gahow/agentic-pd-hybrid
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add transfer queue admission knobs

Browse Source
This commit is contained in:
Gahow Wang
2026-04-25 17:29:15 +00:00
parent fe583fb413
commit c928c7db23
3 changed files with 102 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/agentic_pd_hybrid/benchmark.py
View File

@@ -38,6 +38,8 @@ class BenchmarkConfig:
kvcache_seed_only_multiturn_sessions: bool = False kvcache_seed_only_multiturn_sessions: bool = False
kvcache_prefill_backup_policy: str = "release-after-transfer" kvcache_prefill_backup_policy: str = "release-after-transfer"
kvcache_seed_max_inflight_decode: int | None = 3 kvcache_seed_max_inflight_decode: int | None = 3
kvcache_seed_max_decode_transfer_queue_reqs: int | None = None
kvcache_direct_max_decode_transfer_queue_reqs: int | None = None
kvcache_prefill_priority_eviction: bool = False kvcache_prefill_priority_eviction: bool = False
kvcache_prefill_direct_priority: int = -100 kvcache_prefill_direct_priority: int = -100
kvcache_prefill_normal_priority: int = 100 kvcache_prefill_normal_priority: int = 100
@@ -174,6 +176,12 @@ def run_live_benchmark(config: BenchmarkConfig) -> BenchmarkArtifacts:
kvcache_seed_max_inflight_decode=( kvcache_seed_max_inflight_decode=(
config.kvcache_seed_max_inflight_decode config.kvcache_seed_max_inflight_decode
), ),
kvcache_seed_max_decode_transfer_queue_reqs=(
config.kvcache_seed_max_decode_transfer_queue_reqs
),
kvcache_direct_max_decode_transfer_queue_reqs=(
config.kvcache_direct_max_decode_transfer_queue_reqs
),
kvcache_prefill_priority_eviction=( kvcache_prefill_priority_eviction=(
config.kvcache_prefill_priority_eviction config.kvcache_prefill_priority_eviction
), ),
@@ -220,6 +228,12 @@ def run_live_benchmark(config: BenchmarkConfig) -> BenchmarkArtifacts:
"kvcache_seed_max_inflight_decode": ( "kvcache_seed_max_inflight_decode": (
config.kvcache_seed_max_inflight_decode config.kvcache_seed_max_inflight_decode
), ),
"kvcache_seed_max_decode_transfer_queue_reqs": (
config.kvcache_seed_max_decode_transfer_queue_reqs
),
"kvcache_direct_max_decode_transfer_queue_reqs": (
config.kvcache_direct_max_decode_transfer_queue_reqs
),
"kvcache_prefill_priority_eviction": ( "kvcache_prefill_priority_eviction": (
config.kvcache_prefill_priority_eviction config.kvcache_prefill_priority_eviction
), ),

64
src/agentic_pd_hybrid/cli.py
View File

@@ -197,6 +197,26 @@ def main() -> None:
"Use a negative value to disable this filter." "Use a negative value to disable this filter."
), ),
) )
replay.add_argument(
"--kvcache-seed-max-decode-transfer-queue-reqs",
type=int,
default=None,
help=(
"For kvcache-centric worker admission, skip seed/reseed when the "
"target D reports more transfer-queue requests than this value. "
"Use a negative value to disable this filter."
),
)
replay.add_argument(
"--kvcache-direct-max-decode-transfer-queue-reqs",
type=int,
default=None,
help=(
"For kvcache-centric worker admission, skip direct-to-D append when "
"the target D reports more transfer-queue requests than this value. "
"Use a negative value to disable this filter."
),
)
replay.add_argument( replay.add_argument(
"--kvcache-prefill-priority-eviction", "--kvcache-prefill-priority-eviction",
action="store_true", action="store_true",
@@ -389,6 +409,26 @@ def main() -> None:
"Use a negative value to disable this filter." "Use a negative value to disable this filter."
), ),
) )
benchmark.add_argument(
"--kvcache-seed-max-decode-transfer-queue-reqs",
type=int,
default=None,
help=(
"For kvcache-centric worker admission, skip seed/reseed when the "
"target D reports more transfer-queue requests than this value. "
"Use a negative value to disable this filter."
),
)
benchmark.add_argument(
"--kvcache-direct-max-decode-transfer-queue-reqs",
type=int,
default=None,
help=(
"For kvcache-centric worker admission, skip direct-to-D append when "
"the target D reports more transfer-queue requests than this value. "
"Use a negative value to disable this filter."
),
)
benchmark.add_argument( benchmark.add_argument(
"--kvcache-prefill-priority-eviction", "--kvcache-prefill-priority-eviction",
action="store_true", action="store_true",
@@ -456,6 +496,18 @@ def main() -> None:
if args.kvcache_seed_max_inflight_decode < 0 if args.kvcache_seed_max_inflight_decode < 0
else args.kvcache_seed_max_inflight_decode else args.kvcache_seed_max_inflight_decode
), ),
kvcache_seed_max_decode_transfer_queue_reqs=(
None
if args.kvcache_seed_max_decode_transfer_queue_reqs is None
or args.kvcache_seed_max_decode_transfer_queue_reqs < 0
else args.kvcache_seed_max_decode_transfer_queue_reqs
),
kvcache_direct_max_decode_transfer_queue_reqs=(
None
if args.kvcache_direct_max_decode_transfer_queue_reqs is None
or args.kvcache_direct_max_decode_transfer_queue_reqs < 0
else args.kvcache_direct_max_decode_transfer_queue_reqs
),
kvcache_prefill_priority_eviction=( kvcache_prefill_priority_eviction=(
args.kvcache_prefill_priority_eviction args.kvcache_prefill_priority_eviction
), ),
@@ -582,6 +634,18 @@ def main() -> None:
if args.kvcache_seed_max_inflight_decode < 0 if args.kvcache_seed_max_inflight_decode < 0
else args.kvcache_seed_max_inflight_decode else args.kvcache_seed_max_inflight_decode
), ),
kvcache_seed_max_decode_transfer_queue_reqs=(
None
if args.kvcache_seed_max_decode_transfer_queue_reqs is None
or args.kvcache_seed_max_decode_transfer_queue_reqs < 0
else args.kvcache_seed_max_decode_transfer_queue_reqs
),
kvcache_direct_max_decode_transfer_queue_reqs=(
None
if args.kvcache_direct_max_decode_transfer_queue_reqs is None
or args.kvcache_direct_max_decode_transfer_queue_reqs < 0
else args.kvcache_direct_max_decode_transfer_queue_reqs
),
kvcache_prefill_priority_eviction=( kvcache_prefill_priority_eviction=(
args.kvcache_prefill_priority_eviction args.kvcache_prefill_priority_eviction
), ),

24
src/agentic_pd_hybrid/replay.py
View File

@@ -59,6 +59,8 @@ class ReplayConfig:
"release-after-transfer" "release-after-transfer"
) )
kvcache_seed_max_inflight_decode: int | None = 3 kvcache_seed_max_inflight_decode: int | None = 3
kvcache_seed_max_decode_transfer_queue_reqs: int | None = None
kvcache_direct_max_decode_transfer_queue_reqs: int | None = None
kvcache_prefill_priority_eviction: bool = False kvcache_prefill_priority_eviction: bool = False
kvcache_prefill_direct_priority: int = -100 kvcache_prefill_direct_priority: int = -100
kvcache_prefill_normal_priority: int = 100 kvcache_prefill_normal_priority: int = 100
@@ -922,16 +924,29 @@ async def _fetch_decode_load_snapshot(
def _decode_load_backpressure_reason( def _decode_load_backpressure_reason(
snapshot: DecodeLoadSnapshot | None, snapshot: DecodeLoadSnapshot | None,
*, *,
config: ReplayConfig,
routing_mode: Literal["direct", "seed"], routing_mode: Literal["direct", "seed"],
) -> str | None: ) -> str | None:
if snapshot is None: if snapshot is None:
return None return None
if routing_mode == "direct": if routing_mode == "direct":
if (
config.kvcache_direct_max_decode_transfer_queue_reqs is not None
and snapshot.decode_transfer_queue_reqs
> config.kvcache_direct_max_decode_transfer_queue_reqs
):
return "d-transfer-backpressure"
if snapshot.decode_retracted_queue_reqs > 0 and snapshot.token_usage >= 0.99: if snapshot.decode_retracted_queue_reqs > 0 and snapshot.token_usage >= 0.99:
return "d-retracted" return "d-retracted"
if snapshot.token_usage >= 0.992: if snapshot.token_usage >= 0.992:
return "d-token-usage-critical" return "d-token-usage-critical"
else: else:
if (
config.kvcache_seed_max_decode_transfer_queue_reqs is not None
and snapshot.decode_transfer_queue_reqs
> config.kvcache_seed_max_decode_transfer_queue_reqs
):
return "d-transfer-backpressure"
if snapshot.decode_retracted_queue_reqs > 0: if snapshot.decode_retracted_queue_reqs > 0:
return "d-retracted" return "d-retracted"
if snapshot.token_usage >= 0.985: if snapshot.token_usage >= 0.985:
@@ -949,6 +964,7 @@ def _is_decode_backpressure_reason(reason: str | None) -> bool:
"d-retracted", "d-retracted",
"d-token-usage-critical", "d-token-usage-critical",
"d-prealloc-backpressure", "d-prealloc-backpressure",
"d-transfer-backpressure",
} }
@@ -1221,6 +1237,7 @@ async def _reserve_prefill_backup_capacity(
async def _reserve_decode_session_capacity( async def _reserve_decode_session_capacity(
*, *,
client: httpx.AsyncClient, client: httpx.AsyncClient,
config: ReplayConfig,
request: TraceRequest, request: TraceRequest,
server_url: str, server_url: str,
session: DirectSessionState, session: DirectSessionState,
@@ -1233,6 +1250,7 @@ async def _reserve_decode_session_capacity(
if admission_mode == "router": if admission_mode == "router":
return await _reserve_decode_session_capacity_from_router_state( return await _reserve_decode_session_capacity_from_router_state(
client=client, client=client,
config=config,
request=request, request=request,
server_url=server_url, server_url=server_url,
session=session, session=session,
@@ -1315,6 +1333,7 @@ async def _reserve_decode_session_capacity(
residency.capacity_tokens[server_url] = load_snapshot.max_total_num_tokens residency.capacity_tokens[server_url] = load_snapshot.max_total_num_tokens
backpressure_reason = _decode_load_backpressure_reason( backpressure_reason = _decode_load_backpressure_reason(
load_snapshot, load_snapshot,
config=config,
routing_mode=routing_mode, routing_mode=routing_mode,
) )
if backpressure_reason is not None: if backpressure_reason is not None:
@@ -1429,6 +1448,7 @@ async def _reserve_decode_session_capacity(
async def _reserve_decode_session_capacity_from_router_state( async def _reserve_decode_session_capacity_from_router_state(
*, *,
client: httpx.AsyncClient, client: httpx.AsyncClient,
config: ReplayConfig,
request: TraceRequest, request: TraceRequest,
server_url: str, server_url: str,
session: DirectSessionState, session: DirectSessionState,
@@ -1836,6 +1856,7 @@ async def _execute_request(
can_seed, reserved_tokens, _evicted, _p_backed, seed_reason = ( can_seed, reserved_tokens, _evicted, _p_backed, seed_reason = (
await _reserve_decode_session_capacity( await _reserve_decode_session_capacity(
client=client, client=client,
config=config,
request=request, request=request,
server_url=decode_url, server_url=decode_url,
session=decode_session, session=decode_session,
@@ -1911,6 +1932,7 @@ async def _execute_request(
) = ( ) = (
await _reserve_decode_session_capacity( await _reserve_decode_session_capacity(
client=client, client=client,
config=config,
request=request, request=request,
server_url=decode_url, server_url=decode_url,
session=decode_session, session=decode_session,
@@ -1996,6 +2018,7 @@ async def _execute_request(
) = ( ) = (
await _reserve_decode_session_capacity( await _reserve_decode_session_capacity(
client=client, client=client,
config=config,
request=request, request=request,
server_url=decode_url, server_url=decode_url,
session=decode_session, session=decode_session,
@@ -2084,6 +2107,7 @@ async def _execute_request(
) = ( ) = (
await _reserve_decode_session_capacity( await _reserve_decode_session_capacity(
client=client, client=client,
config=config,
request=request, request=request,
server_url=decode_url, server_url=decode_url,
session=decode_session, session=decode_session,