apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 08:16:42 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[XPU]Fixed the issue of performance degradation caused by enabling ENABLE_V1_KVCACHE_SCHEDULER (#3900)

Browse Source 
* fix bug

* fix bug

* update

* udpate

* update
This commit is contained in:
yinwei
2025-09-05 19:17:25 +08:00
committed by GitHub
parent 473cde779f
commit 77c1bd0813
4 changed files with 20 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/worker/xpu_model_runner.py
View File

@@ -383,6 +383,7 @@ class XPUModelRunner(ModelRunnerBase):
req_len = len(req_dicts)
has_prefill_task = False
has_decode_task = False
for i in range(req_len):
request = req_dicts[i]
idx = request.idx
@@ -392,6 +393,9 @@ class XPUModelRunner(ModelRunnerBase):
prefill_end_index = request.prefill_end_index
length = prefill_end_index - prefill_start_index
input_ids = request.prompt_token_ids + request.output_token_ids
logger.debug(
f"Handle prefill request {request} at idx {idx} prefill_start_index {prefill_start_index} prefill_end_index {prefill_end_index} need_prefilled_token_num {len(input_ids)}"
)
self.share_inputs["input_ids"][idx : idx + 1, :length] = np.array(
input_ids[prefill_start_index:prefill_end_index]
)
@@ -401,6 +405,8 @@ class XPUModelRunner(ModelRunnerBase):
self.share_inputs["block_tables"][idx : idx + 1, :encoder_block_num] = np.array(
request.block_tables, dtype="int32"
)
if self.share_inputs["is_block_step"][idx]: # has tasks to continue to decode
has_decode_task = True
self.share_inputs["stop_flags"][idx : idx + 1] = False
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = prefill_start_index
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = length
@@ -474,7 +480,7 @@ class XPUModelRunner(ModelRunnerBase):
self.share_inputs["stop_seqs"][:stop_seqs_num, : len(request.get("stop_token_ids")[0])] = np.array(
request.get("stop_token_ids"), dtype="int64"
)
if has_prefill_task:
if has_prefill_task or has_decode_task:
self.share_inputs["not_need_stop"][0] = True
def process_prefill_inputs(self, req_dicts: List[Request]):