[CharlelieLrt]

Description

This PR adds the optional DistillationStrategy for users who do not need to
write their own distillation loop. It extends Alchemi's general training strategy
to manage phase-specific optimizers, gradient accumulation, AMP, EMA, and
checkpoint restart.

Users install the diffusion-strategy extra and supply public models, a native
or FastGen-backed loss, and TensorDict batches:

strategy = DistillationStrategy(
    models={"student": student}, loss_fn=loss_fn, num_steps=1000,
    optimizer_configs={"student": [{"optimizer_cls": torch.optim.Adam,
                                    "optimizer_kwargs": {"lr": 1e-4}}]},
)
strategy.run(batches)

For explicit MeanFlow times, store t and r as separate batch tensors.
Use training_fn(models, batch) to return loss arguments containing
(t, DiffusionExtraTimes(r=r, batch_size=[B])). Ordinary single-time batches
need no conversion.

The strategy remains optional: expert users can call the same loss from their
own loop. Compilation and distributed setup remain under user control.
This PR also carries the grouped changelog entry for the distillation feature.

Checklist

• This PR follows the contribution guidelines.
• The documentation describes the API and its role in the stack.
• This PR remains a draft.

Dependencies

• Depends on

  • PhysicsNeMo #2098 adds checkpoint bundles for restarting distillation. This is the immediate parent branch.

• Earlier prerequisites

  • PhysicsNeMo #2094 adds named diffusion inputs and outputs.
  • PhysicsNeMo #2095 adds inverse SNR and TrigFlow components.
  • PhysicsNeMo #2096 adds private adapters and public sCM and MeanFlow losses.
  • PhysicsNeMo #2097 extends the loss API to DMD2 and f-distill.

• Parallel FastGen work

  • FastGen #43 allows the NumPy 2 dependency used by PhysicsNeMo. PhysicsNeMo Add sCM and MeanFlow distillation losses #2096 uses an immutable fork pin, so it does not need to wait for this proposal to merge.
  • FastGen #44 provides shared loss hooks for external training loops. The initial PhysicsNeMo integration does not require these hooks.

• Conditional follow-up

  • PhysicsNeMo #2100 replaces local compatibility code with the proposed FastGen loss hooks. It depends on PhysicsNeMo Add optional diffusion distillation strategy #2099 and FastGen [other] Add modulus.models.fcn_mip_plugin #44, not the other way around.

If FastGen maintainers decline FastGen #44, we would close PhysicsNeMo #2100.
The preceding PhysicsNeMo stack would remain available.

Review process

Reviewers can compare this PR with its immediate parent. After a parent merges,
we update the remaining branches and their PR bases to keep each diff focused.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

CODEOWNERS review map

Current for commit bc163960e295. An approval covers every file listed for that owner; one owner is sufficient for shared files.

⏳ @CharlelieLrt — 7 file(s)

• docs/api/diffusion/distillation.rst
• physicsnemo/diffusion/distillation/init.py
• physicsnemo/diffusion/distillation/_alchemi.py
• physicsnemo/diffusion/distillation/_fastgen.py
• physicsnemo/diffusion/distillation/losses.py
• physicsnemo/diffusion/distillation/strategy.py
• test/diffusion/test_distillation_strategy.py

⏳ @coreyjadams — 1 file(s)

• pyproject.toml

⏳ @ktangsali — 1 file(s)

• pyproject.toml

⏳ @megnvidia — 1 file(s)

• docs/api/diffusion/distillation.rst

No CODEOWNER

• CHANGELOG.md
• uv.lock

Comment /codeowners-info to refresh.

[CharlelieLrt]

Continued in PhysicsNeMo #2204, from branch diffusion-distillation-strategy in CharlelieLrt/modulus. The replacement preserves this implementation and includes the requested documentation clarifications. Please continue the review there. This PR retains the earlier discussion.