Add tutorial: DP fine-tuning of causal LMs with LoRA

[immu4989]

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Docs change / refactoring / dependency upgrade

Motivation and Context / Related issue

Closes #827.

Opacus has one LoRA tutorial today (tutorials/building_text_classifier.ipynb), covering BERT for sequence classification. There's no tutorial for the more common contemporary case: causal language model fine-tuning with DP-SGD + LoRA. This PR adds that as tutorials/building_dp_lora_for_llms.ipynb.

The tutorial also bakes in the silent-corruption avoidance pattern surfaced and root-caused in #820. Three independent confirmations across CPU, Kaggle T4, and RTX 5090 settled it as a device-placement ordering issue (model.to(device) before get_peft_model()). The tutorial demonstrates the safe ordering explicitly with inline rationale and links back to #820.

What's in the tutorial

• E2E NLG as the downstream task (small, standard in DP-NLP, used in the DiSK paper benchmark suite)
• GPT-2-small as the base model
• Two configurations compared: non-DP LoRA baseline vs DP LoRA at target epsilon=8
• BLEU, perplexity, throughput (tok/s), peak GPU memory, and final epsilon reported for each
• A dedicated "three safety patterns" section walking through:
  1. model.to(device) before get_peft_model() (Silent corrupted LoRA weight updates with Opacus 1.5.4 + PEFT 0.18.x: training appears normal but models are unusable #820 ordering)
  2. model.train() before make_private_with_epsilon (validator requirement)
  3. poisson_sampling=False (avoids GPT-2 forward-pass failure on empty Poisson batches)
• Honest discussion of when to reach for DP-LoRA, the learning-rate-needs-to-be-higher finding, and the engineering recipe to enable DP-full in a follow-up PR

Canonical results on Kaggle T4 (embedded in the notebook)

Config	BLEU	PPL	tok/s	Peak GB	Time	epsilon
non-DP LoRA	24.72	1.60	4679	2.17	438 s	no DP
DP LoRA	18.24	2.71	4599	2.65	445 s	7.08

DP costs about 26% relative BLEU at epsilon ~= 7 on this task, with about 22% memory overhead and negligible throughput impact.

What's intentionally out of scope

DP-full fine-tuning of GPT-2 with all 125M parameters trainable. Attempted via opacus.validators.ModuleValidator.fix() (Conv1D -> Linear swap) plus grad_sample_mode='functorch'; neither resolved a per-sample-gradient shape mismatch in clip_and_accumulate that appears to stem from GPT-2's tied wte / lm_head weights. The notebook's section 16 documents a recommended engineering recipe for a follow-up PR (untie weights, then re-apply fix and functorch). Felt cleaner to ship the LoRA portion as scoped here than to block on the DP-full work.

How Has This Been Tested

End-to-end run on Kaggle T4 with pinned versions:

• opacus 1.6.0
• peft 0.18.1
• transformers 5.0.0
• torch 2.10.0+cu128

Cell outputs in the committed notebook reflect the actual run. The three safety patterns each addressed a real failure mode encountered during development.

Checklist

• The documentation is up-to-date with the changes I made.
• I have read the CONTRIBUTING document and completed the CLA (CLA on file from PR Fix batch_first paths in DPMultiheadAttention #825).
• All tests passed, and additional code has been covered with new tests.

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D109661283. (Because this pull request was imported automatically, there will not be any future comments.)