Merge upstream/main into vLLM-HUST

[ShuhaoZhangTony]

Summary

Merge upstream/main into the HUST vLLM fork so this fork's commit graph directly contains upstream commits, instead of continuing to rely only on cherry-pick/backport sync batches.

This resolution preserves HUST-local behavior around Ascend host compatibility, Knorm environment/configuration, plugin pooling/offline encode support, routing headers, engine shutdown handling, KV cache management, and selected worker/attention merge points.

Duplicate-work check

AI-assisted duplicate check was performed before opening this PR. Existing open sync PRs include #75, #82, and #87, but those are partial cherry-pick/batch syncs. This PR is materially different because it performs a direct merge of upstream/main into the fork branch so upstream commits become ancestors in the fork graph. No existing open PR was found for head branch sync/upstream-main-20260703.

Commands run:

gh pr list --repo vLLM-HUST/vllm-hust --state open --search "upstream sync main" --json number,title,headRefName,baseRefName,isDraft,url
gh pr list --repo vLLM-HUST/vllm-hust --state open --head sync/upstream-main-20260703 --json number,title,url,baseRefName,isDraft

Validation

git diff --name-only --diff-filter=U
python -m py_compile vllm/entrypoints/llm.py vllm/v1/core/block_pool.py vllm/v1/core/kv_cache_coordinator.py vllm/v1/core/kv_cache_manager.py vllm/v1/core/sched/scheduler.py vllm/v1/core/single_type_kv_cache_manager.py vllm/envs.py vllm/platforms/__init__.py vllm/utils/async_utils.py vllm/v1/worker/gpu_model_runner.py vllm/v1/attention/backends/utils.py vllm/v1/engine/core_client.py vllm/v1/engine/llm_engine.py vllm/entrypoints/openai/engine/serving.py
git diff --check
rg -n "<<<<<<<|>>>>>>>|^=======$" vllm/entrypoints/llm.py vllm/v1/core vllm/platforms/__init__.py vllm/utils/async_utils.py || true

Also ran a container import smoke with the paired vllm-ascend-hust upstream-merge checkout on PYTHONPATH; it loaded vllm and vllm_ascend from /workspace/upstream-sync-20260703 and imported the vLLM CLI successfully.

NPU generation smoke was not run because the host-managed service was active on NPU4-7, while the test constraint for this work was to use only NPU1. The service was not stopped or restarted.

AI assistance

AI assistance was used to resolve conflicts, run local validation, and prepare this PR. The human submitter should review every changed line and run the relevant full CI/NPU tests before merging.

[ShuhaoZhangTony]

Version metadata update pushed.

• upstream/main anchor: 6429d5f (v0.23.1rc0)
• fork distance after this commit: ahead 455 / behind 0 vs upstream/main
• resolved package version scheme: 0.23.1.post1.dev455+g81d6a64e5a (local suffix may vary by commit/dirty state)

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

[github-actions]

Ascend Benchmark Result

• Commit: 30b9c1f6d68f6b985ff4f996e30023bab16fd034
• Scenario: random-online
• Model: Qwen/Qwen2.5-3B-Instruct
• Publish mode: artifact-preview
• Leaderboard publish: skipped
• HF publish: skipped
• Perfgate mode: report
• Baseline source: unavailable
• Scenario mode: unavailable
• Scenario label: none
• Scenario reason: unavailable
• Workflow run: view run
• Raw benchmark result: missing
• Leaderboard entry: missing
• Note: random-online runs stay as preview artifacts unless random preview publish is explicitly enabled.