Started the long-pending rewrite of Easy Diffusion’s server code. v4 intends to replace the Python (and PyTorch) based server with a simple C++ version. The reason for rewriting the server in C++ is to achieve sub-second startup time for the UI, and to reduce the download size (won’t need to distribute Python along with Easy Diffusion) or mess with conda/venv etc. And it’s also something that I want to do for personal taste, i.e. de-bloating what doesn’t need to be bloated.
For Z-Image, the performance of the stock version of chromaForge is poorer than sd.cpp’s. Mainly because chromaForge isn’t able to run the smaller gguf quantized models that sd.cpp is able to run (chromaForge fails with the errors that I was fixing yesterday). If I really want to push through with this, it would be good to fix the remaining issues with gguf models in chromaForge. Only then can the performance be truly compared (in order to decide whether to release this into ED 3.5). I want to compare the performance of the smaller gguf models, because that’s what ED’s users will run typically.
Worked on fixing Z-Image support in ED’s fork of chromaForge (a fork of Forge WebUI). Fixed a number of integration issues. It’s now crashing on a matrix multiplication error, which looks like an incorrectly transposed matrix (mostly due to reading the weights in the wrong order). I’ll try to install a stock version of chromaForge to see its raw performance with Z-Image (and whether it’s worth pursuing the integration), and also use it to help investigate the matrix multiplication error (and any future errors).
Collecting the worklog over the past few weeks. Enabled Flash-Attention and CPU offloading by default in sdkit3 (i.e. Easy Diffusion v4). Added optional VAE tiling (and VAE tile size configuration) via config.yaml in Easy Diffusion v4. Created Easy Diffusion’s fork of Forge WebUI, in order to apply the patches required to run with ED. And also to try adding new features like Z-Image (which are missing in the seemingly-abandoned main Forge repo). Improved the heuristics used for killing and restarting the backend child process, since /ping requests are unreliable if the backend is under heavy load. Merged a few PRs (1 2) for torchruntime that improve support for pinning pre-cu128 torch versions and fix the order of detection of DirectML and CUDA (prefers CUDA). Added progress bars when downloading v4 backend artifacts.
The new engine that’ll power Easy Diffusion’s upcoming v4 release (i.e. sdkit3) has now been integrated into Easy Diffusion. It’s available to test by selecting v4 engine in the Settings tab (after enabling Beta). Please press Save and restart Easy Diffusion after selecting this. It uses stable-diffusion.cpp and ggml under-the-hood, and produces optimized, lightweight builds for the target hardware. The main benefits of Easy Diffusion’s new engine are:
Following up to the deep-dive on ML compilers: sdkit v3 won’t use general-purpose ML compilers. They aren’t yet ready for sdkit’s target platforms, and need a lot of work (well beyond sdkit v3’s scope). But I’m quite certain that sdkit v4 will use them, and sdkit v3 will start making steps in that direction. For sdkit v3, I see two possible paths: Use an array of vendor-specific compilers (like TensorRT-RTX, MiGraphX, OpenVINO etc), one for each target platform. Auto-generate ggml code from onnx (or pytorch), and beat it on the head until it meets sdkit v3’s performance goals. Hand-tune kernels, contribute to ggml, and take advantage of ggml’s multi-backend kernels. Both approaches provide a big step-up from sdkit v2 in terms of install size and performance. So it makes sense to tap into these first, and leave ML compilers for v4 (as another leap forward).
This post concludes (for now) my ongoing deep-dive into ML compilers, while researching for sdkit v3. I’ve linked (at the end) to some of the papers that I read related to graph execution on GPUs. Some final takeaways: ML compilers might break CUDA’s moat (and fix AMD’s ROCm support). A single compiler is unlikely to fit every scenario. The scheduler needs to be grounded in truth. Simulators might be worth exploring more. ML compilers might break CUDA’s moat (and fix AMD’s ROCm support) It’s pretty clear that ML compilers are going to be a big deal. NVIDIA’s TensorRT is also an ML compiler, but it only targets their GPUs. Once the generated machine code (from cross-vendor ML compilers) is comparable in performance to hand-tuned kernels, these compilers are going to break the (in)famous moat of CUDA.
Good post on using MLIR for compiling ML models to GPUs. It gives a good broad overview of a GPU architecture, and how MLIR fits into that. The overall series looks pretty interesting too! Making a note here for future reference - https://www.stephendiehl.com/posts/mlir_gpu/
Wrote a fresh implementation of most of the popular samplers and schedulers used for image generation (Stable Diffusion and Flux) at https://github.com/cmdr2/samplers.cpp. A few other schedulers (like Align Your Steps) have been left out for now, but are pretty easy to implement. It’s still work-in-progress, and is not ready for public use. The algorithmic port has been completed, and the next step is to test the output values against reference values (from another implementation, e.g. Forge WebUI). After that, I’ll translate it to C++.
Some notes on machine-learning compilers, gathered while researching tech for Easy Diffusion’s next engine (i.e. sdkit v3). For context, see the design constraints of the new engine. tl;dr summary The current state is: Vendor-specific compilers are the only performant options on consumer GPUs right now. For e.g. TensorRT-RTX for NVIDIA, MiGraphX for AMD, OpenVINO for Intel. Cross-vendor compilers are just not performant enough right now for Stable Diffusion-class workloads on consumer GPUs. For e.g. like TVM, IREE, XLA. The focus of cross-vendor compilers seems to be either on datacenter hardware, or embedded devices. The performance on desktops and laptops is pretty poor. Mojo doesn’t target this category (and doesn’t support Windows). Probably because datacenters and embedded devices are currently where the attention (and money) is.