[V1] [P/D] Refactor KV Connector Path (#21980)

Signed-off-by: David Ben-David <davidb@pliops.com>
Co-authored-by: David Ben-David <davidb@pliops.com>
This commit is contained in:
David Ben-David
2025-08-03 04:03:40 -07:00
committed by GitHub
co-authored by David Ben-David
parent 24d1dffbeb
commit aefeea0fde
12 changed files with 142 additions and 80 deletions
+10 -20
View File
@@ -69,7 +69,7 @@ from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
from vllm.v1.worker.kv_connector_model_runner_mixin import (
KVConnectorModelRunnerMixin)
KVConnectorModelRunnerMixin, KVConnectorOutput)
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
from ..sample.logits_processor import LogitsProcessorManager
@@ -1423,8 +1423,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
num_scheduled_tokens_np: np.ndarray,
finished_sending: Optional[set[str]],
finished_recving: Optional[set[str]],
kv_connector_output: Optional[KVConnectorOutput],
) -> ModelRunnerOutput:
assert self.input_batch.num_reqs ==\
len(self.input_batch.pooling_params), \
@@ -1459,8 +1458,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
finished_sending=finished_sending,
finished_recving=finished_recving,
kv_connector_output=kv_connector_output,
)
@torch.inference_mode()
@@ -1564,8 +1562,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
num_tokens=num_input_tokens,
num_tokens_across_dp=num_tokens_across_dp,
skip_cuda_graphs=skip_cuda_graphs,
):
self.maybe_setup_kv_connector(scheduler_output)
), self.maybe_get_kv_connector_output(
scheduler_output) as kv_connector_output:
model_output = self.model(
input_ids=input_ids,
@@ -1578,10 +1576,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
),
)
self.maybe_wait_for_kv_save()
finished_sending, finished_recving = (
self.get_finished_kv_transfers(scheduler_output))
if self.use_aux_hidden_state_outputs:
hidden_states, aux_hidden_states = model_output
else:
@@ -1597,20 +1591,17 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
== "external_launcher" and len(get_pp_group().ranks) > 0
if not get_pp_group().is_last_rank:
# For mid-pipeline stages, return the hidden states.
if not broadcast_pp_output:
if finished_sending or finished_recving:
hidden_states.finished_sending = finished_sending
hidden_states.finished_recving = finished_recving
return hidden_states
assert isinstance(hidden_states, IntermediateTensors)
if not broadcast_pp_output:
hidden_states.kv_connector_output = kv_connector_output
return hidden_states
get_pp_group().send_tensor_dict(hidden_states.tensors,
all_gather_group=get_tp_group())
logits = None
else:
if self.input_batch.pooling_params:
return self._pool(hidden_states, num_scheduled_tokens,
num_scheduled_tokens_np, finished_sending,
finished_recving)
num_scheduled_tokens_np, kv_connector_output)
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states, None)
@@ -1760,8 +1751,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
finished_sending=finished_sending,
finished_recving=finished_recving,
kv_connector_output=kv_connector_output,
num_nans_in_logits=num_nans_in_logits,
)
+12 -10
View File
@@ -16,8 +16,7 @@ from vllm.config import VllmConfig
from vllm.distributed import (ensure_model_parallel_initialized,
init_distributed_environment,
set_custom_all_reduce)
from vllm.distributed.kv_transfer import (ensure_kv_transfer_initialized,
has_kv_transfer_group)
from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
from vllm.distributed.parallel_state import get_pp_group, get_tp_group
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
@@ -369,17 +368,20 @@ class Worker(WorkerBase):
assert isinstance(output, IntermediateTensors)
get_pp_group().send_tensor_dict(output.tensors,
all_gather_group=get_tp_group())
if not has_kv_transfer_group():
kv_connector_output = output.kv_connector_output
if not kv_connector_output:
return None
# In case of PP with kv transfer, we need to pass through the
# finished_sending and finished_recving buffers.
new_output = EMPTY_MODEL_RUNNER_OUTPUT
if output.finished_sending or output.finished_recving:
new_output = copy.copy(new_output)
new_output.finished_sending = output.finished_sending
new_output.finished_recving = output.finished_recving
output = new_output
# kv_connector_output
if (not kv_connector_output.finished_sending
and not kv_connector_output.finished_recving):
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.kv_connector_output = kv_connector_output
return output
assert isinstance(output, ModelRunnerOutput)
return output
@@ -4,6 +4,8 @@
Define KV connector functionality mixin for model runners.
"""
import copy
from contextlib import AbstractContextManager, contextmanager, nullcontext
from typing import Generator # noqa: UP035
from typing import TYPE_CHECKING, Optional
from vllm.config import VllmConfig
@@ -12,7 +14,8 @@ from vllm.distributed.kv_transfer import (get_kv_transfer_group,
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.forward_context import get_forward_context, set_forward_context
from vllm.logger import init_logger
from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT, ModelRunnerOutput
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, KVConnectorOutput,
ModelRunnerOutput)
if TYPE_CHECKING:
from vllm.v1.core.sched.output import SchedulerOutput
@@ -53,18 +56,60 @@ class KVConnectorModelRunnerMixin:
scheduler_output.finished_req_ids)
return None, None
def kv_connector_no_forward(self, scheduler_output: "SchedulerOutput",
@staticmethod
def kv_connector_no_forward(scheduler_output: "SchedulerOutput",
vllm_config: VllmConfig) -> ModelRunnerOutput:
# KV send/recv even if no work to do.
with set_forward_context(None, vllm_config):
self.maybe_setup_kv_connector(scheduler_output)
finished_sending, finished_recving = (
self.get_finished_kv_transfers(scheduler_output))
with set_forward_context(
None, vllm_config
), KVConnectorModelRunnerMixin._get_kv_connector_output(
scheduler_output, wait_for_save=False) as kv_connector_output:
pass
if not finished_sending and not finished_recving:
if (not kv_connector_output.finished_sending
and not kv_connector_output.finished_recving):
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.finished_sending = finished_sending
output.finished_recving = finished_recving
output.kv_connector_output = kv_connector_output
return output
@staticmethod
def maybe_get_kv_connector_output(
scheduler_output: "SchedulerOutput"
) -> AbstractContextManager[Optional[KVConnectorOutput]]:
return KVConnectorModelRunnerMixin._get_kv_connector_output(
scheduler_output) if has_kv_transfer_group() else nullcontext()
# This context manager must be used within an active forward context.
# It encapsulates the entire KV conector lifecycle within execute_model
@staticmethod
@contextmanager
def _get_kv_connector_output(
scheduler_output: "SchedulerOutput",
wait_for_save: bool = True
) -> Generator[KVConnectorOutput, None, None]:
output = KVConnectorOutput()
# Update KVConnector with the KVConnector metadata forward().
kv_connector = get_kv_transfer_group()
assert isinstance(kv_connector, KVConnectorBase_V1)
assert scheduler_output.kv_connector_metadata is not None
kv_connector.bind_connector_metadata(
scheduler_output.kv_connector_metadata)
# Background KV cache transfers happen here.
# These transfers are designed to be async and the requests
# involved may be disjoint from the running requests.
# Do this here to save a collective_rpc.
kv_connector.start_load_kv(get_forward_context())
try:
yield output
finally:
if wait_for_save:
kv_connector.wait_for_save()
output.finished_sending, output.finished_recving = (
kv_connector.get_finished(scheduler_output.finished_req_ids))
kv_connector.clear_connector_metadata()
+5 -4
View File
@@ -51,7 +51,7 @@ from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsLists,
from vllm.v1.sample.tpu.metadata import TPUSupportedSamplingMetadata
from vllm.v1.sample.tpu.sampler import Sampler as TPUSampler
from vllm.v1.worker.kv_connector_model_runner_mixin import (
KVConnectorModelRunnerMixin)
KVConnectorModelRunnerMixin, KVConnectorOutput)
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
from vllm.v1.worker.tpu_input_batch import CachedRequestState, InputBatch
@@ -1175,9 +1175,10 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
finished_sending=finished_sending,
finished_recving=finished_recving,
)
kv_connector_output=KVConnectorOutput(
finished_sending=finished_sending,
finished_recving=finished_recving,
))
# Check there are no new graphs compiled - all the graphs should be
# captured and compiled during warm up.