env: add knob to control SWA eviction interval

[happierpig]

Motivation

SWA eviction only runs every sliding_window_size decode steps per request. Between evictions, each request holds up to 2 * sliding_window_size SWA tokens — double what the sliding window needs.

Adding a knob to trade-off between the token waste and per-forward overhead.

Modifications

Accuracy Tests

Speed Tests and Profiling

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]

Code Review

This pull request introduces a configurable multiplier for the Sliding Window Attention (SWA) eviction interval through a new environment variable, allowing for better control over the trade-off between memory waste and eviction overhead. A review comment points out a potential logic error where the eviction condition might never be met if the calculated interval is 1, and provides a suggestion to handle this edge case while maintaining the requirement to skip the first decode batch index.

[gemini-code-assist]

The condition req.decode_batch_idx % eviction_interval == 1 will never be true if eviction_interval is 1. This can happen if page_size is 1 and the multiplier is set to a very small value. In this case, SWA eviction would never trigger during decoding, which contradicts the user's intent of increasing eviction frequency.

Additionally, the original logic implicitly skipped the decode_batch_idx == 0 case (as mentioned in the comment above), which should be preserved.

Suggested change

	if req.decode_batch_idx % eviction_interval == 1:
	if req.decode_batch_idx > 0 and req.decode_batch_idx % eviction_interval == (1 % eviction_interval):