fix(pcg,mm): fix zeroing of input_embeds when replay PCG

[cs-cat]

Motivation

Before PCG replay, there was an error in the buffer clearing logic for multimodal models. The shape of buffers.input_embeds is (num_tokens, num_dim), and existing code incorrectly cleared it using the second dimension index. In some cases (num_dim > static_num_tokens), this could cause input_embeds to retain content from the previous inference, potentially leading to incorrect inference results. This PR uses the correct clearing logic, indexing in the tokens dimension.

For maintainers: I'm unsure if this issue causes a decrease in multimodal model accuracy and is related to some issues. It seems that PCG for multimodal models is not enabled by default for all models. I tested Qwen 3.5-27B on a sglang build that fixes the graph recompile and was able to reproduce the issue mentioned in this PR. The shape of input_embeds is (4096, 5120), and static_num_tokens = 4096.

Modifications

The logic for clearing inputs_embeds has been corrected before replaying the multimodal model PCG.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

[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!

[cs-cat]

@Oasis-Git @ch-wan Could you please review this?

[Oasis-Git]

Thanks for your contribution. It is correct.

[ispobock]

/tag-and-rerun-ci