Skip to content
Open
Show file tree
Hide file tree
Changes from 2 commits
Commits
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
3,446 changes: 3,446 additions & 0 deletions kernels-rs/Cargo.lock

Large diffs are not rendered by default.

22 changes: 22 additions & 0 deletions kernels-rs/Cargo.toml
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
[package]
name = "kernels"
version = "0.1.0"
edition = "2024"
description = "Load and call Hugging Face Hub kernels in Rust"
homepage = "https://github.com/huggingface/kernels"
license = "Apache-2.0"
repository = "https://github.com/huggingface/kernels"

[features]
default = []
candle = ["dep:candle-core"]
candle-cuda = ["candle", "candle-core/cuda", "dep:cudarc"]

[dependencies]
candle-core = { version = "0.10.0", optional = true }
cudarc = { version = "0.19.0", optional = true }
huggingface-hub = { git = "https://github.com/huggingface/huggingface_hub_rust.git", rev = "8cbc662035e04d4be8e829316272893e980f5926", package = "huggingface-hub", features = ["blocking"] }
libc = "0.2"
libloading = "0.8"
thiserror = "1"
walkdir = "2"
101 changes: 101 additions & 0 deletions kernels-rs/src/backend.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,101 @@
use std::fmt;
use std::process::Command;
use std::str::FromStr;

use crate::error::Error;

#[derive(Debug, Clone)]
pub enum Backend {
Cpu,
Cuda { version: String },
Xpu { version: String },
}

impl Backend {
pub fn kind(&self) -> BackendKind {
match self {
Backend::Cpu => BackendKind::Cpu,
Backend::Cuda { .. } => BackendKind::Cuda,
Backend::Xpu { .. } => BackendKind::Xpu,
}
}

pub fn name(&self) -> &str {
self.kind().as_str()
}
}

impl fmt::Display for Backend {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
match self {
Backend::Cpu => write!(f, "cpu"),
Backend::Cuda { version } => write!(f, "cuda {version}"),
Backend::Xpu { version } => write!(f, "xpu {version}"),
}
}
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum BackendKind {
Cpu,
Cuda,
Xpu,
}

impl BackendKind {
pub fn as_str(self) -> &'static str {
match self {
BackendKind::Cpu => "cpu",
BackendKind::Cuda => "cuda",
BackendKind::Xpu => "xpu",
}
}
}

impl fmt::Display for BackendKind {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
f.write_str(self.as_str())
}
}

impl FromStr for BackendKind {
type Err = Error;

fn from_str(s: &str) -> Result<Self, Self::Err> {
match s {
"cpu" => Ok(Self::Cpu),
"cuda" => Ok(Self::Cuda),
"xpu" => Ok(Self::Xpu),
other => Err(Error::Kernel(format!("unknown backend: {other}"))),
}
}
}

pub fn detect_cuda_version() -> Option<String> {
cuda_version_from_smi().or_else(cuda_version_from_nvcc)
}

fn cuda_version_from_smi() -> Option<String> {
let output = Command::new("nvidia-smi").output().ok()?;
if !output.status.success() {
return None;
}
let stdout = String::from_utf8_lossy(&output.stdout);
let rest = stdout.split("CUDA Version:").nth(1)?;
Some(rest.split_whitespace().next()?.to_string())
}

fn cuda_version_from_nvcc() -> Option<String> {
let output = Command::new("nvcc").arg("--version").output().ok()?;
let stdout = String::from_utf8_lossy(&output.stdout);
let after = stdout.split("release ").nth(1)?;
Some(after.split(',').next()?.trim().to_string())
}
Comment on lines +75 to +98

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This may not be the same as the library that a framework is compiled against and dynamically loads. Also, nvidia-smi gives the driver library version, not the CUDA runtime version. We need to get it from cudart, e.g. see:

def _get_cuda() -> Optional[CUDA]:

libloading seems to be the most widely used library for dlopen:

https://github.com/nagisa/rust_libloading/

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch, I've updated to prefer querying via cudaRuntimeGetVersion from cudart in the latest changes. I've tested locally and am not running into any issues - however I'm not 100% sure if we need more logic to search for the cudart like ctypes.util.find_library("cudart") does if its not in the default location


pub fn detect() -> BackendKind {
if detect_cuda_version().is_some() {
BackendKind::Cuda
} else {
BackendKind::Cpu
}
}
224 changes: 224 additions & 0 deletions kernels-rs/src/candle.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,224 @@
use std::ffi::c_void;

use candle_core::{CpuStorage, DType, Device, Storage, Tensor};

use crate::KernelModule;
use crate::backend::BackendKind;
use crate::error::{Error, Result};
use crate::tvm_ffi::{self, DLDataType, DLDevice, DLTensor, TVMFFIAny};

fn err(msg: impl Into<String>) -> Error {
Error::Kernel(msg.into())
}

impl BackendKind {
pub fn candle_device(self) -> Result<Device> {
match self {
BackendKind::Cpu => Ok(Device::Cpu),
#[cfg(feature = "candle-cuda")]
BackendKind::Cuda => Device::new_cuda(0).map_err(Into::into),
#[cfg(not(feature = "candle-cuda"))]
BackendKind::Cuda => Ok(Device::Cpu),
BackendKind::Xpu => Ok(Device::Cpu),
}
}

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this can be TryFrom<BackendDevice> for Device. Not 100% sure if it works with the coherency rules, since it's a different mod in the same crate. But I think it should.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this is much nicer, thanks for the suggestion! updated in latest


pub fn candle_supported(self) -> Self {
match self {
#[cfg(feature = "candle-cuda")]
BackendKind::Cuda => BackendKind::Cuda,
#[cfg(not(feature = "candle-cuda"))]
BackendKind::Cuda => BackendKind::Cpu,
other => other,
}
}

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The function name is not very descriptive, maybe to_candle_supported?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

sound good to me, updated in latest

}

impl From<&Device> for BackendKind {
fn from(device: &Device) -> Self {
match device {
Device::Cpu => BackendKind::Cpu,
#[cfg(feature = "candle-cuda")]
Device::Cuda(_) => BackendKind::Cuda,
#[allow(unreachable_patterns)]
_ => BackendKind::Cpu,

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it would be better to explicitly enumerate the other variants here, so that we can rely on exhaustiveness checking when other variants get added?

Also it seems that as it is, if Candle returns a device type that we don't support, it would result in Cpu, which results in kernels that are not compatible with the device type?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

agreed thats a much better approach. I've updated to enumerate the Devices in latest and throw errors if the device is not supported by the currently tvmffi impl

}
}
}

struct PreparedArg {
data: *mut c_void,
shape: Vec<i64>,
strides: Vec<i64>,
dtype: DLDataType,
}

fn dtype_to_dl(dtype: DType) -> Result<DLDataType> {
let (code, bits) = match dtype {
DType::U8 => (tvm_ffi::DL_UINT, 8),
DType::U32 => (tvm_ffi::DL_UINT, 32),
DType::I64 => (tvm_ffi::DL_INT, 64),
DType::BF16 => (tvm_ffi::DL_BFLOAT, 16),
DType::F16 => (tvm_ffi::DL_FLOAT, 16),
DType::F32 => (tvm_ffi::DL_FLOAT, 32),
DType::F64 => (tvm_ffi::DL_FLOAT, 64),
other => return Err(err(format!("unsupported dtype: {other:?}"))),
};
Ok(DLDataType {
code,
bits,
lanes: 1,
})
}

fn cpu_storage_data_ptr(cpu: &CpuStorage, offset: usize) -> Result<*mut c_void> {
macro_rules! ptr {
($v:expr) => {
Ok(unsafe { $v.as_ptr().add(offset) as *mut c_void })
};
}

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Remove, make explicit.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

removed and opt'ed to add *_slice_data_ptr functions for both cpu and cuda to avoid the macros in both places.

match cpu {
CpuStorage::U8(v) => ptr!(v),
CpuStorage::U32(v) => ptr!(v),
CpuStorage::I64(v) => ptr!(v),
CpuStorage::BF16(v) => ptr!(v),
CpuStorage::F16(v) => ptr!(v),
CpuStorage::F32(v) => ptr!(v),
CpuStorage::F64(v) => ptr!(v),
_ => Err(err("unsupported CpuStorage variant")),
}
}

#[cfg(feature = "candle-cuda")]
fn cuda_storage_data_ptr(cuda: &candle_core::CudaStorage, offset: usize) -> Result<*mut c_void> {
use candle_core::cuda_backend::CudaStorageSlice as S;
use cudarc::driver::DevicePtr;

let stream = cuda.device.cuda_stream();

// SyncOnDrop records a stream event; the pointer stays valid as long
// as the caller holds the storage read-guard.
macro_rules! ptr {
($slice:expr) => {{
let view = $slice.slice(offset..);
let (device_ptr, _sync) = view.device_ptr(&stream);
Ok(device_ptr as *mut c_void)
}};
}

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think rather than a macro, this could be a trait + impl? At least I think with a generic type it should work with one implementation for all cases?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated to remove the macros (comment above) and explored a trait but ended up settling on a generic function like

fn cuda_slice_data_ptr<T>(
    slice: &cudarc::driver::CudaSlice<T>,
    stream: &cudarc::driver::CudaStream,
    offset: usize,
) -> Result<*mut c_void> {

this helped make the cpu and cuda function follow a similar functional pattern

note: the cpu path uses a generic function like

fn cpu_slice_data_ptr<T>(slice: &[T], offset: usize) -> Result<*mut c_void> {

happy to explore another approach if you see any issues with this! thanks!


match &cuda.slice {
S::U8(s) => ptr!(s),
S::U32(s) => ptr!(s),
S::I64(s) => ptr!(s),
S::BF16(s) => ptr!(s),
S::F16(s) => ptr!(s),
S::F32(s) => ptr!(s),
S::F64(s) => ptr!(s),
_ => Err(err("unsupported CudaStorage variant")),
}
}

fn extract_data_ptr(storage: &Storage, offset: usize) -> Result<*mut c_void> {
match storage {
Storage::Cpu(cpu) => cpu_storage_data_ptr(cpu, offset),
#[cfg(feature = "candle-cuda")]
Storage::Cuda(cuda) => cuda_storage_data_ptr(cuda, offset),
#[allow(unreachable_patterns)]
_ => Err(err("unsupported storage backend")),
}
}

pub fn get_kernel(repo_id: &str, version: u32) -> Result<KernelModule> {
let kind = crate::backend::detect().candle_supported();
crate::get_kernel_for_backend(repo_id, version, kind)
}

pub fn get_local_kernel(repo_path: &std::path::Path) -> Result<KernelModule> {
let kind = crate::backend::detect().candle_supported();
crate::get_local_kernel_for_backend(repo_path, kind)
}

impl KernelModule {
pub fn device(&self) -> Result<Device> {
self.backend().kind().candle_device()
}
}

// Tensors are passed to the kernel as DLPack pointers directly into
// candle's storage - no copies for contiguous tensors.
pub trait CallKernel {
fn call(&self, func_name: &str, args: &[&Tensor]) -> Result<()>;

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What if there are non-tensor argument, e.g. option bools, epsilon floats, etc.?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch I originally only tested with a kernel that expected tensors. updated to handle multiple types in the latest changes

}

impl CallKernel for KernelModule {
fn call(&self, func_name: &str, args: &[&Tensor]) -> Result<()> {
let kind = args
.first()
.map(|t| BackendKind::from(t.device()))
.unwrap_or(self.backend().kind());

let symbol = format!("__tvm_ffi_{}_{}", func_name, kind.as_str());
let func = unsafe { self.get_func(symbol.as_bytes()) }?;

let contiguous: Vec<Tensor> = args
.iter()
.map(|t| t.contiguous().map_err(Into::into))
.collect::<Result<_>>()?;

let guards: Vec<_> = contiguous.iter().map(|t| t.storage_and_layout()).collect();

let mut prepared: Vec<PreparedArg> = guards
.iter()
.enumerate()
.map(|(i, (storage, layout))| {
Ok(PreparedArg {
data: extract_data_ptr(&storage, layout.start_offset())?,
shape: layout.dims().iter().map(|&d| d as i64).collect(),
strides: layout.stride().iter().map(|&s| s as i64).collect(),
dtype: dtype_to_dl(contiguous[i].dtype())?,
})
})
.collect::<Result<_>>()?;

let device_type = match kind {
BackendKind::Cpu => tvm_ffi::DL_CPU,
BackendKind::Cuda => tvm_ffi::DL_CUDA,
BackendKind::Xpu => tvm_ffi::DL_ONEAPI,
};

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Seems like this could use a From implementation outside the function?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good point, updated in latest


let mut dl_tensors: Vec<DLTensor> = prepared
.iter_mut()
.map(|p| DLTensor {
data: p.data,
device: DLDevice {
device_type,
device_id: 0,
},
ndim: p.shape.len() as i32,
dtype: p.dtype,
shape: p.shape.as_mut_ptr(),
strides: p.strides.as_mut_ptr(),
byte_offset: 0,
})
.collect();

let tvm_args: Vec<TVMFFIAny> = dl_tensors
.iter_mut()
.map(|dl| TVMFFIAny::from_dltensor(dl as *mut DLTensor))
.collect();
let mut result = TVMFFIAny::none();

let ret = unsafe {
func(
std::ptr::null_mut(),
tvm_args.as_ptr(),
tvm_args.len() as i32,
&mut result,
)
};
if ret != 0 {
return Err(err(format!("TVM FFI call `{symbol}` failed (rc {ret})")));
}
Ok(())
}
}
20 changes: 20 additions & 0 deletions kernels-rs/src/error.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
#[derive(Debug, thiserror::Error)]
pub enum Error {
#[error("{0}")]
Kernel(String),

#[error(transparent)]
Io(#[from] std::io::Error),

#[error(transparent)]
Library(#[from] libloading::Error),

#[error(transparent)]
Hub(#[from] huggingface_hub::HfError),

#[cfg(feature = "candle")]
#[error(transparent)]
Candle(#[from] candle_core::Error),
}

pub type Result<T> = std::result::Result<T, Error>;
Loading
Loading