[3/5] feat: run DeepSeek-V4 on the async GPU connector

[specture724]

Third of a five-PR stack. Based on #333 — the diff shown here is only this
PR's own change; review #332 and #333 first.

1. [1/5] feat: NVSHMEM symmetric window for GPU AFD transport #332 — NVSHMEM symmetric window (transport)
2. [2/5] feat: the async GPU connector on the symmetric window #333 — the async GPU connector
3. this PR — both GPU runners + DeepSeek-V4
4. [4/5] feat: capture the FFN experts at bucketed padded shapes #335 — bucketed padded FFN expert graphs
5. [5/5][WIP] feat: request-aligned DBO ubatch splitting (flag off by default) #336 — request-aligned DBO ubatch splitting (flag off by default)

What this adds

Wires the connector into both AFD roles and gives DeepSeek-V4-Flash — 256
experts at topk 6, where the per-layer control-plane round trip costs more than
the expert compute it guards — a recipe that uses it.

Both GPU runners previously asserted control_plane is not None, and the FFN
worker loop raised NotImplementedError without one, so the connector could not
run at all:

• FFN side pulls one work item at a time from the connector's own receive
  loop, taking the layer index and row counts from the arriving payload, and
  returns on an idle poll so the worker loop still sees its shutdown event.
• Attention side skips vLLM's cross-DP batch agreement. Async AFD lets each
  replica advance alone, so an idle replica never joins that all-reduce and a
  busy one blocks in it forever — which is exactly where a 2A2F run hung before
  reaching the first MoE layer.

The V4 adapter learns the expert-routed dispatch protocol (the gate runs on the
Attention side, so the wire carries topk ids and weights instead of token ids)
and compute_ffn_output takes a device-side group_list so the FFN side runs
only the grouped GEMM over its local experts. The V2 adapter gets the same GPU
entry point, which is what the connector's e2e tests use as their reference.

Behaviour change worth a look

layers.N.ffn.gate.* now loads on both roles. The gate's parameters live
under .ffn so the checkpoint names resolve, but with
compute_gate_on_attention the Attention side is what runs it.
test_v4_gate_loads_on_both_roles pins the new contract.

The rest

Found bringing 2A2F up on real weights: the FFN role must force vLLM's NoDP MoE
prepare/finalize under EP with DP>1, shared experts must be skipped on an empty
shared slice, the SWIGLUOAI clamp has to reach the routed experts, and the DP
coordinator's startup wait needs to be long enough for the second role's weights
to load.

Testing

Full unit suite unchanged against main: same 13 pre-existing failures, none
added. pre-commit clean over the PR range. The recipe is
recipe/gpu/GpuAsyncAFDConnector/deepseek_v4_flash/2a2f_async.sh (4 GPUs).

🤖 Generated with Claude Code