Incorrect output shape for logits output returned from get_output("logits")

Describe the bug
I am trying to check perplexity of Phi3.5 model I exported using Olive (See Example linked below). This is a quantized with QDQ nodes and have static shapes (sliding window of 64) with decoder-pipeline.

When I call get_output("logits"), It always returns tensor with [64, N] regardless of the input prompt length.

The logits shape is correct if I use a regular gen-ai model without sliding window..

To Reproduce

1. Use olive example to export the model (https://github.com/microsoft/Olive/tree/main/examples/phi3_5#qdq-model-with-4-bit-weights--16-bit-activations)
2. Run the model (something like below)

   seqlen = 128
   input_tokens = test_encodings[ : seqlen] 
   print("Input Seq length: ", input_tokens.shape)
   
   generator = og.Generator(model, params)
   generator.append_tokens(input_tokens)
   
   logits =  torch.tensor(generator.get_output("logits"))
   print("- Get Output:", logits.shape)
   

Outputs:

- Get Output: (1, 64, 32064)

Expected behavior

For Phi3.5, the logits shape should be (1, 128, 32064) for input sequence length of 128.

Desktop (please complete the following information):

• OS: [Windows]

Start with the Olive Phi3.5 QDQ export example and the Generator.append_tokens and get_output("logits") entry points described in the report. Reproduce the static sliding-window case with a 128-token prompt, compare it with a regular model, and trace where the returned logits shape is determined. Done means the quantized Phi3.5 output preserves the input sequence length while retaining the expected vocabulary dimension.