[Refactor] Add make_pp_layers so models stop handing their PP position down

[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

48 models build their decoder stack with make_layers(..., pp_rank=self.pp_group.rank_in_group, pp_size=self.pp_group.world_size, ...); one spells it with get_parallel().pp_group. Passing the pair both supplies the stage and switches make_layers to return (layers, start_layer, end_layer), so it is not redundant as such. But every caller passes the placement of the scope it is being built in.

Modifications

• Add make_pp_layers(num_hidden_layers, layer_fn, prefix, ...), which reads the stage from get_parallel() and returns the tuple, and use it at all 48 sites.
• Models are constructed inside the scope that states their placement (a draft inside its pipeline scope), where get_parallel() answers the same stage as self.pp_group.
• test_make_pp_layers.py: each stage builds its own slice from the published placement, and a single-stage scope builds every layer.

Accuracy Tests

H200:

• Qwen3-0.6B with dummy weights, --tp-size 2 and --pp-size 2: greedy output identical with and without this PR and the two before it.
• upstage/solar-pro-preview-instruct, --tp-size 2 --pp-size 2: greedy output identical to the Solar TP2 run earlier in this stack; GSM8K (200 questions) 0.900.
• 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 #36772881908
Latest PR Test (Extra): 🚫 Run #36772880959
Latest PR Test (AMD ROCm 10): ❌ Run #36772881796