Skip to content

Faster MoE kernels #1573

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
wants to merge 11 commits into
base: master
Choose a base branch
from
Open

Faster MoE kernels #1573

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
6bda081
Handle dispatch to slow moe
EricLBuehler Jul 23, 2025
7aba68d
Add cuda kernel
EricLBuehler Jul 23, 2025
3337128
Fix cuda build
EricLBuehler Jul 24, 2025
c94d89b
Runs now
EricLBuehler Jul 24, 2025
64435c7
Fix load
EricLBuehler Jul 24, 2025
4b02895
Fix i32 -> u32
EricLBuehler Jul 24, 2025
e2fcb48
Fix i32 -> u32
EricLBuehler Jul 24, 2025
a7c268e
tmp
EricLBuehler Jul 24, 2025
4abbe03
Fix
EricLBuehler Jul 24, 2025
33277ac
Runs
EricLBuehler Jul 24, 2025
4acc25b
tmp
EricLBuehler Jul 24, 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
2 changes: 1 addition & 1 deletion mistralrs-core/build.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@ fn main() {
use std::{path::PathBuf, vec};
println!("cargo:rerun-if-changed=build.rs");
let build_dir = PathBuf::from(std::env::var("OUT_DIR").unwrap());
let lib_files = vec!["src/cuda/sort.cu"];
let lib_files = vec!["src/cuda/sort.cu", "src/cuda/indexed_matmul.cu", "src/cuda/fused_moe_optimized.cu"];
for lib_file in lib_files.iter() {
println!("cargo:rerun-if-changed={lib_file}");
}
Expand Down
128 changes: 128 additions & 0 deletions mistralrs-core/src/cuda/ffi.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,4 +114,132 @@ extern "C" {
inplace: bool,
stream: i64,
);

// Indexed matrix multiplication for MoE
pub(crate) fn indexed_matmul_f32(
input: *const c_void,
expert_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
out_dim: i32,
num_selected_experts: i32,
num_experts: i32,
stream: i64,
);
pub(crate) fn indexed_matmul_f16(
input: *const c_void,
expert_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
out_dim: i32,
num_selected_experts: i32,
num_experts: i32,
stream: i64,
);
pub(crate) fn indexed_matmul_bf16(
input: *const c_void,
expert_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
out_dim: i32,
num_selected_experts: i32,
num_experts: i32,
stream: i64,
);

// Fused MoE forward pass
pub(crate) fn fused_moe_forward_f32(
input: *const c_void,
gate_weights: *const c_void,
up_weights: *const c_void,
down_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
intermediate_dim: i32,
num_selected_experts: i32,
num_experts: i32,
activation_type: i32,
stream: i64,
);
pub(crate) fn fused_moe_forward_f16(
input: *const c_void,
gate_weights: *const c_void,
up_weights: *const c_void,
down_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
intermediate_dim: i32,
num_selected_experts: i32,
num_experts: i32,
activation_type: i32,
stream: i64,
);
pub(crate) fn fused_moe_forward_bf16(
input: *const c_void,
gate_weights: *const c_void,
up_weights: *const c_void,
down_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
intermediate_dim: i32,
num_selected_experts: i32,
num_experts: i32,
activation_type: i32,
stream: i64,
);

// Optimized fused MoE forward pass
pub(crate) fn fused_moe_forward_optimized_f32(
input: *const c_void,
gate_weights: *const c_void,
up_weights: *const c_void,
down_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
intermediate_dim: i32,
num_selected_experts: i32,
num_experts: i32,
activation_type: i32,
stream: i64,
);

// Chunked fused MoE forward pass for large batches
pub(crate) fn fused_moe_forward_chunked_f32(
input: *const c_void,
gate_weights: *const c_void,
up_weights: *const c_void,
down_weights: *const c_void,
routing_weights: *const c_void,
expert_indices: *const c_void,
output: *mut c_void,
num_tokens: i32,
hidden_dim: i32,
intermediate_dim: i32,
num_selected_experts: i32,
num_experts: i32,
activation_type: i32,
chunk_size: i32,
stream: i64,
);
}
Loading
Loading