Skip to content

Commit 64f2efc

Browse files
committed
Clarify the AFD connector stage recommendation
Distinguish the README's recommended Decode stage from the connector module's documented prefill and decode support. Retain the intake's runtime qualification boundary. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
1 parent f3325c2 commit 64f2efc

1 file changed

Lines changed: 1 addition & 1 deletion

File tree

‎docs/reviews/issue91-afd-plugin-intake.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -31,7 +31,7 @@ All links use the full source revision above. Blob IDs identify the retrieved fi
3131
| [afd_plugin/connectors/gpu/p2p.py:1–42, 145–205, 243–312, 620–698, 930–980](https://github.com/vllm-project/afd-plugin/blob/df8d6a0c403c3ca1b4ed9cf1687e7204908bebec/afd_plugin/connectors/gpu/p2p.py#L243-L312) | `a0be75ed8bacc6fde8734f46e5208676a5d5fa90` | Mapping consumer, distinct process groups, subgroup send/receive arguments and real `PyNcclCommunicator` calls; operator fake registrations are not runtime evidence |
3232
| [tests/unit/distributed/test_topology_partition.py:21–112](https://github.com/vllm-project/afd-plugin/blob/df8d6a0c403c3ca1b4ed9cf1687e7204908bebec/tests/unit/distributed/test_topology_partition.py#L21-L112) (file read in full) | `9fc955dde26f59972671d7349313412c71a520f6` | Literal uneven partitions, complete balanced membership, namespace agreement and preservation of divisible layouts; tests inspected, not run |
3333

34-
The connector module describes both prefill and decode, whereas the README's support table labels this CUDA branch Decode. This intake preserves that source discrepancy and makes no prefill qualification claim.
34+
The README recommends Decode, while the connector module describes support for prefill and decode; neither is locally qualified.
3535

3636
## Ownership, applicability and follow-up
3737

0 commit comments

Comments
 (0)