[junyyang-amd]

Every SGLang+ATOM case that uses dp-attention fails at server start. Each scheduler process except DP rank 0 raises while the plugin generates its ATOM config:

File "/workspace/atom/plugin/config.py", line 616, in _generate_atom_config_from_sglang_config
  sgl_parallel_config = ParallelConfig(
File "/workspace/atom/config.py", line 1139, in __post_init__
ValueError: data_parallel_rank (3) + data_parallel_size_local (8) must not exceed data_parallel_size (8): this node's slice would run off the end of the global DP group

Rank 0 then reports Rank 0 scheduler died during initialization (exit code: -3).

Observed in the Oct 27 nightlies on MI355, for every DP case that got past model download:

• ATOM SGLang Benchmark #259: DeepSeek-V3.2 FP8 TP4 DP4 EP4 and TP8 DP8 EP8, e.g. job 122269872252.
• SGLang accuracy run: DeepSeek-V3.2-FP8 and DeepSeek-R1-FP4, both TP4 DP4 EP4 and TP8 DP8 EP8, e.g. R1-FP4 TP8 DP8 EP8. R1-FP4 does not use E8M0 block scales, so this is not the DeepSeek-V3.2 GEMM failure fixed in fix(ci): use FP32 block scales for DeepSeek-V3.2 SGLang cases #2783.

Cause:

• Under dp-attention SGLang starts one scheduler process per GPU. _normalize_sglang_parallel_config maps that to atom_dp_size = tp_size and atom_dp_rank = tp_rank.
• The plugin passed the process rank as data_parallel_rank and the global size as data_parallel_size_local.
• Since multi-node dp support #1603, ParallelConfig.__post_init__ reads data_parallel_rank as the first global rank of this node's slice and rejects rank + size_local > size. Native ATOM builds the config with the node's start rank and assigns each process's rank afterwards (EngineCoreManager, engine_core_mgr.py), so it is unaffected. The plugin path was not updated, and the existing plugin tests replace ParallelConfig with a fake, so they did not catch it.

Change:

• Add _build_sglang_parallel_config(dp_size, dp_rank) in atom/plugin/config.py. It builds the config for the single-node slice starting at rank 0, then assigns the process's data_parallel_rank and data_parallel_rank_local, mirroring EngineCoreManager. data_parallel_size_local stays at the full DP size, which is what moe.py passes to MoRI as local_ep_size.
• Rank 0 gets the same config as before. Without dp-attention the plugin still builds dp_size=1, rank=0, which is unchanged.
• Add tests/test_plugin_sglang_parallel_config.py. It runs the real _normalize_sglang_parallel_config and the real ParallelConfig for every rank of TP4 and TP8 dp-attention, plus the non-dp-attention cases. It is under tests/ rather than tests/plugin/ because run_unit_tests.sh skips tests/plugin/.

Not handled: multi-node SGLang (nnodes > 1) with dp-attention; the plugin did not support it before this change either.

Validation:

• The new test passes. Building the config the old way raises the error above for ranks 1 and 7 and succeeds for rank 0.
• bash .github/scripts/run_unit_tests.sh on CPU torch: 10133 passed, 1029 skipped, 0 failed.
• ruff check and black --check pass on the changed files.
• The GPU DP jobs have not been rerun with this change.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2785 --add-label <label>

[junyyang-amd]

GPU benchmark reruns on MI355 passed for both DP-attention sizes:

Case	Run	Successful requests	Output throughput	Mean TTFT	Mean TPOT
DeepSeek-R1 FP4 TP4 DP4 EP4	40827702696	640	2139.82 tok/s	393.89 ms	28.64 ms
DeepSeek-R1 FP4 TP8 DP8 EP8	40827705190	640	2688.84 tok/s	282.90 ms	22.76 ms

• Both servers start on every DP rank. The ValueError: data_parallel_rank (...) + data_parallel_size_local (...) must not exceed data_parallel_size (...) from the Oct 27 nightlies does not appear.
• Parameters: 1024/1024, c=64, random range ratio 0.8, one set per case. The DeepSeek-V3.2 DP cases were not run because they depend on fix(ci): use FP32 block scales for DeepSeek-V3.2 SGLang cases #2783.
• Tested commit: e7bf89368 on temporary branches, which is this PR's head 0bd5aa873 plus one CI-only commit that lets the model download step fall back to an already complete cached model when hf download fails. The shared model cache was hitting a write quota at the time. That fallback did not trigger in either run, so the results reflect this PR's code.