[WIP]Add gradient checkpointing support for AutoencoderKLWan

[victolee0]

What does this PR do?

Fixes #11071 (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@a-r-r-o-w

[a-r-r-o-w]

Thanks, very nice!

[a-r-r-o-w]

@bot /style

[victolee0]

Thanks, very nice!

My PR failed to pass the test code. Would you like to look at this link?

test failure case: test_effective_gradient_checkpointing
test: test_models_autoencoder_wan.py

[quickdahuk]

@victolee0 I implemented it the same way you did. It works fine for the forward pass, but during the backward pass, the cache index gets mixed up and goes out of bounds. I might need to use a dictionary for the cache mechanism instead of an index-based list.

[victolee0]

@victolee0 I implemented it the same way you did. It works fine for the forward pass, but during the backward pass, the cache index gets mixed up and goes out of bounds. I might need to use a dictionary for the cache mechanism instead of an index-based list.

@quickdahuk
When I add exception handling as shown below, I get a different error.

Code

def forward(self, x, feat_cache=None, feat_idx=[0]):
        # Apply shortcut connection
        h = self.conv_shortcut(x)

        # First normalization and activation
        x = self.norm1(x)
        x = self.nonlinearity(x)

        # exception handling
        if feat_cache is not None and len(feat_cache) < feat_idx[0]:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)

            x = self.conv1(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv1(x)

        # Second normalization and activation
        x = self.norm2(x)
        x = self.nonlinearity(x)

        # Dropout
        x = self.dropout(x)

        # exception handling
        if feat_cache is not None and len(feat_cache) < feat_idx[0]:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)

            x = self.conv2(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv2(x)

        # Add residual connection
        return x + h

test case error

E           torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint: A different number of tensors was saved during the original forward and recomputation.
E           Number of tensors saved during forward: 8
E           Number of tensors saved during recomputation: 6

[quickdahuk]

@victolee0 I've implemented gradient checkpointing for the decoder. I don't need it for the encoder now. The training is working fine for me. I implemented checkpointing for each frame but didn't put it inside decoder operations.
I've the code here.

[victolee0]

@quickdahuk
After copying and pasting the code you provided, I still get errors when running the test code on the same test.
I'm including the error message here:

a = tensor([-4.6806e-05, -2.9469e-05,  1.2341e-04]), b = tensor([-3.7164e-06, -4.1561e-06,  9.7381e-06]), args = (), kwargs = {'atol': 5e-05}

    def torch_all_close(a, b, *args, **kwargs):
        if not is_torch_available():
            raise ValueError("PyTorch needs to be installed to use this function.")
        if not torch.allclose(a, b, *args, **kwargs):
>           assert False, f"Max diff is absolute {(a - b).abs().max()}. Diff tensor is {(a - b).abs()}."
E           AssertionError: Max diff is absolute 0.00011366714170435444. Diff tensor is tensor([4.3089e-05, 2.5313e-05, 1.1367e-04]).

src/diffusers/utils/testing_utils.py:111: AssertionError

[quickdahuk]

@victolee0 The gradient calculated (<= 1.1367e-04) differs slightly. This difference may not be significant, but it's better to have a much lower difference. @a-r-r-o-w, Do you see anything suspicious in our implementation?

[quickdahuk]

@victolee0 I just found that if I used use_reentry=False, then it matches perfectly. I've updated the code.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[a-r-r-o-w]

The differing result is definitely suspicious. I would actually prefer having a refactor of the VAE so that we don't have to work with cache indexing in the way it's done here (and rather have it behave similar to what's done in CogVideoX and Mochi). If it's not possible to do it without indexing, we can consider removing the cache completely too..

The speed difference from using cache vs not is minimal from my past benchmarks with CogVideoX VAE and removing it, given it's complicated to do and probably doesn't save much time here, is a tradeoff that we could possibly make. cc @hlky @yiyixuxu Would you be able to take a look?