Add SD3 segmented checkpointing support

[bghira]

Summary

Splits the SD3 segmented checkpointing support model integration out of #2925.

• applies the model-family checkpointing hooks and support flags
• adds this family to the relevant safety-check allow-lists
• keeps the shared runtime/docs in the base PR so this diff stays model-specific

Stack

Base branch: agent/segmented-checkpointing-sanavideo

Validation

• .venv/bin/python -m unittest tests.test_segmented_checkpointing_model_support -v
• .venv/bin/python -m unittest tests.test_transformers.test_sd3_transformer -v
• commit hooks: Black, isort, flake8, whitespace checks

[Copilot]

Pull request overview

Adds Stable Diffusion 3 (SD3) to the project’s “segmented checkpointing” and attention-activation-offload feature set, wiring SD3 into the same model-family support and safety gates introduced in the base stacked PR.

Changes:

• Implements SD3 segmented checkpointing via checkpoint_sequential_state, including optional gradient_checkpointing_segment_stride.
• Adds SD3 attention activation offload support (activation_offload_context) and marks the model as supporting it.
• Extends unit tests to cover SD3 segmented checkpointing behavior and offload-context usage, and updates safety-check allow-lists.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/test_transformers/test_sd3_transformer.py	Adds SD3 unit tests ensuring segmented checkpointing uses checkpoint_sequential_state and that attention offload context is invoked.
tests/test_segmented_checkpointing_model_support.py	Registers SD3 in the model-level “checkpointing controls supported” test matrix (interval/stride/offload/attention-offload).
simpletuner/helpers/training/default_settings/safety_check.py	Adds "sd3" to the allow-lists for interval/stride checkpointing and attention activation offload.
simpletuner/helpers/models/sd3/transformer.py	Adds SD3 segmented checkpointing + segment stride plumbing and wraps SD3 attention calls with activation_offload_context.
simpletuner/helpers/models/sd3/expanded.py	Updates _set_gradient_checkpointing signature and behavior to match the recursive enabling pattern used elsewhere, and broadens backend detection to startswith("unsloth").

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.