Skip to content

Fix Qwen3 #646

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

Merged
merged 5 commits into from
Jun 18, 2025
Merged

Fix Qwen3 #646

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8e59633
fix: left padding
kozistr Jun 17, 2025
a91c023
fix: padding
kozistr Jun 17, 2025
fabd845
update: causal attention mask
kozistr Jun 17, 2025
583ec23
fix: last token pooling
kozistr Jun 17, 2025
9ee64c0
update: test case
kozistr Jun 17, 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
184 changes: 99 additions & 85 deletions backends/candle/src/models/qwen3.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@ pub struct Qwen3Config {
pub rope_theta: f32,
pub sliding_window: Option<usize>,
pub use_sliding_window: bool,
pub eos_token_id: usize,
}

struct Qwen3Attention {
Expand Down Expand Up @@ -164,8 +165,8 @@ impl Qwen3Attention {
.concat(),
)?;

let (q, _res) = self.q_norm.forward(&q, None)?;
let (k, _res) = self.k_norm.forward(&k, None)?;
let (q, _) = self.q_norm.forward(&q, None)?;
let (k, _) = self.k_norm.forward(&k, None)?;

let q = q.transpose(1, 2)?;
let k = k.transpose(1, 2)?;
Expand Down Expand Up @@ -355,16 +356,21 @@ impl Qwen3Layer {
) -> Result<Tensor> {
let _enter = self.span.enter();

let (normed_hidden_states, res) = self.input_layer_norm.forward(hidden_states, None)?;
let (normed_hidden_states, residual) =
self.input_layer_norm.forward(hidden_states, None)?;

let attn_output =
self.attention
.forward(&normed_hidden_states, attention_bias, cos, sin)?;

let (normed_attn_res_output, attn_res) = self
.post_attention_layer_norm
.forward(&attn_output, Some(&res))?;
.forward(&attn_output, Some(&residual))?;

let mlp_output = self.mlp.forward(&normed_attn_res_output)?;

let output = (&mlp_output + &attn_res)?;

Ok(output)
}
}
Expand All @@ -378,6 +384,7 @@ pub struct Qwen3Model {
pool: Pool,
pub device: Device,
num_attention_heads: usize,
pad_token_id: u32,

span: tracing::Span,
}
Expand Down Expand Up @@ -427,12 +434,35 @@ impl Qwen3Model {
rotary_cache,
rotary_dim,
pool,
pad_token_id: config.eos_token_id as u32,
device: vb.device().clone(),
num_attention_heads: config.num_attention_heads,
span: tracing::span!(tracing::Level::TRACE, "model"),
})
}

fn get_causal_attention_bias(&self, attention_bias: Tensor) -> Result<Tensor> {
let (bs, dim, seq_len, _) = attention_bias.dims4()?;

let device = attention_bias.device();

let mask: Vec<u8> = (0..seq_len)
.flat_map(|i| (0..seq_len).map(move |j| (j > i) as u8))
.collect();

let causal_mask = Tensor::from_slice(&mask, (seq_len, seq_len), &Device::Cpu)?;
let causal_mask = causal_mask.expand(&[bs, dim, seq_len, seq_len])?;

let negatives = Tensor::full(f32::MIN, attention_bias.shape(), &Device::Cpu)?;
let zeros = Tensor::zeros_like(&attention_bias)?.to_device(&Device::Cpu)?;

let causal_mask = causal_mask
.where_cond(&negatives, &zeros)?
.to_device(device)?;

attention_bias.broadcast_add(&causal_mask)
}

pub fn forward(&self, batch: Batch) -> Result<(Option<Tensor>, Option<Tensor>)> {
let _enter = self.span.enter();

Expand All @@ -441,93 +471,77 @@ impl Qwen3Model {

let shape = (batch_size, max_length);

let (input_ids, position_ids, input_lengths, attention_bias, _attention_mask) =
if batch_size > 1 {
// Prepare padded batch
let elems = batch_size * max_length;

let mut input_ids = Vec::with_capacity(elems);
let mut position_ids = Vec::with_capacity(elems);
let mut attention_mask = Vec::with_capacity(elems);
let mut attention_bias = Vec::with_capacity(elems);
let mut input_lengths = Vec::with_capacity(batch_size);
let mut masking = false;

for i in 0..batch_size {
let start = batch.cumulative_seq_lengths[i] as usize;
let end = batch.cumulative_seq_lengths[i + 1] as usize;
let seq_length = end - start;
input_lengths.push(seq_length);

// Input ids
for j in start..end {
input_ids.push(batch.input_ids[j]);
position_ids.push(batch.position_ids[j]);
attention_mask.push(1.0_f32);
attention_bias.push(0.0);
}
let (input_ids, position_ids, input_lengths, attention_bias) = if batch_size > 1 {
// Prepare padded batch
let elems = batch_size * max_length;

let mut input_ids = Vec::with_capacity(elems);
let mut position_ids = Vec::with_capacity(elems);
let mut attention_bias = Vec::with_capacity(elems);
let mut input_lengths = Vec::with_capacity(batch_size);
let mut masking = false;

for i in 0..batch_size {
let start = batch.cumulative_seq_lengths[i] as usize;
let end = batch.cumulative_seq_lengths[i + 1] as usize;
let seq_length = end - start;
input_lengths.push(seq_length);

for j in start..end {
input_ids.push(batch.input_ids[j]);
position_ids.push(batch.position_ids[j]);
attention_bias.push(0.0);
}

// Pad to max_length
for _ in seq_length..max_length {
input_ids.push(0);
position_ids.push(0);
attention_mask.push(0.0_f32);
attention_bias.push(f32::NEG_INFINITY);
masking = true;
let padding = max_length - seq_length;
if padding > 0 {
masking = true;
for _ in 0..padding {
input_ids.insert(start, self.pad_token_id);
position_ids.insert(start, 0);
attention_bias.insert(start, f32::MIN);
}
}
}

let input_ids = Tensor::from_vec(input_ids, shape, &self.device)?;
let position_ids = Tensor::from_vec(position_ids, shape, &self.device)?;
let attention_mask = if masking {
Some(Tensor::from_vec(attention_mask, shape, &self.device)?)
} else {
None
};

let attention_bias = if masking {
let attention_bias = Tensor::from_vec(
attention_bias,
(batch_size, 1, 1, max_length),
&self.device,
)?;
// Broadcast once instead of at every layer
let attention_bias = attention_bias
.broadcast_as((
batch_size,
self.num_attention_heads,
max_length,
max_length,
))?
.contiguous()?;
Some(attention_bias)
} else {
None
};

(
input_ids,
position_ids,
input_lengths,
attention_bias,
attention_mask,
)
let input_ids = Tensor::from_vec(input_ids, shape, &self.device)?;
let position_ids = Tensor::from_vec(position_ids, shape, &self.device)?;

let attention_bias = if masking {
let attention_bias =
Tensor::from_vec(attention_bias, (batch_size, 1, 1, max_length), &self.device)?;
// Broadcast once instead of at every layer
let attention_bias = attention_bias
.broadcast_as((batch_size, self.num_attention_heads, max_length, max_length))?
.contiguous()?;
Some(attention_bias)
} else {
let input_ids = Tensor::from_vec(
batch.input_ids.clone(),
(1, batch.input_ids.len()),
&self.device,
)?;
let position_ids = Tensor::from_vec(
batch.position_ids.clone(),
(1, batch.position_ids.len()),
&self.device,
)?;
let input_lengths = vec![batch.input_ids.len()];

(input_ids, position_ids, input_lengths, None, None)
None
};

(input_ids, position_ids, input_lengths, attention_bias)
} else {
let input_ids = Tensor::from_vec(
batch.input_ids.clone(),
(1, batch.input_ids.len()),
&self.device,
)?;
let position_ids = Tensor::from_vec(
batch.position_ids.clone(),
(1, batch.position_ids.len()),
&self.device,
)?;
let input_lengths = vec![batch.input_ids.len()];

(input_ids, position_ids, input_lengths, None)
};

let attention_bias = if let Some(attn_bias) = attention_bias {
Some(self.get_causal_attention_bias(attn_bias)?)
} else {
None
};

let mut hidden_states = self.embeddings.forward(&input_ids)?;

let cos = self
Expand Down Expand Up @@ -583,7 +597,7 @@ impl Qwen3Model {
.iter()
.map(|&i| {
let i = i as usize;
let last_token_idx = input_lengths[i] - 1;
let last_token_idx = max_length - 1;
outputs.i((i, last_token_idx))?.unsqueeze(0)
})
.collect();
Expand Down
Loading
Loading