[chenjr97]

Overview

Fix routed-expert output reconstruction when a complete block is still waiting for its hash under EAGLE/MTP lookahead hashing. Include these request-local pending blocks when returning routing rows alongside a cached prefix.

Related: [#61589 — Lookahead block hashes for EAGLE prefix caching](#61589).

Claims

• Prevent assertion failures when requested routing rows span published blocks, unkeyed full blocks, and an incomplete tail.
• Preserve token order and return owned data that remains valid after buffer reuse.
• Add buffer, worker, and LLM regression coverage, and select the EAGLE regression in existing speculative-decoding CI jobs when auxiliary-output connector files change.

Validation

.venv/bin/python -m pytest \
  tests/distributed/aux_output_connector/test_store.py -q

Result: 81 passed

Details

EAGLE/MTP lookahead hashing can delay a full block's hash until subsequent tokens are known. Its routing rows remain in pending_blocks, but the previous read path only consulted the incomplete tail. Reconstructing output after a prefix-cache hit could therefore trigger missing-request or unavailable-range assertions.

Pass the request's pending blocks into RoutedExpertsBuffer.read() and assemble the requested interval from pending blocks and the tail. Preserve complete-range validation and return an independent array so buffer reuse cannot overwrite returned rows. Pending blocks remain available for normal publication when their hashes arrive.

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).
• AI assistance was used during the creation of this PR.
• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
• [] Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.