[Refactor] Enter draft TP scopes by attention ownership and drop ModelRunner.tp_group

[ch-wan]

This PR is part of a stack (oldest at bottom):

• [Feature] Add --attn-dp-size and deprecate --enable-dp-attention #41818
• -> [Refactor] Enter draft TP scopes by attention ownership and drop ModelRunner.tp_group #41817
• [Refactor] Add make_pp_layers so models stop handing their PP position down #41816
• [Refactor] Stop passing models' layers the placement they already read #41815
• [Refactor] Let FusedMoE's weight-loading helpers read the layer's MoE-TP rank #41814
• [Refactor] Keep only the TP and PP groups on the model runner #41813
• [Refactor] Drop model placement attributes and parameters nothing reads #41812
• [Refactor] Drop placement values nothing reads #41811
• [Fix] Make the Solar model constructible and runnable #41810
• [Fix] Size the prefill delayer's gather buffer to its TP group #41809
• [Fix] Pass the draft's attention ownership to DFLASH's eager LiLiCorr scope #41808
• [Fix] Shard MoE WNA16 and Quark INT4-FP8 weights by the MoE placement #41807
• [Fix] State the configured MoE-DP width in the weight-cache fingerprint #41806
• [Fix] Read TensorCast's WORLD placement under its current names #41805

Motivation

Each speculative worker states three times whether its draft owns attention:

• self.draft_owns_attention;
• self.draft_tp_context, which is draft_tp_context under DP attention and a no-op otherwise;
• the TP group every call site passes in, taken off the draft runner.

That last one is the only reason ModelRunner.tp_group has to be a snapshot that outlives the scope the runner is built in.

Modifications

draft_tp_context(owns_attention). It now takes only owns_attention.

• With it, the scope installs the target's attention-TP group, which is the group an attention-owning draft is built on.
• Without it, there is nothing to enter.

The 28 scopes in the EAGLE, DFLASH, frozen-KV MTP, multi-layer EAGLE, Standalone and DSpark workers call it with their ownership flag.

Why this installs the same group as before:

• The scope is always entered from outside any draft scope; patch_tensor_parallel_group does not nest. There, get_parallel().attn_tp_group is the group the owning draft was built on.
• For a draft that does not own attention under DP attention, the old call re-installed the TP group that was already installed. That no-op scope is gone.
• DSpark's two scopes already had this shape and now use the same helper.
• The Standalone worker's attention-backend and CUDA-graph overrides only re-entered scopes that its base class already enters, so they are removed.

The old two-argument form of draft_tp_context is gone. A caller that still passes a group gets a TypeError rather than a different scope.

ModelRunner.tp_group is removed. With drafts entering their scope from their ownership flag, every remaining reader runs while the runner is built, during a graph capture or replay, during a forward, or only on the target. They read get_parallel().tp_group, which in those scopes is the group the runner copied:

• graph-runner barriers;
• the CUDA-graph backends (full, breakable, piecewise, NPU, XPU), which keep the group they capture with;
• the FlashInfer autotune sync group;
• patch_model;
• the elastic EP joiner and EP-join paths;
• the PP + speculative seed broadcast in TpModelWorker.

ModelRunner keeps only its PP group. Code outside the tree that reads model_runner.tp_group should read get_parallel().tp_group in the scope it runs in.

Two readers can run on a draft outside its scope:

• The HiSparse coordinator, built with the pools, reads the TP group only without DP attention. There no draft owns attention, and a draft's TP group is the target's.
• Recapturing graphs after a weight update (recapture_cuda_graph=True) runs with no draft scope entered. The draft forward there already issues its collectives on the target's TP group. Its capture barriers now use that group too, and every rank of the group runs the recapture.

The DFLASH and DSpark rank-0 gates that run inside the draft scope used to ask the target runner's TP group for the target's rank; inside the scope, the context answers the draft's. Those workers now record the target's TP rank in their constructors, outside any draft scope.

Tests.

• test_draft_tp_context.py covers both ownership cases.
• The speculative attention test kits state their TP group on the context.
• The full CUDA-graph backend, DFLASH sampler, FlashInfer autotune and DP speculative-prefill tests no longer fake a runner's TP group.

Accuracy Tests

H200, Qwen3-8B, --tp-size 2, 4 greedy prompts × 64 tokens and GSM8K (200 questions), compared with main at the top of this stack:

draft	config	greedy text + spec_verify_ct	GSM8K main / this stack
EAGLE3 (AngelSlim/Qwen3-8B_eagle3)	TP2	identical	0.955 / 0.955
EAGLE3	TP2 + DP attention (attention-owning draft)	identical	0.945 / 0.955
Standalone (Qwen/Qwen3-0.6B)	TP2	text identical	0.950 / 0.955

For the Standalone draft, one prompt's spec_verify_ct alternates between 20 and 21 across repeated runs of main itself, and the same happens on every layer of this stack.

test/registered/unit at this PR's head, compared with main: no new failures.

Speed Tests and Profiling

Not applicable.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): 🚫 Run #36772917062
Latest PR Test (Extra): 🚫 Run #36772916590
Latest PR Test (AMD ROCm 10): ❌ Run #36772917133