[Bugfix] Guard spec decode slot mapping padding rows

[ashishpatel26]

Summary

Refs #40756 and #39295.

Guard the fused EAGLE/MTP draft-step slot-mapping kernel against inactive
padding rows with seq_lens == 0. These rows can carry block_table entries of
-1; advancing them as real requests can produce invalid slot mappings and
surface later as CUDA illegal memory access/device-side asserts in speculative
decoding.

This change returns early for those rows, writes PADDING_SLOT_ID, and leaves
their sequence length at zero. It also adds a regression test that creates a
zero-length row with a -1 block table.

Duplicate-work check

• Checked issue [Bug]: MTP speculative decoding crash with illegal memory access on long sequences (Qwen3.6-27B-FP8, v0.19.1) #40756 and its comments.
• gh pr list --repo vllm-project/vllm --state open --search "40756 in:body" returned no open PRs.
• Area searches for MTP speculative decoding illegal memory access, eagle_step_slot_mapping_metadata_kernel seq_lens padding, and CUDA-graph padded-position wording found no open PR covering this guard.
• [Bugfix] Fix race condition in MTP speculative decoding draft loop #42603 is a closed PR for [Bug]: MTP speculative decoding crash with illegal memory access on long sequences (Qwen3.6-27B-FP8, v0.19.1) #40756, but it used a stream synchronization workaround and was rejected by maintainers as not root-caused. This PR does not add synchronization.
• [Bugfix] Skip CUDA-graph padded positions in eagle draft step kernel #43682 is a closed, unmerged PR that proposed the same kernel guard for related issues. There is currently no open PR for that fix; this PR keeps the scope narrow and includes the regression coverage and local checks below.

Tests

.venv/bin/ruff check vllm/v1/spec_decode/utils.py tests/v1/spec_decode/test_eagle_step_kernel.py
# All checks passed

.venv/bin/ruff format --check vllm/v1/spec_decode/utils.py tests/v1/spec_decode/test_eagle_step_kernel.py
# 2 files already formatted

.venv/bin/python -m py_compile vllm/v1/spec_decode/utils.py tests/v1/spec_decode/test_eagle_step_kernel.py
# passed

git diff --check -- vllm/v1/spec_decode/utils.py tests/v1/spec_decode/test_eagle_step_kernel.py
# passed

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/python -m pytest tests/v1/spec_decode/test_eagle_step_kernel.py -q
# 1 skipped, 16 warnings

The targeted pytest is a Triton GPU kernel test and was skipped locally because
CUDA/XPU is not available in this environment.

AI assistance disclosure

This change was prepared with AI assistance from OpenAI Codex. The submitter
should review the two changed files and validate the GPU kernel test in a CUDA
environment before merge.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[ashishpatel26]

👋 Hi maintainers — could someone please add the ready label to trigger CI? This contributor account has no merged PRs in vllm-project/vllm yet, so the pre-run-check gate requires a maintainer label. Fix reviewed locally, tests pass, DCO signed. Thank you!

[github-actions]

This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!

[ashishpatel26]

Rebased onto latest main (no conflicts). Fix and test still apply cleanly. Still waiting on the ready label to trigger CI — contributor account has no merged PRs yet, so the pre-run-check gate blocks CI without a maintainer label. Happy to address any review feedback.