apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

support logprob in v1 for release/2.1 (#3446)

Browse Source
This commit is contained in:
chenjian
2025-08-17 08:16:00 +08:00
committed by GitHub
parent 1065406ed3
commit edd13aad66
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
fastdeploy/output/token_processor.py
View File

@@ -315,6 +315,11 @@ class TokenProcessor:
scores = self.output_scores[: batch * (K + 1)].numpy().reshape([batch, K + 1])[:, : (K + 1)] scores = self.output_scores[: batch * (K + 1)].numpy().reshape([batch, K + 1])[:, : (K + 1)]
ranks = self.output_ranks[:batch].numpy() ranks = self.output_ranks[:batch].numpy()
batch_result = list() batch_result = list()
if envs.ENABLE_V1_KVCACHE_SCHEDULER:
need_to_be_reschedule_req_ids = list(self.resource_manager.to_be_rescheduled_request_id_set)
for request_id in need_to_be_reschedule_req_ids:
if self.resource_manager.requests[request_id].idx >= (batch - 1): # No more token generated for preempted request
self.resource_manager.reschedule_preempt_task(request_id)
for i in range(batch): for i in range(batch):
if self.resource_manager.stop_flags[i]: if self.resource_manager.stop_flags[i]:
continue continue
@@ -326,6 +331,9 @@ class TokenProcessor:
if recovery_stop: if recovery_stop:
llm_logger.info(f"recovery stop signal found at task {task_id}") llm_logger.info(f"recovery stop signal found at task {task_id}")
if not recovery_stop and token_id < 0: if not recovery_stop and token_id < 0:
if envs.ENABLE_V1_KVCACHE_SCHEDULER:
if task_id in self.resource_manager.to_be_rescheduled_request_id_set:
self.resource_manager.reschedule_preempt_task(task_id)
continue continue
if task.get("prefill_chunk_info", None) is not None: if task.get("prefill_chunk_info", None) is not None:
@@ -383,6 +391,7 @@ class TokenProcessor:
self.tokens_counter[task_id] += 1 self.tokens_counter[task_id] += 1
if token_id != RECOVERY_STOP_SIGNAL: if token_id != RECOVERY_STOP_SIGNAL:
result.outputs.token_ids.append(token_id) result.outputs.token_ids.append(token_id)
task.output_token_ids.append(token_id)
result.outputs.logprob = float(scores[i, 0]) result.outputs.logprob = float(scores[i, 0])
# Construct top_logprobs # Construct top_logprobs
topk_token_ids = tokens[i, :].tolist() topk_token_ids = tokens[i, :].tolist()