Skip to content

add faster index select kernel #773

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
alvarobartt merged 2 commits into from
Dec 15, 2025
Merged

add faster index select kernel #773

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
d9848c9
add index_select
michaelfeil Dec 10, 2025
2f868a2
add faster index select
michaelfeil Dec 10, 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 @@ -50,6 +50,7 @@ candle-transformers = { version = "0.8" }
candle-flash-attn = { version = "0.8" }
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 }
candle-rotary = { version = "0.0.1" }
candle-flash-attn-v1 = { version = "0.0.1" }
half = { version = "2.3.1", features = ["num-traits"] }
Expand Down
3 changes: 2 additions & 1 deletion backends/candle/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ candle-transformers = { workspace = true }
candle-flash-attn = { workspace = true, optional = true}
candle-flash-attn-v1 = { 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 }
candle-rotary = { workspace = true, optional = true }
nohash-hasher = { workspace = true }
Expand All @@ -41,6 +42,6 @@ anyhow = { version = "1", features = ["backtrace"] }
accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate"]
metal = ["candle/metal", "candle-nn/metal"]
mkl = ["dep:intel-mkl-src", "candle/_mkl"]
cuda = ["candle/_cuda", "candle-nn/_cuda", "dep:candle-cublaslt", "dep:candle-layer-norm", "dep:candle-rotary"]
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"]
18 changes: 18 additions & 0 deletions backends/candle/src/layers/index_select.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,18 @@
// SPDX-License-Identifier: MIT or Apache-2.0
// First Published under RadixMLP and https://github.com/michaelfeil/candle-index-select-cu by Michael Feil

use candle::{Result, Tensor};
#[cfg(feature = "cuda")]
use candle_index_select_cu;

#[inline]
pub fn index_select(tensor: &Tensor, ids: &Tensor, dim: usize) -> Result<Tensor> {
#[cfg(not(feature = "cuda"))]
{
tensor.index_select(ids, dim)
}
#[cfg(feature = "cuda")]
{
candle_index_select_cu::index_select(tensor, ids, dim)
}
}
3 changes: 3 additions & 0 deletions backends/candle/src/layers/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
#[allow(dead_code, unused)]
mod cublaslt;
mod index_select;
mod layer_norm;
mod linear;
#[allow(dead_code, unused)]
Expand All @@ -11,4 +12,6 @@ pub use layer_norm::{LayerNorm, LayerNormNoBias};
pub use linear::{HiddenAct, Linear};
#[allow(unused_imports)]
pub use rms_norm::RMSNorm;
#[allow(unused_imports)]
pub use index_select::index_select;
pub use rotary::{apply_rotary, get_cos_sin, get_inv_freqs, RopeScaling};
8 changes: 4 additions & 4 deletions backends/candle/src/models/flash_bert.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{LayerNorm, Linear};
use crate::layers::{index_select, LayerNorm, Linear};
use crate::models::bert::{
BertClassificationHead, BertConfig, BertEmbeddings, BertSpladeHead, ClassificationHead,
PositionEmbeddingType, RobertaClassificationHead,
Expand Down Expand Up @@ -419,11 +419,11 @@ impl FlashBertModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -514,7 +514,7 @@ impl FlashBertModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
8 changes: 4 additions & 4 deletions backends/candle/src/models/flash_distilbert.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{LayerNorm, Linear};
use crate::layers::{index_select, LayerNorm, Linear};
use crate::models::distilbert::{
DistilBertConfig, DistilBertEmbeddings, DistilBertMLP, DistilBertSpladeHead,
};
Expand Down Expand Up @@ -290,11 +290,11 @@ impl FlashDistilBertModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
outputs.index_select(&indices, 0)?
index_select(&outputs, &indices, 0)?
} else {
match self.pool {
Pool::Cls => outputs.i(0)?,
Expand Down Expand Up @@ -384,7 +384,7 @@ impl FlashDistilBertModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
12 changes: 6 additions & 6 deletions backends/candle/src/models/flash_gte.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{get_cos_sin, get_inv_freqs, LayerNorm, Linear};
use crate::layers::{get_cos_sin, get_inv_freqs, index_select, LayerNorm, Linear};
use crate::models::gte::{GTEClassificationHead, GTEConfig, GTEMLP};
use crate::models::{Model, PositionEmbeddingType};

Expand Down Expand Up @@ -291,8 +291,8 @@ impl FlashGTEModel {
.embeddings_norm
.forward(&word_embeddings, token_type_embeddings.as_ref())?;

let cos = self.cos_cache.index_select(&position_ids, 0)?;
let sin = self.sin_cache.index_select(&position_ids, 0)?;
let cos = index_select(&self.cos_cache, &position_ids, 0)?;
let sin = index_select(&self.sin_cache, &position_ids, 0)?;

for layer in &self.layers {
let h = layer.forward(
Expand Down Expand Up @@ -336,11 +336,11 @@ impl FlashGTEModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -407,7 +407,7 @@ impl FlashGTEModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
8 changes: 4 additions & 4 deletions backends/candle/src/models/flash_jina.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
use crate::alibi::alibi_head_slopes;
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{HiddenAct, LayerNorm, Linear};
use crate::layers::{index_select, HiddenAct, LayerNorm, Linear};
use crate::models::bert::PositionEmbeddingType;
use crate::models::jina::{ClassificationHead, JinaBertClassificationHead, JinaEmbeddings};
use crate::models::{BertConfig, Model};
Expand Down Expand Up @@ -349,11 +349,11 @@ impl FlashJinaBertModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -420,7 +420,7 @@ impl FlashJinaBertModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
8 changes: 4 additions & 4 deletions backends/candle/src/models/flash_jina_code.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
use crate::alibi::alibi_head_slopes;
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{HiddenAct, LayerNorm, Linear};
use crate::layers::{index_select, HiddenAct, LayerNorm, Linear};
use crate::models::bert::PositionEmbeddingType;
use crate::models::jina::JinaEmbeddings;
use crate::models::{BertConfig, Model};
Expand Down Expand Up @@ -395,11 +395,11 @@ impl FlashJinaCodeBertModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -466,7 +466,7 @@ impl FlashJinaCodeBertModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
12 changes: 6 additions & 6 deletions backends/candle/src/models/flash_mistral.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{get_cos_sin, get_inv_freqs, HiddenAct, Linear, RMSNorm};
use crate::layers::{get_cos_sin, get_inv_freqs, index_select, HiddenAct, Linear, RMSNorm};
use crate::models::{MistralConfig, Model};
use candle::{DType, Device, IndexOp, Result, Tensor};
use candle_nn::{Embedding, Module, VarBuilder};
Expand Down Expand Up @@ -307,8 +307,8 @@ impl FlashMistralModel {

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

let cos = self.cos_cache.index_select(&position_ids, 0)?;
let sin = self.sin_cache.index_select(&position_ids, 0)?;
let cos = index_select(&self.cos_cache, &position_ids, 0)?;
let sin = index_select(&self.sin_cache, &position_ids, 0)?;

let mut residual = None;
for layer in &self.layers {
Expand Down Expand Up @@ -355,11 +355,11 @@ impl FlashMistralModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -426,7 +426,7 @@ impl FlashMistralModel {
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

// Select the tokens with final indices
Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
12 changes: 6 additions & 6 deletions backends/candle/src/models/flash_modernbert.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
use std::collections::HashMap;

use crate::flash_attn::flash_attn_varlen;
use crate::layers::{get_cos_sin, get_inv_freqs, LayerNormNoBias, Linear};
use crate::layers::{get_cos_sin, get_inv_freqs, index_select, LayerNormNoBias, Linear};
use crate::models::modernbert::{
ClassificationHead, ModernBertClassificationHead, ModernBertConfig, ModernBertEmbeddings,
ModernBertMLP,
Expand Down Expand Up @@ -343,8 +343,8 @@ impl FlashModernBertModel {
for use_local_attention in [true, false] {
let (cos, sin) = &self.rotary_cache[&use_local_attention];

let cos = cos.index_select(&position_ids, 0)?;
let sin = sin.index_select(&position_ids, 0)?;
let cos = index_select(&cos, &position_ids, 0)?;
let sin = index_select(&sin, &position_ids, 0)?;

rotary_cache.insert(use_local_attention, (cos, sin));
}
Expand Down Expand Up @@ -378,10 +378,10 @@ impl FlashModernBertModel {
&self.device,
)?;

indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down Expand Up @@ -441,7 +441,7 @@ impl FlashModernBertModel {
let final_indices =
Tensor::from_vec(final_indices, final_indices_length, &self.device)?;

Some(outputs.index_select(&final_indices, 0)?)
Some(index_select(&outputs, &final_indices, 0)?)
} else {
Some(outputs)
}
Expand Down
32 changes: 15 additions & 17 deletions backends/candle/src/models/flash_nomic.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
use crate::flash_attn::flash_attn_varlen;
use crate::layers::{get_cos_sin, get_inv_freqs, LayerNorm, Linear};
use crate::layers::{get_cos_sin, get_inv_freqs, index_select, LayerNorm, Linear};
use crate::models::nomic::{NomicBertEmbeddings, NomicMLP};
use crate::models::{Model, NomicConfig};
use candle::{DType, Device, IndexOp, Result, Tensor, D};
Expand Down Expand Up @@ -285,22 +285,20 @@ impl FlashNomicBertModel {
let (cos, sin) = if self.scaled_rotary_cache.is_some()
&& batch.max_length > self.max_trained_positions
{
let cos = self
.scaled_rotary_cache
.as_ref()
.unwrap()
.0
.index_select(&position_ids, 0)?;
let sin = self
.scaled_rotary_cache
.as_ref()
.unwrap()
.1
.index_select(&position_ids, 0)?;
let cos = index_select(
&self.scaled_rotary_cache.as_ref().unwrap().0,
&position_ids,
0,
)?;
let sin = index_select(
&self.scaled_rotary_cache.as_ref().unwrap().1,
&position_ids,
0,
)?;
(cos, sin)
} else {
let cos = self.rotary_cache.0.index_select(&position_ids, 0)?;
let sin = self.rotary_cache.1.index_select(&position_ids, 0)?;
let cos = index_select(&self.rotary_cache.0, &position_ids, 0)?;
let sin = index_select(&self.rotary_cache.1, &position_ids, 0)?;
(cos, sin)
};

Expand Down Expand Up @@ -343,11 +341,11 @@ impl FlashNomicBertModel {
)?;

// Only select indices that requires pooling
indices = indices.index_select(&pooled_indices, 0)?
indices = index_select(&indices, &pooled_indices, 0)?
}

// Select tokens
Some(outputs.index_select(&indices, 0)?)
Some(index_select(&outputs, &indices, 0)?)
} else {
Some(
match self.pool {
Expand Down
Loading