[Tracker]: afd-plugin active development

Purpose

This issue is the execution-level index for active afd-plugin work.

• Project direction, support policy, and long-term workstreams remain in #155.
• Detailed design and acceptance criteria remain in the linked RFCs and issues.
• The linked issue or PR is the source of truth for status; this checklist is a
  maintainer-facing snapshot.

Last reviewed: 2026-09-14

Upstream alignment

• GPU backend alignment with vLLM 0.26.0 — #167
• GPU backend alignment with vLLM 0.28.0 — #310
• NPU backend alignment with vLLM/vLLM-Ascend 0.26.0 — #168
• NPU backend alignment with vLLM 0.28.0 and a pinned compatible vLLM-Ascend runtime — #341

ModelRunnerV2

• GPU ModelRunnerV2 support — #223 @jiaran-king
  • Implementation: #245
  • DBO support (https://github.com/vllm-project/vllm/pull/50945) https://github.com/vllm-project/vllm/pull/51700
• Ascend NPU ModelRunnerV2 support — #240 @lirx-pd
  • Implementation: #257
  • DBO support #265

Native MoERunner forward refactor

• Refactor AFD MoE forward around native MoERunner injection — #225

Phase 1: Attention-side native MoERunner reuse

• Replace handwritten GPU and CAMP2p MoE forward proxies with
  role-specific runners that share one remote-experts implementation.
• Reuse native Ascend routing for CAMAsync and remove the copied routing
  selection path.
• Preserve connector payloads, ordering, FFN-driver ownership, dense-layer
  behavior, and role-aware weight allocation.
• Validate GPU P2P NCCL, NPU CAMP2p, and NPU CAMAsync with pre-/post-refactor
  parity across their currently supported execution modes.

Phase 2: Model-independent FFN MoE computation

• Introduce reusable GPU and NPU FFN MoE runners instead of
  model-specific FFN helper methods.
• Keep connector communication state in the FFN driver.
• Demonstrate reuse with at least one non-DeepSeek native FusedMoE model.
• Document and fail closed for dense, sequence-parallel, and non-factory
  MoE exceptions.

Detailed architecture, validation requirements, and acceptance criteria remain
in #225.

Model support

DeepSeek V4

• DeepSeek V4 support tracker — #169 @bjf-frz
  • GPU P2pNcclAFDConnector implementation — #191
  • NPU CAMAsyncAFDConnector — #227
    • Implementation: #300
  • NPU CAMP2pAFDConnector — #228

GLM5.3-flash

• NPU CAMAsyncAFDConnector(prioritize this) @yujuancao07
• NPU CAMP2pAFDConnector
• GPU P2pNcclAFDConnector

Qwen3 MoE

• CUDA AFD adapter on vLLM 0.26.0 — #215, #216
  • GPU PyNcclAFDConnector support — #215, #216
  • NPU CAMP2PConnector support (help wanted)
  • NPU CAMAsyncConnector support (help wanted)

Qwen3.5 / Qwen3.6 MoE

• CUDA text-only adapter for the native
  Qwen3_5MoeForConditionalGeneration family — #179, #181, #256
  • GPU PyNcclAFDConnector support — #179, #181, #256
  • NPU CAMP2PConnector support (help wanted)
  • NPU CAMAsyncConnector support (help wanted)

Model performance

• DeepSeek V3.2 prefill performance and token-balanced TP/SP dual batching
  — #170 @ShwStone

Connectors and runtime optimization

• Async GPU connector — #233 @specture724
  • Implementation: #239
• CAMAsync multi-stream execution — #150
• CAMP2p A2E/E2A multi-stream execution on the FFN side — #68
  • Implementation: #161
• Derive CAM HCCL buffer sizes per role — #234

Reliability and CI

Merged reliability fixes (FFN graph replay, NPU teardown, and multi-host
initialization): #270, #279, #328.

CI now includes L4 Kubernetes jobs, weekly CUDA E2E, and strengthened
pre-commit checks: #299, #268, #271.

NPU runtime qualification for vLLM 0.28.0 remains tracked in #341.

Platform and contributor ecosystem

• Experimental Moore Threads MUSA backend — #173
• Repository-backed agent workflow — #196
• Maintain the Attention–FFN disaggregation paper index — #32

This is a maintainer-facing tracker, not a standalone implementation task, and it names no files or tests. Start with #155 for project direction, then choose a linked issue or RFC; for the MoERunner work, #225 contains the architecture and acceptance criteria. Done means the selected work item meets its own criteria and this checklist reflects its source-of-truth status.