apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-12 20:11:20 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Sync Code] Update vs branch (#3403)

Browse Source 
* Pre ce modified (#3335) (#3360)

* Pre ce modified (#3335)

* update

* update

* fix

* fix

* update

* update

* update

* fix

* update

* update

* update

* add ut fix pr(3367)

* [Bug Fix] Fix V1 video bug (#3387)

* fix stopseq error info (#3342)

Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>

* [BugFix] Fix default log level of paddleformers (#3377)

Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>

* [Polish Code] Remove useless notes

* feat(log):add_request_and_response_log (#3392)

* Optimize CI execution workflow. (#3371) (#3384)

* fix

* [BugFix] fix control signal release failed (#3374)

* [BugFix]

* [BugFix]

* [BugFix]

* [BugFix]

* fix

* fix

---------

Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com>

---------

Co-authored-by: YUNSHEN XIE <1084314248@qq.com>
Co-authored-by: ming1753 <61511741+ming1753@users.noreply.github.com>
Co-authored-by: JYChen <zoooo0820@qq.com>
Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
Co-authored-by: Jiang-Jia-Jun <jiangjiajun@baidu.com>
Co-authored-by: xiaolei373 <zley373@gmail.com>
Co-authored-by: ltd0924 <32387785+ltd0924@users.noreply.github.com>
This commit is contained in:
Jiang-Jia-Jun
2025-08-14 17:14:45 +08:00
committed by GitHub
parent 81092c0fe3
commit e11331927f
16 changed files with 329 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
fastdeploy/engine/sched/resource_manager_v1.py
View File

@@ -97,13 +97,13 @@ class ResourceManagerV1(ResourceManager):
def _prepare_preempt_task(self, request):
return ScheduledPreemptTask(idx=request.idx, request_id=request.request_id)
def reschedule_preempt_task(self, request_id):
with self.lock:
if request_id in self.to_be_rescheduled_request_id_set and request_id in self.requests:
request = self.requests[request_id]
self.waiting.appendleft(request)
self.to_be_rescheduled_request_id_set.remove(request_id)
self.to_be_rescheduled_request_id_set.remove(request_id)
def _trigger_preempt(self, request, num_new_blocks, preempted_reqs, scheduled_reqs):
can_schedule = True
@@ -142,26 +142,31 @@ class ResourceManagerV1(ResourceManager):
input_ids_lst = request.prompt_token_ids + request.output_token_ids
input_ids = paddle.to_tensor(input_ids_lst, dtype="int64")
grid_thw = []
for one in inputs["grid_thw"]:
if one[0] == 1:
grid_thw.append(one)
else:
grid_thw.extend([[2, one[1], one[2]]] * (one[0] // 2))
input_ids = paddle.to_tensor(input_ids_lst, dtype="int64")
image_patch_id = inputs["image_patch_id"]
grid_thw = paddle.to_tensor(grid_thw, dtype="int64")
if request.multimodal_img_boundaries is None:
grid_thw = []
for one in inputs["grid_thw"]:
if one[0] == 1:
grid_thw.append(one)
else:
grid_thw.extend([[2, one[1], one[2]]] * (one[0] // 2))
grid_thw = paddle.to_tensor(grid_thw, dtype="int64")
from fastdeploy.model_executor.ops.gpu import get_img_boundaries
request.multimodal_img_boundaries = get_img_boundaries(
task_input_ids=input_ids, grid_thw=grid_thw, image_patch_id=image_patch_id
).numpy()
grid_thw = grid_thw.numpy().reshape([-1, 3])
inputs["grid_thw"] = grid_thw
grid_thw = inputs["grid_thw"]
img_boundaries_idx = request.multimodal_img_boundaries[0]
img_num_per_boundary = request.multimodal_img_boundaries[1]
ori_prompt_len = img_boundaries_idx[-1].item()
grid_thw = grid_thw.numpy().reshape([-1, 3])
pre_end_idx = request.num_computed_tokens
new_end_idx = pre_end_idx + num_new_tokens
if new_end_idx < ori_prompt_len and input_ids[new_end_idx - 1] == image_patch_id:
@@ -421,9 +426,15 @@ class ResourceManagerV1(ResourceManager):
self.running.remove(request)
request.status = RequestStatus.FINISHED
self._free_blocks(request)
if request.request_id in self.to_be_rescheduled_request_id_set: # finished after preempted, blocks have been recycled.
self.to_be_rescheduled_request_id_set.remove(request.request_id) # just remove from to_be_rescheduled_request_id_set
if request in self.waiting: # after finished, this request still scheduled from preempted to waiting, unexpected error, should not be here
if (
request.request_id in self.to_be_rescheduled_request_id_set
): # finished after preempted, blocks have been recycled.
self.to_be_rescheduled_request_id_set.remove(
request.request_id
) # just remove from to_be_rescheduled_request_id_set
if (
request in self.waiting
): # after finished, this request still scheduled from preempted to waiting, unexpected error, should not be here
raise RuntimeError(f"request {request.request_id} scheduled into waiting list, after finished")
self.tasks_list[request.idx] = None
self.stop_flags[request.idx] = True