[New Model] Add LensPipeline （microsoft/Lens）

[RuixiangMa]

What does this PR do?

Add support for Lens , text-to-image models from Microsoft.
Repoid: microsoft/Lens

import torch
from diffusers import LensPipeline

pipe = LensPipeline.from_pretrained("microsoft/Lens", torch_dtype=torch.bfloat16)
pipe.to("cuda")

prompt = "A steampunk floating sky-city built on massive gear-driven platforms, 
brass and copper towers connected by chain bridges, steam-powered airships and 
hot air balloons docking at various levels, sunset clouds below the city, detailed concept art"
generator = torch.Generator("cuda").manual_seed(42)
image = pipe(prompt, height=1440, width=1440, num_inference_steps=20, guidance_scale=5.0, generator=generator).images[0]
image.save("lens.png")

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.

cc @yiyixuxu

[dxqb]

thanks for the PR!
I am not associated with diffusers and they'll have to review it, but I might have some useful comments and questions because I have implemented Lens for https://github.com/Nerogar/OneTrainer using the Microsoft upstream. As soon as your PR is merged, I'd like to change it to diffusers as upstream.

• GPT-OSS is a 20B model that comes pre-quantized to mxfp4 for most of its parameters. If you load it to CPU using the Microsoft/transformer upstream code, it gets automatically dequantized, using 40 GB of RAM - but then it also requires 40 GB of VRAM when you move it to GPU.
  If you load it to GPU, it materializes on GPU as 10 GB of VRAM using the kernels library - but it cannot be moved to CPU then.
  It doesn't seem to be possible to load the model quantized to CPU, and move it on demand.
  How do you handle this in this PR? diffusers might need some new infrastructure for this case, otherwise the vram-saving optimizations that diffusers have will fail.
  I have settled for on-demand loading and discard-after-use for now, to avoid the 40 GB ram.

• I think your PR does this correctly, but it's something to watch out for:
  Lens uses a few selected hidden states of the text encoder, one of which is the last layer. The Microsoft upstream extends the text encoder for that. First I wondered why they don't just used output_hidden_states=True, but this is not the same thing: output_hidden_states returns the hidden states, but the last selected layer is the actual last layer of the model: The output then runs through a norm-layer before it's returned - it does not run through the norm-layer using the upstream text encoder.
  I'm not sure if output_hidden_states is intended/documented by transformers to work like that, but using it as it is now is an embedding mismatch.

[RuixiangMa]

thanks for the PR! I am not associated with diffusers and they'll have to review it, but I might have some useful comments and questions because I have implemented Lens for https://github.com/Nerogar/OneTrainer using the Microsoft upstream. As soon as your PR is merged, I'd like to change it to diffusers as upstream.

• GPT-OSS is a 20B model that comes pre-quantized to mxfp4 for most of its parameters. If you load it to CPU using the Microsoft/transformer upstream code, it gets automatically dequantized, using 40 GB of RAM - but then it also requires 40 GB of VRAM when you move it to GPU.
  If you load it to GPU, it materializes on GPU as 10 GB of VRAM using the kernels library - but it cannot be moved to CPU then.
  It doesn't seem to be possible to load the model quantized to CPU, and move it on demand.
  How do you handle this in this PR? diffusers might need some new infrastructure for this case, otherwise the vram-saving optimizations that diffusers have will fail.
  I have settled for on-demand loading and discard-after-use for now, to avoid the 40 GB ram.
• I think your PR does this correctly, but it's something to watch out for:
  Lens uses a few selected hidden states of the text encoder, one of which is the last layer. The Microsoft upstream extends the text encoder for that. First I wondered why they don't just used output_hidden_states=True, but this is not the same thing: output_hidden_states returns the hidden states, but the last selected layer is the actual last layer of the model: The output then runs through a norm-layer before it's returned - it does not run through the norm-layer using the upstream text encoder.
  I'm not sure if output_hidden_states is intended/documented by transformers to work like that, but using it as it is now is an embedding mismatch.

Thx, this is very helpful. This PR does not add any special handling beyond the current diffusers loading/device-placement behavior, so agreed this is something to watch for with CPU offload / VRAM-saving paths.

For the hidden-state selection: this PR intentionally does not use output_hidden_states=True. Instead it uses LensGptOssEncoder to capture the selected decoder-layer outputs directly, before the final norm, to match the Microsoft upstream behavior.