[Feature][PP] Support Dynamic Chunk for Chunked Pipeline Parallelism by gjc0824 · Pull Request #7896 · vllm-project/vllm-ascend

gjc0824 · 2026-04-01T07:14:23Z

What this PR does / why we need it?

Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency.

We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time.

Specific modifications:

Patch modifications

Patch Scheduler to add dynamic chunking logic
Minimal patch to EngineCore to add logic for calling Profile

Changes on the vllm-ascend side

Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile
Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions

vLLM version: v0.19.0
vLLM main: vllm-project/vllm@620e892

Performance Results:

Fixed-Length Scene
256k Qwen3-235B chunk32k
TTFT
PP 61.4s -> CPP 53.5s, 15% benefit
Variable-length scene
1k~32k
DeepSeek V3.1
CPP
TTFT 9.9s, Throughput 23635 tps
PCP
TTFT 11.4s, Throughput 20734 tps

gemini-code-assist · 2026-04-01T07:14:34Z

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request implements a dynamic chunk sizing feature for Chunked Pipeline Parallelism (CPP) to mitigate bubbles in long-sequence scenarios. By profiling prefill latency at startup and fitting a quadratic model, the scheduler can now predict and apply optimal chunk sizes at runtime. The solution includes online refinement capabilities to continuously improve prediction accuracy based on actual execution times, significantly enhancing load balancing for variable-length sequences.

Highlights

Dynamic Chunking Scheduler: Introduced ProfilingChunkScheduler, which dynamically adjusts chunk sizes based on prefill latency profiling to optimize load balancing in Chunked Pipeline Parallelism.
Profiling and Modeling: Added a profiling mechanism that fits a quadratic latency model to prefill execution times, enabling runtime prediction of optimal chunk sizes.
Online Refinement: Implemented online model refinement by recording execution timing after each model step, allowing the predictor to adapt to changing workload characteristics.
Infrastructure Support: Updated EngineCore and Worker components to support startup profiling and execution timing feedback, ensuring compatibility with multiprocessing environments.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

github-actions · 2026-04-01T07:14:40Z

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

A PR should do only one thing, smaller PRs enable faster reviews.
Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

gemini-code-assist

Code Review

Suggested PR Title:\n\nmarkdown\n[Scheduler][Feature] Implement profiling-based dynamic chunk sizing for Ascend\n\n\nSuggested PR Summary:\n\nmarkdown\n### What this PR does / why we need it?\nThis pull request introduces a profiling-based dynamic chunk sizing strategy for the Ascend backend to optimize prefill performance. It implements a `ProfilingChunkScheduler` that utilizes a quadratic latency model to predict optimal chunk sizes based on request history length. The system includes a `ProfilingChunkManager` for model fitting and online refinement using real-time execution data. A critical bug was identified in the `ProfilingChunkManager` where the `_set_time_done` attribute was not initialized, which would result in an `AttributeError` during the first batch execution.\n\n### Does this PR introduce _any_ user-facing change?\nYes, it adds a new `profiling_chunk_config` option within the engine's additional configuration, allowing users to enable and tune the dynamic chunking behavior.\n\n### How was this patch tested?\nThe implementation extends the vLLM v1 scheduler and engine core via monkey-patching and configuration updates; however, no new automated tests were included in this PR.\n

gemini-code-assist · 2026-04-01T07:17:35Z

+        self._profiling_done = False
+


The attribute _set_time_done is used in vllm_ascend/patch/platform/patch_profiling_chunk.py (line 106) to guard the target latency calibration, but it is not initialized in the ProfilingChunkManager.__init__ method. This will cause an AttributeError when the first batch execution timing is recorded in the engine core process.

Suggested change

self._profiling_done = False

self._profiling_done = False

self._set_time_done = False

yiz-liu

Add or modify performance test cases.

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

MengqingCao

Thanks for your great work! Could you post more performance results on dynamic cpp and the design doc, usage doc, ~~test cases~~ should also be supplemented, I'm okay with adding them in a follow-up pr.

MengqingCao · 2026-04-15T01:08:04Z

+        L = np.array(seq_lens, dtype=np.float64)
+        T = np.array(latencies, dtype=np.float64)
+
+        if len(L) < 8:


why we constraint the length of L to 8 here?

These values are all judgments of the amount of fitting data. For fitting the relationship between sequence length and time, a small amount of data can lead to inaccurate fitting. All are experience value. For clarity, I will create a variable name to represent.

MengqingCao · 2026-04-15T01:12:15Z

+            True if fitting succeeded, False otherwise
+        """
+        num_points = len(chunked_data)
+        if num_points < 5:


Is num_points < 5 an experience value?

MengqingCao · 2026-04-15T01:13:44Z

+                num_points,
+            )
+            return False
+        if num_points > 30:


ditto, why when num_points > 30 means fitted here?

ditto, addtionally, this is the online calibration stage, and users need to use several pieces of real data for preheating. It is not advisable to be too long here to affect the user experience, so the amount of fitting data is limited. The actual measurement of 30 data points has shown good fitting effect, and online calibration can be completed by consuming approximately 4 data points

MengqingCao · 2026-04-15T01:15:07Z

+        if fitted_a < 0:
+            logger.warning("Fitted a=%.2e is not positive. Setting a=1e-9.", fitted_a)
+            fitted_a = 1e-9
+
+        if fitted_b < 0:
+            logger.warning("Fitted b=%.2e is not positive. Setting b=0.0.", fitted_b)
+            fitted_b = 0.0


it seems the post process of a and b could be extract into a function.

Sure. We will extract it into a tool function.

MengqingCao · 2026-04-15T01:33:45Z

@@ -0,0 +1,179 @@
+#


plz add the details of this patch like others in https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/patch/__init__.py

MengqingCao · 2026-04-15T01:43:13Z

+                and request.num_computed_tokens > 0
+            ):
+                predicted_chunk = self.profiling_chunk_manager.predict_chunk_size(
+                    history_len=request.num_computed_tokens,


I think it would be better to use consistent naming conventions for num_computed_tokens and history_len.

gjc0824 · 2026-04-15T03:32:24Z

Thanks for your great work! Could you post more performance results on dynamic cpp and the design doc, usage doc, ~~test cases~~ should also be supplemented, I'm okay with adding them in a follow-up pr.

Sure. We plan to collect more test results and include them in the subsequent document PR. The subsequent PR will include interface specifications, user guides, and more performance data.

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

MengqingCao

LGTM now, thx!

… values Signed-off-by: gaojc <gaojingchun1@huawei.com>

MengqingCao · 2026-04-17T14:15:58Z

+    # The method below is based on the upstream Scheduler.schedule()
+    # with profiling-based chunk sizing applied to both RUNNING requests
+    # (chunked prefill continuation) and WAITING requests (new prefill).
+    # Modified sections are marked with ">>> PROFILING CHUNK" comments.


This makes the code very clear! 👍

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com> Signed-off-by: xulei_ict <xulei292@huawei.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com> Signed-off-by: tfhddd <2272751277@qq.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com> Signed-off-by: guxin108 <1252896542@qq.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com> Signed-off-by: zouyida2052 <zouyida2002@gmail.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com>

…llm-project#7896) ### What this PR does / why we need it? Chunked Pipeline Parallelism (CPP), with its relatively low communication overhead, has become a key feature for ultra-long sequences and variable-length sequence scenarios. However, ordinary CPP faces the problem of empty slots in long sequences as the KVCache grows, reducing computational efficiency. We control the execution time of each scheduling batch to be equal by using a dynamic chunk strategy. Specifically, we profile requests of different lengths and fit the functional relationship between sequence length and execution time. During the scheduling process, we calculate the appropriate chunk size based on the given target execution time. Specific modifications: Patch modifications 1. Patch Scheduler to add dynamic chunking logic 2. Minimal patch to EngineCore to add logic for calling Profile Changes on the vllm-ascend side 1. Add timing logic in worker & modelrunner, minor adaptation of dummy_run to CPP Profile 2. Create ProfileChunkManager to uniformly manage dynamic chunking resolution, function fitting, and other utility functions - vLLM version: v0.19.0 - vLLM main: vllm-project/vllm@620e892 --------- Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com> Signed-off-by: Xiaochao Wang <wangxiaochao6@hisilicon.com> Signed-off-by: gaojc <gaojingchun1@huawei.com> Co-authored-by: gaojc <gaojc0714@163.com> Signed-off-by: yangzhe-2026 <yangzhe@isrc.iscas.ac.cn>

gjc0824 requested review from LCAIZJ, MengqingCao and wangxiyuan as code owners April 1, 2026 07:14

github-actions Bot added the module:core label Apr 1, 2026

gemini-code-assist Bot reviewed Apr 1, 2026

View reviewed changes

yiz-liu reviewed Apr 1, 2026

View reviewed changes

gjc0824 force-pushed the refactor/profiling-chunk branch 8 times, most recently from 417025d to 1b3bc22 Compare April 8, 2026 08:40

gjc0824 added 7 commits April 9, 2026 17:20

support profile cpp in ascend

3ea97c6

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

add running dynamic

6a76eca

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

fix enginecore spawn patch

3b83ef9

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

fix fc1 & multi-batch sched & log output

469613b

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

remove env variable

9b5cffa

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Zeng Shu <shuzeng@huawei.com>

fix lint

7c2669c

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

fix ascend_config and profile bugs

4ca0e41

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch 3 times, most recently from 7fb08e6 to f5b8ce1 Compare April 10, 2026 11:24

add ut/e2e

8be8209

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch 3 times, most recently from 5b071c6 to a9bfeec Compare April 12, 2026 08:46

gjc0824 force-pushed the refactor/profiling-chunk branch from 32d2a6c to 49061cd Compare April 12, 2026 09:31

Merge branch 'main' into refactor/profiling-chunk

9c91a14

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch 2 times, most recently from f973c57 to 84a24a3 Compare April 13, 2026 02:39

fix ut

389d2b1

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch from 84a24a3 to 389d2b1 Compare April 13, 2026 02:49

Merge branch 'main' into refactor/profiling-chunk

82cc643

MengqingCao reviewed Apr 15, 2026

View reviewed changes

Merge branch 'main' into refactor/profiling-chunk

5da477f

Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch 2 times, most recently from 6503a4b to b7b096c Compare April 15, 2026 06:21

MengqingCao approved these changes Apr 16, 2026

View reviewed changes

MengqingCao added ready read for review ready-for-test start test by label for PR labels Apr 16, 2026

create experience value name && extract clamp function && rename some…

99e3184

… values Signed-off-by: gaojc <gaojingchun1@huawei.com>

gjc0824 force-pushed the refactor/profiling-chunk branch from b7b096c to 99e3184 Compare April 16, 2026 15:46

MengqingCao reviewed Apr 17, 2026

View reviewed changes

MengqingCao merged commit bec6314 into vllm-project:main Apr 17, 2026
52 checks passed

This was referenced Apr 28, 2026

[Bug]: pipeline parallelism errors when applying flashcomm1/sp #6283

Closed

[Bugfix]: fix pp errors when applying flashcomm1 #6282

Closed

	self._profiling_done = False
	self._profiling_done = False
	self._set_time_done = False

Conversation

gjc0824 commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Uh oh!

gemini-code-assist Bot commented Apr 1, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

github-actions Bot commented Apr 1, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

yiz-liu left a comment

Choose a reason for hiding this comment

Uh oh!

MengqingCao left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

gjc0824 commented Apr 15, 2026

Uh oh!

MengqingCao left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

gjc0824 commented Apr 1, 2026 •

edited

Loading

MengqingCao left a comment •

edited

Loading