k-l-lambda_-_Llama-3.2-1B-vocab32k-exl2

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Llama-3.2-1B-vocab32k - EXL2

• Model creator: https://huggingface.co/k-l-lambda/
• Original model: https://huggingface.co/k-l-lambda/Llama-3.2-1B-vocab32k/

Available sizes

| Branch | Bits | Description | | ----- | ---- | ------- | ------ | ------ | ------ | ------ | ------------ | | 8_0 | 8.0 | Maximum quality that ExLlamaV2 can produce, near unquantized performance. | | 6_5 | 6.5 | Very similar to 8.0, good tradeoff of size vs performance, recommended. | | 5_0 | 5.0 | Slightly lower quality vs 6.5, but usable on 8GB cards. | | 4_25 | 4.25 | GPTQ equivalent bits per weight, slightly higher quality. | | 3_5 | 3.5 | Lower quality, only use if you have to. |

Download instructions

With git:

git clone --single-branch --branch 6_5 https://huggingface.co/k-l-lambda_-_Llama-3.2-1B-vocab32k-exl2 Llama-3.2-1B-vocab32k-6_5

With huggingface hub:

pip3 install huggingface-hub

To download a specific branch, use the --revision parameter. For example, to download the 6.5 bpw branch: Linux:

huggingface-cli download k-l-lambda_-_Llama-3.2-1B-vocab32k-exl2 --revision 6_5 --local-dir Llama-3.2-1B-vocab32k-6_5 --local-dir-use-symlinks False

Windows (which apparently doesn't like _ in folders sometimes?):

huggingface-cli download k-l-lambda_-_Llama-3.2-1B-vocab32k-exl2 --revision 6_5 --local-dir Llama-3.2-1B-vocab32k-6.5 --local-dir-use-symlinks False

Original model description:

license: mit language:

• en library_name: transformers pipeline_tag: text-generation

---

Model Information

This is a vocab pruned variant of Llama-3.2-1B-Instruct. The vocabulary size is pruned from 128256 to 32256. The total parameter size is: 1,039,214,756, ~200M parameters pruned from origin.

How to use

This is a code example:

import torch
from transformers import pipeline


pipe = pipeline(
    "text-generation",
    model='k-l-lambda/Llama-3.2-1B-vocab32k',
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]
outputs = pipe(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Another:

from transformers import AutoModelForCausalLM, AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("k-l-lambda/Llama-3.2-1B-vocab32k")
model = AutoModelForCausalLM.from_pretrained("k-l-lambda/Llama-3.2-1B-vocab32k")

input_ids = tokenizer.encode("Hello, ", return_tensors="pt")
output = model.generate(input_ids)
print(tokenizer.decode(output[0]))

Tokens conversion

You can map an ID value in 32k vocab to the ID value in original 128k vocab, by the tensor in token_indices.pt and inv_token_indices.pt.

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer


tokenizer128k = AutoTokenizer.from_pretrained('meta-llama/Llama-3.2-1B-Instruct')
tokenizer32k = AutoTokenizer.from_pretrained('k-l-lambda/Llama-3.2-1B-vocab32k')

indices_path = hf_hub_download(repo_id='k-l-lambda/Llama-3.2-1B-vocab32k', filename='token_indices.pt')
inv_indices_path = hf_hub_download(repo_id='k-l-lambda/Llama-3.2-1B-vocab32k', filename='inv_token_indices.pt')
token_indices = torch.load(indices_path)
inv_token_indices = torch.load(inv_indices_path)

ids_32k = tokenizer32k.encode('This is an example sentence.')
ids_128k = [token_indices[id].item() for id in ids_32k]
print(f'{ids_32k=}')
print(f'{ids_128k=}')

print(tokenizer128k.decode(ids_128k))


ids_128k = tokenizer128k.encode('This is another example sentence.')
ids_32k = [inv_token_indices[id].item() for id in ids_128k]
print(f'{ids_128k=}')
print(f'{ids_32k=}')	# non-exist tokens in 32k vocab will map to -1

print(tokenizer32k.decode(ids_32k))