This is an implementation of language model inference, aiming to get maximum single-GPU single-batch hardware utilization for LLM architectures with a minimal implementation and no dependencies¹.

The goal of this project is experimentation and prototyping; it does not aim to be production ready or stable.

If you need support for a wide range of models, computing devices or quantization methods, you're probably looking for llama.cpp or 🤗 Transformers. If you need to run inference for multiple batches, you're probably looking for vLLM.

Parts of this code are based on Andrej Karpathy's llama2.c.

To build and run calm, you need to download and convert a model, build the code using make² and run it:

git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
python tools/convert.py mistral-7b-instruct.calm Mistral-7B-Instruct-v0.2/
make && ./build/run mistral-7b-instruct.calm -i "Q: What is the meaning of life?" -t 0

You can also run the model in chat mode (for models like Mistral/Mixtral you might want to increase context size via -c from the default 4096):

make && ./build/run mistral-7b-instruct.calm -y "You are a helpful AI assistant."

Before running Python you may want to install the dependencies via pip install -r tools/requirements.txt. When using git to download models, git-lfs is required and the download size may be larger than necessary; you can use tools/download.py instead (assumes models use Safetensors by default):

python tools/download.py Mistral-7B-Instruct-v0.2/ mistralai/Mistral-7B-Instruct-v0.2

calm supports a subset of decoder-only transformer architectures:

• Llama-like baseline (pre/post normalization, gated FFN, sequential attention mixing and FFN, RoPE)
• RoPE enhancements (partial rotary dimension, independent head dimension)
• SiLU or GELU FFN gate activation
• RMSNorm or LayerNorm* normalization (no bias support)
• Optional minor variations (QKV bias, QKV clipping, tied embeddings)
• Optional mixture of experts (with top-k expert selection)

It has been tested on following models:

Model weights support fp16, fp8 and gf4 formats; the weight type is specified at conversion time via --dtype argument to convert.py, and defaults to fp8.

fp16 corresponds to 16-bit floating point (e5m10). Note that some models store weights in bf16 which will be automatically converted.

fp8 corresponds to 8-bit floating point (e5m2). Using fp8 carries a ~0.5% perplexity penalty at almost double the inference speed and half the model size. e4m3 variant of fp8 would result in a much smaller perplexity penalty (~0.1%) with basic tensor scaling, but it's currently not used because of performance issues wrt floating-point conversion.

gf4 corresponds to 4-bit grouped floating point (8 values are stored in 32 bits using 3 bit quantized scale per value and one fp8 group scale). Using gf4 currently carries a perplexity penalty but increases inference speed by ~75% and halves the model size compared to fp8. Unlike llama.cpp's K-quants, gf4 quantization is pure and uniform - all layers are quantized to exactly 4 bits per weight.

KV cache is using fp16 by default; when using longer contexts (> 4096), CUDA implementation automatically switches to fp8 to improve memory/performance. This comes at a small perplexity cost.

Auto-regressive prediction for a single sequence needs to read the entire model and the entire KV cache (until current token) for every token. As such, given an optimal implementation we'd expect the process to be bandwidth bound. Note that the cost of token generation at the beginning of the sequence should be smaller than the cost at the end of the sequence due to the need to read data from KV cache.

When using NVidia GeForce RTX 4090, calm gets the following performance on a few models; each model is measured with fp16, fp8 and gf4 weights at the beginning of the context window (first 32 tokens) and at the end (last 32 tokens with an offset 2000 for 2048 contexts, 4000 for 4096 contexts and 16000 for 16384 contexts):

Currently prompts are processed serially, one token at a time; in the future, prompt processing will need to be parallelized to avoid the bandwidth bottleneck.

With smaller weights on small models, getting closer to bandwidth limit becomes more difficult. Future optimizations may increase the gap here for small models, although smaller weights are most valuable to be able to infer larger models.

RTX 4090 has a peak bandwidth of ~1008 GB/s, however it's unclear if a peak higher than ~950 GB/s is attainable in practice³.

calm can run on A100/H100 accelerators (but is mostly tuned for H100 fp8 weights). When using Mixtral 8x7B (fp8) on 1xH100 SXM, it runs at ~200 tok/s (2550 GB/s) for 256-token outputs.

calm uses 🤗 Safetensors to store model files. Note that the models require conversion (see below), because calm stores model hyperparameters in .safetensors metadata and may expect a particular set of tensor names or weight order within tensors that is not always compatible with the source. Tokenizer data is stored as tensors inside the model file as well.