[P/D] NIXL Integration (#17751)

Signed-off-by: ApostaC <yihua98@uchicago.edu>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>
Signed-off-by: Robert Shaw <rshaw@neuralmagic.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Brent Salisbury <bsalisbu@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: ApostaC <yihua98@uchicago.edu>
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
Co-authored-by: Tyler Michael Smith <tysmith@redhat.com>
Co-authored-by: Brent Salisbury <bsalisbu@redhat.com>
This commit is contained in:
Robert Shaw
2025-05-12 09:46:16 -07:00
committed by GitHub
co-authored by Tyler Michael Smith ApostaC Robert Shaw mgoin Nick Hill Tyler Michael Smith Brent Salisbury
parent 05a4324f8e
commit d19110204c
34 changed files with 2723 additions and 108 deletions
-21
View File
@@ -11,10 +11,6 @@ import torch.distributed as dist
import vllm.envs as envs
from vllm.config import VllmConfig
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
has_kv_transfer_group,
is_v1_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.logger import init_logger
if TYPE_CHECKING:
@@ -106,16 +102,6 @@ def set_forward_context(attn_metadata: Any,
attn_metadata=attn_metadata,
dp_metadata=dp_metadata)
# KVConnector: trigger (possibly async) load before forward.
# Each attn layer will block until the reading is complete.
trigger_kv_transfer = (attn_metadata is not None
and has_kv_transfer_group()
and is_v1_kv_transfer_group())
if trigger_kv_transfer:
kv_connector = get_kv_transfer_group()
assert isinstance(kv_connector, KVConnectorBase_V1)
kv_connector.start_load_kv(_forward_context)
try:
yield
finally:
@@ -152,11 +138,4 @@ def set_forward_context(attn_metadata: Any,
"(batchsize, count, median_time(ms)): %s"),
forward_stats)
# KVConnector: each attn layer triggers (possibly async) save.
# Ensure all those operations complete before forward() is done.
if trigger_kv_transfer:
kv_connector = get_kv_transfer_group()
assert isinstance(kv_connector, KVConnectorBase_V1)
kv_connector.wait_for_save()
_forward_context = prev_context