Skip to content

Fix Qwen3 Embedding Float16 DType #663

New issue

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

#

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

Already on GitHub? # to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+7 −2
Open

Fix Qwen3 Embedding Float16 DType #663

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
912b370
Convert attention_bias to the right dtype.
tpendragon Jun 27, 2025
3e32246
Fix min value.
tpendragon Jun 27, 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
9 changes: 7 additions & 2 deletions backends/candle/src/models/qwen3.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -455,8 +455,13 @@ impl Qwen3Model {
let causal_mask = Tensor::from_slice(&mask, (seq_len, seq_len), device)?;
let causal_mask = causal_mask.expand(&[bs, dim, seq_len, seq_len])?;

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

let negatives =
Tensor::full(f32::MIN, attention_bias.shape(), device)?.to_dtype(self.dtype)?;
Tensor::full(min_value, attention_bias.shape(), device)?.to_dtype(self.dtype)?;
let zeros = Tensor::zeros_like(&attention_bias)?.to_dtype(self.dtype)?;

let causal_mask = causal_mask
Expand Down Expand Up @@ -514,7 +519,7 @@ impl Qwen3Model {

let attention_bias = if masking {
let attention_bias =
Tensor::from_vec(attention_bias, (batch_size, 1, 1, max_length), &self.device)?;
Tensor::from_vec(attention_bias, (batch_size, 1, 1, max_length), &self.device)?.to_dtype(self.dtype)?;
// Broadcast once instead of at every layer
let attention_bias = attention_bias
.broadcast_as((batch_size, self.num_attention_heads, max_length, max_length))?
Expand Down