zygi

Add key/value caching for autoregressive generation

6

Modifies model and generation code to support key/value caching. This should mostly be a no-op in terms of behavior except for some numerical stability related changes. Tested locally. Semantic array...

Easier way to do mixed-mode matrix multiplication

6

**Describe the bug** In deep learning, people often use fp16 matmuls with fp32 accumulation (cuBLAS compute type) as a balance between performance and preserving numerical accuracy. In Torch, if you...

### Describe the feature Hi! Thanks for all the work, after the 04/15 patch I can now reproduce most of the SWE-bench instances using the default harness. However, I'm still...

Modeling mappings as child elements?

1

Imagine I have the following xml: ``` Hello text_content_1 text_content_2 ... ``` That is, the metadata consists of a dynamic number of elements with dynamic tags and no attributes, each...