apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-07 09:31:35 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

After enabling "top_logprobs supports passing 0 and fix max_completion_tokens", an incorrect finish_reason was returned. (#2815)

Browse Source 
* /v1/chat/completions endpoint now supports max_completion_tokens and fixes the return value of finish_reason

* top_logprobs supports passing 0
This commit is contained in:
zhenwenDang
2025-07-11 16:53:12 +08:00
committed by GitHub
parent 4596dd7248
commit 87203ec87b
1 changed files with 17 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
fastdeploy/entrypoints/openai/serving_chat.py
View File

@@ -198,7 +198,8 @@ class OpenAIServingChat:
sampled_token_ranks=raw_top_logprobs[2],
)
logprobs_res = self.build_logprobs_response(
logprobs=top_logprobs,
request_logprobs= request.logprobs,
response_logprobs=top_logprobs,
request_top_logprobs=request.top_logprobs,
)
@@ -215,7 +216,9 @@ class OpenAIServingChat:
if res["finished"]:
num_choices -= 1
work_process_metrics.e2e_request_latency.observe(time.time() - res["metrics"]["request_start_time"])
if request.max_tokens is None or previous_num_tokens != request.max_tokens:
has_no_token_limit = request.max_tokens is None and request.max_completion_tokens is None
max_tokens = request.max_completion_tokens or request.max_tokens
if has_no_token_limit or previous_num_tokens != max_tokens:
choice.finish_reason = "stop"
if self.engine_client.reasoning_parser == "ernie_x1" and \
output.get("finish_reason", "") == "tool_calls":
@@ -323,7 +326,8 @@ class OpenAIServingChat:
sampled_token_ranks=raw_top_logprobs[2],
)
logprobs_res = self.build_logprobs_response(
logprobs=top_logprobs,
request_logprobs=request.logprobs,
response_logprobs=top_logprobs,
request_top_logprobs=request.top_logprobs,
)
if logprobs_res and logprobs_res.content is not None:
@@ -355,7 +359,9 @@ class OpenAIServingChat:
logprobs=logprobs_full_res,
finish_reason=None
)
if request.max_tokens is None or previous_num_tokens != request.max_tokens:
has_no_token_limit = request.max_tokens is None and request.max_completion_tokens is None
max_tokens = request.max_completion_tokens or request.max_tokens
if has_no_token_limit or previous_num_tokens != max_tokens:
choice.finish_reason = "stop"
if self.engine_client.reasoning_parser == "ernie_x1" and \
output.get("finish_reason", "") == "tool_calls":
@@ -386,7 +392,8 @@ class OpenAIServingChat:
def build_logprobs_response(
self,
logprobs: Optional[LogprobsLists],
request_logprobs: bool,
response_logprobs: Optional[LogprobsLists],
request_top_logprobs: int,
) -> Optional[LogProbs]:
"""
@@ -396,17 +403,17 @@ class OpenAIServingChat:
# Parameter validation
if (
logprobs is None
response_logprobs is None
or not request_logprobs
or request_top_logprobs is None
or request_top_logprobs <= 0
or len(logprobs.logprob_token_ids) == 0
or request_top_logprobs < 0
):
return None
try:
# The top-k candidates for the current token
topk_token_ids = logprobs.logprob_token_ids[0][:request_top_logprobs + 1]
topk_logprobs = logprobs.logprobs[0][:request_top_logprobs + 1]
topk_token_ids = response_logprobs.logprob_token_ids[0][:request_top_logprobs + 1]
topk_logprobs = response_logprobs.logprobs[0][:request_top_logprobs + 1]
# Construct the candidate token structure (LogProbEntry) of topk
top_logprob_entries: List[LogProbEntry] = []