):
# request is still running.
scheduler.schedule()
# Model output of the second request.
model_runner_output = ModelRunnerOutput(
req_ids=[requests[1].request_id],
req_id_to_index={requests[1].request_id: 0},
sampled_token_ids=[[0]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=[requests[1].request_id],
req_id_to_index={requests[1].request_id: 0},
sampled_token_ids=[[0]],
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_runner_output = ModelRunnerOutput(
req_ids=[requests[1].request_id],
req_id_to_index={requests[1].request_id: 0},
sampled_token_ids=[[0]],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
scheduler.update_from_output(scheduler_output1,
model_runner_output)
@@ -579,19 +670,29 @@ class TestAscendScheduler(TestBase):
req_id = requests[i].request_id
self.assertEqual(output.num_scheduled_tokens[req_id], 1)
self.assertNotIn(req_id, output.scheduled_spec_decode_tokens)
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=[[0] for _ in range(len(requests))],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
draft_token_ids = DraftTokenIds(req_ids, spec_tokens)
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=[[0] for _ in range(len(requests))],
logprobs=None,
prompt_logprobs_dict={},
spec_token_ids=spec_tokens,
pooler_output=[])
else:
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=[[0] for _ in range(len(requests))],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
draft_token_ids = DraftTokenIds(req_ids, spec_tokens)
engine_core_outputs = scheduler.update_from_output(
output, model_runner_output)
scheduler.update_draft_token_ids(draft_token_ids)
if not vllm_version_is("0.10.1.1"):
scheduler.update_draft_token_ids(draft_token_ids)
for i in range(len(requests)):
running_req = scheduler.running[i]
@@ -627,14 +728,23 @@ class TestAscendScheduler(TestBase):
else:
self.assertNotIn(req_id,
output.scheduled_spec_decode_tokens)
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=output_tokens,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=output_tokens,
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
else:
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_to_index,
sampled_token_ids=output_tokens,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[])
engine_core_outputs = scheduler.update_from_output(
output, model_runner_output)

32
tests/ut/kv_connector/utils.py
View File

@@ -200,12 +200,26 @@ def create_model_runner_output(
kv_connector_output = KVConnectorOutput(finished_sending=finished_sending,
finished_recving=finished_recving)
extra_args = {"kv_connector_output": kv_connector_output}
return ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_id_to_index,
sampled_token_ids=sampled_token_ids,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[],
**extra_args,
)
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_id_to_index,
sampled_token_ids=sampled_token_ids,
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[],
**extra_args,
)
else:
model_runner_output = ModelRunnerOutput(
req_ids=req_ids,
req_id_to_index=req_id_to_index,
sampled_token_ids=sampled_token_ids,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=[],
**extra_args,
)
return model_runner_output