clebert / llama2.zig Goto Github PK

Model: https://huggingface.co/georgesung/llama2_7b_chat_uncensored

See:

• karpathy/llama2.c#277
• karpathy/llama2.c#298
• karpathy/llama2.c#312
• karpathy/llama2.c#364
• ggerganov/llama.cpp#397
• https://arxiv.org/pdf/2101.01321v3.pdf

Running with just checkpoint and -t 0 procudes

❯ zig-out/bin/llama2 stories15M.bin -t 0
Error: To control the diversity of samples use either the temperat
ure or the top-p value, but not both.

Usage: llama2 <checkpoint_path> [options]

Options:
  -t <float>  temperature  = 1
  -p <float>  top_p        = 0.9; 0 == off
  -s <int>    random_seed  = milli_timestamp
  -n <int>    n_steps      = 256; 0 == max_seq_len
  -i <string> input_prompt = ""

Example: llama2 model.bin -i "Once upon a time"

Also I dont think there is anything wrong with using temperature and top p together.

I think the vector widths you are using are too large. 32x4 is 128 bytes or 1024 bits which is much larger than the width of any vector type that is on cpus today.

std.simd.suggestVectorSize(f32) will return the current cpus maximum vector size for a given type. I think seeing an improvement by making it larger means something else is going on that is not related.

Also there is performance to be had from reworking your summing by using a vector of sums. Currently the CPU struggles to optimize because of a data dependency as each sum needs to depend on the previous lanes value. If you instead keep N sums and then reduce them at the end you should see a speedup.

var value: @Vector(v_len, f32) = @splat(0.0);
... loop
value += a * b;
...
xoutptr.* = @reduce(.Add, value);