[KVConnector][NIXL] Support pipeline-parallel prefill in push mode (#45880)
Signed-off-by: zixi-qi <zixi@inferact.ai> Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -185,6 +185,7 @@ class NixlBaseConnectorScheduler:
|
||||
host = params.get("remote_host")
|
||||
port = params.get("remote_port")
|
||||
tp_size = params.get("tp_size")
|
||||
pp_size = params.get("pp_size", 1)
|
||||
if (
|
||||
remote_engine_id is None
|
||||
or remote_request_id is None
|
||||
@@ -199,6 +200,7 @@ class NixlBaseConnectorScheduler:
|
||||
host=host,
|
||||
port=port,
|
||||
tp_size=tp_size,
|
||||
pp_size=pp_size,
|
||||
)
|
||||
self._heartbeat_by_engine[remote_engine_id].req_ids.add(remote_request_id)
|
||||
self._heartbeat_req_engine[request.request_id] = (
|
||||
@@ -244,7 +246,7 @@ class NixlBaseConnectorScheduler:
|
||||
)
|
||||
|
||||
def set_xfer_handshake_metadata(
|
||||
self, metadata: dict[int, KVConnectorHandshakeMetadata]
|
||||
self, metadata: dict[tuple[int, int], KVConnectorHandshakeMetadata]
|
||||
) -> None:
|
||||
"""
|
||||
Set the KV connector handshake metadata for this connector.
|
||||
@@ -252,19 +254,20 @@ class NixlBaseConnectorScheduler:
|
||||
Args:
|
||||
metadata (dict): the handshake metadata to set.
|
||||
"""
|
||||
encoded_data: dict[int, bytes] = {}
|
||||
encoded_data: dict[tuple[int, int], bytes] = {}
|
||||
encoder = msgspec.msgpack.Encoder()
|
||||
for tp_rank, rank_metadata in metadata.items():
|
||||
for (pp_rank, tp_rank), rank_metadata in metadata.items():
|
||||
if not isinstance(rank_metadata, NixlHandshakePayload):
|
||||
raise ValueError(
|
||||
"NixlConnectorScheduler expects NixlHandshakePayload for "
|
||||
"handshake metadata."
|
||||
)
|
||||
encoded_data[tp_rank] = encoder.encode(rank_metadata)
|
||||
encoded_data[(pp_rank, tp_rank)] = encoder.encode(rank_metadata)
|
||||
logger.debug(
|
||||
"Tp rank %d: encoded NixlHandshakePayload size: %s bytes",
|
||||
"PP rank %d, TP rank %d: encoded NixlHandshakePayload size: %s bytes",
|
||||
pp_rank,
|
||||
tp_rank,
|
||||
str(len(encoded_data[tp_rank])),
|
||||
str(len(encoded_data[(pp_rank, tp_rank)])),
|
||||
)
|
||||
|
||||
# Only start the listener when we have metadata to serve.
|
||||
@@ -287,7 +290,7 @@ class NixlBaseConnectorScheduler:
|
||||
|
||||
@staticmethod
|
||||
def _nixl_handshake_listener(
|
||||
encoded_data: dict[int, Any],
|
||||
encoded_data: dict[tuple[int, int], Any],
|
||||
ready_event: threading.Event,
|
||||
stop_event: threading.Event,
|
||||
host: str,
|
||||
@@ -310,15 +313,18 @@ class NixlBaseConnectorScheduler:
|
||||
if stop_event.is_set():
|
||||
break
|
||||
continue
|
||||
# Decode the message which contains (GET_META_MSG, rank)
|
||||
msg, target_tp_rank = msgspec.msgpack.decode(msg)
|
||||
# Decode (GET_META_MSG, pp_rank, tp_rank).
|
||||
msg, target_pp_rank, target_tp_rank = msgspec.msgpack.decode(msg)
|
||||
logger.debug(
|
||||
"Received message for tp rank %s",
|
||||
"Received message for pp rank %s, tp rank %s",
|
||||
target_pp_rank,
|
||||
target_tp_rank,
|
||||
)
|
||||
if msg != GET_META_MSG:
|
||||
logger.warning("Connection listener got unexpected message %s", msg)
|
||||
sock.send_multipart((identity, b"", encoded_data[target_tp_rank]))
|
||||
sock.send_multipart(
|
||||
(identity, b"", encoded_data[(target_pp_rank, target_tp_rank)])
|
||||
)
|
||||
|
||||
def _get_remote_prefill_token_count(self, num_prompt_tokens: int) -> int:
|
||||
"""D-side only. Returns N-1 for Mamba models since the decoder
|
||||
|
||||
@@ -2,6 +2,7 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Base worker-side logic for the NIXL connector."""
|
||||
|
||||
import itertools
|
||||
import logging
|
||||
import os
|
||||
import queue
|
||||
@@ -336,8 +337,11 @@ class NixlBaseConnectorWorker:
|
||||
)
|
||||
|
||||
self.nixl_wrapper = nixl_wrapper_cls(str(uuid.uuid4()), config)
|
||||
# Map of engine_id -> {rank0: agent_name0, rank1: agent_name1..}.
|
||||
self._remote_agents: dict[EngineId, dict[int, str]] = defaultdict(dict)
|
||||
# Map of engine_id -> {(pp_rank, tp_rank): agent_name, ...}.
|
||||
# non-PP remote uses pp_rank 0, i.e. (0, tp_rank).
|
||||
self._remote_agents: dict[EngineId, dict[tuple[int, int], str]] = defaultdict(
|
||||
dict
|
||||
)
|
||||
|
||||
# Metadata.
|
||||
self.engine_id: EngineId = engine_id
|
||||
@@ -411,6 +415,25 @@ class NixlBaseConnectorWorker:
|
||||
# (so 1 per layer for MLA, otherwise 2 per layer)
|
||||
self.num_regions = 0
|
||||
|
||||
# PP>1 (push mode): this worker holds a contiguous layer slice and
|
||||
# transfers into the matching sub-range of a PP=1 remote's regions.
|
||||
self.pp_size = vllm_config.parallel_config.pipeline_parallel_size
|
||||
self._remote_region_offset = 0
|
||||
# PP push slices regions per layer (uniform count); HMA breaks that.
|
||||
if self.pp_size > 1 and self._is_hma_required:
|
||||
raise NotImplementedError(
|
||||
"NixlPushConnector does not support pipeline_parallel_size > 1 "
|
||||
"with hybrid KV cache layouts (HMA) yet."
|
||||
)
|
||||
# Decode-side PP is unsupported (completions counted per consumer rank).
|
||||
if vllm_config.kv_transfer_config.kv_role == "kv_consumer" and self.pp_size > 1:
|
||||
raise NotImplementedError(
|
||||
"NixlPushConnector consumer (decode) does not support "
|
||||
"pipeline_parallel_size > 1."
|
||||
)
|
||||
# Keep heartbeat handshakes to a PP-sharded producer notif-only.
|
||||
self._hb_handshake_notif_only = False
|
||||
|
||||
# nixl_prepped_dlist_handle.
|
||||
self.src_xfer_handles_by_block_size: dict[int, int] = {}
|
||||
# Populated dynamically during handshake based on remote configuration.
|
||||
@@ -449,7 +472,7 @@ class NixlBaseConnectorWorker:
|
||||
thread_name_prefix="vllm-nixl-handshake-initiator",
|
||||
)
|
||||
self._ready_requests = queue.Queue[tuple[ReqId, ReqMeta]]()
|
||||
self._handshake_futures: dict[EngineId, Future[dict[int, str]]] = {}
|
||||
self._handshake_futures: dict[EngineId, Future[dict[tuple[int, int], str]]] = {}
|
||||
# Protects _handshake_futures and _remote_agents.
|
||||
self._handshake_lock = threading.RLock()
|
||||
|
||||
@@ -536,7 +559,9 @@ class NixlBaseConnectorWorker:
|
||||
port: int,
|
||||
remote_tp_size: int,
|
||||
expected_engine_id: str,
|
||||
) -> dict[int, str]:
|
||||
remote_pp_size: int = 1,
|
||||
notif_agents_only: bool = False,
|
||||
) -> dict[tuple[int, int], str]:
|
||||
"""Do a NIXL handshake with a remote instance."""
|
||||
|
||||
# the first time we connect to a remote agent.
|
||||
@@ -558,20 +583,25 @@ class NixlBaseConnectorWorker:
|
||||
# this happens to be the same single rank_i.
|
||||
assert self.transfer_topo is not None
|
||||
p_remote_ranks = self.transfer_topo.handshake_target_ranks(remote_tp_size)
|
||||
remote_rank_to_agent_name = {}
|
||||
remote_rank_to_agent_name: dict[tuple[int, int], str] = {}
|
||||
path = make_zmq_path("tcp", host, port)
|
||||
|
||||
with zmq_ctx(zmq.REQ, path) as sock:
|
||||
for remote_rank in p_remote_ranks:
|
||||
for remote_pp_rank, remote_rank in itertools.product(
|
||||
range(remote_pp_size), p_remote_ranks
|
||||
):
|
||||
logger.debug(
|
||||
"Querying metadata on path: %s at remote tp rank %s",
|
||||
"Querying metadata on path: %s at remote pp rank %s, tp rank %s",
|
||||
path,
|
||||
remote_pp_rank,
|
||||
remote_rank,
|
||||
)
|
||||
|
||||
start_time = time.perf_counter()
|
||||
# Send query for the request.
|
||||
msg = msgspec.msgpack.encode((GET_META_MSG, remote_rank))
|
||||
msg = msgspec.msgpack.encode(
|
||||
(GET_META_MSG, remote_pp_rank, remote_rank)
|
||||
)
|
||||
# Set receive timeout to 5 seconds to avoid hanging on dead server
|
||||
sock.setsockopt(zmq.RCVTIMEO, 5000) # milliseconds
|
||||
sock.send(msg)
|
||||
@@ -638,17 +668,44 @@ class NixlBaseConnectorWorker:
|
||||
)
|
||||
|
||||
# Register Remote agent.
|
||||
remote_agent_name = self.add_remote_agent(
|
||||
metadata, remote_rank, remote_tp_size
|
||||
)
|
||||
if notif_agents_only:
|
||||
remote_agent_name = self._add_notif_only_remote_agent(
|
||||
metadata, remote_tp_size
|
||||
)
|
||||
else:
|
||||
remote_agent_name = self.add_remote_agent(
|
||||
metadata, remote_rank, remote_tp_size
|
||||
)
|
||||
setup_agent_time = time.perf_counter()
|
||||
logger.debug(
|
||||
"NIXL handshake: add agent took: %s",
|
||||
setup_agent_time - got_metadata_time,
|
||||
)
|
||||
remote_rank_to_agent_name[remote_rank] = remote_agent_name
|
||||
remote_ranks = (remote_pp_rank, remote_rank)
|
||||
remote_rank_to_agent_name[remote_ranks] = remote_agent_name
|
||||
return remote_rank_to_agent_name
|
||||
|
||||
def _add_notif_only_remote_agent(
|
||||
self, metadata: NixlAgentMetadata, remote_tp_size: int
|
||||
) -> str:
|
||||
"""Load a remote agent for notifs only on the push-mode decode side.
|
||||
|
||||
Skips descriptor setup but records engine info for block accounting.
|
||||
"""
|
||||
assert self.transfer_topo is not None
|
||||
self.transfer_topo.register_remote_engine(
|
||||
metadata.engine_id,
|
||||
EngineTransferInfo(
|
||||
remote_tp_size=remote_tp_size,
|
||||
remote_block_size=metadata.block_size,
|
||||
remote_block_len=metadata.block_lens[0],
|
||||
remote_physical_blocks_per_logical=(
|
||||
metadata.physical_blocks_per_logical_kv_block
|
||||
),
|
||||
),
|
||||
)
|
||||
return self.nixl_wrapper.add_remote_agent(metadata.agent_metadata)
|
||||
|
||||
def initialize_host_xfer_buffer(self, kv_caches: dict[str, torch.Tensor]) -> None:
|
||||
"""
|
||||
Initialize transfer buffer in CPU mem for accelerators
|
||||
@@ -761,7 +818,9 @@ class NixlBaseConnectorWorker:
|
||||
host: str,
|
||||
port: int,
|
||||
tp_size: int,
|
||||
) -> Future[dict[int, str]] | None:
|
||||
pp_size: int = 1,
|
||||
notif_agents_only: bool = False,
|
||||
) -> Future[dict[tuple[int, int], str]] | None:
|
||||
"""
|
||||
Ensure a handshake is in-flight (or already done) for *engine_id*.
|
||||
|
||||
@@ -784,10 +843,12 @@ class NixlBaseConnectorWorker:
|
||||
port,
|
||||
tp_size,
|
||||
engine_id,
|
||||
pp_size,
|
||||
notif_agents_only,
|
||||
)
|
||||
self._handshake_futures[engine_id] = fut
|
||||
|
||||
def done_callback(f: Future[dict[int, str]], eid=engine_id):
|
||||
def done_callback(f: Future[dict[tuple[int, int], str]], eid=engine_id):
|
||||
with self._handshake_lock:
|
||||
del self._handshake_futures[eid]
|
||||
try:
|
||||
@@ -1114,6 +1175,15 @@ class NixlBaseConnectorWorker:
|
||||
self.kv_caches_base_addr[self.engine_id][self.tp_rank] = seen_base_addresses
|
||||
self.num_regions = len(caches_data)
|
||||
|
||||
if self.pp_size > 1:
|
||||
start_layer, end_layer = self.model_config.get_layers_start_end_indices(
|
||||
self.vllm_config.parallel_config
|
||||
)
|
||||
num_local_layers = end_layer - start_layer
|
||||
assert num_local_layers > 0 and self.num_regions % num_local_layers == 0
|
||||
regions_per_layer = self.num_regions // num_local_layers
|
||||
self._remote_region_offset = regions_per_layer * start_layer
|
||||
|
||||
if self.transfer_topo.virtually_split_kv_in_blocks:
|
||||
# NOTE (NickLucche) When FlashInfer is used, memory is registered
|
||||
# with joint KV for each block. This minimizes the overhead in
|
||||
@@ -1464,14 +1534,32 @@ class NixlBaseConnectorWorker:
|
||||
""" # noqa: E501
|
||||
engine_id = nixl_agent_meta.engine_id
|
||||
# TODO re-evaluate refreshing for scaling/recovery
|
||||
if remote_tp_rank in self._remote_agents.get(engine_id, {}):
|
||||
if (0, remote_tp_rank) in self._remote_agents.get(engine_id, {}):
|
||||
logger.debug(
|
||||
"Remote agent with engine_id %s and rank"
|
||||
"%s already exchanged metadata, skip handshake.",
|
||||
engine_id,
|
||||
remote_tp_rank,
|
||||
)
|
||||
return self._remote_agents[engine_id][remote_tp_rank]
|
||||
return self._remote_agents[engine_id][(0, remote_tp_rank)]
|
||||
|
||||
# Compare physical regions, not self.num_regions (doubled by
|
||||
# FlashInfer's virtual K/V split).
|
||||
num_local_regions = len(self.block_len_per_layer)
|
||||
if (
|
||||
self.pp_size > 1
|
||||
and len(nixl_agent_meta.kv_caches_base_addr) > num_local_regions
|
||||
):
|
||||
# This worker holds a PP layer-slice; the PP=1 remote registered
|
||||
# the full model. Slice its regions to our layer window so the
|
||||
# logic below sees congruent local/remote lists.
|
||||
start = self._remote_region_offset
|
||||
end = start + num_local_regions
|
||||
assert len(nixl_agent_meta.kv_caches_base_addr) >= end
|
||||
nixl_agent_meta.kv_caches_base_addr = nixl_agent_meta.kv_caches_base_addr[
|
||||
start:end
|
||||
]
|
||||
nixl_agent_meta.block_lens = nixl_agent_meta.block_lens[start:end]
|
||||
|
||||
### Register remote engine in TransferTopology (idempotent).
|
||||
assert self.transfer_topo is not None
|
||||
@@ -2098,7 +2186,12 @@ class NixlBaseConnectorWorker:
|
||||
# the **next** heartbeat for this remote can go through.
|
||||
if (
|
||||
self._ensure_handshake(
|
||||
engine_id, hb_info.host, hb_info.port, hb_info.tp_size
|
||||
engine_id,
|
||||
hb_info.host,
|
||||
hb_info.port,
|
||||
hb_info.tp_size,
|
||||
hb_info.pp_size,
|
||||
self._hb_handshake_notif_only and hb_info.pp_size > 1,
|
||||
)
|
||||
is not None
|
||||
):
|
||||
@@ -2374,19 +2467,23 @@ class NixlBaseConnectorWorker:
|
||||
"""
|
||||
assert engine_id in self._remote_agents
|
||||
|
||||
for handle in self.dst_xfer_side_handles.pop(engine_id).values():
|
||||
# Notif-only engines (push-mode D side) have no descriptor state.
|
||||
for handle in self.dst_xfer_side_handles.pop(engine_id, {}).values():
|
||||
self.nixl_wrapper.release_dlist_handle(handle)
|
||||
for agent_name in self._remote_agents.pop(engine_id).values():
|
||||
self.nixl_wrapper.remove_remote_agent(agent_name)
|
||||
|
||||
del self.kv_caches_base_addr[engine_id]
|
||||
del self.dst_num_blocks[engine_id]
|
||||
del self.tp_mappings[engine_id]
|
||||
self.kv_caches_base_addr.pop(engine_id, None)
|
||||
self.dst_num_blocks.pop(engine_id, None)
|
||||
self.tp_mappings.pop(engine_id, None)
|
||||
if self.transfer_topo is not None:
|
||||
self.transfer_topo.unregister_remote_engine(engine_id)
|
||||
|
||||
last_active = self._engine_last_active.pop(engine_id)
|
||||
if log_eviction:
|
||||
# Push P-side engines are tracked in _remote_agents but not in
|
||||
# _engine_last_active (they don't participate in stale eviction), so
|
||||
# tolerate a missing entry.
|
||||
last_active = self._engine_last_active.pop(engine_id, None)
|
||||
if log_eviction and last_active is not None:
|
||||
logger.info(
|
||||
"Evicted stale remote engine %s (inactive for %.1fs).",
|
||||
engine_id,
|
||||
|
||||
@@ -207,11 +207,12 @@ class NixlBaseConnector(KVConnectorBase_V1, SupportsHMA):
|
||||
assert self.connector_scheduler is not None
|
||||
return self.connector_scheduler.request_finished(request, block_ids)
|
||||
|
||||
def set_xfer_handshake_metadata(
|
||||
self, metadata: dict[int, KVConnectorHandshakeMetadata]
|
||||
def set_xfer_handshake_metadata_pp_aware(
|
||||
self, metadata: dict[tuple[int, int], KVConnectorHandshakeMetadata]
|
||||
) -> None:
|
||||
"""
|
||||
Set the KV connector handshake metadata for this connector.
|
||||
Set handshake metadata keyed by (pp_rank, tp_rank) so the side
|
||||
channel can serve every PP stage's agent metadata.
|
||||
|
||||
Args:
|
||||
metadata (dict): the handshake metadata to set.
|
||||
|
||||
@@ -147,6 +147,7 @@ class HeartbeatInfo:
|
||||
host: str
|
||||
port: int
|
||||
tp_size: int
|
||||
pp_size: int = 1
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -167,6 +168,8 @@ class ReqMeta:
|
||||
remote: RemoteMeta | None = None
|
||||
# Remote block size, discovered during NIXL handshake (push mode).
|
||||
remote_block_size: int | None = None
|
||||
# Remote producer pipeline-parallel size (push mode, D side).
|
||||
pp_size: int = 1
|
||||
|
||||
|
||||
class NixlConnectorMetadata(KVConnectorMetadata):
|
||||
@@ -196,6 +199,7 @@ class NixlConnectorMetadata(KVConnectorMetadata):
|
||||
# P workers don't need to receive tp_size from proxy here.
|
||||
tp_size=kv_transfer_params.get("tp_size", 1),
|
||||
remote_block_size=kv_transfer_params.get("remote_block_size"),
|
||||
pp_size=kv_transfer_params.get("pp_size", 1),
|
||||
)
|
||||
|
||||
def add_new_req_to_save(
|
||||
|
||||
@@ -183,7 +183,7 @@ class NixlPullConnectorWorker(NixlBaseConnectorWorker):
|
||||
notif_id = f"{meta.remote.request_id}:{self.world_size}".encode()
|
||||
remote_agents = self._remote_agents[meta.remote.engine_id]
|
||||
for rank_to_notify, agent in remote_agents.items():
|
||||
if rank_to_notify != read_specs[0].remote_rank:
|
||||
if rank_to_notify != (0, read_specs[0].remote_rank):
|
||||
self.nixl_wrapper.send_notif(agent, notif_msg=notif_id)
|
||||
|
||||
def _read_blocks(
|
||||
@@ -250,7 +250,7 @@ class NixlPullConnectorWorker(NixlBaseConnectorWorker):
|
||||
# just notify P worker that we have the blocks we need.
|
||||
if len(local_block_ids) == 0:
|
||||
# A full prefix cache hit is indicated with an empty list.
|
||||
agent_name = self._remote_agents[dst_engine_id][remote_rank]
|
||||
agent_name = self._remote_agents[dst_engine_id][(0, remote_rank)]
|
||||
try:
|
||||
self.nixl_wrapper.send_notif(agent_name, notif_msg=notif_id)
|
||||
except Exception as e:
|
||||
|
||||
@@ -186,6 +186,7 @@ class NixlPushConnectorScheduler(NixlBaseConnectorScheduler):
|
||||
"remote_host": params["remote_host"],
|
||||
"remote_port": params["remote_port"],
|
||||
"remote_tp_size": params["tp_size"],
|
||||
"remote_pp_size": params.get("pp_size", 1),
|
||||
}
|
||||
self._push_registration_deadlines[request.request_id] = (
|
||||
time.perf_counter() + self._push_registration_timeout
|
||||
@@ -238,6 +239,8 @@ class NixlPushConnectorScheduler(NixlBaseConnectorScheduler):
|
||||
# serving layer via abort_immediately. To keep P from
|
||||
# stranding the prefill blocks, we still register an empty
|
||||
# recv so the worker emits a notif that lets P free them.
|
||||
# Seed remote_block_ids so add_new_req_to_recv won't KeyError.
|
||||
params["remote_block_ids"] = ()
|
||||
self._reqs_need_recv[request.request_id] = (request, [])
|
||||
params["do_remote_prefill"] = False
|
||||
return False, None
|
||||
@@ -285,6 +288,7 @@ class NixlPushConnectorScheduler(NixlBaseConnectorScheduler):
|
||||
remote_host=self.side_channel_host,
|
||||
remote_port=self.side_channel_port,
|
||||
tp_size=self.vllm_config.parallel_config.tensor_parallel_size,
|
||||
pp_size=self.vllm_config.parallel_config.pipeline_parallel_size,
|
||||
remote_num_tokens=remote_num_tokens,
|
||||
)
|
||||
|
||||
|
||||
@@ -80,6 +80,10 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
):
|
||||
super().__init__(vllm_config, engine_id, kv_cache_config)
|
||||
|
||||
# Heartbeat handshakes to a PP-sharded producer must be notif-only,
|
||||
# like the PUSH_REG path.
|
||||
self._hb_handshake_notif_only = True
|
||||
|
||||
# Push-specific state.
|
||||
# P-side: outgoing WRITE handles awaiting completion, keyed by
|
||||
# request_id. Mutated by writer (submit) and main thread
|
||||
@@ -275,18 +279,22 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
and the request is re-queued onto ``_reg_send_inbox`` once it
|
||||
completes (at which point ``_ensure_handshake`` returns ``None`` and we
|
||||
send directly)."""
|
||||
remote_pp_size = reg_data.get("remote_pp_size", 1)
|
||||
fut = self._ensure_handshake(
|
||||
reg_data["remote_engine_id"],
|
||||
reg_data["remote_host"],
|
||||
reg_data["remote_port"],
|
||||
reg_data["remote_tp_size"],
|
||||
pp_size=remote_pp_size,
|
||||
# D never addresses P memory in push mode; just load P's agents.
|
||||
notif_agents_only=remote_pp_size > 1,
|
||||
)
|
||||
if fut is None:
|
||||
self._do_send_reg_notif(req_id, reg_data)
|
||||
return
|
||||
|
||||
def _on_handshake(
|
||||
f: Future[dict[int, str]],
|
||||
f: Future[dict[tuple[int, int], str]],
|
||||
rid: str = req_id,
|
||||
rd: dict[str, Any] = reg_data,
|
||||
) -> None:
|
||||
@@ -500,28 +508,43 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
remote_block_ids = meta.remote.block_ids
|
||||
local_block_ids = meta.local_physical_block_ids
|
||||
num_groups = len(local_block_ids)
|
||||
read_specs = [
|
||||
ReadSpec(
|
||||
remote_rank=rank,
|
||||
local_block_ids=[
|
||||
list(local_block_ids[g])
|
||||
if rank in plan.source_ranks_per_group[g]
|
||||
else []
|
||||
for g in range(num_groups)
|
||||
],
|
||||
remote_block_ids=[
|
||||
list(remote_block_ids[g])
|
||||
if rank in plan.source_ranks_per_group[g]
|
||||
else []
|
||||
for g in range(num_groups)
|
||||
],
|
||||
)
|
||||
for rank in plan.all_source_ranks
|
||||
]
|
||||
|
||||
if self.use_mla and tp_ratio < 0:
|
||||
assert len(read_specs) == 1
|
||||
# MLA latent is replicated across D's TP ranks: the tp-mapping
|
||||
# collapses to one rank (fine for reads), but push must WRITE every
|
||||
# D rank or the rest decode stale KV; only the dst differs per rank.
|
||||
assert len(plan.all_source_ranks) == 1
|
||||
mla_local_ids = [list(ids) for ids in local_block_ids]
|
||||
mla_remote_ids = [list(ids) for ids in remote_block_ids]
|
||||
read_specs = [
|
||||
ReadSpec(
|
||||
remote_rank=rank,
|
||||
local_block_ids=mla_local_ids,
|
||||
remote_block_ids=mla_remote_ids,
|
||||
)
|
||||
for rank in self.dst_xfer_side_handles[engine_id]
|
||||
]
|
||||
else:
|
||||
read_specs = [
|
||||
ReadSpec(
|
||||
remote_rank=rank,
|
||||
local_block_ids=[
|
||||
list(local_block_ids[g])
|
||||
if rank in plan.source_ranks_per_group[g]
|
||||
else []
|
||||
for g in range(num_groups)
|
||||
],
|
||||
remote_block_ids=[
|
||||
list(remote_block_ids[g])
|
||||
if rank in plan.source_ranks_per_group[g]
|
||||
else []
|
||||
for g in range(num_groups)
|
||||
],
|
||||
)
|
||||
for rank in plan.all_source_ranks
|
||||
]
|
||||
|
||||
handles: list[int] = []
|
||||
for i, spec in enumerate(read_specs):
|
||||
remote_block_size = remote_info.remote_block_size
|
||||
logger.debug(
|
||||
@@ -544,7 +567,7 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
spec.remote_rank
|
||||
]
|
||||
|
||||
self._xfer_blocks(
|
||||
handle = self._xfer_blocks(
|
||||
read_spec=spec,
|
||||
request_id=req_id,
|
||||
dst_engine_id=meta.remote.engine_id,
|
||||
@@ -552,13 +575,15 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
local_xfer_side_handle=local_xfer_side_handle,
|
||||
remote_xfer_side_handle=remote_xfer_side_handle,
|
||||
)
|
||||
if handle is not None:
|
||||
handles.append(handle)
|
||||
|
||||
if self.use_mla and tp_ratio < 0 and read_specs:
|
||||
notif_id = f"{meta.remote.request_id}:{self.world_size}".encode()
|
||||
remote_agents = self._remote_agents[meta.remote.engine_id]
|
||||
for rank_to_notify, agent in remote_agents.items():
|
||||
if rank_to_notify != read_specs[0].remote_rank:
|
||||
self.nixl_wrapper.send_notif(agent, notif_msg=notif_id)
|
||||
# Publish all the request's WRITE handles in one locked update: a
|
||||
# partial set would let ``_pop_done_transfers`` finish the request
|
||||
# early, then double-report it as the remaining writes land.
|
||||
if handles:
|
||||
with self._sending_transfers_lock:
|
||||
self._sending_transfers[req_id].extend(handles)
|
||||
|
||||
def _xfer_blocks(
|
||||
self,
|
||||
@@ -568,8 +593,12 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
remote_request_id: str,
|
||||
local_xfer_side_handle: int,
|
||||
remote_xfer_side_handle: int,
|
||||
):
|
||||
"""Post a WRITE point-to-point xfer request."""
|
||||
) -> int | None:
|
||||
"""Post a WRITE point-to-point xfer request.
|
||||
|
||||
Returns the in-flight transfer handle (so the caller can track all of
|
||||
a request's handles atomically), or ``None`` if nothing was submitted.
|
||||
"""
|
||||
assert self.transfer_topo is not None
|
||||
remote_rank = read_spec.remote_rank
|
||||
local_block_ids = read_spec.local_block_ids
|
||||
@@ -597,7 +626,7 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
|
||||
if len(local_block_ids) == 0:
|
||||
logger.warning("No blocks to push for request %s", request_id)
|
||||
return
|
||||
return None
|
||||
|
||||
# Align per-group block counts for push.
|
||||
local_block_ids = list(local_block_ids)
|
||||
@@ -637,9 +666,9 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
notif_msg=notif_id,
|
||||
)
|
||||
self.nixl_wrapper.transfer(handle)
|
||||
# Track push WRITE handles so P can free blocks once done.
|
||||
with self._sending_transfers_lock:
|
||||
self._sending_transfers[request_id].append(handle)
|
||||
# Caller tracks the handle (atomically with the request's other
|
||||
# writes) so P can free blocks once all of them are done.
|
||||
return handle
|
||||
except Exception as e:
|
||||
self._log_failure(
|
||||
failure_type="transfer_setup_failed",
|
||||
@@ -656,6 +685,7 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
if handle is not None:
|
||||
self.nixl_wrapper.release_xfer_handle(handle)
|
||||
self.xfer_stats.record_failed_transfer()
|
||||
return None
|
||||
|
||||
# --- Notification handling on engine main thread ------------------ #
|
||||
|
||||
@@ -683,12 +713,20 @@ class NixlPushConnectorWorker(NixlBaseConnectorWorker):
|
||||
|
||||
# Not tracked as a P-side send/process for this notif.
|
||||
if req_id not in self._reqs_to_send and req_id not in self._reqs_to_process:
|
||||
if req_id in self._recving_metadata:
|
||||
# D-side: P signalled push completion. The transfer was
|
||||
# driven entirely by P (we don't own a NIXL handle here),
|
||||
# so materialise an empty entry in ``_recving_transfers``
|
||||
# and let ``_pop_done_transfers`` report it done on the
|
||||
# next ``get_finished``.
|
||||
if (meta := self._recving_metadata.get(req_id)) is not None:
|
||||
# Consumer waits for one notif per producer rank writing
|
||||
# here: pp_size stages * producers-per-consumer (>1 when
|
||||
# producer TP > consumer TP; tp_size is the producer TP).
|
||||
producers_per_consumer = max(1, int(tp_size) // self.world_size)
|
||||
expected_notifs = meta.pp_size * producers_per_consumer
|
||||
self.consumer_notification_counts_by_req[req_id] += 1
|
||||
notifs = self.consumer_notification_counts_by_req[req_id]
|
||||
if notifs < expected_notifs:
|
||||
continue
|
||||
del self.consumer_notification_counts_by_req[req_id]
|
||||
# P drove the transfer (we own no NIXL handle), so
|
||||
# materialise an empty ``_recving_transfers`` entry for
|
||||
# ``_pop_done_transfers`` to report done.
|
||||
self._recving_transfers.setdefault(req_id, [])
|
||||
else:
|
||||
# Not tracked on either side (lease may have expired
|
||||
|
||||
Reference in New Issue
Block a user