[BugFix] Make PD work with Ray (#21072)

Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
This commit is contained in:
kourosh hakhamaneshi
2025-07-19 08:46:50 -07:00
committed by GitHub
parent 6a971ed692
commit 9f414a12ad
11 changed files with 330 additions and 222 deletions
+43 -6
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import copy
import gc
import time
from contextlib import contextmanager
@@ -1270,6 +1271,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
num_scheduled_tokens_np: np.ndarray,
finished_sending: Optional[set[str]],
finished_recving: Optional[set[str]],
) -> ModelRunnerOutput:
assert self.input_batch.num_reqs ==\
len(self.input_batch.pooling_params), \
@@ -1304,6 +1307,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
finished_sending=finished_sending,
finished_recving=finished_recving,
)
@torch.inference_mode()
@@ -1314,12 +1319,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
) -> Union[ModelRunnerOutput, IntermediateTensors]:
self._update_states(scheduler_output)
if not scheduler_output.total_num_scheduled_tokens:
if has_kv_transfer_group():
with set_forward_context(None, self.vllm_config):
self.maybe_setup_kv_connector(scheduler_output)
if not has_kv_transfer_group():
# Return empty ModelRunnerOutput if there's no work to do.
return EMPTY_MODEL_RUNNER_OUTPUT
# Return empty ModelRunnerOutput if there's no work to do.
return EMPTY_MODEL_RUNNER_OUTPUT
return self.kv_connector_no_forward(scheduler_output)
# Prepare the decoder inputs.
(attn_metadata, attention_cuda_graphs, logits_indices,
@@ -1412,6 +1416,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
)
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
@@ -1429,6 +1435,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
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)
get_pp_group().send_tensor_dict(hidden_states.tensors,
@@ -1437,7 +1446,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
else:
if self.input_batch.pooling_params:
return self._pool(hidden_states, num_scheduled_tokens,
num_scheduled_tokens_np)
num_scheduled_tokens_np, finished_sending,
finished_recving)
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states, None)
@@ -1587,6 +1597,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
finished_sending=finished_sending,
finished_recving=finished_recving,
num_nans_in_logits=num_nans_in_logits,
)
@@ -1711,6 +1723,31 @@ class GPUModelRunner(LoRAModelRunnerMixin):
if has_kv_transfer_group():
get_kv_transfer_group().wait_for_save()
@staticmethod
def get_finished_kv_transfers(
scheduler_output: "SchedulerOutput",
) -> tuple[Optional[set[str]], Optional[set[str]]]:
if has_kv_transfer_group():
return get_kv_transfer_group().get_finished(
scheduler_output.finished_req_ids)
return None, None
def kv_connector_no_forward(
self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput:
# KV send/recv even if no work to do.
with set_forward_context(None, self.vllm_config):
self.maybe_setup_kv_connector(scheduler_output)
finished_sending, finished_recving = (
self.get_finished_kv_transfers(scheduler_output))
if not finished_sending and not finished_recving:
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.finished_sending = finished_sending
output.finished_recving = finished_recving
return output
def propose_ngram_draft_token_ids(
self,
sampled_token_ids: list[list[int]],