[draft] support ATOM plugin for qwen3.5 DPxTPx/DPxEPx

[zovonoir]

Summary

Add Qwen3.5 SGLang plugin support for DP-attention layouts, including DP+TP and DP+EP execution.

This change makes the Qwen3.5-specific language-model adapter use the same SGLangPluginRuntime lifecycle as the generic ATOM SGLang wrapper, and fixes the nested GDN context handling that previously caused rank-divergent collectives.

Problem

Qwen3.5 uses a dedicated SGLang language-model adapter instead of the generic ATOM causal-LM wrapper. As a result:

• idle DP ranks could pass zero-token tensors into AITER GEMM kernels (M=0);
• Qwen3.5-VL idle forwards could omit positions and input_ids;
• mRoPE positions shaped [3, num_tokens] were interpreted as three tokens;
• each rank could build a different physical DP token-count vector after idle batches were materialized;
• the nested GDN binder rebuilt ForwardContext/DPMetadata and issued a second CPU all-reduce only on ranks with valid GDN metadata, while idle ranks skipped it and later waited in a different collective.

The last issue produced a deterministic distributed deadlock: active ranks waited in GDN DP metadata synchronization while idle ranks progressed to the logits all-gather.

Implementation

Qwen3.5 SGLang adapter

• Wrap the Qwen3.5 language-model forward with plugin_runtime_scope and SGLangPluginRuntime.
• Build SGLang metadata from the runtime-normalized batch.
• Forward the normalized input_ids, positions, and input_embeds to the ATOM decoder.
• Trim physical dummy-token outputs back to the logical SGLang token count.

SGLang forward-context bridge

• Materialize an idle batch from any available tensor (positions, input_ids, or input_embeds).
• Create fallback dummy positions/token IDs when Qwen3.5-VL does not provide them.
• Normalize every zero entry in the cross-DP token-count vector to one physical dummy token so all ranks use identical DPMetadata.
• Read the token dimension correctly for mRoPE [3, num_tokens] positions.
• Propagate SGLang's global forward mode into dp_uniform_decode so mixed EXTEND/IDLE steps use the variable-length DP path.

GDN context handling

• Reuse the cross-rank-consistent ForwardContext and DPMetadata already created by SGLangPluginRuntime.
• Temporarily inject only GDN attention metadata and KV-cache data.
• Restore the previous metadata/cache on exit instead of resetting the complete context.
• Retain the previous standalone fallback when no outer context exists.

This prevents GDN from creating a second, rank-local DP synchronization point.

Parallel layouts covered

For the full-DP SGLang configurations used here:

• TPN: attention/dense TP=N, expert TP=N.
• DPNTPN: attention DP=N with physical attention TP=1; experts use logical TP=N over the DP ranks.
• DPNEPN: attention DP=N with physical attention TP=1; experts use EP=N and TP=1. When MORI is available, expert tokens use MORI dispatch/combine.

[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 1690 --add-label <label>