[Refactor] Keep only the TP and PP groups on the model runner

[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

ModelRunner.init_torch_distributed copies fifteen placement values onto the runner. Only two of them have to outlive the scope the runner is built in:

• speculative workers re-enter a draft's scope through the draft runner's TP group;
• draft forwards read its PP group without entering the pipeline scope.

The other thirteen are second copies of values that get_parallel() already answers wherever they are read.

Modifications

• Keep tp_group and pp_group. Drop the other thirteen: attention_tp_group, tp_rank, tp_size, dp_size, attn_dp_size, pp_rank, pp_size, attn_cp_rank, attn_cp_size, attn_dcp_rank, attn_dcp_size, moe_ep_size, dp_rank.
• Each reader takes the value from where it runs:
  • Readers that run while the runner is built, during a forward or a graph capture, or only on the target read get_parallel(). In those scopes it answers what the runner copied.
  • Readers that can run on a draft runner outside its scope read the runner's own groups. The KV cache configurator is built while pools are allocated, and the PP proxy hidden size is also reached when graphs are recaptured; both use pp_group.world_size / rank_in_group. The DFLASH / DSpark rank-0 log gates inside the draft scope read the target runner's tp_group.rank_in_group.
  • The HiSparse coordinator is built with the pools and reads get_parallel().attn_tp_group. Outside the scope that is the target's attention-TP group, which is also the group an attention-owning draft runs on; a draft that does not own attention never narrows it.
• The runner-owned tp_size / tp_rank arguments to build_load_config, dist_barrier_after_load, the tensor-dump hook, apply_torch_tp and the quantized-MoE check come from the context too. All of these run in the runner's constructor, except the barrier after an overlapped startup load, which the scheduler runs on the target's runners outside any draft scope.
• Tests that stood in for the removed fields follow:
  • the MLX runner stub tests publish their attention-DP width;
  • the DFLASH sampler and phase-1 tests give their fake a TP group and publish a topology;
  • fake runners in the attention test kits and backend tests drop placement fields nothing reads;
  • the placement-freeze census asserts that it finds the TP and PP groups.

No value a reader sees changes. Code outside the tree that reads, for example, model_runner.tp_rank should read get_parallel().tp_rank.

Accuracy Tests

H200:

• Qwen3-0.6B with dummy weights, --random-seed 42, --tp-size 2 and --tp-size 2 --dp-size 2 --enable-dp-attention: 4 greedy prompts × 32 tokens are identical to the parent commit.
• Qwen3-8B with an EAGLE3 draft, TP2 and TP2 + DP attention (the attention-owning draft path): greedy text and spec_verify_ct are identical to main. Measured at the top of this stack, which contains this PR.
• The MLX runner stub tests pass (21) with a stand-in mlx module, on main and on this PR.
• 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 #36772756644
Latest PR Test (Extra): 🚫 Run #36772756255
Latest PR Test (AMD ROCm 10): 🚫 Run #36772756706