[Feature]Optimization of Thinking Pattern Framework by luukunn · Pull Request #4302 · PaddlePaddle/FastDeploy

luukunn · 2025-09-28T01:51:21Z

1.Add the get_model_status method in the reasoning_parser to obtain the model's status, and determine how to perform parsing based on the status.
2.Remove the judgment of enable_thinking.

This reverts commit 300f446.

paddle-bot · 2025-09-28T01:51:26Z

Thanks for your contribution!

…into think

LiqinruiG · 2025-09-30T06:45:10Z

@@ -153,6 +154,10 @@ def process_request(self, request, max_model_len=None, **kwargs):
            request.set("top_p", _SAMPLING_EPS)
        if self.reasoning_parser and self.reasoning_parser.__class__.__name__ == "ErnieX1ReasoningParser":


这个判断可以去掉了

LiqinruiG · 2025-09-30T06:46:39Z

@@ -231,7 +236,10 @@ def process_request_dict(self, request, max_model_len=None):
            request["top_p"] = _SAMPLING_EPS
        if self.reasoning_parser and self.reasoning_parser.__class__.__name__ == "ErnieX1ReasoningParser":
            request["enable_thinking"] = True


LiqinruiG · 2025-09-30T07:36:56Z

        if response_dict.outputs.text == "" and response_dict.outputs.reasoning_content == "":
            return None
+        if req_id in self.model_status_dict:
+            del self.model_status_dict[req_id]


要在上面的 return之前做del 的判断。

LiqinruiG · 2025-09-30T07:46:07Z

            request.enable_thinking = True
+        if self.reasoning_parser:
+            model_status = self.reasoning_parser.get_model_status(request.prompt_token_ids)
+            self.model_status_dict[request.request_id] = model_status


在 n 参数的场景， request_id = XX_m，在后处理的 response 中取到的 req_id，是 XX_m_n，会造成取不到这个 model_status

LiqinruiG · 2025-09-30T07:53:12Z

-                enable_thinking=enable_thinking,
                **kwargs,
            )



还有一个 qwen_vl 的processor，看看是不是需要修改。

LiqinruiG · 2025-09-30T08:22:06Z

+                reasoning_content = delta_text[:end_index]
+                content = delta_text[end_index + len(self.think_end_token) :]
+                return DeltaMessage(reasoning_content=reasoning_content, content=content)
+            elif self.think_end_token_id in previous_token_ids:


直接用 if 就可以，不用 elif。下面也不需要 else

LiqinruiG · 2025-09-30T09:23:31Z

+            self.think_end_token_id,
+            self.response_start_token_id,
+            self.response_end_token_id,
+        ]:


如果遇到的是 tool_call 的special, 也可以直接返回 none 了

…into think

LiqinruiG · 2025-11-26T07:07:38Z

+            n = request.get("n", 1)
+            for idx in range(n):
+                self.model_status_dict[f"{real_req_id}_{idx}"] = model_status
+            request["enable_thinking"] = model_status == "think_start"


real_req_id 没法区分多个 prompt

LiqinruiG · 2025-11-26T07:08:27Z

+            reasoning_content, text = self.reasoning_parser.extract_reasoning_content(
+                full_text,
+                response_dict,
+                self.model_status_dict[req_id],


idx // (1 if request.n is None else request.n) 用这种方式，获取 status

codecov-commenter · 2025-12-03T11:40:58Z

Codecov Report

❌ Patch coverage is 73.55623% with 87 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@b0cf2c4). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/input/text_processor.py	25.00%	21 Missing and 6 partials ⚠️
...reasoning/ernie_45_vl_thinking_reasoning_parser.py	75.38%	10 Missing and 6 partials ⚠️
fastdeploy/input/ernie4_5_processor.py	58.82%	12 Missing and 2 partials ⚠️
fastdeploy/reasoning/qwen3_reasoning_parsers.py	78.68%	9 Missing and 4 partials ⚠️
fastdeploy/reasoning/ernie_x1_reasoning_parsers.py	80.32%	3 Missing and 9 partials ⚠️
fastdeploy/reasoning/ernie_vl_reasoning_parsers.py	86.48%	3 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #4302   +/-   ##
==========================================
  Coverage           ?   60.12%           
==========================================
  Files              ?      329           
  Lines              ?    41066           
  Branches           ?     6244           
==========================================
  Hits               ?    24689           
  Misses             ?    14492           
  Partials           ?     1885

Flag	Coverage Δ
GPU	`60.12% <73.55%> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

LiqinruiG

思考机制优化 ok， parser 细节可以进一步优化

* add model status in vl * add x1 parser * add model_status * fix parser * fix parser * fix parser * fix parser * Revert "fix parser" This reverts commit 300f446. * fix parser * fix * fix * fix * fix * fix parser * fix unit test * fix unit test * add unit test * fix * fix * add unit test * fix unit test * add unit test * add unit test * fix unit test * fix unit test * fix bug * fix unit test * x1 tool parser * fix unit test * fix unit test * fix unit test * fix n * fix unit test * add unit test * add unit test * remove pring

luukunn added 13 commits September 23, 2025 16:03

add model status in vl

234ef92

add x1 parser

671a4dc

add model_status

8bbe39d

fix parser

d087afb

fix parser

2f6f063

fix parser

41f1418

fix parser

300f446

Revert "fix parser"

3b93672

This reverts commit 300f446.

fix parser

dae8419

fix

e49676c

fix

2c92f6f

fix

c433e05

fix

bfdec9f

paddle-bot Bot added the contributor External developers label Sep 28, 2025

luukunn added 2 commits September 28, 2025 19:59

fix parser

bd192b2

fix parser

ae24353

luukunn force-pushed the think branch from d296ec1 to ae24353 Compare September 28, 2025 12:01

luukunn added 3 commits September 29, 2025 00:33

fix unit test

dd30110

fix unit test

31d639d

Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …

13ab4d4

…into think

luukunn force-pushed the think branch from beeef52 to 13ab4d4 Compare September 29, 2025 02:28

add unit test

46e3c13

LiqinruiG reviewed Sep 30, 2025

View reviewed changes

luukunn added 2 commits October 14, 2025 19:09

Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …

bc4a71e

…into think

pull

ede57ff

luukunn force-pushed the think branch from 212063a to ede57ff Compare October 16, 2025 04:12

LiqinruiG changed the title ~~[Feature]add model status~~ [Feature]Optimization of Thinking Pattern Framework Oct 17, 2025

fix

d159f27

luukunn force-pushed the think branch from 66c0c40 to d159f27 Compare October 20, 2025 02:25

luukunn added 2 commits October 22, 2025 19:52

x1 tool parser

1cb6205

update

f19c727

luukunn force-pushed the think branch from a970ada to f19c727 Compare November 3, 2025 07:45

fix unit test

4ef4df1

LiqinruiG mentioned this pull request Nov 19, 2025

【Hackathon 9th Sprint No.60】【RFC】为 FastDeploy 新增支持 DeepSeek 模型的 Reasoning Parser & Tool Parser PaddlePaddle/community#1185

Merged

luukunn added 2 commits November 20, 2025 15:36

Resolved merge conflicts

13ba41f

fix unit test

7c17812

luukunn force-pushed the think branch from 54e1f15 to 7c17812 Compare November 20, 2025 11:36

fix unit test

d3171a2

LiqinruiG mentioned this pull request Nov 24, 2025

[Feature] The 45VL supports prompt_token_ids + messages input. #5148

Merged

5 tasks

fix n

4317e15

LiqinruiG reviewed Nov 26, 2025

View reviewed changes

conflicts

71b99d3

fix unit test

ad7cb7a

luukunn force-pushed the think branch from 5bb15ed to ad7cb7a Compare December 4, 2025 09:14

fix unit test

ea424c7

luukunn force-pushed the think branch from 5cb46ab to ea424c7 Compare December 4, 2025 11:11

luukunn added 3 commits December 4, 2025 22:11

add unit test

44bb5f7

add unit test

397df60

merge develop

638c3ab

luukunn force-pushed the think branch from e563bdd to 638c3ab Compare December 8, 2025 06:36

remove pring

569457c

luukunn force-pushed the think branch from 5e182c4 to 569457c Compare December 9, 2025 08:50

Merge branch 'develop' into think

6725702

LiqinruiG approved these changes Dec 10, 2025

View reviewed changes

LiqinruiG merged commit fbc9bce into PaddlePaddle:develop Dec 10, 2025
13 of 17 checks passed

luukunn deleted the think branch February 5, 2026 06:11

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature]Optimization of Thinking Pattern Framework#4302

[Feature]Optimization of Thinking Pattern Framework#4302
LiqinruiG merged 49 commits into
PaddlePaddle:developfrom
luukunn:think

luukunn commented Sep 28, 2025 •

edited

Loading

Uh oh!

paddle-bot Bot commented Sep 28, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Sep 30, 2025

Uh oh!

LiqinruiG Nov 26, 2025

Uh oh!

LiqinruiG Nov 26, 2025

Uh oh!

codecov-commenter commented Dec 3, 2025 •

edited

Loading

Uh oh!

LiqinruiG left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		@@ -153,6 +154,10 @@ def process_request(self, request, max_model_len=None, **kwargs):
		request.set("top_p", _SAMPLING_EPS)
		if self.reasoning_parser and self.reasoning_parser.__class__.__name__ == "ErnieX1ReasoningParser":

Conversation

luukunn commented Sep 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

paddle-bot Bot commented Sep 28, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented Dec 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

LiqinruiG left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

luukunn commented Sep 28, 2025 •

edited

Loading

codecov-commenter commented Dec 3, 2025 •

edited

Loading