Skip to content

llama.cpp clone with additional SOTA quants and improved CPU performance

License

Notifications You must be signed in to change notification settings

ikawrakow/ik_llama.cpp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

llama.cpp clone with better CPU performance

License: MIT

TL;DR

This repository is a clone of llama.cpp with the following improvements

  • Better implementation of CPU matrix multiplications (AVX2 and ARM_NEON) for fp16/fp32 and all k-, i-, and legacy llama.cpp quants, that leads to a significant improvement in prompt processing (PP) speed, typically in the range of 2X, but up to 4X for some quantization types. Token generation (TG) also benefits, but to a lesser extent due to TG being memory bound
  • Faster CPU inference for MoE models with similar performance gains
  • Implementation of the Bitnet b1.58 model for the CPU (AVX2 and ARM_NEON) and GPU (CUDA and Metal). This implementation is much faster than the unmerged llama.cpp PR-8151

If you are not already familiar with llama.cpp, it is better to start there. For those familiar with llama.cpp, everything here works the same as in llama.cpp (or at least the way llama.cpp worked when I last synced on Aug 12 2024).

Note that I have published some, but not all, of the code in this repository in a series of llamafile PRs (394, 405, 428, 435, 453, and 464)

The implementation of matrix-matrix and matrix-vector multiplications is in a single C++ source file (iqk_mul_mat.cpp) with just two interface functions iqk_mul_mat (fp16/fp32 and quantized matrix multiplications) and iqk_mul_mat_moe (as iqk_mul_mat but meant to be used for the FFN part of a MoE model). Under the hood iqk_mul_mat_moe uses the same implementation as iqk_mul_mat, with the only difference being where results are stored in memory. Bitnet quantization related stuff is in iqk-quantize.cpp.

Why?

Mostly out of curiosity:

  • Justine Tunney's tinyBLAS, which she contributed to llama.cpp in PR 6414, only works for Q4_0, Q8_0 and fp16/bf16 models. In the surrounding discussion about possibly extending tinyBLAS to k- and i-quants, she felt that k-quants are not amenable to block-tiling, which is required to improve performance. This statement piqued my curiosity, so here we are.
  • Bitnet-1.58b has been one of the most discussed topics in the llama.cpp project, so eventually I decided to see how efficiently one can implement a ternary model

Curiosity aside, improved CPU performance may be (or may become) important in practice. According to The Register, 70% of AI inference is done on the CPU of mobile phones, at least in the Android world (but I haven't come around to actually comparing performance on a phone). With ever increasing number of LLM model parameters, and with Meta's 400B model just released, the CPU may become the only viable option for people not willing (or not able to) rent/buy uber expensive GPU instances capable of running such models. Granted, one would need a pretty beefy computer to run a 400B model, and inference speed will be sluggish, but at least one will not need to spend the equivalent of a luxury apartment in the downtown of the city where I live to buy the GPU system capable of running the model.

Performance comparison to llama.cpp

The results in the following tables are obtained with these parameters:

  • Model is LLaMA-v3-8B for AVX2 and LLaMA-v2-7B for ARM_NEON
  • The AVX2 CPU is a 16-core Ryzen-7950X
  • The ARM_NEON CPU is M2-Max
  • tinyBLAS is enabled in llama.cpp
  • llama.cpp results are for build: 081fe431 (3441), which was the current llama.cpp master branch when I pulled on July 23 2024.
  • The projects are built without CUDA support, no BLAS, and Accelerate framework disabled

Prompt processing

Here I set the number of threads to be equal to the number of (performance) cores of the CPU, so 16 threads for the Ryzen-7950X and 8 threads for the M2-Max. The following table summarizes the results. To not make the table too long, I have listed only quantized models containing predominantly one quantization type (i.e., excluded the QX_K - Medium/Large variants, which are typically a mix of QX_K and Q(X+1)_K, as well as IQ2_S and IQ3_XS).

The command line to generate the benchmark data is

./bin/llama-bench -m $model -p 512 -n 0 -t $num_threads -ngl 0
Quantization size backend threads t/s (llama.cpp) t/s (iqk_mul_mat) Speedup
8B F16 14.96 GiB AVX2 16 112.37 ± 0.40 131.27 ± 0.38 1.168
7B F16 12.55 GiB NEON 8 90.28 ± 1.25 95.34 ± 0.15 1.056
8B Q8_0 7.95 GiB AVX2 16 118.07 ± 0.53 134.00 ± 0.47 1.135
7B Q8_0 6.67 GiB NEON 8 77.25 ± 1.81 94.14 ± 1.15 1.219
8B Q4_0 4.35 GiB AVX2 16 104.46 ± 0.33 130.20 ± 0.29 1.246
7B Q4_0 3.57 GiB NEON 8 65.46 ± 0.79 76.22 ± 0.71 1.164
8B Q4_1 4.77 GiB AVX2 16 57.83 ± 0.24 160.69 ± 0.49 2.779
7B Q4_1 3.95 GiB NEON 8 37.40 ± 0.50 65.83 ± 0.98 1.760
8B Q5_0 5.22 GiB AVX2 16 53.50 ± 0.35 122.62 ± 0.48 2.292
7B Q5_0 4.34 GiB NEON 8 29.31 ± 0.51 67.51 ± 1.17 2.303
8B Q5_1 5.64 GiB AVX2 16 50.85 ± 0.36 147.15 ± 0.47 2.894
7B Q5_1 4.72 GiB NEON 8 26.02 ± 0.37 58.49 ± 0.85 2.248
8B Q2_K_S 2.78 GiB AVX2 16 110.11 ± 0.28 192.47 ± 1.35 1.748
7B Q2_K_S 2.16 GiB NEON 8 35.44 ± 0.06 77.93 ± 1.64 2.199
8B Q3_K_S 3.41 GiB AVX2 16 77.42 ± 0.36 181.64 ± 0.44 2.346
7B Q3_K_S 2.75 GiB NEON 8 26.79 ± 0.03 59.38 ± 1.08 2.216
8B Q4_K_S 4.36 GiB AVX2 16 98.92 ± 0.34 185.35 ± 0.39 1.874
7B Q4_K_S 3.59 GiB NEON 8 46.55 ± 0.67 76.31 ± 0.38 1.639
8B Q5_K_S 5.21 GiB AVX2 16 69.44 ± 0.31 179.62 ± 0.69 2.587
7B Q5_K_S 4.33 GiB NEON 8 30.18 ± 0.23 65.34 ± 0.79 2.165
8B Q6_K 6.14 GiB AVX2 16 74.89 ± 0.26 181.86 ± 0.55 2.428
7B Q6_K 5.15 GiB NEON 8 28.12 ± 1.24 60.75 ± 1.15 2.160
8B IQ2_XXS 2.23 GiB AVX2 16 42.57 ± 0.16 126.63 ± 0.55 2.975
7B IQ2_XXS 1.73 GiB NEON 8 20.87 ± 0.20 64.29 ± 1.12 3.080
8B IQ2_XS 2.42 GiB AVX2 16 46.45 ± 0.27 125.46 ± 0.43 2.701
7B IQ2_XS 1.89 GiB NEON 8 22.77 ± 0.21 51.15 ± 0.24 2.246
8B IQ2_M 2.74 GiB AVX2 16 40.76 ± 0.18 113.07 ± 0.48 2.774
7B IQ2_M 2.20 GiB NEON 8 14.95 ± 0.26 44.87 ± 0.50 3.001
8B IQ3_XXS 3.04 GiB AVX2 16 31.95 ± 0.20 109.86 ± 0.45 3.438
7B IQ3_XXS 2.41 GiB NEON 8 14.40 ± 0.10 53.58 ± 0.85 3.721
8B IQ3_S 3.42 GiB AVX2 16 28.04 ± 0.08 96.28 ± 0.45 3.434
7B IQ3_S 2.75 GiB NEON 8 12.08 ± 0.30 49.72 ± 0.06 4.116
8B IQ4_XS 4.13 GiB AVX2 16 68.98 ± 0.31 180.34 ± 0.55 2.614
7B IQ4_XS 3.37 GiB NEON 8 40.67 ± 1.97 75.11 ± 1.97 1.847
8B IQ4_NL 4.35 GiB AVX2 16 59.94 ± 0.21 129.06 ± 0.43 2.153
7B IQ4_NL 3.56 GiB NEON 8 34.36 ± 0.81 76.02 ± 1.36 2.212

We see that llama.cpp achieves respectable performance for fp16, Q8_0, and Q4_0, being only up to 25% slower than this implementation. This is thanks to the use of Justine Tunney's tinyBLAS, which is utilized for these quantization types. For all other quants we observe performance gains in the 1.75X - 4X range, which is not a small feat considering that the ggml matrix multiplication functions has been rewritten several times since llama.cpp was first published. Performance gains are larger for i-quants due to the higher quant unpacking cost (see discussion in "To tile or not to tile")

Token generation

On the Ryzen-7950X TG is memory bound, and for many quantization types peak performance is achieved at just 4 threads. Hence, only results for 2 and 4 threads are shown for AVX2. The M2-Max has a much more capable memory subsystem and as a result performance keep increasing up to 8 threads. Thus, results are given for up to 8 threads for ARM_NEON.

The command line to generate the data was

./bin/llama-bench -m $model -p 0 -n 128 -t $num_threads -ngl 0
Quantization size backend threads t/s (llama.cpp) t/s (iqk_mul_mat) Speedup
8B F16 14.96 GiB AVX2 1 2.20 ± 0.00 2.25 ± 0.00 1.023
2 3.63 ± 0.00 3.68 ± 0.00 1.014
4 4.20 ± 0.00 4.20 ± 0.00 1.000
7B F16 12.55 GiB NEON 2 6.94 ± 0.27 7.40 ± 0.01 1.066
4 8.73 ± 0.01 8.83 ± 0.01 1.011
6 9.05 ± 0.02 9.05 ± 0.01 1.000
8B Q8_0 7.95 GiB AVX2 2 5.03 ± 0.00 7.87 ± 0.00 1.565
4 7.40 ± 0.00 7.82 ± 0.00 1.057
7B Q8_0 6.67 GiB NEON 2 8.29 ± 0.44 12.07 ± 0.10 1.456
4 13.53 ± 0.03 15.77 ± 0.08 1.166
8 16.24 ± 0.10 16.94 ± 0.04 1.043
8B Q4_0 4.35 GiB AVX2 2 6.36 ± 0.00 10.28 ± 0.00 1.616
4 10.97 ± 0.06 13.55 ± 0.07 1.235
7B Q4_0 3.57 GiB NEON 2 9.77 ± 0.02 13.69 ± 0.03 1.401
4 17.82 ± 0.06 23.98 ± 0.11 1.346
8 26.63 ± 0.41 29.86 ± 0.04 1.121
8B Q4_1 4.77 GiB AVX2 2 5.11 ± 0.00 11.45 ± 0.00 2.241
4 9.08 ± 0.02 12.58 ± 0.00 1.385
7B Q4_1 3.95 GiB NEON 2 9.11 ± 0.06 14.62 ± 0.04 1.605
4 17.04 ± 0.09 24.08 ± 0.28 1.413
8 25.26 ± 0.24 27.23 ± 0.14 1.078
8B Q5_0 5.22 GiB AVX2 2 5.31 ± 0.01 8.30 ± 0.01 1.563
4 9.40 ± 0.01 11.47 ± 0.00 1.220
7B Q5_0 4.34 GiB NEON 2 7.26 ± 0.06 7.52 ± 0.00 1.036
4 13.63 ± 0.18 14.16 ± 0.10 1.039
8 22.55 ± 0.35 24.34 ± 0.22 1.079
8B Q5_1 5.64 GiB AVX2 2 4.52 ± 0.00 8.86 ± 0.00 1.960
4 7.72 ± 0.05 10.68 ± 0.03 1.383
7B Q5_1 4.72 GiB NEON 2 6.51 ± 0.01 6.42 ± 0.03 0.986
4 12.26 ± 0.18 12.21 ± 0.14 0.996
8 20.33 ± 0.52 21.85 ± 0.22 1.075
8B Q2_K_S 2.78 GiB AVX2 2 11.30 ± 0.00 13.06 ± 0.01 1.156
4 18.70 ± 0.00 19.04 ± 0.65 1.014
7B Q2_K_S 2.16 GiB NEON 2 8.42 ± 0.05 11.97 ± 0.10 1.422
4 15.74 ± 0.01 22.09 ± 0.08 1.403
8 27.35 ± 0.05 38.32 ± 0.05 1.401
8B Q3_K_S 3.41 GiB AVX2 2 8.58 ± 0.00 10.82 ± 0.00 1.261
4 15.26 ± 0.01 16.25 ± 0.01 1.065
7B Q3_K_S 2.75 GiB NEON 2 6.40 ± 0.02 9.12 ± 0.09 1.425
4 12.17 ± 0.00 17.11 ± 0.03 1.406
8 22.04 ± 0.08 31.39 ± 0.31 1.424
8B Q4_K_S 4.36 GiB AVX2 2 9.61 ± 0.00 10.72 ± 0.01 1.116
4 13.24 ± 0.31 13.28 ± 0.01 1.003
7B Q4_K_S 3.59 GiB NEON 2 11.15 ± 0.05 12.93 ± 0.09 1.160
4 20.24 ± 0.16 23.49 ± 0.29 1.161
8 25.76 ± 0.07 28.31 ± 0.22 1.099
8B Q5_K_S 5.21 GiB AVX2 2 7.45 ± 0.00 9.73 ± 0.00 1.306
4 11.05 ± 0.33 11.43 ± 0.02 1.034
7B Q5_K_S 4.33 GiB NEON 2 7.20 ± 0.04 8.81 ± 0.04 1.224
4 13.62 ± 0.15 16.81 ± 0.16 1.234
8 20.56 ± 0.19 23.96 ± 0.14 1.165
8B Q6_K 6.14 GiB AVX2 2 7.53 ± 0.00 9.42 ± 0.00 1.251
4 9.74 ± 0.00 9.97 ± 0.01 1.024
7B Q6_K 5.15 GiB NEON 2 6.85 ± 0.04 8.30 ± 0.06 1.212
4 13.03 ± 0.05 15.47 ± 0.17 1.187
8 18.52 ± 0.07 20.67 ± 0.08 1.116
8B IQ2_XXS 2.23 GiB AVX2 2 5.33 ± 0.01 6.40 ± 0.00 1.201
4 10.06 ± 0.03 11.76 ± 0.03 1.169
7B IQ2_XXS 1.73 GiB NEON 2 5.07 ± 0.04 5.22 ± 0.05 1.030
4 9.63 ± 0.00 9.91 ± 0.07 1.029
8 17.40 ± 0.50 18.65 ± 0.22 1.072
8B IQ2_XS 2.42 GiB AVX2 2 5.83 ± 0.00 6.55 ± 0.00 1.123
4 10.88 ± 0.09 12.07 ± 0.07 1.109
7B IQ2_XS 1.89 GiB NEON 2 5.52 ± 0.01 5.60 ± 0.00 1.014
4 10.50 ± 0.01 11.15 ± 0.00 1.062
8 18.19 ± 1.30 20.94 ± 0.19 1.151
8B IQ2_M 2.74 GiB AVX2 2 5.12 ± 0.01 5.17 ± 0.00 1.010
4 9.60 ± 0.28 9.68 ± 0.16 1.008
7B IQ2_M 2.20 GiB NEON 2 3.73 ± 0.02 4.53 ± 0.00 1.214
4 7.14 ± 0.05 8.70 ± 0.06 1.218
8 11.99 ± 0.48 16.41 ± 0.05 1.369
8B IQ3_XXS 3.04 GiB AVX2 2 4.06 ± 0.01 5.00 ± 0.00 1.232
4 7.75 ± 0.02 9.13 ± 0.45 1.178
7B IQ3_XXS 2.41 GiB NEON 2 3.53 ± 0.00 3.82 ± 0.00 1.082
4 6.74 ± 0.04 7.42 ± 0.07 1.103
8 11.96 ± 0.40 13.19 ± 0.29 1.103
8B IQ3_S 3.42 GiB AVX2 2 3.62 ± 0.00 4.06 ± 0.00 1.122
4 6.80 ± 0.01 7.62 ± 0.10 1.121
7B IQ3_S 2.75 GiB NEON 2 2.96 ± 0.01 3.21 ± 0.03 1.084
4 5.68 ± 0.01 6.25 ± 0.05 1.100
8 10.32 ± 0.25 11.11 ± 0.37 1.077
8B IQ4_XS 4.13 GiB AVX2 2 8.08 ± 0.00 11.35 ± 0.00 1.405
4 13.36 ± 0.72 14.32 ± 0.24 1.072
7B IQ4_XS 3.37 GiB NEON 2 9.87 ± 0.03 12.06 ± 0.00 1.222
4 17.78 ± 0.23 22.06 ± 0.28 1.241
8 27.62 ± 0.09 29.70 ± 0.39 1.075
8B IQ4_NL 4.35 GiB AVX2 2 5.52 ± 0.00 10.26 ± 0.00 1.859
4 10.78 ± 0.01 13.69 ± 0.08 1.270
7B IQ4_NL 3.56 GiB NEON 2 8.32 ± 0.01 13.54 ± 0.01 1.627
4 15.89 ± 0.00 24.28 ± 0.29 1.528
8 26.56 ± 0.36 29.87 ± 0.08 1.125

Here gains are generally lower compared to PP due to TG performance being limited by memory bandwidth. Nevertheless, for some quants/architectures/threads the speedup is quite remarkable (e.g., almost a factor of 2 for Q5_1 on AVX2 with 2 threads).

MoE models

There is PR-6840 from Justine Tunney in llama.cpp, but it has not been merged since April 23, so I'll compare performance to the master branch for Mixtral-8x7B. As Mixtral8x7B quantization is quite a lengthy process, the following table shows data only for Q4_K_S (a commonly used k-quant, 4 bit), Q5_0 (a legacy quant, 5 bit), and IQ4_XXS (a 3-bit i-quant)

model size backend threads test t/s (llama.cpp) t/s (iqk_mul_mat) Speedup
8x7B Q4_K_S 48.75 GiB AVX2 16 pp512 54.92 ± 0.23 102.94 ± 0.37 1.874
NEON 8 pp512 23.54 ± 1.56 38.32 ± 0.54 1.628
AVX2 4 tg128 7.80 ± 0.07 7.83 ± 0.09 1.004
NEON 8 tg128 14.95 ± 0.25 15.28 ± 0.24 2.022
8x7B IQ3_XXS 33.07 GiB AVX2 16 pp512 17.58 ± 0.04 68.45 ± 0.22 3.894
NEON 8 pp512 7.75 ± 0.04 34.67 ± 0.40 4.474
AVX2 4 tg128 4.60 ± 0.01 5.45 ± 0.09 1.185
AVX2 8 tg128 8.04 ± 0.65 9.83 ± 0.06 1.223
AVX2 16 tg128 10.42 ± 0.01 10.57 ± 0.01 1.014
NEON 8 tg128 6.19 ± 1.16 7.27 ± 0.14 1.174
8x7B Q5_0 59.11 GiB AVX2 16 pp512 29.06 ± 0.43 62.67 ± 0.32 2.157
NEON 8 pp512 15.17 ± 0.51 27.36 ± 1.03 1.804
AVX2 4 tg128 5.44 ± 0.10 6.81 ± 0.06 1.252
NEON 8 tg128 12.03 ± 0.77 12.41 ± 1.27 1.032

Bitnet-1.58B

Two implementations are provided

  • IQ1_BN - uses 1.625 bits-per-weight (bpw)
  • IQ2_BN - uses 2.0 bpw

IQ2_BN is faster for PP (CPU and GPU, although the PP performance difference on CUDA is very minor). IQ1_BN can arrive at a higher TG performance on the Ryzen-7950X (given enough threads) because of the smaller model size, but it is always slower on the GPU and on the M2-Max CPU.

There is the unmerged PR 8151 in llama.cpp that implements Bitnet-1.58B for the CPU (AVX and ARM_NEON, no GPU implementation). The following table compares performance between this repo and PR-8151 in llama.cpp. The CUDA results were obtained on an RTX-4080, the Metal results on a 30-core M2-Max GPU.

model size backend threads test t/s (llama.cpp) t/s (this repo) Speedup
3B - IQ1_BN 729.64 MiB AVX2 16 pp512 120.61 ± 0.48 423.19 ± 1.28 3.509
NEON 8 pp512 46.64 ± 0.02 205.90 ± 0.88 4.415
CUDA 8 pp512 - 10660 ± 170 -
Metal 8 pp512 - 698.25 ± 1.91 -
AVX2 2 tg128 15.79 ± 0.01 22.13 ± 0.02 1.402
AVX2 4 tg128 28.64 ± 1.72 40.14 ± 0.04 1.402
AVX2 8 tg128 48.91 ± 0.08 61.79 ± 0.09 1.263
AVX2 16 tg128 57.73 ± 0.05 60.79 ± 0.05 1.053
NEON 2 tg128 11.43 ± 0.04 16.87 ± 0.02 1.476
NEON 4 tg128 21.11 ± 0.05 30.66 ± 0.11 1.452
NEON 8 tg128 37.36 ± 0.07 55.21 ± 0.16 1.478
CUDA 8 tg128 - 301.44 ± 0.12 -
Metal 8 tg128 - 76.70 ± 0.07 -
3B - IQ2_BN 873.65 MiB AVX2 16 pp512 151.39 ± 0.35 540.82 ± 2.48 3.572
NEON 8 pp512 46.54 ± 0.03 242.05 ± 0.34 5.201
CUDA 8 pp512 - 10800 ± 160 -
Metal 8 pp512 - 723.19 ± 0.53 -
AVX2 2 tg128 18.93 ± 0.02 38.34 ± 0.08 2.026
AVX2 4 tg128 34.54 ± 0.06 56.29 ± 0.07 1.630
AVX2 8 tg128 52.97 ± 0.07 53.44 ± 0.08 1.009
AVX2 16 tg128 51.84 ± 0.25 53.46 ± 0.07 1.031
NEON 2 tg128 11.40 ± 0.02 32.01 ± 0.27 2.808
NEON 4 tg128 20.99 ± 0.00 56.45 ± 0.11 2.689
NEON 8 tg128 37.28 ± 0.08 89.77 ± 0.70 2.408
CUDA 8 tg128 - 322.10 ± 0.07 -
Metal 8 tg128 - 110.39 ± 0.13 -

We can make the following observations:

  • For prompt processing this Bitnet-1.58b implementation is massively better than PR-8151 in llama.cpp, with gains between 3.4X and 5.2X!
  • We get PP-512 = 520 t/s for the 2.0 bpw variant on the Ryzen-7950X, which costs less than $500. Hey, who needs a GPU?
  • For low number of threads (2), this implementation is also much faster than PR-8151 for TG, where speed gains are between 1.4X and 2.8X. As we become memory bound on the Ryzen-7950X, the speed advantage goes away there for sufficiently high number of threads. But on the M2-Max this implementation is 1.4X (1.625 bpw) or 2.4X faster even at 8 threads
  • Looking at TG on the M2-Max, the GPU looks a bit like wasted silicon (90 vs 110 t/s for TG-128 and the 2.0 bpw variant). If the GPU transistors had been spent to double the M2 number of CPU cores (and all memory bandwidth is given to the CPU), the CPU would be wiping the floor with the GPU.
  • I'm of course kidding with the above. Still, it seems there are massive inefficiencies in the llama.cpp Metal implementation that start showing up when matrix multiplications become very fast as is the case here. The difference between CPU and GPU prompt processing speed is typically at least a factor of 7 in favor of the GPU on the M2-Max, but it is only around a factor of 3 here.
  • It is worth noting that one needs to offload the token embeddings tensor to the GPU, else performance on CUDA/Metal is significantly lower. Bitnet uses the same tensor for token embeddings and for output. Mainline llama.cpp currently puts the token embeddings tensor on the CPU, and this results in running the matrix multiplication with the output tensor on the CPU. This most likely affects other models as well (e.g., Gemma), but I haven't yet looked into this.

To reproduce these results:

  • Clone https://huggingface.co/1bitLLM/bitnet_b1_58-3B
  • Run python3 --outtype f16 path_to_bitnet to convert to GGUF
  • Run ./bin/llama-quantize path_to_bitnet/ggml-model-f16.gguf quantized.gguf [iq1_bn | iq2_bn]. Note: no imatrix is required (and, if you provide one, it is ignored)
  • Caveat: only the 3B Bitnet variant works. The smaller Bitnet models contain tensors with number of columns that are not even a multiple of 32, so basically no llama.cpp quant will work for these.

To tile or not to tile

The common wisdom for efficient matrix multiplications is to use block tiling, and this is also used here for fp16/fp32 matrices. But block tiling does not somehow magically reduce the amount of computation that needs to get done. Performance gains are simply due to the better utilization of memory caches. When dealing with quantized matrix multiplications, there is an additional factor that comes into play: the quantized data needs to be unpacked to 8-bit integers before being used in the matrix multiplication multiply-add operations. Depending on quantization type, this unpacking can represent a significant fraction of the overall computation cost. Hence, for best performance, one would want to reuse the unpacked quants as much as possible, thus spending some fraction of the available vector registers to hold the unpacked data. But when using block tiling, one also needs a certain number of vector registers for accumulating results. For instance, on AVX2 (16 vector registers available), for fp16/fp32 models best performance is achieved with 2 x 6 tiles (where the 2 refers to rows in the left matrix and is measured in units of the vector register size, so 16/8 floats for fp16/fp32, and 6 is for the number of columns in the right matrix). Unpacking quantized data works best when done in blocks of 128 or 256 quants so that, if we wanted to keep unpacked quants for 2 rows, we would need at least 8 vector registers, thus being left with less than 8 registers for result accumulation, so at best 2 x 3 tiles. In practice one needs addition vector registers for various constants that are typically needed for de-quantization, so that, at the end, it becomes better to use 1 x N "tiles", i.e., a row-wise multiplication where each row in the left matrix is multiplied with N columns in the right matrix, thus reusing the unpacked data N times. This (i.e., amortizing de-quantization cost) is the main mechanism for seeding up quantized matrix multiplications. Having started with quantized matrices, and having gone from tiles to a row-wise implementation after some experimentation, I did try row-wise multiplication for float matrices first. Performance was not quite as good as for block-tiling, but I did get up to 90-95% of the speed of tinyBLAS that way before switching the fp16/fp32 implementation to 2 x 6 (AVX2) or 5 x 5 (AVX512 and ARM_NEON) block-tiles. But even for for Q8_0 x Q8_0 multiplications, where there is basically no de-quantization cost, row-wise multiplication is faster than tiling (and hence this implemeintation beats tinyBLAS, which uses block-tiling also for Q8_0).

About

llama.cpp clone with additional SOTA quants and improved CPU performance

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published