Add profiling multimodal model step and fix the OOM bug when profilin…

[ChenTaoyu-SJTU]

What this PR does / why we need it?

This PR is to add the multimodal profile functionality. It also fixes the out-of-memory bug (using a solution that might be temporary).

The reason this approach is temporary is that on the CUDA platform in the upstream vLLM, the Qwen2.5 VL's ViT layers (i.e., the encoder layers) do not run out of memory even without adding @torch.inference_mode(). However, under the current PyTorch and Ascend hardware environment, during the forward propagation through the multiple blocks of the ViT (over 30 layers in total), the memory increases by approximately 1.2 GB per layer, which is unacceptable. Yet, after adding @torch.inference_mode(), this issue disappears, and the memory increase behavior becomes almost similar to the growth pattern observed on the CUDA platform. But the underlying cause currently remains unknown, need more research.

Does this PR introduce any user-facing change?

no

How was this patch tested?

By pass the CI

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 34.14%. Comparing base (c30ddb8) to head (e285e6c).
Report is 56 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1408      +/-   ##
==========================================
+ Coverage   27.39%   34.14%   +6.75%     
==========================================
  Files          56       63       +7     
  Lines        6191     7315    +1124     
==========================================
+ Hits         1696     2498     +802     
- Misses       4495     4817     +322     

Flag	Coverage Δ
unittests	34.14% <ø> (+6.75%)	⬆️

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.

[Yikun]

please do a rebase

[ChenTaoyu-SJTU]

need code review @Yikun @wangxiyuan @yiz-liu @shen-shanshan . By the way, why didn't the accuracy test trigger after I added the label?

[Yikun]

You should first add vl-accuracy-test and then label ready-for-test otherwise will trigger twice.

The previous job link:
https://github.com/vllm-project/vllm-ascend/actions/runs/15869325461

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[shen-shanshan]

LGTM.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.