huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 10:38:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

Softmax numerical stability. (#267)

Browse Source 
* Softmax numerical stability.

* Fix the flash-attn test.
This commit is contained in:
Laurent Mazare
2023-07-28 13:13:01 +01:00
committed by GitHub
parent 68eab38de6
commit 3eb2bc6d07
28 changed files with 117 additions and 188 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
candle-wasm-examples/llama2-c/src/model.rs
View File

@ -158,7 +158,7 @@ impl CausalSelfAttention {
let att = (q.matmul(&k.t()?)? / (self.head_dim as f64).sqrt())?;
let mask = self.cache.mask(seq_len)?.broadcast_as(att.shape())?;
let att = masked_fill(&att, &mask, f32::NEG_INFINITY)?;
let att = att.softmax(D::Minus1)?;
let att = candle_nn::ops::softmax(&att, D::Minus1)?;
// Convert to contiguous as matmul doesn't support strided vs for now.
let y = att.matmul(&v.contiguous()?)?;
let y = y.transpose(1, 2)?.reshape(&[b_sz, seq_len, n_embd])?;

4
candle-wasm-examples/llama2-c/src/worker.rs
View File

@ -1,7 +1,7 @@
use crate::model::{Cache, Config, Llama};
use byteorder::{LittleEndian, ReadBytesExt};
use candle::{DType, Device, IndexOp, Result, Shape, Tensor, D};
use candle_nn::VarBuilder;
use candle_nn::{ops::softmax, VarBuilder};
use rand::{distributions::Distribution, SeedableRng};
use serde::{Deserialize, Serialize};
use wasm_bindgen::prelude::*;
@ -88,7 +88,7 @@ impl LogitsProcessor {
pub fn sample(&mut self, logits: &Tensor) -> Result<u32> {
let logits = logits.to_dtype(DType::F32)?;
let next_token = if let Some(temperature) = self.temperature {
let prs = (&logits / temperature)?.softmax(D::Minus1)?;
let prs = softmax(&(&logits / temperature)?, D::Minus1)?;
let prs: Vec<f32> = prs.to_vec1()?;
let distr =
rand::distributions::WeightedIndex::new(prs).map_err(candle::Error::wrap)?;

2
candle-wasm-examples/whisper/src/model.rs
View File

@ -200,7 +200,7 @@ impl MultiHeadAttention {
}
let w = {
let _timer = crate::Timer::new("qk::softmax");
qk.softmax(candle::D::Minus1)?
candle_nn::ops::softmax(&qk, candle::D::Minus1)?
};
let wv = {
let _timer = crate::Timer::new("wv::matmul");

11
candle-wasm-examples/whisper/src/worker.rs
View File

@ -1,7 +1,7 @@
use crate::model::{Config, Whisper};
use anyhow::Error as E;
use candle::{safetensors::Load, DType, Device, Tensor};
use candle_nn::VarBuilder;
use candle_nn::{ops::softmax, VarBuilder};
use rand::{distributions::Distribution, rngs::StdRng, SeedableRng};
use serde::{Deserialize, Serialize};
use tokenizers::Tokenizer;
@ -127,9 +127,7 @@ impl Decoder {
// Extract the no speech probability on the first iteration by looking at the first
// token logits and the probability for the according token.
if i == 0 {
no_speech_prob = logits
.get(0)?
.softmax(0)?
no_speech_prob = softmax(&logits.get(0)?, 0)?
.get(NO_SPEECH_TOKEN as usize)?
.to_scalar::<f32>()? as f64;
}
@ -139,7 +137,7 @@ impl Decoder {
.get(seq_len - 1)?
.broadcast_add(&self.suppress_tokens)?;
let next_token = if t > 0f64 {
let prs = (&logits / t)?.softmax(0)?;
let prs = softmax(&(&logits / t)?, 0)?;
let logits_v: Vec<f32> = prs.to_vec1()?;
let distr = rand::distributions::WeightedIndex::new(&logits_v)?;
distr.sample(rng) as u32
@ -153,8 +151,7 @@ impl Decoder {
.unwrap()
};
tokens.push(next_token);
let prob = logits
.softmax(candle::D::Minus1)?
let prob = softmax(&logits, candle::D::Minus1)?
.get(next_token as usize)?
.to_scalar::<f32>()? as f64;
if next_token == EOT_TOKEN || tokens.len() > model.config.max_target_positions {