[Oke Iyanuoluwa Enoch (Iyanuoluwa007)]

Describe your changes

After Rtn, Gptq or KQuant with moe=True, reloading the checkpoint and calling model(...) crashes with:

RuntimeError: Shape-movement / view ops (transpose, reshape, view, permute, flatten, expand, …) are not supported on an Olive QuantTensor ...

transformers selects the grouped_mm fused-experts forward by default (even on CPU), and its _grouped_linear calls weight.transpose(-2, -1), which the storage-only QuantTensor deliberately rejects. The eager implementation works, but users had to know to call model.set_experts_implementation("eager") themselves.

finalize() (shared by all three passes) now writes "experts_implementation": "eager" to the saved config.json for moe=True runs, so from_pretrained loads the checkpoint with the eager forward by default. save_pretrained does not persist the implementation choice, and save_metadata rewrites config.json, so the key is written after both.

The quantization docs section that described the manual workaround now explains this default instead.

This only addresses the inference crash (item 1 of the issue). Making QuantTensor support a transpose (and whether batched_mm is affected) stays with the transposed-layout design in #2621.

Checklist before requesting a review

• Add unit tests for this change. (test_rtn_moe_checkpoint_runs_with_default_experts_implementation: quantizes a tiny local Qwen3-MoE whose config uses the default implementation, reloads it and runs a forward pass. Without the fix it fails with the QuantTensor RuntimeError above.)
• Make sure all tests can pass. (test_rtn.py, test_kquant.py, test_gptq_moe.py, test_selective_mixed_precision.py: 278 passed, 3 skipped)
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a (ruff 0.15.11 check and format on changed files)
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.
  Checkpoints quantized with moe=True now load with the eager experts implementation, so native PyTorch inference works without calling set_experts_implementation("eager") manually.

(Optional) Issue link

Fixes #2619

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟢 Approval recommended

The focused shared-path fix addresses the reported crash with regression coverage and no unresolved findings.

No open issues 🎉

What changed in this PR

Fixes #2619 by making MoE-quantized checkpoints default to eager experts during native PyTorch inference.

Changes:

• Persists the eager setting after shared quantization finalization saves metadata.
• Adds a regression test covering checkpoint reload and inference.
• Updates documentation to explain the new default.

File	Description
test/​passes/​pytorch/​test_rtn.py	Tests default reload and inference for an MoE-quantized checkpoint.
olive/​passes/​pytorch/​quant_utils.py	Pins eager experts in saved configuration for moe=True.
docs/​source/​features/​quantization.md	Documents the eager default and explicit overrides.

_{Review effort: Balanced}