pszemraj / perSLIMmon-8b-base

perSLIMmon-8b-base

persimmon-8b went to the vocab lipo clinic

A slimmed-down version of persimmon-8b-base which removes the ~70,000 unused entries in the model vocabulary and tokenizer (see the safetensors layer overview). Should be slightly faster.

Credit: fine-tune-fuyu ( scripts/surgery.py was adapted for persimmon)

inference

install required pkgs:

pip install -U transformers accelerate bitsandbytes sentencepiece

load in 4bit & run inference:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("pszemraj/perSLIMmon-8b-base")
model = AutoModelForCausalLM.from_pretrained(
    "pszemraj/perSLIMmon-8b-base",
    load_in_4bit=True, # GPU required
    torch_dtype="auto",
    device_map="auto",
)
inputs = tokenizer("The weather is always wonderful", return_tensors="pt").to(
    model.device
)
tokens = model.generate(
    **inputs,
    max_new_tokens=64,
    temperature=0.75,
    top_p=0.95,
    epsilon_cutoff=1e-5,
    repetition_penalty=1.05,
    renormalize_logits=True,
    do_sample=True,
) # adapt inference params as needed

print(tokenizer.decode(tokens[0], skip_special_tokens=True))

inference is decently fast on a colab T4:

CPU times: user 6.01 s, sys: 138 ms, total: 6.15 s
Wall time: 6.23 s