mirror of
https://github.com/huggingface/candle.git
synced 2025-06-20 04:00:28 +00:00
Use BF16 for llama v3 by default.
This commit is contained in:
@ -120,7 +120,10 @@ fn main() -> Result<()> {
|
||||
Some("bf16") => DType::BF16,
|
||||
Some("f32") => DType::F32,
|
||||
Some(dtype) => bail!("Unsupported dtype {dtype}"),
|
||||
None => DType::F16,
|
||||
None => match args.which {
|
||||
Which::V3 | Which::V3Instruct => DType::BF16,
|
||||
Which::V1 | Which::V2 | Which::Solar10_7B | Which::TinyLlama1_1BChat => DType::F16,
|
||||
},
|
||||
};
|
||||
let (llama, tokenizer_filename, mut cache, config) = {
|
||||
let api = Api::new()?;
|
||||
|
Reference in New Issue
Block a user