Need help porting PyTorch code to GGML

[tbtlr]

Hi Community.

I'm struggling porting the following PyTorch code get the same tensor using GGML:

# 'encoder_output' now contains extracted audio features from the whisper model
# Shape is (batch_size, T, C) for the encoder output
features = encoder_output.squeeze(0)  # Remove batch dimension for (T, C) shape

# Ensure features are on the same device
features = features.to(model.device)

features_cpu = features.cpu()  # Move to CPU
features = features_cpu.numpy()  # Convert to NumPy array

# Convert to PyTorch tensor and set device
features = torch.tensor(features, dtype=torch.float32).to(model.device)
#print("shape of features:", features.shape)

# Reshape features for unfolding operation
features = features.view(-1, C)
features = features.permute(1, 0).contiguous()  # Shape becomes (C, T)
features = features.view(1, C, -1, 1)  # Shape becomes (1, C, T, 1)

#print("shape of features:", features.shape)
# Unfold the features using a sliding window
unfolded_features = F.unfold(features, kernel_size=(window_size, 1), padding=(padding, 0), stride=(2, 1))
# Reshape unfolded features to desired shape
unfolded_features = unfolded_features.view(C, window_size, -1).permute(2, 1, 0).contiguous()

Getting pretty frustrated at this point since everything I've tried so far does not lead to a desired output tensor of (PyTorch) shape (750, 16, 384).

Any help is highly appreciated!