[BUGFIX] Fix dp size > 1 for qwen3 vl model

[zju-stu-lizheng]

Question

In my local tests, the Qwen3-VL service fails to start when both
--mm-enable-dp-encoder
and
--enable-dp-attention
are enabled.

I came across the related PR: pr17157
After applying this PR and setting tp = dp, enabling both options works without any issues.

However, when using configurations where tp != dp (for example, tp=8 and dp=4), enabling both options leads to precision/accuracy problems[sometimes ima bug].

Environment:
Model: Qwen3-VL
Command-line options: --mm-enable-dp-encoder --enable-dp-attention
Observed: service startup failure (before PR), precision issue (after PR) in tp != dp setups

Fix bug

This branch is intended to fix the problem mentioned in this issue.

Many thanks to yizhang2077 for identifying and fixing the mrope_positions padding issue.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[yhyang201]

/tag-and-rerun-ci

[yizhang2077]

move it into header

[zju-stu-lizheng]

move it into header has some bug:
ImportError: cannot import name 'get_global_server_args' from partially initialized module 'sglang.srt.server_args' (most likely due to a circular import)

[yizhang2077]

it should be compatible with self._pad_tensor_to_size, maybe you could write like
self.mrope_positions = self._pad_tensor_to_size(self.mrope_positions.t(), bs).t()

[zju-stu-lizheng]

yes, it can refactor as

self.mrope_positions = self._pad_tensor_to_size(
                self.mrope_positions.transpose(0, 1), num_tokens
            ).transpose(0, 1)

[yhyang201]

/rerun-failed-ci

[yhyang201]

I’m using this PR and running the following command:

python -m sglang.launch_server \
  --model-path Qwen/Qwen3-VL-30B-A3B-Instruct \
  --tp 4 \
  --enable-dp-attention \
  --dp 4 \
  --ep 4 \
  --disable-cuda-graph

However, the server fails to launch. Could you please help check what might be wrong? Thank you!

[zju-stu-lizheng]

Thanks for the report — this is indeed a known issue with the current PR.
After adding support for DP attention, the problem should be resolved.
I will work on this and include the fix in an upcoming commit.

[yhyang201]

I tested the following command with four different DP configurations (by varying --dp and whether --mm-enable-dp-encoder is enabled), and all of them worked fine:

python -m sglang.launch_server \
  --model-path Qwen/Qwen3-VL-30B-A3B-Instruct \
  --tp 4 \
  --enable-dp-attention \
  --ep 4 \
  [--dp 4 | --dp 2 | --dp 2 --mm-enable-dp-encoder | --dp 4 --mm-enable-dp-encoder]

No issues on my side. Many thanks for the fix!

[yizhang2077]

/rerun-failed-ci

[yhyang201]

While making the vision part compatible with DP attention for kimi-k2.5, I modified the vision.py file, so it may now need to be rebased.
I think the approach in this PR is better, so we can revert the changes to vision.py in https://github.com/sgl-project/sglang/pull/17789/changes and then apply the changes from the current PR.