Trian Xylouris

[BUG] DeepSpeed non-deterministic inference with HF GPT2 when `replace_with_kernel_inject=True`

2

**Describe the bug** https://github.com/microsoft/DeepSpeed/issues/1950 describes a bug by which running inference twice on the same input leads to different outputs. It was supposedly fixed in version `0.6.5`, but I am...

**Describe the bug** For the models GPT-neo-1.3B, Bloom 1b7, Pythia 1.4b, GPT2-xl, I get non-deterministic model outputs when using context length 1 and `engine = deepspeed.init_inference(model, dtype=torch.float16, replace_with_kernel_inject=True)`. The context...

[BUG] Non-deterministic model output if `replace_with_kernel_inject=True` for GPT-neo-1.3B/Bloom 1b7 and others

2

**Describe the bug** For the models GPT-neo-1.3B, Bloom 1b7, Pythia 1.4b, GPT2-xl, I get non-deterministic model outputs when using context length 1 and `engine = deepspeed.init_inference(model, dtype=torch.float16, replace_with_kernel_inject=True)`. The context...

Add efficient validation on HellaSwag

5

This PR adds optional validation using HellaSwag accuracy (on 10042 validation examples). It is controlled by setting the hyperparameter `hellaswag=True`. *Motivation* I am trying to make the model smarter by...

PyMuPDF Pro 1.25.0: Some graphical elements of a `.docx` file not visible with `get_pixmap()`

2

### Description of the bug At a `.docx` file, PyMuPDF Pro did not display certain graphical purple elements: With PyMuPDF Pro:  With Google Docs or Mac's Pages app: ...

PyMuPDF Pro 1.25.0: A 1-page .docx file is split into 4 pages

2

### Description of the bug For some documents, PyMuPDF Pro splits the document into many more pages than if I open the document with Google Docs (or Mac Pages/libreoffice). This...

PyMuPDF Pro 1.25.0: Image in `.doc` file unexpectedly overlaps with text when using `get_pixmap()`

3

### Description of the bug Sometimes, an embedded image inside a `.doc` file overlaps with the text when creating an image of the document using `get_pixmap()`, although at other software...