/calm

CUDA/Metal accelerated language model inference

Primary LanguageCMIT LicenseMIT

😌 calm

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.

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

Running

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

Supported models

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:

Architecture Models
Llama Llama2 7B, Llama2 13B, Llama3 8B
Llama-like TinyLlama 1.1B, Cosmo 1B, LLaMA Pro 8B, H2O Danube 1.8B, DeepSeekMath 7B, LargeWorldModel 7B 1M, Xverse 7B, LLM360 K2
Yi Yi 1.5 6B, Yi 1.5 9B, Yi 1.5 34B
Mistral Mistral 7B, Mistral Nemo 12B, Codestral 22B, Mistral Pro 8B, SOLAR 10.7B, GritLM 7B, Starling 7B
Qwen2 Qwen1.5 0.5B, Qwen1.5 1.8B, Qwen1.5 4B, Qwen1.5 7B, Qwen1.5 14B, Qwen2 0.5B, Qwen2 1.5B, Qwen2 7B
Mixtral Mixtral 8x7B, Mixtral 8x22B, GritLM 8x7B
OLMo OLMo 1B, OLMo 7B, OLMo 1.7 7B
Gemma Gemma 2B, Gemma 7B (note: 7B version has issues with fp8 quantization)
MiniCPM MiniCPM 2B, MiniCPM 2B 128K, MiniCPM MoE 8x2B
Cohere Command-R, Aya 23 8B, Aya 23 35B
InternLM InternLM2-1.8B, InternLM2-7B, InternLM2-20B
DBRX DBRX 132B
Phi3 Phi3 Mini 3.8B, Phi3 Medium 14B

Supported formats

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.

Model files

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.

Performance

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.

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.

NVidia

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):

Model (context) Performance (first 32) Performance (last 32)
Llama3 8B (4096), fp16 61 tok/s (923 GB/s) 59 tok/s (919 GB/s)
Llama3 8B (4096), fp8 120 tok/s (903 GB/s) 110 tok/s (889 GB/s)
Llama3 8B (4096), gf4 225 tok/s (846 GB/s) 194 tok/s (830 GB/s)
Llama2 7B (4096), fp16 69 tok/s (919 GB/s) 60 tok/s (921 GB/s)
Llama2 7B (4096), fp8 135 tok/s (893 GB/s) 103 tok/s (899 GB/s)
Llama2 7B (4096), gf4 246 tok/s (815 GB/s) 158 tok/s (857 GB/s)
Llama2 13B (4096), fp8 70 tok/s (910 GB/s) 56 tok/s (907 GB/s)
Llama2 13B (4096), gf4 131 tok/s (848 GB/s) 88 tok/s (863 GB/s)
Mistral 7B (4096), fp16 65 tok/s (925 GB/s) 62 tok/s (916 GB/s)
Mistral 7B (4096), fp8 127 tok/s (902 GB/s) 116 tok/s (888 GB/s)
Mistral 7B (4096), gf4 237 tok/s (843 GB/s) 203 tok/s (832 GB/s)
Mixtral 8x7B (4096), gf4 137 tok/s (875 GB/s) 125 tok/s (862 GB/s)
Mixtral 8x7B (16384), gf4 137 tok/s (879 GB/s) 105 tok/s (781 GB/s)
Yi 34B (4096), gf4 52 tok/s (884 GB/s) 47 tok/s (851 GB/s)

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.

Apple

When using Apple Silicon (Metal), calm gets the following performance; 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):

Chip Model (context) Performance (first 32) Performance (last 32)
M2 (100 GB/s) Llama3 8B (4096), fp8 12 tok/s (90 GB/s) 11 tok/s (89 GB/s)
M2 (100 GB/s) Llama3 8B (4096), gf4 23 tok/s (89 GB/s) 20 tok/s (85 GB/s)
M2 Pro (200 GB/s) Llama3 8B (4096), fp8 24 tok/s (180 GB/s) 21 tok/s (172 GB/s)
M2 Pro (200 GB/s) Llama3 8B (4096), gf4 45 tok/s (169 GB/s) 36 tok/s (157 GB/s)
M1 Max (400 GB/s) Llama3 8B (4096), fp8 44 tok/s (332 GB/s) 38 tok/s (306 GB/s)
M1 Max (400 GB/s) Llama3 8B (4096), gf4 73 tok/s (274 GB/s) 58 tok/s (248 GB/s)

Note: on higher end chips calm currently doesn't reach peak performance; some of this is due to limitations of the chips in other areas, and some is due to the author not having hardware access to the high end models to profile and optimize for. Hardware donations are welcome ;)

Footnotes

  1. CUDA runtime and compiler is used for GPU acceleration, but no CUDA or C libraries are used. Python conversion scripts use safetensors and torch, see tools/requirements.txt.

  2. Linux is the main supported OS at the moment; calm also works on macOS (on CPU) and has experimental Metal support.

  3. Based on testing a specific Gigabyte GeForce RTX 4090 where both individual kernels from this repository and cuBLAS peak at about ~955 GB/s.