Upgrade Npu

Upgrade the AFD NPU backend

Run this workflow as a gated state machine. Resolve both runtime histories, finish architectural analysis, and pass the architecture gate before editing code. Preserve AFD behavior on the target upstream architecture; do not preserve obsolete implementation shapes merely because they existed at the old version.

Read the repository AGENTS.md before acting. Read references/upgrade-workbook.md for the required evidence tables, commands, and report templates. List the available references/*lessons.md files and read every one before analysis; each is short, historical upgrade evidence whose relevance may not be obvious from its filename. In particular, apply the failure patterns in references/v0191-to-v026-lessons.md when the upgrade affects MoE construction, W8A8, PCP, DBO/uBatch, CAM, rank/device mapping, or commit boundaries.

Hard boundaries

• Treat the current and target runtime as four immutable revisions: current vLLM, current vLLM-Ascend, target vLLM, and target vLLM-Ascend. Resolve every tag or branch to a full commit SHA before diffing or editing.
• Require the user to provide both target refs and the local source repository directories for vLLM and vLLM-Ascend. Infer missing current refs from repository evidence. If either target ref or source directory is absent, ask for it and stop.
• Do not modify, switch, or check out another revision in either user-provided source directory. Resolve its HEAD before comparison. If it is already at the target SHA, use it as the target source tree. Otherwise create a separate detached worktree from that repository at the target SHA. Preserve a provided checkout at the current SHA as the current source tree; if its HEAD is neither the current nor target SHA, leave it untouched and read the current revision with git show while using the new worktree for the target revision.
• Do not edit AFD code until the architecture gate passes. If a major architecture change is found, stop this upgrade and report it; do not silently remove a supported feature or invent a large compatibility layer.
• Preserve unrelated dirty changes. Before editing, record branch, HEAD, and status. Stop if an existing user change overlaps the planned files. Never reset, discard, or include unrelated changes in an upgrade commit.
• Use the repository's patch, inheritance, and composition rules. Never modify vLLM or vLLM-Ascend source to make AFD tests pass.
• Do not add reflection-heavy compatibility. Access known upstream members directly, avoid Any and object unless necessary, and allow upstream errors to reveal incompatible contracts as required by AGENTS.md.
• Make one local signed-off commit for each coherent adaptation or test-failure root cause. Include code and its focused tests together. Do not push, publish, merge, or open a PR unless the user separately asks.
• Run all non-accuracy NPU E2E gates before accuracy. Never use a limited accuracy run as final qualification.

Subagent orchestration

Use distinct subagents when available to separate analysis, implementation, testing, and review. The primary agent owns the phase state, architecture-gate decision, task graph, integration, staged diff, commits, and final report. Subagent conclusions are evidence, not automatic gate decisions.

• Analysis agents: keep them read-only. Split work by upstream history or AFD surface, such as vLLM, vLLM-Ascend, patches, models, NPU runtime, and connectors. Require exact paths, symbols, SHAs, classifications, affected invariants, and proposed tests. Allow independent analysis agents to run in parallel.
• Implementation agents: dispatch them only after the architecture gate passes. Give each agent one coherent root cause and an explicit, non-overlapping AFD file set. They may edit only those files and their focused tests. Prohibit staging, committing, pushing, changing branches, or modifying either upstream source tree or target worktree. Serialize implementation in a shared AFD checkout. Parallelize only in primary-created, fixed-base AFD worktrees with independent file ownership; return each raw diff to the primary agent for one-at-a-time integration and validation in the main upgrade checkout.
• Test agents: prohibit changes to tracked or source files. Allow writes only to handoff-designated temporary, log, cache, and artifact paths. Assign exact commands and validation cells, require complete logs, counts, first root traceback, skips, and cleanup, and return failures to the primary agent for classification. Make agents that run NPU E2E read ../run-e2e/SKILL.md. Serialize jobs unless device IDs, port ranges, writable cache paths, temporary and log directories, process/service ownership, and cleanup ownership are all explicitly isolated. Permit shared read-only model access.
• Review agents: keep them read-only and independent from the implementation task. Give them the raw diff plus exact target upstream trees and require review first for minimal and simple implementation, compliance with AGENTS.md and established local style, and duplicated, obsolete, dead, or unnecessary compatibility code. Then review exact upstream signatures, patch markers, architecture, state ownership, execution order, and test coverage. Require actionable findings with priority and file/line evidence; do not ask the reviewer to fix its findings.

Before dispatching any subagent, record its role, objective, current and target SHAs, upstream source/worktree paths, assigned AFD checkout/worktree and fixed base, allowed AFD files, allowed commands, prohibited actions, dependencies, and required output using the handoff template in the workbook. Because agents share one workspace, record the starting diff and never give concurrent implementation agents the same checkout or overlapping paths. Require every edit, test, and review command to run in the assigned AFD path. Never run testing or review in a workspace while an implementation agent is writing it. Freeze the diff first, then inspect every returned diff and command result directly.

Use this loop for every coherent adaptation or test-failure root cause:

ANALYZE -> PRIMARY GATE DECISION -> IMPLEMENT -> FOCUSED TEST
        -> INDEPENDENT REVIEW -> PRIMARY STAGE/COMMIT -> REGRESSION GATE

If review or testing finds a new major architecture change, return to the architecture gate and stop production edits. If it finds a local defect, assign one bounded implementation task, repeat focused validation and independent review, then let only the primary agent create the signed-off commit.

Required phase record

At each phase boundary record:

phase: <name>
status: PASS | FAIL | BLOCKED | STOPPED | SKIPPED
evidence: <resolved SHAs, files, commands, tests, or logs>
blocker: <none or exact blocker>
next_allowed_action: <one action>

Do not enter a later phase when an earlier required phase is FAIL, BLOCKED, or STOPPED.

FREEZE_IDENTITY
  -> INVENTORY_AFD_CONTRACTS
  -> DIFF_BOTH_UPSTREAMS
  -> ARCHITECTURE_GATE
  -> IMPLEMENT_ATOMIC_ADAPTATIONS
  -> STATIC_AND_UNIT_GATE
  -> TARGET_NATIVE_NPU_CONTROL
  -> NPU_BASIC_E2E_GATE
  -> NPU_ACCURACY_GATE
  -> DOCUMENTATION_REFRESH_GATE
  -> FINAL_AUDIT
  -> CAPTURE_LESSONS_AND_REPORT

Phase 0 — Freeze identity

Record the AFD checkout, branch, HEAD, dirty status, Python environment, and available NPU hardware. Validate the user-provided vLLM and vLLM-Ascend source directories as Git repositories; record their paths, HEADs, branches, and dirty status. Then build the four-revision ledger and create any required detached target worktrees without changing the provided checkouts.

For a missing current ref, use evidence in this order:

1. exact dependency pins and lockfiles in AFD;
2. version constants, patch source annotations, support matrices, installation docs, recipes, containers, and CI configuration in AFD;
3. the current vLLM-Ascend revision's .github/vllm-main-verified.commit or .github/vllm-release-tag.commit;
4. for older vLLM-Ascend revisions, VLLM_TAG or VLLM_COMMIT in that revision's Dockerfile, then its compatibility matrix;
5. installed package versions and adjacent checkout state as corroboration only.

Do not assume equal version strings imply compatibility. Resolve annotated tags to commits and record both the user-facing ref and SHA. If authoritative evidence conflicts, mark identity BLOCKED and ask the user which current pair is authoritative.

Validate the target pair against vLLM-Ascend's commit locks, release matrix, and complete software tuple: Python, PyTorch, torch_npu, CANN, Triton Ascend, ABI, and any required communication/operator package. If the target refs are not a supported pair, stop before code changes.

Phase 1 — Inventory AFD contracts

Build an inventory before reading upstream diffs. Start with these high-risk surfaces:

• afd_plugin/compat/patches/**, especially patches/npu/**;
• afd_plugin/compat/npu/**;
• afd_plugin/model_executor/models/**, including NPU overlays, native model factories, role construction, forward paths, and weight loading;
• afd_plugin/v1/worker/npu/** and the shared workers/model runners they extend;
• forward context, Attention metadata, graph capture, DBO/uBatch, scheduler output, input batch, sampler, speculative decode, and device selection;
• NPU connectors, payloads, process topology, rank/group/device mapping, CAM packages, configuration validation, and plugin initialization order;
• dependency pins, tests, recipes, support matrices, and versioned docs.

For every copied or patched function, record its AFD target, current upstream source revision and symbol, exact signature, AFD marker blocks, behavioral invariant, state/lifecycle ownership, direct dependencies, and focused tests. Use source grep as authority; documentation is supporting evidence.

Also freeze the supported feature matrix. Include every discovered combination of runner generation, model, eager/graph, DBO/uBatch, PCP, sync/async CAM, Attention/FFN gate placement, TP/DP/EP, quantization/W8A8/EPLB, and speculative/MTP behavior. Do not exclude a feature merely because its path is hard to upgrade.

Phase 2 — Diff both upstreams

Analyze both histories independently:

• current vLLM SHA to target vLLM SHA;
• current vLLM-Ascend SHA to target vLLM-Ascend SHA.

Start with name-status, rename-aware stats, first-parent logs, and focused function-context diffs. Map every relevant upstream change to one or more AFD contracts. Classify each item:

• MECHANICAL: import move, rename, or signature change with equivalent behavior;
• BEHAVIORAL: defaults, metadata schema, state lifetime, call order, device mapping, or ownership changed while an equivalent AFD invariant still exists;
• ARCHITECTURAL: an execution generation, feature, stable seam, ownership boundary, or cross-component protocol was removed or conceptually replaced.

Use a method-level reconstruction for copied upstream logic:

1. take the target pinned upstream function as the new skeleton;
2. take current AFD patch markers and tests as the local-difference ledger;
3. take current AFD model/runtime behavior as the semantic authority;
4. replay only still-required AFD differences into the new skeleton;
5. verify call order, state ownership, and behavior, not only imports.

Do not mechanically apply the old-to-new upstream diff to an AFD copy.

Phase 3 — Architecture gate

Stop without editing production code if any of these is true:

• the target vLLM and vLLM-Ascend refs or software stack are incompatible;
• the NPU worker/model-runner base, execution generation, or lifecycle is replaced and AFD would require a redesign rather than local adaptation;
• an AFD patch seam is removed with no equivalent stable target, or the patch can no longer be expressed as target upstream logic plus small marked AFD differences;
• upstream removes or explicitly disables an AFD-supported feature, such as a runner-generation PCP path or DBO path, and continuing would require deleting the feature, bypassing upstream validation, or owning a new execution engine;
• model split, remote-experts ownership, connector payload, rank/group topology, scheduler/KV-cache/request-output contract, or core state lifecycle changes across multiple subsystems without a one-to-one invariant mapping;
• preserving behavior requires broad core rewrites, new public abstractions, extensive unknown-version copies, exception swallowing, or reflection-based version probing;
• the required Python/PyTorch/torch_npu/CANN/operator ABI transition cannot be built and tested as one supported target environment.

A moved implementation is not automatically a stop. Continue only when the old and new inputs, outputs, lifecycle, and invariant map cleanly and the adaptation remains narrow. For example, a patch moving from a layer to its quantization method requires architectural review, but may pass if ownership and behavior are still provably equivalent.

When stopping, report the old and new architecture, exact removed or replaced interfaces, affected AFD files/features, why a local compatibility patch is not safe, and the smallest design decision needed from maintainers. Leave the worktree unchanged.

Phase 4 — Implement atomic adaptations

Create or use a dedicated upgrade branch. Before each edit batch, write a small ledger of files, upstream cause, preserved invariant, focused tests, and patch removal/upstream plan.

Implement in this order unless the dependency graph proves another order:

1. dependency/version contract and CPU-safe imports;
2. compatibility patches and initialization order;
3. model construction, factory binding, forward, and weight policy;
4. NPU worker/model runner, forward context, graph, DBO/uBatch, and device mapping;
5. connectors and payload/topology integration;
6. focused tests and E2E scenarios required to qualify the adaptations.

Defer repository-wide recipes, support matrices, and documentation refreshes to Phase 9 so their claims are based on completed NPU validation evidence. Update a focused test fixture or narrowly coupled comment earlier only when the code adaptation requires it.

For each patch function:

• copy it from the exact target pinned source revision;
• match the target signature exactly, including defaults and return annotation;
• add comments immediately above it stating source revision, reason, changed behavior, and whether parameters differ;
• surround only AFD-specific differences with paired # ### PATCH START: ... and # ### PATCH END: ... markers;
• re-evaluate whether upstream absorbed the workaround; remove obsolete patches instead of rebasing them indefinitely;
• add tests for enabled and disabled paths, initialization order, config lifetime, state ownership, topology, and device remapping as applicable.

Use inheritance or composition when the target exposes a stable seam. Do not change common runtime code to conceal an incomplete NPU adaptation. Preserve observable execution order where state is consumed or mutated.

After each coherent adaptation passes its focused static/unit tests, inspect the exact staged diff and create a signed-off local commit. Do not combine separate root causes. Use the commit template in the workbook and include the old/target SHAs when the change is caused by vLLM or vLLM-Ascend.

Phase 5 — Static and unit gate

Before NPU execution, run the repository's actual commands for:

• git diff --check, Ruff check and format check, and compile/import checks;
• target-runtime import and plugin registration;
• exact patch signature and marker checks;
• affected focused unit tests;
• the complete CPU/unit suite;
• python -m pip check in the target runtime environment.

Treat collection/import failure as a version or environment failure until the four-revision tuple and installed packages are proven. Do not call import success, construction, dummy execution, or server startup full validation.

Phase 6 — Target native NPU control

Before AFD E2E, prove that the exact target vLLM/vLLM-Ascend runtime can load the selected checkpoint and serve one real NPU request without AFD. Record runtime SHAs, package versions, model, quantization, topology, eager/graph mode, command, readiness, output, and cleanup.

If the native control fails, preserve the root traceback and stop AFD parity testing. Diagnose native runtime, environment, model, and operator-package failures separately from AFD compatibility. A fallback mode is a new validation cell and must not silently replace the requested target.

Phase 7 — NPU basic E2E gate

Read and follow ../run-e2e/SKILL.md for hardware detection, provisioning, marker selection, live output, and cleanup. Its tests-only boundary governs this execution phase; return to this skill only after the test result is captured.

Run the complete marker-based NPU non-accuracy suite before accuracy:

1. features;
2. models, including every configured sync/async connector scenario;
3. any repository-owned NPU upgrade scenario not covered by those categories.

Use enough NPUs to avoid hardware-capacity skips when claiming a full upgrade. Every skip must map to a pre-existing documented unsupported combination or an explicitly out-of-scope test in run-e2e; no new or unexplained skip is a pass. Send real requests and validate cleanup of processes, ports, and NPU allocation.

For each failure:

1. save the exact command, environment, runtime SHAs, failing test ID, first useful root traceback, logs, resource state, and cleanup result;
2. classify it as environment, native runtime, AFD patch/model/runtime, connector/topology, accuracy, or unknown;
3. trace it to the smallest upstream contract change and reproduce it with the narrowest focused test;
4. if it reveals a major architecture change, return to the architecture gate and stop;
5. otherwise implement one minimal root-cause fix with its regression test;
6. run focused static/unit validation, create one signed-off commit with the failure and upstream cause in the body, then rerun the failed E2E and the complete non-accuracy gate.

Do not commit environment-only changes as AFD compatibility fixes. Do not change multiple runtime knobs and code in one retry.

Phase 8 — NPU accuracy gate

Enter only after the complete non-accuracy gate passes. Use run-e2e to run all NPU accuracy cases, including eager and graph. Leave AFD_GSM8K_LIMIT unset for final qualification so the full dataset runs. A limited run may be used only as a post-basic diagnostic and must be labeled non-final.

On an accuracy failure, preserve metrics and artifacts, compare the exact native control and AFD validation cells, make only a proven minimal fix, add a focused regression test, create one root-cause commit, rerun the complete non-accuracy gate, and then rerun full accuracy. Never lower the threshold or widen tolerance to make an upgrade pass unless the user explicitly approves a separately justified policy change.

Phase 9 — Documentation refresh gate

Enter only after the complete NPU basic and full accuracy gates pass. Before editing any upgrade-related documentation in this phase, read references/documentation-refresh.md completely and follow its evidence, coverage, history-preservation, consistency, and completion requirements.

Refresh the public runtime baseline, connector support statements and guides, target recipes and runnable scripts, internal design and patch contracts, contributor templates, and applicable generated metadata. Base every support claim on the exact target validation ledger. Distinguish hardware-validated, unit-only, implemented-but-unvalidated, and unsupported combinations. Do not mechanically replace intentionally historical version references or relabel old benchmark results as target evidence.

Run the reference's repository-wide consistency searches and inspect every hit in context. Create one or more focused signed-off documentation commits after the documentation checks pass. Record changed documents, intentionally preserved history, reconciled claims, validation commands, and commit SHAs in the phase evidence.

Do not enter final audit while documentation claims conflict, target runnable examples still use an obsolete stack, or a validated/unsupported status lacks evidence.

Phase 10 — Final audit

Audit all version pins, source annotations, support matrices, generated metadata, recipes, container/operator packages, historical-branch notes, and docs. Do not replace historical version references that remain intentionally scoped to old branches. Confirm every changed patch has a focused test and an upstream/removal plan.

Phase 11 — Capture lessons and report

Enter only after validation, documentation refresh, and the final audit pass. Before declaring the workflow complete, create a new version-specific Markdown file under .agents/skills/upgrade-npu/references/ using the lesson template in the workbook. Name it <YYYYMMDD>-<current-vllm>-to-<target-vllm>-lessons.md; sanitize refs for a portable filename and use a short resolved SHA when no immutable release tag is available. Record both vLLM-Ascend refs and SHAs inside the file. Never overwrite or rewrite a prior upgrade's lesson file.

Record the exact four revisions and stack, architecture decision, adaptations, patches removed or retained, test failures and root causes, misleading attempts, effective fixes, validation evidence, documentation inconsistencies, remaining debt, and concrete checks for the next upgrade. Separate reusable principles from facts specific to this version pair. Link evidence instead of copying large logs, and exclude credentials, private endpoints, and machine-specific secrets.

When available, have an independent review agent check the lesson for factual evidence, duplication, and actionable next-upgrade guidance. Otherwise freeze the diff and have the primary agent perform a separate read-only review pass before editing again. If review causes any change, freeze the new diff and repeat review. Create a focused signed-off documentation commit only after the final lesson diff passes review. Include the new lesson and any direct index/reference update it requires, then record its path and commit SHA in the completion report.

Report:

• all four refs and resolved SHAs plus the target software stack;
• the upstream-to-AFD impact matrix and architecture-gate decision;
• changed components, preserved invariants, patch inventory, and feature matrix;
• every created commit SHA/title and the root cause it isolates;
• static/unit, native control, basic E2E, and full accuracy commands/results;
• pass/fail/error/skip counts, skip reasons, metrics, logs, and cleanup status;
• every unvalidated, unsupported, removed, or explicitly excluded combination;
• the new upgrade lesson path and its signed-off commit;
• whether the branch is only local and that nothing was pushed.

Declare the upgrade complete only when the target pair is immutable and compatible, the architecture gate passed before implementation, all code changes are committed atomically, the complete non-accuracy NPU gate passed with no new unexplained skips, full accuracy passed as the final test gate, the documentation refresh gate passed, the final version/documentation audit is clean, and a new reviewed version-specific lesson has been committed for the next upgrade.