Enable PyTorch's FakeTensorMode for EulerDiscreteScheduler scheduler

[thiagocrepaldi]

PyTorch's FakeTensorMode does not support .numpy() or numpy.array() calls.

This PR replaces sigmas numpy tensor by a PyTorch tensor equivalent

Repro

with torch._subclasses.FakeTensorMode() as fake_mode, ONNXTorchPatcher():
    fake_model = DiffusionPipeline.from_pretrained(model_name, low_cpu_mem_usage=False)

that otherwise would fail with
RuntimeError: .numpy() is not supported for tensor subclasses.

Fixes #7152

[thiagocrepaldi]

@hlky @patil-suraj @anton-l

[hlky]

torch.tensor here introduces a warning UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).

self.alphas_cumprod is already a torch.tensor and already in float32 so (((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5).flip(0) will be sufficient.

This fix could also be applied to other schedulers.

[yiyixuxu]

@thiagocrepaldi
I think @hlky's recommendation here makes sense, no?

[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.

[thiagocrepaldi]

@thiagocrepaldi I think @hlky's recommendation here makes sense, no?

yes, will do now. thanks

[thiagocrepaldi]

For other schedulers, we need a way to workaround the absence of torch.interp to fix lines such as sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)

Any ideas?

Maybe use the following snippet instead? (from the link above)

def interpolate(x: torch.Tensor, xp: torch.Tensor, fp: torch.Tensor) -> torch.Tensor:
    """One-dimensional linear interpolation for monotonically increasing sample
    points.

    Returns the one-dimensional piecewise linear interpolant to a function with
    given discrete data points :math:`(xp, fp)`, evaluated at :math:`x`.

    Args:
        x: the :math:`x`-coordinates at which to evaluate the interpolated
            values.
        xp: the :math:`x`-coordinates of the data points, must be increasing.
        fp: the :math:`y`-coordinates of the data points, same length as `xp`.

    Returns:
        the interpolated values, same size as `x`.
    """
    m = (fp[:,1:] - fp[:,:-1]) / (xp[:,1:] - xp[:,:-1])  #slope
    b = fp[:, :-1] - (m.mul(xp[:, :-1]) )

    indicies = torch.sum(torch.ge(x[:, :, None], xp[:, None, :]), -1) - 1  #torch.ge:  x[i] >= xp[i] ? true: false
    indicies = torch.clamp(indicies, 0, m.shape[-1] - 1)

    line_idx = torch.linspace(0, indicies.shape[0], 1, device=indicies.device).to(torch.long)
    line_idx = line_idx.expand(indicies.shape)
    # idx = torch.cat([line_idx, indicies] , 0)
    return m[line_idx, indicies].mul(x) + b[line_idx, indicies]

[thiagocrepaldi]

Hi folks, from the discussion at #7151 , do you think we can merge this one?

[yiyixuxu]

thanks! merged