huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 03:54:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update to cudarc 0.14 (breaking change). (#2858)

Browse Source 
* Start updating to cudarc 0.14.

* Adapt a couple more things.

* And a couple more fixes.

* More tweaks.

* And a couple more fixes.

* Bump the major version number.

* Proper module system for the cuda kernels.

* Proper ptx loading.

* Launch the sort kernel.

* Custom op.

* Start using the builder pattern.

* More builder.

* More builder.

* Get candle-core to compile.

* Get the tests to pass.

* Get candle-nn to work too.

* Support for custom cuda functions.

* cudnn fixes.

* Get flash attn to run.

* Switch the crate versions to be alpha.

* Bump the ug dependency.
This commit is contained in:
Laurent Mazare
2025-04-03 09:12:19 +02:00
committed by GitHub
parent d6db305829
commit d9904a3baf
18 changed files with 924 additions and 590 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
candle-core/src/sort.rs
View File

@ -56,7 +56,7 @@ impl ArgSort {
mod cuda {
use super::*;
use crate::cuda_backend::cudarc::driver::{
CudaSlice, DeviceRepr, LaunchAsync, LaunchConfig, ValidAsZeroBits,
CudaSlice, DeviceRepr, LaunchConfig, ValidAsZeroBits,
};
use crate::cuda_backend::{kernel_name, kernels, CudaStorageSlice as S, WrapErr};
use crate::{CudaDevice, WithDType};
@ -69,6 +69,8 @@ mod cuda {
layout: &crate::Layout,
_wrap: W,
) -> Result<S> {
use cudarc::driver::PushKernelArg;
let slice = match layout.contiguous_offsets() {
None => crate::bail!("input has to be contiguous"),
Some((o1, o2)) => src.slice(o1..o2),
@ -76,20 +78,24 @@ mod cuda {
let elem_count = layout.shape().elem_count();
let dst = unsafe { dev.alloc::<u32>(elem_count) }.w()?;
let func = if self.asc {
dev.get_or_load_func(&kernel_name::<T>("asort_asc"), kernels::SORT)?
dev.get_or_load_func(&kernel_name::<T>("asort_asc"), &kernels::SORT)?
} else {
dev.get_or_load_func(&kernel_name::<T>("asort_desc"), kernels::SORT)?
dev.get_or_load_func(&kernel_name::<T>("asort_desc"), &kernels::SORT)?
};
let ncols = self.last_dim;
let nrows = elem_count / ncols;
let ncols_pad = next_power_of_2(ncols);
let params = (&slice, &dst, ncols as i32, ncols_pad as i32);
let cfg = LaunchConfig {
grid_dim: (1, nrows as u32, 1),
block_dim: (ncols_pad as u32, 1, 1),
shared_mem_bytes: (ncols_pad * std::mem::size_of::<u32>()) as u32,
};
unsafe { func.launch(cfg, params) }.w()?;
let stream = dev.cuda_stream();
let mut builder = stream.launch_builder(&func);
let ncols = ncols as i32;
let ncols_pad = ncols_pad as i32;
builder.arg(&slice).arg(&dst).arg(&ncols).arg(&ncols_pad);
unsafe { builder.launch(cfg) }.w()?;
Ok(S::U32(dst))
}
}