[CharlelieLrt]

Description

This PR extends DistillationLoss to DMD2 and f-distill. Users supply their
student, teacher, and auxiliary models, then choose the training phase explicitly
in their optimizer loop:

loss_fn = DistillationLoss(
    student, scheduler, method="DMD2", teacher=teacher,
    auxiliary_models={"fake_score": fake_score},
)
losses = loss_fn(x0, phase="auxiliary", step=step)
losses["total_loss"].backward()

The loss manages method-specific state, including f-distill ratio statistics,
while the caller controls phase order and optimizer updates. For adversarial
training, configure the teacher's feature_names and a
feature_fn(prediction, extras) callback that selects named features from its
output. The teacher always returns its prediction, even when the discriminator
uses only its features.

Weighted losses exclude masked values from residuals and applicable statistics.
For masked adversarial training, users provide an adversarial_fn callback
that excludes corrupted inputs during feature
extraction and weights the resulting discriminator scores. A pixel mask alone
cannot determine which discriminator scores are valid.

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 #2096 adds private adapters and public sCM and MeanFlow losses. This is the immediate parent branch.

• Earlier prerequisites

  • PhysicsNeMo #2094 adds named diffusion inputs and outputs.
  • PhysicsNeMo #2095 adds inverse SNR and TrigFlow components.

• Depends on this PR

  • PhysicsNeMo #2098 adds checkpoint bundles for restarting distillation. It follows this PR directly in the stack.

• Later follow-ups

  • PhysicsNeMo #2099 adds the optional training strategy and the grouped changelog entry.

• 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 6d301d6f2512. An approval covers every file listed for that owner; one owner is sufficient for shared files.

⏳ @CharlelieLrt — 8 file(s)

• docs/api/diffusion/distillation.rst
• physicsnemo/diffusion/distillation/losses.py
• physicsnemo/diffusion/distillation/methods/dmd2/init.py
• physicsnemo/diffusion/distillation/methods/dmd2/_fastgen.py
• physicsnemo/diffusion/distillation/methods/f_distill/init.py
• physicsnemo/diffusion/distillation/methods/f_distill/_fastgen.py
• test/diffusion/helpers.py
• test/diffusion/test_distribution_distillation.py

⏳ @megnvidia — 1 file(s)

• docs/api/diffusion/distillation.rst

No CODEOWNER

• CHANGELOG.md

Comment /codeowners-info to refresh.

[CharlelieLrt]

Continued in PhysicsNeMo #2202, from branch diffusion-distribution-matching 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.