Skip to content

Update the local attention mask logic to work on MPS and CUDA in ModernBERT #561

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
kozistr wants to merge 5 commits into from
Closed

Update the local attention mask logic to work on MPS and CUDA in ModernBERT #561

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
7a7e0c0
update: not to calculate window_mask with candle
kozistr Apr 5, 2025
8be6ed1
refactor: get_attention_mask
kozistr Apr 5, 2025
0d09fa3
refactor: window_size
kozistr Apr 5, 2025
8dc4c8b
Merge branch 'main' into refactor/local-attention
kozistr Apr 5, 2025
e2d97f2
update: optimize the speed of get_window_mask()
kozistr Apr 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
75 changes: 39 additions & 36 deletions backends/candle/src/models/modernbert.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -466,7 +466,7 @@ pub struct ModernBertModel {
pool: Pool,
classifier: Option<Box<dyn ClassificationHead + Send>>,

local_attention: usize,
window_size: usize,
global_inv_freqs: Tensor,
local_inv_freqs: Tensor,
rotary_dim: usize,
Expand Down Expand Up @@ -541,7 +541,7 @@ impl ModernBertModel {
final_norm,
pool,
classifier,
local_attention: config.local_attention,
window_size: config.local_attention / 2,
global_inv_freqs,
local_inv_freqs,
rotary_dim: attention_head_size,
Expand Down Expand Up @@ -577,32 +577,47 @@ impl ModernBertModel {
Ok(extended_attention_mask)
}

fn get_local_attention_mask(&self, attention_mask: &Tensor) -> Result<Tensor> {
let dev = attention_mask.device();
let attention_mask = attention_mask
.to_device(&Device::Cpu)?
.to_dtype(DType::U8)?;
fn get_window_mask(&self, seq_len: usize) -> Result<Tensor> {
let mut inverted_window_mask = vec![0.0_f32; seq_len * seq_len];

let mask_shape = attention_mask.shape();
let (_, _, seq_len, _) = mask_shape.dims4()?;
for i in 0..seq_len {
let start = i.saturating_sub(self.window_size);
let end = (i + self.window_size + 1).min(seq_len);

let rows = Tensor::arange(0, seq_len as i64, attention_mask.device())?.unsqueeze(0)?;
let rows = rows.broadcast_as((seq_len, seq_len))?;
inverted_window_mask[(i * seq_len)..(i * seq_len + start)].fill(1.0);
inverted_window_mask[(i * seq_len + end)..((i + 1) * seq_len)].fill(1.0);
}

let inverted_window_mask =
Tensor::from_slice(&inverted_window_mask, (seq_len, seq_len), &self.device)?;

Ok(inverted_window_mask)
}

fn get_attention_mask(
&self,
attention_mask: Option<&Tensor>,
input_shape: &(usize, usize),
) -> Result<(Tensor, Tensor)> {
let global_attention_mask = self
.get_global_attention_mask(attention_mask, input_shape)?
.to_dtype(self.dtype)?;

let min_value = match self.dtype {
DType::F32 => f32::MIN as f64,
_ => -65504.0, // f16 minimum value
};

let distance = (&rows - &rows.t()?)?.abs()?;
let global_attention_mask = ((1.0 - global_attention_mask)? * min_value)?;
let global_attention_mask = global_attention_mask.to_dtype(self.dtype)?;

let window_size = (self.local_attention / 2) as i64;
let window_mask = distance
.le(window_size)?
.unsqueeze(0)?
.unsqueeze(0)?
.broadcast_as(mask_shape)?;
let seq_len = global_attention_mask.dim(2)?;
let window_mask = self.get_window_mask(seq_len)?;
let window_mask = (window_mask * min_value)?.to_dtype(self.dtype)?;

let zero_tensor = Tensor::zeros_like(&attention_mask)?;
let local_attention_mask = attention_mask.where_cond(&window_mask, &zero_tensor)?;
let local_attention_mask = local_attention_mask.to_device(dev)?;
let local_attention_mask = global_attention_mask.broadcast_add(&window_mask)?;
Comment thread
kozistr marked this conversation as resolved.

Ok(local_attention_mask)
Ok((global_attention_mask, local_attention_mask))
}

fn forward(&self, batch: Batch) -> Result<(Option<Tensor>, Option<Tensor>)> {
Expand Down Expand Up @@ -675,20 +690,8 @@ impl ModernBertModel {
let mut input_lengths =
Tensor::from_vec(input_lengths, (batch_size, 1), &self.device)?.to_dtype(self.dtype)?;

let global_attention_mask = self
.get_global_attention_mask(attention_mask.as_ref(), &shape)?
.to_dtype(self.dtype)?;
let local_attention_mask = self
.get_local_attention_mask(&global_attention_mask)?
.to_dtype(self.dtype)?;

let min_value = match self.dtype {
DType::F32 => f32::MIN as f64,
_ => -65504.0, // f16 minimum value
};

let global_attention_mask = ((1.0 - global_attention_mask)? * min_value)?;
let local_attention_mask = ((1.0 - local_attention_mask)? * min_value)?;
let (global_attention_mask, local_attention_mask) =
self.get_attention_mask(attention_mask.as_ref(), &shape)?;

let global_rotary_cache =
get_cos_sin(max_length, &self.global_inv_freqs, self.dtype, true)?;
Expand Down