[Refactor][1/N Scheduler]Remove duplicated AR/generation scheduler plumbing and establish explicit shared lifecycle contracts. (#5461)
Signed-off-by: R2-Y <ruiruyang2@gmail.com>
This commit is contained in:
1 parent
d5d47e9be1
commit
5215e03a91
7 files changed
+655
-584
No files matched your search
@@ -2,7 +2,9 @@ from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
from vllm.v1.core.sched.output import NewRequestData
|
||||
|
||||
from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
|
||||
from vllm_omni.core.sched.output import OmniNewRequestData
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
@@ -52,3 +54,80 @@ def test_omni_new_request_data_allows_missing_payloads():
|
||||
|
||||
assert data.prompt_embeds is None
|
||||
assert data.additional_information is None
|
||||
|
||||
|
||||
def test_from_base_preserves_every_upstream_field_and_adds_omni_payloads():
|
||||
prompt_embeds = torch.randn(2, 3)
|
||||
base = NewRequestData(
|
||||
req_id="req-base",
|
||||
prompt_token_ids=[1, 2],
|
||||
mm_features=[],
|
||||
sampling_params=None,
|
||||
pooling_params=None,
|
||||
block_ids=([3, 4],),
|
||||
num_computed_tokens=1,
|
||||
lora_request=None,
|
||||
prompt_embeds=prompt_embeds,
|
||||
prompt_is_token_ids=False,
|
||||
prefill_token_ids=[1],
|
||||
)
|
||||
request = SimpleNamespace(
|
||||
external_req_id="external-base",
|
||||
additional_information={"payload": "value"},
|
||||
model_intermediate_buffer={"hidden": "state"},
|
||||
)
|
||||
|
||||
data = OmniNewRequestData.from_base(base, request)
|
||||
|
||||
for field_name in NewRequestData.__dataclass_fields__:
|
||||
assert getattr(data, field_name) is getattr(base, field_name)
|
||||
assert data.external_req_id == "external-base"
|
||||
assert data.additional_information == {"payload": "value"}
|
||||
assert data.model_intermediate_buffer == {"hidden": "state"}
|
||||
|
||||
|
||||
def test_rewrap_converts_fallback_entries_and_keeps_fast_path_entries():
|
||||
base = NewRequestData(
|
||||
req_id="req-fallback",
|
||||
prompt_token_ids=[1],
|
||||
mm_features=[],
|
||||
sampling_params=None,
|
||||
pooling_params=None,
|
||||
block_ids=([2],),
|
||||
num_computed_tokens=0,
|
||||
lora_request=None,
|
||||
prompt_embeds=None,
|
||||
prompt_is_token_ids=True,
|
||||
prefill_token_ids=[1],
|
||||
)
|
||||
fast_path = OmniNewRequestData(
|
||||
req_id="req-fast",
|
||||
prompt_token_ids=[3],
|
||||
mm_features=[],
|
||||
sampling_params=None,
|
||||
pooling_params=None,
|
||||
block_ids=([4],),
|
||||
num_computed_tokens=0,
|
||||
lora_request=None,
|
||||
prompt_embeds=None,
|
||||
prompt_is_token_ids=True,
|
||||
prefill_token_ids=[3],
|
||||
additional_information={"fast": True},
|
||||
)
|
||||
scheduler = OmniSchedulerMixin()
|
||||
scheduler.requests = {
|
||||
"req-fallback": SimpleNamespace(
|
||||
external_req_id="external-fallback",
|
||||
additional_information={"fallback": True},
|
||||
model_intermediate_buffer=None,
|
||||
)
|
||||
}
|
||||
output = SimpleNamespace(scheduled_new_reqs=[base, fast_path])
|
||||
|
||||
scheduler._rewrap_scheduled_new_reqs(output)
|
||||
|
||||
fallback_data, fast_path_data = output.scheduled_new_reqs
|
||||
assert isinstance(fallback_data, OmniNewRequestData)
|
||||
assert fallback_data.prefill_token_ids == [1]
|
||||
assert fallback_data.additional_information == {"fallback": True}
|
||||
assert fast_path_data is fast_path
|
||||
Reference in new issue
Block a user