[Diffusion] Prepare HunyuanImage3 for Scheduler-managed paged KV cache (#5541)

Signed-off-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com>
Co-authored-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
authored and GitHub committed 2026-08-10 08:48:56 +08:00
1 parent 615de4e041
commit 6ae7ff7846
24 files changed
+1728 -277

No files matched your search

@@ -10,6 +10,7 @@ from dataclasses import fields
from vllm.logger import init_logger
from vllm_omni.diffusion.data import OmniDiffusionConfig
from vllm_omni.diffusion.diffusion_kv.config import DiffusionKVCacheMode
from vllm_omni.diffusion.request import OmniDiffusionRequest
from vllm_omni.diffusion.sched.interface import (
CachedRequestData,
@@ -47,6 +48,7 @@ class BaseScheduler(ABC):
self._finished_req_ids: set[str] = set()
self.max_num_running_reqs: int = 1
self._prefetch_enabled: bool = False
self._diffusion_kv_enabled = False
def initialize(self, od_config: OmniDiffusionConfig) -> None:
self.od_config = od_config
@@ -63,6 +65,10 @@ class BaseScheduler(ABC):
self.max_num_running_reqs = 1
omni_kv = getattr(od_config, "omni_kv_config", None) or {}
self._prefetch_enabled = bool(omni_kv.get("enable_kv_async_prefetch", False))
self._diffusion_kv_enabled = (
getattr(od_config, "diffusion_kv_mode", DiffusionKVCacheMode.DENSE_LEGACY)
is DiffusionKVCacheMode.PAGED_SCHEDULER
)
self._reset_scheduler_state()
def add_request(self, request: OmniDiffusionRequest) -> str:
@@ -251,12 +257,43 @@ class BaseScheduler(ABC):
"""Remove subclass-owned per-request state before popping request state."""
def _make_request_state(self, request_id: str, request: OmniDiffusionRequest) -> SchedulerRequestState:
kv_requests = request.diffusion_kv_requests or ()
if self._diffusion_kv_enabled:
self._reject_legacy_dense_kv(request)
if not kv_requests:
raise ValueError("paged_scheduler request preprocessing did not produce DiffusionKVRequest state")
elif kv_requests:
raise ValueError("dense_legacy request unexpectedly contains Scheduler Diffusion KV requests")
# DiffusionKVRequest objects are mutable Scheduler/native-KVCacheManager
# state and must never ride the normal request payload to a Worker.
request.diffusion_kv_requests = None
return SchedulerRequestState(
request_id=request_id,
req=request,
sampling_params_key=self._build_sampling_params_key(request),
diffusion_kv_requests=kv_requests,
)
@staticmethod
def _reject_legacy_dense_kv(request: OmniDiffusionRequest) -> None:
"""Keep dense injected KV out of the Scheduler-owned paged path."""
populated_fields: list[str] = []
for owner_name, owner in (
("request", request),
("sampling_params", request.sampling_params),
):
for field_name, value in vars(owner).items():
if value is not None and (field_name == "past_key_values" or field_name.endswith("_past_key_values")):
populated_fields.append(f"{owner_name}.{field_name}")
if populated_fields:
fields_text = ", ".join(sorted(populated_fields))
raise ValueError(
"paged_scheduler Diffusion KV does not accept legacy dense KV payloads; "
f"clear these fields before admission: {fields_text}"
)
def _can_schedule_waiting(self, state: SchedulerRequestState) -> bool:
if not self._running:
return True
+5 -1
View File
@@ -6,10 +6,13 @@ from __future__ import annotations
import enum
from dataclasses import dataclass
from functools import cached_property
from typing import Any, TypedDict
from typing import TYPE_CHECKING, Any, TypedDict
from vllm_omni.diffusion.request import OmniDiffusionRequest
if TYPE_CHECKING:
from vllm_omni.diffusion.diffusion_kv.request import DiffusionKVRequest
class DiffusionRequestStatus(enum.IntEnum):
"""Request status tracked by diffusion scheduler."""
@@ -129,6 +132,7 @@ class SchedulerRequestState:
request_id: str
req: OmniDiffusionRequest
sampling_params_key: StepBatchSamplingParamsKey | RequestBatchSamplingParamsKey | None = None
diffusion_kv_requests: tuple[DiffusionKVRequest, ...] = ()
status: DiffusionRequestStatus = DiffusionRequestStatus.WAITING
error: str | None = None