[Bugfix] Fix sequence parallelism bug when enable pipeline parallelism (#24021)

Signed-off-by: cascade812 <cascade812@outlook.com>
This commit is contained in:
cascade
2025-09-16 04:32:32 +00:00
committed by GitHub
parent 759ef49b15
commit 17871983a2
6 changed files with 135 additions and 42 deletions
+15 -3
View File
@@ -19,6 +19,7 @@ from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.worker.cpu_model_runner import CPUModelRunner
from vllm.v1.worker.gpu_worker import (Worker,
init_worker_distributed_environment)
from vllm.v1.worker.utils import is_residual_scattered_for_sp
logger = init_logger(__name__)
@@ -107,18 +108,29 @@ class CPUWorker(Worker):
scheduler_output: "SchedulerOutput",
) -> Optional[ModelRunnerOutput]:
intermediate_tensors = None
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
num_input_tokens = self.model_runner._get_num_input_tokens(
num_scheduled_tokens)
all_gather_tensors = {
"residual":
not is_residual_scattered_for_sp(self.vllm_config,
num_input_tokens)
}
if not get_pp_group().is_first_rank:
intermediate_tensors = IntermediateTensors(
get_pp_group().recv_tensor_dict(
all_gather_group=get_tp_group()))
all_gather_group=get_tp_group(),
all_gather_tensors=all_gather_tensors))
output = self.model_runner.execute_model(scheduler_output,
intermediate_tensors)
if not get_pp_group().is_last_rank:
assert isinstance(output, IntermediateTensors)
get_pp_group().send_tensor_dict(output.tensors,
all_gather_group=get_tp_group())
get_pp_group().send_tensor_dict(
output.tensors,
all_gather_group=get_tp_group(),
all_gather_tensors=all_gather_tensors)
return None
assert isinstance(output, ModelRunnerOutput)