Unify LoRA target selection with HF model layouts

[yueming-yuan]

Stacked on #3620 (LoRA arguments/utilities and HF config relocation only). Review this PR against that base for the HF target-selection changes.

ci-sglang-pr: #40242

Use HF module names as the source of truth for LoRA targets. Builds on #3291 and SGLang #40242.

• Share HF target groups through --target-modules attn,mlp,unembed; ordinary LoRA can mix groups with explicit HF targets, while Tinker accepts only whole groups. Exclude DSA indexers.
• Convert HF → checkpoint → Megatron names at the backend boundary; reject unsupported partial fused targets.
• Build adapter configs from exported names; preserve Inkling's native adapter format.

Review fixes

• Emit accepted canonical aliases for fully selected expert FC1 targets; grouped-expert adapters retain their fused layout.

• Allow fully selected GDN input projections in canonical mode; reject partial expert/GDN selections.

• Replace Tinker training flags with --target-modules attn,mlp,unembed; default to all three groups and derive SDK flags from the selection. Reject explicit module names and exclusions in Tinker.

• Exclude packed gate_up_proj in both GLM launchers when KEEP_MOE_LORA=0.

• Apply scoped HF exclusions to resolved module names before retaining the final target selection; ordinary HF selection does not initialize Bridge.

At ba82dee23a, pre-commit run --all-files and git diff --check pass. A CPU A/B probe against 62b70f94fe, using the real launcher functions and Transformers meta-models, distinguishes both bugs from their fixes; default launcher commands remain unchanged. Added CPU cases cover both launchers and all-layer/single-layer exclusions. The full local launcher pytest gate was blocked during collection by ModuleNotFoundError: No module named 'sglang'; no new GPU validation was performed.

Model validation and Multi-LoRA support fixes are tracked in #3584.

At 53e27ae330, a CPU A/B probe against 294fe27e766 confirms that five explicit-target/exclusion cases change from accepted to rejected, while all eight training-group flag combinations preserve their behavior. pre-commit run --all-files and git diff --check pass. No regression test files were added, as requested; no GPU validation was run for this Tinker argument fix.

At c005f80efd, CPU A/B probes against 53e27ae330 reproduce and distinguish the canonical expert/GDN initialization failures for Qwen3-MoE, GPT-OSS, Qwen3-Next, and Qwen3.5 text. Probes use native Transformers 5.12.1 meta-model names and the canonical matcher method from pinned Bridge 2e09c234a3272285140224d0d698593418b55ba5. Standard LoRA outputs and dense canonical split selections are preserved; partial expert/GDN selections remain rejected. pre-commit run --all-files and git diff --check pass. No regression test files were added, as requested. Full GPU injection, training, and adapter export were not validated by these CPU probes.

Kimi K3 integration

Define Kimi K3 HF defaults to match its existing native training recipe. The launcher now uses all-linear instead of maintaining its own target list; omitting targets or selecting attn,mlp produces the same selection. Explicit targets can still omit routed-expert w2, and the native backend rejects unsupported layouts. Shared-outer adapter names drop the expert-index wildcard; SGLang #40242 resolves exported w1/w2/w3 names to its LoRA pool modules.

Validation: an extracted argument-path A/B probe reproduced the original undefined-layout error; the updated defaults were compared against the previous launcher's exact target set for K3, K3-4layer, and K3-4layer-64experts. The HF loader/model introspection were stubbed; these checks do not exercise backend initialization. Kimi K3 manual launcher snapshot checks: 10 passed. No new regression tests or GPU validation.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[Zhichenzzz]

LGTHumans

[yueming-yuan]

/rerun-test tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_lora_ci.py

[github-actions]

❌ tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_lora_ci.py failed on stage-c-8-gpu-h200 in 11m24s — workflow run

The execution job failed; inspect the workflow run for the failing step.

[yueming-yuan]

/rerun-test tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_lora_ci.py

[github-actions]

✅ tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_lora_ci.py passed on stage-c-8-gpu-h200 in 16m13s — workflow run