[MRV2] Switch CI tests still running on MRV1 unnecessarily (#61014)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -352,6 +352,12 @@ def test_async_scheduling_pp_allows_rescheduling_with_output_placeholders():
|
||||
_ = scheduler.schedule()
|
||||
assert req.num_output_placeholders > 0
|
||||
|
||||
if scheduler.use_v2_model_runner:
|
||||
# MRV2 staggers same-request decodes by pp_size steps to match the
|
||||
# worker-side sampled-token broadcast cadence.
|
||||
output = scheduler.schedule()
|
||||
assert req.request_id not in output.num_scheduled_tokens
|
||||
|
||||
# before any update_from_output, we still expect the request can be
|
||||
# scheduled again (multi-step in-flight).
|
||||
output = scheduler.schedule()
|
||||
|
||||
@@ -238,7 +238,7 @@ def create_scheduler(
|
||||
structured_output_manager=StructuredOutputManager(vllm_config),
|
||||
)
|
||||
if use_v2_model_runner is None:
|
||||
use_v2_model_runner = bool(envs.VLLM_USE_V2_MODEL_RUNNER)
|
||||
use_v2_model_runner = envs.VLLM_USE_V2_MODEL_RUNNER is not False
|
||||
scheduler.use_v2_model_runner = use_v2_model_runner
|
||||
return scheduler
|
||||
|
||||
|
||||
Reference in New Issue
Block a user