huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-21 12:20:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add the SmolLM2 models. (#2595)

Browse Source 
* Add the SmolLM2 models.

* More SmolLM2 support.
This commit is contained in:
Laurent Mazare
2024-11-03 17:11:12 +01:00
committed by GitHub
parent 530ab96036
commit 3fba2b5fc4
3 changed files with 73 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
candle-transformers/src/models/quantized_llama.rs
View File

@ -351,13 +351,16 @@ impl ModelWeights {
let (cos, sin) = precomput_freqs_cis(rope_dim, rope_freq_base, device)?;
let neg_inf = Tensor::new(f32::NEG_INFINITY, device)?;
let tok_embeddings = ct.tensor(reader, "token_embd.weight", device)?;
let tok_embeddings = tok_embeddings.dequantize(device)?;
let tok_embeddings_q = ct.tensor(reader, "token_embd.weight", device)?;
let tok_embeddings = tok_embeddings_q.dequantize(device)?;
let norm = RmsNorm::from_qtensor(
ct.tensor(reader, "output_norm.weight", device)?,
rms_norm_eps,
)?;
let output = ct.tensor(reader, "output.weight", device)?;
let output = match ct.tensor(reader, "output.weight", device) {
Ok(tensor) => tensor,
Err(_) => tok_embeddings_q,
};
let mut layers = Vec::with_capacity(block_count);
for layer_idx in 0..block_count {
let prefix = format!("blk.{layer_idx}");