[jiangkuaixue123]

Summary

This PR brings the latest fork main changes into upstream/main.

Commits included from jiangkuaixue123/afd-plugin-1:main:

• 273317882ef27e740afd0502f6c01049623c061c Internalize DP metadata sender selection
• e689c11d850e32b5313764fd7de4bf3ef121688e Remove connector recv handle list
• fe87bb1 Type connector-specific metadata data
• 8c20fae Align DP metadata state payload

Validation

Validated DeepSeekV2-Lite 2A2F functionality using model path:

/home/jcz/models/deepseekv2-lite

Eager 2A2F

Command shape:

• tests/e2e/runner.py
• --device-backend gpu
• --num-attention-ranks 2
• --num-ffn-ranks 2
• --attention-gpus 0,1
• --ffn-gpus 2,3
• --max-tokens 8
• --common-vllm-arg=--trust-remote-code

Result:

• Passed
• Runner exited with code 0
• Attention API became ready
• 2 completion requests returned HTTP 200
• Generated completion text was stable: city of neighborhoods, and each one has

DBO + CUDA graph 2A2F

Command shape:

• tests/e2e/runner.py
• --device-backend gpu
• --num-attention-ranks 2
• --num-ffn-ranks 2
• --attention-gpus 0,1
• --ffn-gpus 2,3
• --max-tokens 8
• --common-vllm-arg=--trust-remote-code
• --cuda-graph-full-decode-only
• --cudagraph-capture-size 8
• --num-requests 8
• --request-concurrency 8
• --enable-dbo
• --dbo-decode-token-threshold 1
• --dbo-prefill-token-threshold 8

Result:

• Passed
• Runner exited with code 0
• Attention API became ready
• 8 concurrent completion requests returned HTTP 200
• Logs confirmed DBO was enabled
• Logs confirmed CUDA graph full-decode capture completed
• Generated completion text was stable: city of neighborhoods, and each one has

Post-validation cleanup check:

• Validation ports were released
• GPU 0/1 had no residual memory usage
• GPU 2/3 only had minor 4 MiB runtime residual usage

[gemini-code-assist]

Code Review

This pull request refactors the AFD connector interfaces and metadata handling. It introduces a new AFDConnectorData base class for backend-specific connector metadata payloads and updates update_state_from_dp_metadata to accept a single AFDDPMetadataPayload object instead of multiple arguments. Additionally, it removes recv_handle_list from AFDConnectorMetadata along with its associated wait logic in the FFN model runners, simplifies rank-checking methods, and updates the unit tests to align with these API changes. I have no feedback to provide as there are no review comments.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.