Mirror SGLang rollout architecture for vLLM - #66

Closed

aoshen02 wants to merge 12 commits into

mainfrom

aoshen/vllm-mirror-sglang-arch

Commits on May 28, 2026

Mirror SGLang rollout architecture for vLLM
aoshen02
committed
872adb9

Commits on May 29, 2026

examples/multinode_validation: reproducible cross-host sweep harness for PR #66

aoshen02
and
claude
committed
9fb4fec
tests/: move multi-node validation harness from examples/ into tests/

aoshen02
and
claude
committed
25472d2
tests/test_qwen3_30B_A3B_pr66_sweep: run 2 rollout steps, not 1

aoshen02
and
claude
committed
3706853
tests/test_qwen3_30B_A3B_pr66_sweep: run vLLM PP>1 configs eager

aoshen02
and
claude
committed
4ad42b1
vllm DP+EP: derive TP as engine_gpus // (pp * dp); add DP+EP sweep configs

aoshen02
and
claude
committed
948d7fc
tests: add combined-dimension topology configs s14-s19 (PP+DP, PP=4, TP+PP+DP)

aoshen02
and
claude
committed
f69b3c2
vllm PD disaggregation: NIXL engine wiring + PD proxy (1P1D proven on H200)

aoshen02
and
claude
committed
000dd8a
vllm PD: dynamic worker registration in pd_proxy + rollout wiring

aoshen02
and
claude
committed
b63035a
vllm PD: end-to-end proven through pd_proxy + /inference/v1/generate

aoshen02
and
claude
committed
a32ebb2
tests: Qwen3-4B GRPO convergence recipe (dapo-math-17k -> AIME, DP16 colocate)

aoshen02
and
claude
committed
dcf8082
tests/qwen3_4B converge: guard hf download (skip when pre-staged)

aoshen02
and
claude
committed
741ade8