Skip to content

flash-attention-3 support in candle. #785

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

Open
michaelfeil wants to merge 2 commits into
base: main
Choose a base branch
from
Open

flash-attention-3 support in candle. #785

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c092919
add fa3
michaelfeil Dec 21, 2025
0422d92
add fa3 - public crate
michaelfeil Dec 21, 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
14 changes: 14 additions & 0 deletions Cargo.lock
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

1 change: 1 addition & 0 deletions Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,7 @@ candle = { version = "0.8", package = "candle-core" }
candle-nn = { version = "0.8" }
candle-transformers = { version = "0.8" }
candle-flash-attn = { version = "0.8" }
candle-flash-attn-3 = { version = "0.0.1", features = ["cuda-11"], default-features = false }
candle-cublaslt = { version = "0.0.1" }
candle-layer-norm = { version = "0.0.1" }
candle-index-select-cu = { version = "0.0.1", features = ["cuda-11"], default-features = false }
Expand Down
1 change: 1 addition & 0 deletions backends/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,3 +29,4 @@ mkl = ["text-embeddings-backend-candle?/mkl"]
accelerate = ["text-embeddings-backend-candle?/accelerate"]
flash-attn = ["text-embeddings-backend-candle?/flash-attn"]
flash-attn-v1 = ["text-embeddings-backend-candle?/flash-attn-v1"]
flash-attn-v3 = ["text-embeddings-backend-candle?/flash-attn-v3"]
2 changes: 2 additions & 0 deletions backends/candle/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@ candle-nn = { workspace = true }
candle-transformers = { workspace = true }
candle-flash-attn = { workspace = true, optional = true}
candle-flash-attn-v1 = { workspace = true, optional = true }
candle-flash-attn-3 = { workspace = true, optional = true }
candle-cublaslt = { workspace = true, optional = true }
candle-index-select-cu = { workspace = true, optional = true, features = ["cuda-11"], default-features = false}
candle-layer-norm = { workspace = true, optional = true }
Expand Down Expand Up @@ -45,3 +46,4 @@ mkl = ["dep:intel-mkl-src", "candle/_mkl"]
cuda = ["candle/_cuda", "candle-nn/_cuda", "dep:candle-cublaslt", "dep:candle-layer-norm", "dep:candle-rotary", "dep:candle-index-select-cu"]
flash-attn-v1 = ["dep:candle-flash-attn-v1", "cuda"]
flash-attn = ["dep:candle-flash-attn", "cuda"]
flash-attn-v3 = ["dep:candle-flash-attn-3", "cuda"]
60 changes: 57 additions & 3 deletions backends/candle/src/flash_attn.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,14 @@ pub fn get_runtime_compute_cap() -> usize {
}
}

static PRINT_ONCE: Once = Once::new();

fn print_version_once(version: &str) {
PRINT_ONCE.call_once(|| {
println!("Using michaelfeil/candle-flash-attn-3 v{}", version);
});
}

#[allow(clippy::too_many_arguments, unused)]
pub(crate) fn flash_attn_varlen(
q: &Tensor,
Expand Down Expand Up @@ -62,7 +70,53 @@ pub(crate) fn flash_attn_varlen(
#[cfg(not(feature = "flash-attn-v1"))]
candle::bail!("Flash attention v1 is not installed. Use `flash-attn-v1` feature.")
} else if (80..90).contains(&runtime_compute_cap) || runtime_compute_cap == 90 {
#[cfg(feature = "flash-attn")]
#[cfg(feature = "flash-attn-v3")]
{
use candle_flash_attn_v3::{flash_attn_varlen_alibi_windowed, flash_attn_varlen_windowed};
print_version_once("0.0.1");
let window_size_right = if causal {
Some(0)
} else if window_size_right.is_some() {
window_size_right
} else {
None
};

let attention = if let Some(alibi_slopes) = alibi_slopes {
flash_attn_varlen_alibi_windowed(
q,
k,
v,
alibi_slopes,
seqlens_q,
seqlens_k,
max_seqlen_q,
max_seqlen_k,
softmax_scale,
window_size_left,
window_size_right,
false, // use_gqa_packing - set to false for now
)
} else {
flash_attn_varlen_windowed(
q,
k,
v,
seqlens_q,
seqlens_k,
max_seqlen_q,
max_seqlen_k,
softmax_scale,
window_size_left,
window_size_right,
false, // use_gqa_packing - set to false for now
)
};

return attention;
}

#[cfg(all(not(feature = "flash-attn-v3"), feature = "flash-attn"))]
{
use candle_flash_attn::{flash_attn_varlen_alibi_windowed, flash_attn_varlen_windowed};

Expand Down Expand Up @@ -105,8 +159,8 @@ pub(crate) fn flash_attn_varlen(

return attention;
}
#[cfg(not(feature = "flash-attn"))]
candle::bail!("Flash attention is not installed. Use `flash-attn` feature.")
#[cfg(not(any(feature = "flash-attn-v3", feature = "flash-attn")))]
candle::bail!("Flash attention is not installed. Use `flash-attn-v3` or `flash-attn` feature.")
}
candle::bail!(
"GPU with CUDA capability {} is not supported",
Expand Down