fix(api): use torch.no_grad() instead of torch.inference_mode() on DepthAnything3.forward()

[jaewilson07]

Summary

Swaps @torch.inference_mode() → @torch.no_grad() on the public DepthAnything3.forward() method.

Why

torch.inference_mode() is strictly stricter than torch.no_grad() — it strips version counters from tensors, which autograd still expects downstream. DA3's cls_token (src/depth_anything_3/model.py:583) is registered as nn.Parameter(torch.zeros(1, 1, embed_dim)). Under @torch.inference_mode(), that parameter gets stamped as an "inference tensor" on first call; subsequent calls fail at prepare_cls_token with:

RuntimeError: Inference tensors do not track version counter.
  File ".../depth_anything_3/model.py", line 668, in prepare_cls_token
    cls_token = self.cls_token.expand(B, S, -1)

The failure becomes catastrophic when the model is wrapped by another torch.inference_mode() context (e.g. comfy_env's persistent worker) — composition makes the version-counter poisoning permanent for the worker's lifetime.

This is the PyTorch #90882 canonical fix: any user-facing inference decorator should be no_grad, not inference_mode, unless the model itself is known to never re-use persistent buffers across calls.

Diff

```diff

• @torch.inference_mode()

• @torch.no_grad()
  def forward(self, ...):
  ```

(src/depth_anything_3/api.py:99)

Trade-off

Marginal (~10-15%) overhead vs torch.inference_mode() in inference-heavy paths. Same perf profile for view tracking, but no version-counter poisoning on persistent parameters. For DA3 specifically, the actual compute happens inside with torch.no_grad(): already (line 127), so the outer decorator is essentially a perf hint — the cost is negligible.

Verification

• DA3 + comfy_env persistent worker
• 4/4 batch renders succeeded after this change (sex, spreadass, handjob, feetfocus at 896×1536, ~30s each)
• 0/4 succeeded before (every render errored at prepare_cls_token)

Alternative considered

Remove the decorator entirely and rely solely on the inner with torch.no_grad(): at line 127. That works too — but the outer decorator serves as documentation of intent ("this is the public inference entrypoint, autograd is off"). Keeping @torch.no_grad() preserves that intent without the version-counter hazard.

Closes

Closes #266.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Updates the Depth Anything 3 public DepthAnything3.forward() inference entrypoint to avoid PyTorch inference_mode “inference tensor” side effects on persistent nn.Parameter buffers when inference contexts are nested, which can break subsequent calls.

Changes:

• Replace @torch.inference_mode() with @torch.no_grad() on DepthAnything3.forward() to prevent version-counter issues under repeated/nested inference contexts.

---

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