[Test] Add PD disagg + SD acceptance tests

[ZhanqiuHu]

Summary

Follow-up of #35158.
Add integration tests for PD disaggregation + speculative decoding via NixlConnector.

Currently covers two configs: Qwen3-8B + EAGLE3 (FLASH_ATTN, acceptance length 2.245 vs 2.260 expected) and GPT-OSS-20B + EAGLE3 (TRITON_ATTN, 2.549 vs 2.560 expected), both validated on H100.

Not wired into default CI — run manually with bash tests/v1/kv_connector/nixl_integration/pd_spec_decode_eagle3_test.sh --all. Future work includes attention backend sweeps and MTP model configs once #35158 merges.

Purpose

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request introduces a new test script (pd_spec_decode_eagle3_test.sh) and a corresponding pytest file (test_pd_spec_decode_eagle3.py) to validate the acceptance length of PD disaggregation with EAGLE3 speculative decoding via NixlConnector. The changes aim to ensure that the acceptance lengths match standalone SD baselines for specific model configurations. The review focuses on identifying potential issues related to script logic, error handling, and the correctness of the acceptance length validation.

[gemini-code-assist]

The which command can return an empty string if the command is not found, which might lead to unexpected behavior. It's safer to provide a default value directly within the command substitution to ensure SMI_BIN always has a valid value, even if the binaries are not found.

Suggested change

	SMI_BIN=$(which nvidia-smi || which rocm-smi || echo "")
	SMI_BIN=$(which nvidia-smi || which rocm-smi || echo "/usr/bin/nvidia-smi")

[gemini-code-assist]

Using kill $(jobs -pr) might not always terminate all background processes, especially if they are not direct children of the script. Consider using pkill with a more specific pattern to ensure all relevant processes are terminated, or storing the PIDs of the started processes and killing them directly.

Suggested change

	kill $(jobs -pr) 2>/dev/null || true
	kill "$PREFILL_PID" "$DECODE_PID" "$PROXY_PID" 2>/dev/null || true

[gemini-code-assist]

The curl command lacks error handling. If curl fails (e.g., due to network issues or the server not being ready), the script will continue, potentially leading to incorrect test results. Add a check for the curl exit code to ensure the server is properly running.

Suggested change

	while [ $elapsed -lt $deadline ]; do
	if ! curl -s "localhost:${port}/v1/completions" > /dev/null 2>&1; then
	echo "curl failed on port ${port}"
	return 1
	fi

[gemini-code-assist]

The pytest command does not include a timeout. If the tests hang, the script will not terminate, potentially causing CI failures. Add a timeout to the pytest command to ensure it completes within a reasonable time.

Suggested change

	pytest -xvs \
	"${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/test_pd_spec_decode_eagle3.py"
	pytest -xvs --timeout=600 "${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/test_pd_spec_decode_eagle3.py"

[mergify]

Documentation preview: https://vllm--35760.org.readthedocs.build/en/35760/

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ZhanqiuHu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork