[Refactor] Extract model specific logics in export lib

[h-guo18]

What does this PR do?

Type of change: Refactor / code health

Extracts model-specific logic out of the generic export code into a new model-family registry, modelopt/torch/export/modeling/. Each family is described declaratively (per-model data + optional behavioral hooks); the export engine resolves a family's ModelSpec and reads from it instead of branching on model names. An unmatched lookup returns None, so the engine falls back to its original path — the migration is incremental and behavior-preserving.

Migrated so far:

• Data: MoE expert names, iterable-expert flag, activation override, shared-embedding flag, MLP keyword roles, and AWQ pre_quant_scale fusion rules.
• Hooks (ModelHooks + ExportContext, dependency-injected to keep modeling/ cycle-free): decoder sub-module placement for Gemma2/3 layernorms, Gemma3 q/k-norms, and MLlama self/cross attention.

Shared tables (HF_CONFIG_MAP) and generic algorithms stay in the engine. See modelopt/torch/export/MODEL_SPECIFIC_REFACTOR.md for the inventory, design, and migration plan.

Not yet migrated

Planned for follow-ups (still uses the engine's default path):

• build_moe — per-family MoE router / experts / shared-expert construction (Llama/Phi3, DBRX, DeepSeek, Qwen).
• unwrap_decoder_layer — DBRX/ExaOne/Deci module-tree unwrap, plus the head_is_first_dim flag (Bloom/Falcon/Phi3Small/InternLM) as data.
• TRT-LLM target-config extras (tensorrt_llm_utils) — positional-embedding type, MPT alibi, RecurrentGemma, DBRX clip_qkv, Phi3-MoE sparse-mixer (mostly data).
• HF path (unified_export_hf / moe_utils) — MoE expert export branches (Llama4/GptOss/DBRX/iterable) are a separate engine and would need HF-side seams; VLM language-tower extraction currently lives in model_utils.

Deferred (low value): embed √-scale (Gemma1-only + version-gated) and norm+1 (mixes the generic LayerNorm1P).

Intentionally kept in the engine: HF_CONFIG_MAP (a shared alias table, not a per-model branch), generic algorithms (_GATE_UP_PAIRS, expert-amax fallback), the ChatGLM/Phi3 fused gate/fc chunk-swap (a fused-weight reshape), and speculative-decoding export (already modular under plugins/hf_spec_*).

Usage

# Adding a model family is a small declarative file; no engine edits.
# modelopt/torch/export/modeling/families/<family>.py
register(
    ModelSpec(
        name="qwen_moe",
        moe_block_names=("Qwen3MoeSparseMoeBlock",),
        expert_linear_names=("gate_proj", "down_proj", "up_proj"),
        has_iterable_experts=True,
    )
)

Testing

• pytest tests/unit/torch/export/ passes (78 passed; test_export_diffusers.py skipped — pre-existing CUDA/glibc collection error unrelated to this change).
• Each migration step verified for behavioral equivalence against the legacy branches; fallback-first means un-migrated families are byte-for-byte unaffected.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅ (behavior-preserving; fallback-first)
• If you copied code from any other sources or added a new PIP dependency: N/A
• Did you write any new necessary tests?: N/A — behavior-preserving refactor covered by existing tests/unit/torch/export/
• Did you update Changelog?: N/A (internal refactor, no public API change)
• Did you get Claude approval on this PR?: ❌ (pending)

Additional Information

Refactor only — no change to exported checkpoints. Design/rationale: modelopt/torch/export/MODEL_SPECIFIC_REFACTOR.md.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 1f7520a8-570a-432d-a5dd-0b18214bd98f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch haoguo/export-modelinglib

---

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

❌ Patch coverage is 62.96296% with 70 lines in your changes missing coverage. Please review.
✅ Project coverage is 77.16%. Comparing base (4093664) to head (8ddd26c).
⚠️ Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
modelopt/torch/export/layer_utils.py	14.28%	24 Missing ⚠️
modelopt/torch/export/modeling/families/gemma.py	33.33%	16 Missing ⚠️
modelopt/torch/export/modeling/registry.py	50.00%	16 Missing ⚠️
modelopt/torch/export/modeling/families/mllama.py	53.84%	6 Missing ⚠️
modelopt/torch/export/modeling/hooks.py	72.72%	3 Missing ⚠️
modelopt/torch/export/quant_utils.py	25.00%	3 Missing ⚠️
modelopt/torch/export/model_config_export.py	33.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1828      +/-   ##
==========================================
- Coverage   77.36%   77.16%   -0.21%     
==========================================
  Files         513      534      +21     
  Lines       56889    57365     +476     
==========================================
+ Hits        44012    44265     +253     
- Misses      12877    13100     +223     

Flag	Coverage Δ
unit	54.69% <62.96%> (+0.09%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[h-guo18]

trtllm export path is deprecated.

[h-guo18]

HF export only.