apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[fix]Modify follow-up push parameters and Modify the verification method for thinking length (#4086)

Browse Source 
* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* add completion_token_ids

* add logger

* fix reasoning_max_tokens ParameterError

* add unittest

* add unittest

* add unittest

* add unittest

* add unittest

* add unit test
This commit is contained in:
luukunn
2025-09-19 14:26:01 +08:00
committed by GitHub
parent 66a98b44ed
commit ee9d8a840a
6 changed files with 75 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py
View File

@@ -241,10 +241,8 @@ class Ernie4_5_VLProcessor(Ernie4_5Processor):
else:
raise ValueError(f"Request must contain 'prompt', or 'messages': {request}")
metadata = request.get("metadata")
# 如果metadata包含之前输出的token，将这些token添加到input_ids末尾
if metadata and metadata.get("generated_token_ids"):
self.append_generated_tokens(outputs, metadata["generated_token_ids"])
if request.get("completion_token_ids"):
self.append_completion_tokens(outputs, request["completion_token_ids"])
outputs = self.pack_outputs(outputs)
request["prompt_token_ids"] = outputs["input_ids"].tolist()
request["prompt_token_ids_len"] = len(request["prompt_token_ids"])
@@ -259,11 +257,11 @@ class Ernie4_5_VLProcessor(Ernie4_5Processor):
return request
def append_generated_tokens(self, multimodal_inputs, generated_token_ids):
"append already generated tokens"
def append_completion_tokens(self, multimodal_inputs, completion_token_ids):
"append already completion tokens"
num_tokens = len(generated_token_ids)
multimodal_inputs["input_ids"].extend(generated_token_ids)
num_tokens = len(completion_token_ids)
multimodal_inputs["input_ids"].extend(completion_token_ids)
multimodal_inputs["token_type_ids"].extend([IDS_TYPE_FLAG["text"]] * num_tokens)
start = multimodal_inputs["cur_position"]

16
fastdeploy/input/qwen_vl_processor/qwen_vl_processor.py
View File

@@ -245,15 +245,11 @@ class QwenVLProcessor(TextProcessor):
else:
raise ValueError(f"Request must contain 'prompt', or 'messages': {request}")
metadata = request.get("metadata")
# Handle continuation of previous generation by appending existing tokens
if metadata and metadata.get("generated_token_ids"):
self.append_generated_tokens(outputs, metadata["generated_token_ids"])
if request.get("completion_token_ids"):
self.append_completion_tokens(outputs, request["completion_token_ids"])
enable_thinking = False
if metadata:
enable_thinking = metadata.get("enable_thinking", False)
if request.get("chat_template_kwargs"):
chat_template_kwargs = request.get("chat_template_kwargs")
enable_thinking = chat_template_kwargs.get("enable_thinking", False)
@@ -278,16 +274,16 @@ class QwenVLProcessor(TextProcessor):
return request
def append_generated_tokens(self, outputs, generated_token_ids):
def append_completion_tokens(self, outputs, completion_token_ids):
"""
Append generated tokens to existing outputs.
Append completion tokens to existing outputs.
Args:
outputs: Current model outputs
generated_token_ids: Generated tokens to append
completion_token_ids: completion tokens to append
"""
out = {"input_ids": [], "token_type_ids": [], "position_ids": [], "cur_position": outputs["cur_position"]}
self.processor._add_text(generated_token_ids, out)
self.processor._add_text(completion_token_ids, out)
outputs["input_ids"] = np.concatenate(
[outputs["input_ids"], np.array(out["input_ids"], dtype=np.int64)], axis=0