Birch-san

Results 29 issues of Birch-san

Spent a very long time wondering why I had to "end()" `stdin` to send a message to Python. Thanks to issue #18 and this [StackOverflow answer about flushing Python output](http://stackoverflow.com/a/18849329),...

I've been trying to get dalle-playground running performantly on M1, but there's a lot of work remaining to make the JAX model work via IREE/Vulkan. so, I tried out your...

The "Building against system LLDB" instructions didn't work out-of-the-box on M1 macOS Monterey 12.4 with XCode 13.2.1. Here's a known-good sequence of instructions. Pretty similar to the existing "Building against...

Improves support for diffusion models with discrete time-steps (such as Stable Diffusion's DDIM). I have some questions though, so this may need some iterating. The user would invoke like so:...

should help with https://github.com/crowsonkb/k-diffusion/issues/25 (giving users more control over determinism), at least for Karras samplers at the default `churn=0`. this also eliminates allocation of an unused tensor on every sampler...

When I build, the makefile detects my M1 Max as 86_64. This is because I have GNU coreutils `uname` on my `PATH`, which announces my architecture as `arm64` (whereas the...

bug
hardware
build

I encounter `Data didn't contain opus stream: /assets/converted/sounds/6042.opus`. I checked and confirmed that the absolute path to the file was correct, and pointed to a real file on my computer....

bug
os: macos
lang: c++
area: assets

In codegen, we frequently do a `CREATE_MEMSET()` to zero out any memory we've allocated with AllocaInst: > // initializing memory needed for older kernels > // memset so verifier doesn't...

enhancement
discussion-needed

- loads OpenAI guided-diffusion UNet - assigns to `inner_model_ema` only - constructs 256x256 guided diffusion UNet, with config.json overrides - prepares `model_ema` with accelerate - wraps it with k-diffusion wrapper...

[Grouped Query Attention](https://arxiv.org/abs/2305.13245) improves parameter-efficiency of attention KV projections and reduces IO at inference-time, making inference faster. It can be implemented naïvely by just repeating K and V along the...

feature-request