More cleanup.

Fix NaN errors for Gelu in Metal

Cargo.toml

@@ -61,7 +61,7 @@ tracing-subscriber = "0.3.7"
 wav = "1.0.0"
 yoke = { version = "0.7.2", features = ["derive"] }
 zip = { version = "0.6.6", default-features = false }
-metal = { version = "0.27.1", features = ["mps"], package="candle-metal" }
+metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

candle-core/src/metal_backend.rs

@@ -4,9 +4,7 @@ use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
 use crate::{CpuStorage, DType, Layout, Result, Shape};
 use candle_metal_kernels;
 use candle_metal_kernels::Kernels;
-use half::f16;
 use metal;
-use metal::mps::matrix::{Matrix, MatrixDescriptor, MatrixMultiplication};
 use metal::{Buffer, CommandBuffer, CommandQueue, MTLResourceOptions, NSUInteger};
 use std::collections::HashMap;
 use std::path::Path;
@@ -38,7 +36,9 @@ impl From<String> for MetalError {
 pub struct MetalDevice {
     device: metal::Device,
     command_queue: metal::CommandQueue,
-    command_buffer: Arc<RwLock<metal::CommandBuffer>>,
+    command_buffers: Arc<RwLock<Vec<metal::CommandBuffer>>>,
+    command_buffer_index: Arc<RwLock<usize>>,
+    fence: metal::Fence,
     kernels: Arc<candle_metal_kernels::Kernels>,
     buffers: Arc<RwLock<HashMap<(NSUInteger, MTLResourceOptions), Vec<Arc<Buffer>>>>>,
 }
@@ -70,38 +70,34 @@ impl MetalDevice {
         &self.command_queue
     }
 
-    pub fn command_buffer(&self) -> std::sync::RwLockReadGuard<CommandBuffer> {
+    pub fn command_buffer(&self) -> CommandBuffer {
-        self.command_buffer.try_read().unwrap()
+        let mut command_buffers = self.command_buffers.try_write().unwrap();
-    }
+        let mut command_buffer = command_buffers[0].to_owned();
-
+        let mut index = self.command_buffer_index.try_write().unwrap();
-    pub fn commit(&self) {
+        if *index > 20 {
-        let mut old = self.command_buffer.try_write().unwrap();
+            command_buffer.commit();
-        match old.status() {
+            command_buffer = self.command_queue.new_command_buffer().to_owned();
-            metal::MTLCommandBufferStatus::NotEnqueued
+            *command_buffers = vec![command_buffer.clone()];
-            | metal::MTLCommandBufferStatus::Enqueued => {
+            *index = 0;
-                old.commit();
-                let command_buffer = self.command_queue.new_command_buffer().to_owned();
-                *old = command_buffer;
-            }
-            _ => {}
         }
+        *index += 1;
+        command_buffer
     }
 
     pub fn wait_until_completed(&self) {
-        let mut old = self.command_buffer.try_write().unwrap();
+        let mut command_buffers = self.command_buffers.try_write().unwrap();
-        match old.status() {
+        let command_buffer = &command_buffers[0];
-            metal::MTLCommandBufferStatus::NotEnqueued
+        match command_buffer.status() {
-            | metal::MTLCommandBufferStatus::Enqueued => {
+            metal::MTLCommandBufferStatus::Committed
-                old.commit();
+            | metal::MTLCommandBufferStatus::Scheduled
-                old.wait_until_completed();
+            | metal::MTLCommandBufferStatus::Completed => {
-            }
+                panic!("Already committed");
-            metal::MTLCommandBufferStatus::Committed | metal::MTLCommandBufferStatus::Scheduled => {
-                old.wait_until_completed();
             }
             _ => {}
         }
-        let command_buffer = self.command_queue.new_command_buffer().to_owned();
+        command_buffer.commit();
-        *old = command_buffer;
+        command_buffer.wait_until_completed();
+        *command_buffers = vec![self.command_queue.new_command_buffer().to_owned()];
     }
 
     pub fn kernels(&self) -> &Kernels {
@@ -112,12 +108,17 @@ impl MetalDevice {
         &self.device
     }
 
-    pub fn new_buffer(&self, element_count: usize, dtype: DType) -> Arc<Buffer> {
+    pub fn new_buffer(&self, element_count: usize, dtype: DType, name: &str) -> Arc<Buffer> {
         let size = (element_count * dtype.size_in_bytes()) as NSUInteger;
-        self._new_buffer(size, MTLResourceOptions::StorageModePrivate)
+        self._new_buffer(size, MTLResourceOptions::StorageModePrivate, name)
     }
 
-    fn _new_buffer(&self, size: NSUInteger, option: MTLResourceOptions) -> Arc<Buffer> {
+    fn _new_buffer(
+        &self,
+        size: NSUInteger,
+        option: MTLResourceOptions,
+        _name: &str,
+    ) -> Arc<Buffer> {
         let mut buffers = self.buffers.try_write().unwrap();
         let subbuffers = buffers.entry((size, option)).or_insert(vec![]);
 
@@ -129,11 +130,20 @@ impl MetalDevice {
         let new_buffer = self.device.new_buffer(size as NSUInteger, option);
         let new_buffer = Arc::new(new_buffer);
         subbuffers.push(new_buffer.clone());
+        for subbuffers in buffers.values_mut() {
+            let newbuffers = subbuffers
+                .iter()
+                .filter(|s| Arc::strong_count(s) > 1)
+                .map(|s| Arc::clone(s))
+                .collect();
+            *subbuffers = newbuffers;
+        }
+
         new_buffer
     }
 
     pub fn new_buffer_managed(&self, size: NSUInteger) -> Arc<Buffer> {
-        self._new_buffer(size, MTLResourceOptions::StorageModeManaged)
+        self._new_buffer(size, MTLResourceOptions::StorageModeManaged, "managed")
     }
 
     pub fn new_buffer_with_data<T>(&self, data: &[T]) -> Arc<Buffer> {
@@ -143,13 +153,20 @@ impl MetalDevice {
             size,
             metal::MTLResourceOptions::StorageModeManaged,
         );
-        let real = self._new_buffer(size, metal::MTLResourceOptions::StorageModePrivate);
+        let real = self._new_buffer(
-        {
+            size,
-            let command = self.command_buffer();
+            metal::MTLResourceOptions::StorageModePrivate,
-            let blit = command.new_blit_command_encoder();
+            "with_data",
+        );
+        let command_buffer = self.command_buffer();
+        command_buffer.set_label("with_data");
+        let blit = command_buffer.new_blit_command_encoder();
+        blit.wait_for_fence(&self.fence);
+        blit.set_label("with_data_blit");
         blit.copy_from_buffer(&tmp, 0, &real, 0, tmp.length());
+        blit.update_fence(&self.fence);
         blit.end_encoding();
-        }
+
         // This is necessary, for mmaped safetensors
         // Because of the unsafe slice cast we're doing.
         // The slice might not live long enough for metal
@@ -161,25 +178,6 @@ impl MetalDevice {
         real
     }
 
-    pub fn new_matrix(
-        &self,
-        (b, m, n): (NSUInteger, NSUInteger, NSUInteger),
-        size: NSUInteger,
-        type_id: u32,
-        dtype: DType,
-    ) -> Result<(Matrix, Arc<Buffer>)> {
-        let elem_count = (b * m * n) as usize;
-        let out_buffer = self.new_buffer(elem_count, dtype);
-
-        let result_descriptor =
-            MatrixDescriptor::init_multiple(m, n, b, n * size, m * n * size, type_id);
-        let result_matrix = Matrix::init_with_buffer_descriptor(&out_buffer, 0, &result_descriptor)
-            .ok_or_else(|| {
-                MetalError::from("Failed to create matrix multiplication kernel".to_string())
-            })?;
-        Ok((result_matrix, out_buffer))
-    }
-
     pub fn capture<P: AsRef<Path>>(&self, path: P) -> Result<()> {
         let capture = metal::CaptureManager::shared();
         let descriptor = metal::CaptureDescriptor::new();
@@ -197,22 +195,6 @@ impl MetalDevice {
 #[derive(Debug, Clone)]
 pub struct MetalStorage {
     buffer: Arc<metal::Buffer>,
-    matrices: Arc<
-        RwLock<
-            HashMap<
-                (
-                    NSUInteger,
-                    NSUInteger,
-                    NSUInteger,
-                    bool,
-                    NSUInteger,
-                    NSUInteger,
-                    u32,
-                ),
-                Matrix,
-            >,
-        >,
-    >,
     device: MetalDevice,
     dtype: DType,
 }
@@ -241,23 +223,27 @@ impl BackendStorage for MetalStorage {
                 self.dtype
             );
         }
-
         let buffer = self.device.new_buffer_managed(self.buffer.length());
+        {
             let command_buffer = self.device.command_buffer();
+            command_buffer.set_label("to_cpu");
             let blit = command_buffer.new_blit_command_encoder();
+            blit.set_label("blit_to_cpu");
+            blit.wait_for_fence(&self.device.fence);
             blit.copy_from_buffer(&self.buffer, 0, &buffer, 0, self.buffer.length());
+            blit.update_fence(&self.device.fence);
             blit.end_encoding();
-        drop(command_buffer);
+        }
         self.device.wait_until_completed();
 
         match self.dtype {
-            DType::U8 => Ok(CpuStorage::U8(buffer.read_to_vec(length / size))),
+            DType::U8 => Ok(CpuStorage::U8(read_to_vec(&buffer, length / size))),
-            DType::U32 => Ok(CpuStorage::U32(buffer.read_to_vec(length / size))),
+            DType::U32 => Ok(CpuStorage::U32(read_to_vec(&buffer, length / size))),
-            DType::I64 => Ok(CpuStorage::I64(buffer.read_to_vec(length / size))),
+            DType::I64 => Ok(CpuStorage::I64(read_to_vec(&buffer, length / size))),
-            DType::F16 => Ok(CpuStorage::F16(buffer.read_to_vec(length / size))),
+            DType::F16 => Ok(CpuStorage::F16(read_to_vec(&buffer, length / size))),
-            DType::BF16 => Ok(CpuStorage::BF16(buffer.read_to_vec(length / size))),
+            DType::BF16 => Ok(CpuStorage::BF16(read_to_vec(&buffer, length / size))),
-            DType::F32 => Ok(CpuStorage::F32(buffer.read_to_vec(length / size))),
+            DType::F32 => Ok(CpuStorage::F32(read_to_vec(&buffer, length / size))),
-            DType::F64 => Ok(CpuStorage::F64(buffer.read_to_vec(length / size))),
+            DType::F64 => Ok(CpuStorage::F64(read_to_vec(&buffer, length / size))),
         }
     }
 
@@ -268,7 +254,7 @@ impl BackendStorage for MetalStorage {
         let el = shape.elem_count();
         let dtype = self.dtype;
 
-        let buffer = device.new_buffer(el, self.dtype);
+        let buffer = device.new_buffer(el, self.dtype, "affine");
         let command_buffer = self.device.command_buffer();
         if layout.is_contiguous() && layout.start_offset() == 0 {
             let name = match self.dtype {
@@ -312,12 +298,102 @@ impl BackendStorage for MetalStorage {
         Ok(Self::new(buffer, device.clone(), dtype))
     }
 
-    fn powf(&self, _: &Layout, _: f64) -> Result<Self> {
+    fn powf(&self, layout: &Layout, pow: f64) -> Result<Self> {
-        crate::bail!("powf metal")
+        let device = self.device().clone();
+
+        let shape = layout.shape();
+        let el = shape.elem_count();
+        let dtype = self.dtype;
+
+        let buffer = device.new_buffer(el, self.dtype, "powf");
+        let command_buffer = self.device.command_buffer();
+        if layout.is_contiguous() && layout.start_offset() == 0 {
+            let name = match self.dtype {
+                DType::F32 => "powf_float",
+                DType::F16 => "powf_half",
+                dtype => crate::bail!("Powf {dtype:?}"),
+            };
+            candle_metal_kernels::call_powf(
+                &device.device,
+                &command_buffer,
+                &device.kernels,
+                name,
+                el,
+                &self.buffer,
+                &buffer,
+                pow as f32,
+            )
+            .map_err(MetalError::from)?;
+        } else {
+            let name = match self.dtype {
+                DType::F32 => "powf_float_strided",
+                DType::F16 => "powf_half_strided",
+                dtype => crate::bail!("Powf {dtype:?}"),
+            };
+            candle_metal_kernels::call_powf_strided(
+                &device.device,
+                &command_buffer,
+                &device.kernels,
+                name,
+                layout.dims(),
+                &self.buffer,
+                layout.stride(),
+                layout.start_offset() * dtype.size_in_bytes(),
+                &buffer,
+                pow as f32,
+            )
+            .map_err(MetalError::from)?;
+        }
+        Ok(Self::new(buffer, device.clone(), dtype))
     }
 
-    fn elu(&self, _: &Layout, _: f64) -> Result<Self> {
+    fn elu(&self, layout: &Layout, alpha: f64) -> Result<Self> {
-        crate::bail!("elu metal")
+        let device = self.device().clone();
+
+        let shape = layout.shape();
+        let el = shape.elem_count();
+        let dtype = self.dtype;
+
+        let buffer = device.new_buffer(el, self.dtype, "elu");
+        let command_buffer = self.device.command_buffer();
+        if layout.is_contiguous() && layout.start_offset() == 0 {
+            let name = match self.dtype {
+                DType::F32 => "elu_float",
+                DType::F16 => "elu_half",
+                dtype => crate::bail!("Powf {dtype:?}"),
+            };
+            candle_metal_kernels::call_elu(
+                &device.device,
+                &command_buffer,
+                &device.kernels,
+                name,
+                el,
+                &self.buffer,
+                &buffer,
+                alpha as f32,
+            )
+            .map_err(MetalError::from)?;
+        } else {
+            let name = match self.dtype {
+                DType::F32 => "elu_float_strided",
+                DType::F16 => "elu_half_strided",
+                dtype => crate::bail!("Powf {dtype:?}"),
+            };
+            candle_metal_kernels::call_elu_strided(
+                &device.device,
+                &command_buffer,
+                &device.kernels,
+                name,
+                layout.dims(),
+                &self.buffer,
+                layout.stride(),
+                layout.start_offset() * dtype.size_in_bytes(),
+                &buffer,
+                alpha as f32,
+            )
+            .map_err(MetalError::from)?;
+        }
+        Ok(Self::new(buffer, device.clone(), dtype))
     }
 
     fn reduce_op(&self, op: ReduceOp, layout: &Layout, sum_dims: &[usize]) -> Result<Self> {
@@ -365,7 +441,7 @@ impl BackendStorage for MetalStorage {
         if dtype == DType::U32 {
             crate::bail!("Implement return index reduce op");
         }
-        let buffer = device.new_buffer(dst_el, dtype);
+        let buffer = device.new_buffer(dst_el, dtype, "reduce");
         let command_buffer = self.device.command_buffer();
         candle_metal_kernels::call_reduce_contiguous(
             &device.device,
@@ -391,9 +467,9 @@ impl BackendStorage for MetalStorage {
         let device = self.device();
         let shape = layout.shape();
         let el_count = shape.elem_count();
-        let buffer = device.new_buffer(el_count, dtype);
+        let buffer = device.new_buffer(el_count, dtype, "todtype");
         let command_buffer = device.command_buffer();
-        if layout.is_contiguous() {
+        if layout.is_contiguous() && layout.start_offset() == 0 {
             let kernel_name = match (self.dtype, dtype) {
                 (DType::U32, DType::F32) => "cast_u32_f32",
                 (DType::U32, DType::U8) => "cast_u32_u8",
@@ -435,7 +511,7 @@ impl BackendStorage for MetalStorage {
             )
             .map_err(MetalError::from)?;
         }
-
+        command_buffer.set_label("to_dtype");
         Ok(Self::new(buffer, device.clone(), dtype))
     }
 
@@ -444,8 +520,9 @@ impl BackendStorage for MetalStorage {
         let dtype = self.dtype;
         let shape = layout.shape();
         let el_count = shape.elem_count();
-        let buffer = device.new_buffer(el_count, dtype);
+        let buffer = device.new_buffer(el_count, dtype, B::KERNEL);
         let command_buffer = device.command_buffer();
+        command_buffer.set_label(B::KERNEL);
         if layout.is_contiguous() && layout.start_offset() == 0 {
             use candle_metal_kernels::unary::contiguous;
 
@@ -463,6 +540,7 @@ impl BackendStorage for MetalStorage {
                 ("uceil", DType::F32) => contiguous::ceil::FLOAT,
                 ("ufloor", DType::F32) => contiguous::floor::FLOAT,
                 ("uround", DType::F32) => contiguous::round::FLOAT,
+                ("utanh", DType::F32) => contiguous::tanh::FLOAT,
                 ("ucos", DType::F16) => contiguous::cos::HALF,
                 ("usin", DType::F16) => contiguous::sin::HALF,
                 ("usqr", DType::F16) => contiguous::sqr::HALF,
@@ -476,6 +554,7 @@ impl BackendStorage for MetalStorage {
                 ("uceil", DType::F16) => contiguous::ceil::HALF,
                 ("ufloor", DType::F16) => contiguous::floor::HALF,
                 ("uround", DType::F16) => contiguous::round::HALF,
+                ("utanh", DType::F16) => contiguous::tanh::HALF,
                 (name, dtype) => crate::bail!("Match {name} - {dtype:?}"),
             };
             candle_metal_kernels::call_unary_contiguous(
@@ -533,9 +612,6 @@ impl BackendStorage for MetalStorage {
             )
             .map_err(MetalError::from)?;
         }
-        command_buffer.set_label("unary");
-        drop(command_buffer);
-        self.device.commit();
         Ok(Self::new(buffer, device.clone(), dtype))
     }
 
@@ -549,30 +625,23 @@ impl BackendStorage for MetalStorage {
         let dtype = self.dtype;
         let shape = lhs_l.shape();
         let el_count = shape.elem_count();
-        let buffer = device.new_buffer(el_count, dtype);
+        let buffer = device.new_buffer(el_count, dtype, B::KERNEL);
         let command_buffer = device.command_buffer();
         if (lhs_l.is_contiguous() && lhs_l.start_offset() == 0)
             && (rhs_l.is_contiguous() && rhs_l.start_offset() == 0)
+            && &B::KERNEL[..1] != "b"
         {
             use candle_metal_kernels::binary::contiguous;
 
             let kernel_name = match (B::KERNEL, dtype) {
                 ("add", DType::F32) => contiguous::add::FLOAT,
-                ("badd", DType::F32) => contiguous::add::FLOAT,
                 ("sub", DType::F32) => contiguous::sub::FLOAT,
-                ("bsub", DType::F32) => contiguous::sub::FLOAT,
                 ("mul", DType::F32) => contiguous::mul::FLOAT,
-                ("bmul", DType::F32) => contiguous::mul::FLOAT,
                 ("div", DType::F32) => contiguous::div::FLOAT,
-                ("bdiv", DType::F32) => contiguous::div::FLOAT,
                 ("add", DType::F16) => contiguous::add::HALF,
-                ("badd", DType::F16) => contiguous::add::HALF,
                 ("sub", DType::F16) => contiguous::sub::HALF,
-                ("bsub", DType::F16) => contiguous::sub::HALF,
                 ("mul", DType::F16) => contiguous::mul::HALF,
-                ("bmul", DType::F16) => contiguous::mul::HALF,
                 ("div", DType::F16) => contiguous::div::HALF,
-                ("bdiv", DType::F16) => contiguous::div::HALF,
                 (name, dtype) => crate::bail!("Match {name} - {dtype:?}"),
             };
             candle_metal_kernels::call_binary_contiguous(
@@ -617,8 +686,6 @@ impl BackendStorage for MetalStorage {
             .map_err(MetalError::from)?;
         }
         command_buffer.set_label("binary");
-        drop(command_buffer);
-        self.device.commit();
         Ok(Self::new(buffer, device.clone(), dtype))
     }
 
@@ -635,7 +702,7 @@ impl BackendStorage for MetalStorage {
         let dims = shape.dims();
         let el = shape.elem_count();
         let dtype = t.dtype;
-        let buffer = self.device.new_buffer(el, dtype);
+        let buffer = self.device.new_buffer(el, dtype, "where");
         let command_buffer = self.device.command_buffer();
         if t.dtype() != f.dtype() {
             crate::bail!("Invalid ternary different dtypes for values");
@@ -752,7 +819,7 @@ impl BackendStorage for MetalStorage {
         let dst_el = ids_el * left_size * right_size;
         let dtype = self.dtype;
         let device = self.device();
-        let buffer = device.new_buffer(dst_el, dtype);
+        let buffer = device.new_buffer(dst_el, dtype, "index_select");
         let name = match (ids.dtype, self.dtype) {
             (DType::U32, DType::F32) => "is_u32_f32",
             (DType::U32, DType::F16) => "is_u32_f16",
@@ -786,7 +853,6 @@ impl BackendStorage for MetalStorage {
     ) -> Result<Self> {
         crate::bail!("index_add metal")
     }
-
     fn matmul(
         &self,
         rhs: &Self,
@@ -794,81 +860,33 @@ impl BackendStorage for MetalStorage {
         lhs_l: &Layout,
         rhs_l: &Layout,
     ) -> Result<Self> {
-        // Create descriptors
+        let buffer = self.device.new_buffer(b * m * n, self.dtype, "matmul");
-        let (type_id, size, name) = match self.dtype {
+        let name = match self.dtype {
-            DType::F32 => (
+            DType::F32 => "sgemm",
-                metal::mps::MPS_FLOATBIT_ENCODING | 32,
+            DType::F16 => "hgemm",
-                core::mem::size_of::<f32>() as NSUInteger,
+            dtype => {
-                "sgemm",
+                return Err(MetalError::Message(format!("matmul doesn't support {dtype:?}")).into())
-            ),
+            }
-            DType::F16 => (
-                metal::mps::MPS_FLOATBIT_ENCODING | 16,
-                core::mem::size_of::<f16>() as NSUInteger,
-                "hgemm",
-            ),
-            dtype => todo!("Dtype for matmul {dtype:?} is not supported"),
         };
 
-        let lhs_stride = lhs_l.stride();
-        let rhs_stride = rhs_l.stride();
-        let rhs_m1 = rhs_stride[rhs_stride.len() - 1];
-        let rhs_m2 = rhs_stride[rhs_stride.len() - 2];
-        let lhs_m1 = lhs_stride[lhs_stride.len() - 1];
-        let lhs_m2 = lhs_stride[lhs_stride.len() - 2];
-        // The a tensor has dims batching, k, n (rhs)
-        let transpose_left = if lhs_m1 == 1 && lhs_m2 == k {
-            false
-        } else if lhs_m1 == m && lhs_m2 == 1 {
-            true
-        } else {
-            Err(MetalError::MatMulNonContiguous {
-                lhs_stride: lhs_stride.to_vec(),
-                rhs_stride: rhs_stride.to_vec(),
-                mnk: (m, n, k),
-            })?
-        };
-        let transpose_right = if rhs_m1 == 1 && rhs_m2 == n {
-            false
-        } else if rhs_m1 == k && rhs_m2 == 1 {
-            true
-        } else {
-            Err(MetalError::MatMulNonContiguous {
-                lhs_stride: lhs_stride.to_vec(),
-                rhs_stride: rhs_stride.to_vec(),
-                mnk: (m, n, k),
-            })?
-        };
-
-        let result_buffer = self.device.new_buffer(b * m * n, self.dtype);
-
         let command_buffer = self.device.command_buffer();
-
+        command_buffer.set_label("matmul");
-        command_buffer.set_label("mfa gemm");
+        candle_metal_kernels::call_gemm(
-
-        candle_metal_kernels::call_mfa_gemm(
             &self.device.device,
             &command_buffer,
             &self.device.kernels,
             name,
-            &self.buffer,
-            lhs_l.shape().dims(),
-            &rhs.buffer,
-            rhs_l.shape().dims(),
-            &result_buffer,
             (b, m, n, k),
-            transpose_left,
+            &lhs_l.stride(),
-            transpose_right,
+            lhs_l.start_offset() * self.dtype.size_in_bytes(),
+            &self.buffer,
+            &rhs_l.stride(),
+            rhs_l.start_offset() * rhs.dtype.size_in_bytes(),
+            &rhs.buffer,
+            &buffer,
         )
         .map_err(MetalError::from)?;
-
+        Ok(Self::new(buffer, self.device.clone(), self.dtype()))
-        drop(command_buffer);
-        self.device.commit();
-
-        Ok(Self::new(
-            self.buffer.clone(),
-            self.device.clone(),
-            self.dtype(),
-        ))
     }
 
     fn copy_strided_src(&self, dst: &mut Self, dst_offset: usize, src_l: &Layout) -> Result<()> {
@@ -876,15 +894,11 @@ impl BackendStorage for MetalStorage {
         if src_l.is_contiguous() && self.dtype == dst.dtype() {
             command_buffer.set_label("copy_contiguous");
             let blit = command_buffer.new_blit_command_encoder();
+            blit.set_label("copy_contiguous");
             let src_offset = (src_l.start_offset() * self.dtype.size_in_bytes()) as NSUInteger;
+            let length = (src_l.shape().elem_count() * self.dtype.size_in_bytes()) as NSUInteger;
             let dst_offset = (dst_offset * dst.dtype().size_in_bytes()) as NSUInteger;
-            blit.copy_from_buffer(
+            blit.copy_from_buffer(&self.buffer, src_offset, dst.buffer(), dst_offset, length);
-                &self.buffer,
-                src_offset,
-                dst.buffer(),
-                dst_offset,
-                self.buffer.length() - src_offset,
-            );
             blit.end_encoding();
         } else {
             let src_shape = src_l.shape();
@@ -915,54 +929,22 @@ impl BackendStorage for MetalStorage {
             .map_err(MetalError::from)?;
             command_buffer.set_label("copy_strided");
         }
-        drop(command_buffer);
-        self.device.commit();
         Ok(())
     }
 }
 
 impl MetalStorage {
     pub fn new(buffer: Arc<Buffer>, device: MetalDevice, dtype: DType) -> Self {
-        let matrices = Arc::new(RwLock::new(HashMap::new()));
         Self {
             buffer,
             device,
             dtype,
-            matrices,
         }
     }
 
     pub fn buffer(&self) -> &Buffer {
         &self.buffer
     }
-
-    fn matrix(
-        &self,
-        (b, m, n): (NSUInteger, NSUInteger, NSUInteger),
-        transpose: bool,
-        size: NSUInteger,
-        offset: NSUInteger,
-        type_id: u32,
-    ) -> Result<Matrix> {
-        let key = (b, m, n, transpose, size, offset, type_id);
-
-        let mut matrices = self.matrices.try_write().unwrap();
-        if let Some(matrix) = matrices.get(&key) {
-            Ok(matrix.clone())
-        } else {
-            let descriptor = if transpose {
-                MatrixDescriptor::init_multiple(n, m, b, m * size, m * n * size, type_id)
-            } else {
-                MatrixDescriptor::init_multiple(m, n, b, n * size, m * n * size, type_id)
-            };
-            let matrix = Matrix::init_with_buffer_descriptor(&self.buffer, offset, &descriptor)
-                .ok_or_else(|| {
-                    MetalError::from("Failed to create matrix multiplication kernel".to_string())
-                })?;
-            matrices.insert(key, matrix.clone());
-            Ok(matrix)
-        }
-    }
 }
 
 impl BackendDevice for MetalDevice {
@@ -971,14 +953,28 @@ impl BackendDevice for MetalDevice {
     fn new(ordinal: usize) -> Result<Self> {
         let device = metal::Device::all().swap_remove(ordinal);
 
+        let n = 1;
         let command_queue = device.new_command_queue();
-        let command_buffer = Arc::new(RwLock::new(command_queue.new_command_buffer().to_owned()));
+
-        let kernels = Arc::new(Kernels::new());
+        let command_buffers = (0..n)
+            .map(|i| {
+                let command_buffer = command_queue.new_command_buffer().to_owned();
+                command_buffer.enqueue();
+                command_buffer.set_label(&format!("num {i}"));
+                command_buffer
+            })
+            .collect();
+        let command_buffers = Arc::new(RwLock::new(command_buffers));
+        let command_buffer_index = Arc::new(RwLock::new(0));
+        let fence = device.new_fence();
+        let kernels = Arc::new(Kernels::new(fence.clone()));
         let buffers = Arc::new(RwLock::new(HashMap::new()));
         Ok(Self {
             device,
+            fence,
             command_queue,
-            command_buffer,
+            command_buffers,
+            command_buffer_index,
             buffers,
             kernels,
         })
@@ -999,7 +995,21 @@ impl BackendDevice for MetalDevice {
     }
 
     fn zeros_impl(&self, shape: &Shape, dtype: DType) -> Result<MetalStorage> {
-        let buffer = self.new_buffer(shape.elem_count(), dtype);
+        let buffer = self.new_buffer(shape.elem_count(), dtype, "zeros");
+        let command_buffer = self.command_buffer();
+        command_buffer.set_label("zeros");
+        let blit = command_buffer.new_blit_command_encoder();
+        blit.wait_for_fence(&self.fence);
+        blit.fill_buffer(
+            &buffer,
+            metal::NSRange {
+                location: 0,
+                length: buffer.length(),
+            },
+            0,
+        );
+        blit.update_fence(&self.fence);
+        blit.end_encoding();
         Ok(MetalStorage::new(buffer, self.clone(), dtype))
     }
 
@@ -1050,3 +1060,10 @@ impl BackendDevice for MetalDevice {
         self.storage_from_cpu_storage(&cpu_storage)
     }
 }
+
+fn read_to_vec<T: Clone>(buffer: &Buffer, n: usize) -> Vec<T> {
+    let ptr = buffer.contents() as *const T;
+    assert!(!ptr.is_null());
+    let slice = unsafe { std::slice::from_raw_parts(ptr, n) };
+    slice.to_vec()
+}

candle-core/src/tensor.rs

@@ -1864,7 +1864,7 @@ impl Tensor {
                 }
                 (Storage::Cuda(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
                 (Storage::Metal(storage), Device::Cpu) => {
-                    println!("{storage:?} - {:?}", storage.to_cpu_storage()?);
+                    // println!("{storage:?} - {:?}", storage.to_cpu_storage()?);
                     Storage::Cpu(storage.to_cpu_storage()?)
                 }
                 (Storage::Cuda(storage), Device::Cuda(cuda)) => {

candle-core/tests/tensor_tests.rs

@@ -900,7 +900,9 @@ fn matmul(device: &Device) -> Result<()> {
     let b = Tensor::from_slice(&data, (2, 2), device)?;
 
     let c = a.matmul(&b)?;
+    let d = a.matmul(&c)?;
     assert_eq!(c.to_vec2::<f32>()?, &[[7.0f32, 10.0], [15.0, 22.0]]);
+    assert_eq!(d.to_vec2::<f32>()?, &[[37.0, 54.0], [81.0, 118.0]]);
 
     let data = vec![1.0f32, 2.0];
     let a = Tensor::from_slice(&data, (2, 1), device)?;

candle-metal-kernels/Cargo.toml

@@ -10,8 +10,7 @@ categories = ["science"]
 license = "MIT OR Apache-2.0"
 
 [dependencies]
-metal = { version = "0.27.1", features = ["mps"], package="candle-metal" }
+metal = { version = "0.27.0", features = ["mps"]}
-metal-flash-attention = { path = "../../../metal-flash-attention" }
 once_cell = "1.18.0"
 thiserror = "1"
 tracing = "0.1.37"

candle-metal-kernels/src/affine.metal

@@ -29,9 +29,7 @@ kernel void FN_NAME( \
     if (id >= dim) { \
         return; \
     } \
-    const TYPENAME m = TYPENAME(mul); \
+    output[id] = TYPENAME(float(input[id]) * mul + add); \
-    const TYPENAME a = TYPENAME(add); \
-    output[id] = input[id] * m + a; \
 } \
 kernel void FN_NAME##_strided( \
     constant size_t &dim, \
@@ -47,15 +45,80 @@ kernel void FN_NAME##_strided( \
     if (id >= dim) { \
         return; \
     } \
-    const TYPENAME m = TYPENAME(mul); \
+    output[id] = TYPENAME(float(input[get_strided_index(id, num_dims, dims, strides)]) * mul + add); \
-    const TYPENAME a = TYPENAME(add); \
+}
-    output[id] = input[get_strided_index(id, num_dims, dims, strides)] * m + a; \
+
+#define POWF(FN_NAME, TYPENAME) \
+kernel void FN_NAME( \
+    constant size_t &dim, \
+    constant float &mul, \
+    device const TYPENAME *input,  \
+    device TYPENAME *output, \
+    uint id [[ thread_position_in_grid ]] \
+) { \
+    if (id >= dim) { \
+        return; \
     } \
+    output[id] = TYPENAME(pow(input[id], TYPENAME(mul))); \
+} \
+kernel void FN_NAME##_strided( \
+    constant size_t &dim, \
+    constant size_t &num_dims, \
+    constant size_t *dims, \
+    constant size_t *strides, \
+    constant float &mul, \
+    device const TYPENAME *input,  \
+    device TYPENAME *output, \
+    uint id [[ thread_position_in_grid ]] \
+) { \
+    if (id >= dim) { \
+        return; \
+    } \
+    output[id] = TYPENAME(pow(input[get_strided_index(id, num_dims, dims, strides)], TYPENAME(mul))); \
+}
+
+#define ELU(FN_NAME, TYPENAME) \
+kernel void FN_NAME( \
+    constant size_t &dim, \
+    constant float &mul, \
+    device const TYPENAME *input,  \
+    device TYPENAME *output, \
+    uint id [[ thread_position_in_grid ]] \
+) { \
+    if (id >= dim) { \
+        return; \
+    } \
+    const TYPENAME x = input[id]; \
+    output[id] = TYPENAME((x > 0)?x: mul * exp(x - 1)); \
+} \
+kernel void FN_NAME##_strided( \
+    constant size_t &dim, \
+    constant size_t &num_dims, \
+    constant size_t *dims, \
+    constant size_t *strides, \
+    constant float &mul, \
+    device const TYPENAME *input,  \
+    device TYPENAME *output, \
+    uint id [[ thread_position_in_grid ]] \
+) { \
+    if (id >= dim) { \
+        return; \
+    } \
+    const TYPENAME x = input[get_strided_index(id, num_dims, dims, strides)]; \
+    output[id] = TYPENAME((x > 0)?x: mul * exp(x - 1)); \
+} \
+
 
 AFFINE(affine_float, float)
 AFFINE(affine_half, half)
+POWF(powf_float, float)
+POWF(powf_half, half)
+ELU(elu_float, float)
+ELU(elu_half, half)
 
 
 #if __METAL_VERSION__ >= 310
 AFFINE(affine_bfloat, bfloat);
+POWF(powf_bfloat, bfloat);
+ELU(elu_bfloat, bfloat);
 #endif

candle-metal-kernels/src/reduce.metal

@@ -18,7 +18,7 @@ METAL_FUNC uint get_strided_index(
     return strided_i;
 }
 
-constant int THREADGROUP_SIZE = 1024;
+constant int THREADGROUP_SIZE = 2048;
 
 # define REDUCE(FN, NAME, T) \
 kernel void NAME( \
@@ -32,7 +32,7 @@ kernel void NAME( \
     uint block_dim [[ threads_per_threadgroup ]] \
 ) { \
      \
-   threadgroup float shared_memory[THREADGROUP_SIZE]; \
+   threadgroup T shared_memory[THREADGROUP_SIZE]; \
       \
    shared_memory[tid] = 0; \
    /* \
@@ -93,12 +93,13 @@ kernel void NAME(
     size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);            \
     size_t idx = start_idx + tid;                                                 \
                                                                                   \
-    threadgroup_barrier(mem_flags::mem_threadgroup);                              \
                                                                                   \
+    float tmp = -INFINITY; \
     while (idx < stop_idx) {                                                      \
-        shared_memory[tid] = MAX(shared_memory[tid], src[idx]);                   \
+        tmp = MAX(tmp, float(src[idx]));                   \
         idx += block_dim;                                                         \
     }                                                                             \
+    shared_memory[tid] = tmp; \
                                                                                   \
     threadgroup_barrier(mem_flags::mem_threadgroup);                              \
                                                                                   \
@@ -106,21 +107,26 @@ kernel void NAME(
         if (tid < s) {                                                            \
             shared_memory[tid] = MAX(shared_memory[tid], shared_memory[tid + s]); \
         }                                                                         \
+        threadgroup_barrier(mem_flags::mem_threadgroup);                              \
     }                                                                             \
                                                                                   \
+    /* wait for shared_memory[0] to be filled */ \
     threadgroup_barrier(mem_flags::mem_threadgroup);                              \
                                                                                   \
     float _max = shared_memory[0];                                                    \
                                                                                   \
+    /* prevent tid=0 from overwriting _max before other threads have written it */ \
+    threadgroup_barrier(mem_flags::mem_threadgroup);                              \
     shared_memory[tid] = 0;                                                       \
                                                                                   \
     idx = start_idx + tid;                                                        \
     while (idx < stop_idx) {                                                      \
-        const T val = T(exp(src[idx] - _max));                                    \
+        const float val = exp(float(src[idx]) - _max);                                    \
-        dst[idx] = val;                                                           \
+        dst[idx] = T(val);                                                           \
         shared_memory[tid] += val;                                                \
         idx += block_dim;                                                         \
     }                                                                             \
+    threadgroup_barrier(mem_flags::mem_threadgroup);                              \
     for (uint s = block_dim / 2; s > 0; s >>= 1) {                                \
         if (tid < s) {                                                            \
             shared_memory[tid] += shared_memory[tid + s];                         \
@@ -128,7 +134,7 @@ kernel void NAME(
         threadgroup_barrier(mem_flags::mem_threadgroup);                              \
     }                                                                             \
                                                                                   \
-    const T inv_acc = T(1/shared_memory[0]);                                         \
+    const T inv_acc = T(1.0/shared_memory[0]);                                         \
     idx = start_idx + tid;                                                        \
     while (idx < stop_idx) {                                                      \
         dst[idx] *= inv_acc;                                                      \

candle-metal-kernels/src/test.swift

@@ -0,0 +1,209 @@
+
+import Metal
+import MetalPerformanceShadersGraph
+
+
+
+let type = MTLDataType.float;
+let dataType = type;
+var B = 2;
+var M = 2;
+var N = 2;
+var K = 2;
+var A_trans = false;
+var B_trans = false;
+var D_trans = false;
+var alpha = Float(1.0);
+var beta = Float(0.0);
+var batched = B > 1;
+var fused_activation = false;
+var fused_bias = false;
+let constants = MTLFunctionConstantValues()
+constants.setConstantValue(&M, type: .uint, index: 0)
+constants.setConstantValue(&N, type: .uint, index: 1)
+constants.setConstantValue(&K, type: .uint, index: 2)
+constants.setConstantValue(&A_trans, type: .bool, index: 10)
+constants.setConstantValue(&B_trans, type: .bool, index: 11)
+constants.setConstantValue(&D_trans, type: .bool, index: 13)
+constants.setConstantValue(&alpha, type: .float, index: 20)
+constants.setConstantValue(&beta, type: .float, index: 21)
+constants.setConstantValue(&batched, type: .bool, index: 100)
+constants.setConstantValue(&fused_activation, type: .bool, index: 101)
+constants.setConstantValue(&fused_bias, type: .bool, index: 50001)
+
+
+var M_simd = UInt16(16)
+var N_simd = UInt16(16)
+var K_simd = UInt16(32)
+var M_splits = UInt16(2)
+var N_splits = UInt16(2)
+constants.setConstantValue(&M_simd, type: .ushort, index: 200)
+constants.setConstantValue(&N_simd, type: .ushort, index: 201)
+constants.setConstantValue(&K_simd, type: .ushort, index: 202)
+constants.setConstantValue(&M_splits, type: .ushort, index: 210)
+constants.setConstantValue(&N_splits, type: .ushort, index: 211)
+
+let M_group = M_simd * M_splits
+let N_group = N_simd * N_splits
+
+// Satisfy Metal API validation.
+#if DEBUG
+do {
+  var garbage: SIMD4<UInt64> = .zero
+  constants.setConstantValue(&garbage, type: .bool, index: 102)
+  constants.setConstantValue(&garbage, type: .bool, index: 103)
+  constants.setConstantValue(&garbage, type: .bool, index: 113)
+  constants.setConstantValue(&garbage, type: .bool, index: 50000)
+}
+#endif
+
+let device = MTLCopyAllDevices().first!
+device.shouldMaximizeConcurrentCompilation = true
+
+var libraryURL = URL.init(string: "/Users/nicolas/src/candle/candle-metal-kernels/")!;
+libraryURL.append(component: "src")
+libraryURL.append(component: "libMetalFlashAttention.metallib")
+let library = try! device.makeLibrary(URL: libraryURL)
+
+var name: String
+    switch dataType {
+    case .half: name = "hgemm"
+    case .float: name = "sgemm"
+    default: fatalError()
+    }
+let function = try! library.makeFunction(
+  name: name, constantValues: constants)
+
+let A_block_length = M_group * K_simd
+let B_block_length = K_simd * N_group
+
+var blockElements = A_block_length + B_block_length;
+if (M % 8 != 0) && (N % 8 != 0) {
+  let C_block_length = M_group * N_group;
+  blockElements = max(C_block_length, blockElements)
+}
+if fused_bias {
+  if D_trans {
+    blockElements = max(blockElements, M_group)
+  } else {
+    blockElements = max(blockElements, N_group)
+  }
+}
+// let blockBytes = blockElements * UInt16(dataType.size)
+let elementSize = 4
+let blockBytes = blockElements * UInt16(elementSize)
+
+func ceilDivide(target: Int, granularity: UInt16) -> Int {
+  (target + Int(granularity) - 1) / Int(granularity)
+}
+var gridSize = MTLSize(
+  width: ceilDivide(target: N, granularity: N_group),
+  height: ceilDivide(target: M, granularity: M_group),
+  depth: 1)
+let groupSize = MTLSize(
+  width: Int(32 * M_splits * N_splits),
+  height: 1,
+  depth: 1)
+
+let commandQueue = device.makeCommandQueue()!
+
+let threadgroupMemoryLength = blockBytes;
+
+let rowsA = M;
+let columnsA = K;
+let rowsB = K;
+let columnsB = N;
+let rowsC = M;
+let columnsC = N;
+var arrayA = [Float](repeating: 0, count: B * rowsA * columnsA)
+
+var arrayB = [Float](repeating: 0, count: B * rowsB * columnsB)
+
+var arrayC = [Float](repeating: 0, count: B * rowsC * columnsC)
+var arrayD = [Float](repeating: 0, count: B * rowsC * columnsC)
+for i in 0..<arrayA.count {
+  arrayA[i] = Float(i)
+}
+
+for i in 0..<arrayB.count {
+  arrayB[i] = Float(i)
+}
+
+let bufferA = device.makeBuffer(bytes: arrayA, length: B * rowsA * columnsA * MemoryLayout<Float>.stride, options: [])!
+
+let bufferB = device.makeBuffer(bytes: arrayB, length: B * rowsB * columnsB * MemoryLayout<Float>.stride, options: [])!
+
+let bufferC = device.makeBuffer(length: B * rowsC * columnsC * MemoryLayout<Float>.stride, options: [])!
+let bufferD = device.makeBuffer(length: B * rowsC * columnsC * MemoryLayout<Float>.stride, options: [])!
+
+
+let pipeline = try device.makeComputePipelineState(function: function)
+
+func call(bufferA: MTLBuffer, bufferB: MTLBuffer, bufferC: MTLBuffer){ 
+    let encoder = commandBuffer.makeComputeCommandEncoder(dispatchType: MTLDispatchType.serial)!
+    encoder.setComputePipelineState(pipeline)
+    encoder.setThreadgroupMemoryLength(Int(threadgroupMemoryLength), index: 0)
+
+    encoder.setBuffer(bufferA, offset: 0, index: 0)
+    encoder.setBuffer(bufferB, offset: 0, index: 1)
+    encoder.setBuffer(bufferC, offset: 0, index: 2)
+    let gridZ: Int = B
+    if batched{
+      func byteStride(shape: [Int]) -> Int {
+        let rank = shape.count
+        var output = elementSize * shape[rank - 2] * shape[rank - 1]
+        if shape.dropLast(2).reduce(1, *) == 1 {
+          output = 0
+        }
+        return output
+      }
+      let byteStrideA = M*K*elementSize
+      let byteStrideB = N*K*elementSize
+      let byteStrideC = M*N*elementSize
+      
+      let byteStrideD = 0
+      withUnsafeTemporaryAllocation(
+        of: SIMD4<UInt64>.self, capacity: gridZ
+      ) { buffer in
+        for i in 0..<buffer.count {
+          buffer[i] = SIMD4(
+            UInt64(truncatingIfNeeded: i * byteStrideA),
+            UInt64(truncatingIfNeeded: i * byteStrideB),
+            UInt64(truncatingIfNeeded: i * byteStrideC),
+            UInt64(truncatingIfNeeded: i * byteStrideD))
+        }
+        
+        let bufferLength = buffer.count * MemoryLayout<SIMD4<UInt64>>.stride
+        assert(MemoryLayout<SIMD4<UInt64>>.stride == 8 * 4)
+        encoder.setBytes(buffer.baseAddress!, length: bufferLength, index: 10)
+      }
+    }
+    gridSize.depth = gridZ
+
+
+    encoder.dispatchThreadgroups(
+      gridSize, threadsPerThreadgroup: groupSize
+    )
+    encoder.endEncoding()
+}
+
+var commandBuffer = commandQueue.makeCommandBuffer()!
+call(bufferA:bufferA, bufferB:bufferB, bufferC:bufferC)
+commandBuffer.commit()
+commandBuffer = commandQueue.makeCommandBuffer()!
+commandBuffer.encodeWaitForEvent(event, value: 2)
+call(bufferA:bufferA, bufferB:bufferC, bufferC:bufferD)
+commandBuffer.commit()
+
+commandBuffer.waitUntilCompleted()
+var contents = bufferC.contents();
+var count = B * rowsA * columnsB;
+var typedPointer = contents.bindMemory(to: Float.self, capacity: count)
+var bufferedPointer = UnsafeBufferPointer(start: typedPointer, count: count)
+print("First matmul is OK", Array(bufferedPointer))
+
+contents = bufferD.contents();
+count = B * rowsA * columnsB;
+typedPointer = contents.bindMemory(to: Float.self, capacity: count)
+bufferedPointer = UnsafeBufferPointer(start: typedPointer, count: count)
+print("This should be filled", Array(bufferedPointer))

candle-metal-kernels/src/tests.rs

@@ -2,6 +2,13 @@ use super::*;
 use half::{bf16, f16};
 use metal::{CompileOptions, Device, MTLResourceOptions, MTLSize, NSUInteger};
 
+fn read_to_vec<T: Clone>(buffer: &Buffer, n: usize) -> Vec<T> {
+    let ptr = buffer.contents() as *const T;
+    assert!(!ptr.is_null());
+    let slice = unsafe { std::slice::from_raw_parts(ptr, n) };
+    slice.to_vec()
+}
+
 fn new_buffer<T>(device: &Device, data: &[T]) -> Buffer {
     let options = MTLResourceOptions::StorageModeManaged;
     let ptr = data.as_ptr() as *const core::ffi::c_void;
@@ -30,7 +37,8 @@ fn approx_bf16(v: Vec<bf16>, digits: i32) -> Vec<f32> {
 
 fn run<T: Clone>(v: &[T], name: unary::contiguous::Kernel) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let input = new_buffer(&device, v);
@@ -47,12 +55,13 @@ fn run<T: Clone>(v: &[T], name: unary::contiguous::Kernel) -> Vec<T> {
     .unwrap();
     command_buffer.commit();
     command_buffer.wait_until_completed();
-    output.read_to_vec::<T>(v.len())
+    read_to_vec(&output, v.len())
 }
 
 fn run_binary<T: Clone>(x: &[T], y: &[T], name: binary::contiguous::Kernel) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let options = MTLResourceOptions::StorageModeManaged;
@@ -72,7 +81,7 @@ fn run_binary<T: Clone>(x: &[T], y: &[T], name: binary::contiguous::Kernel) -> V
     .unwrap();
     command_buffer.commit();
     command_buffer.wait_until_completed();
-    output.read_to_vec::<T>(x.len())
+    read_to_vec(&output, x.len())
 }
 
 fn run_strided<T: Clone>(
@@ -87,7 +96,8 @@ fn run_strided<T: Clone>(
     let command_buffer = command_queue.new_command_buffer();
     let input = new_buffer(&device, v);
     let output = new_buffer(&device, v);
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     call_unary_strided(
         &device,
         command_buffer,
@@ -103,7 +113,7 @@ fn run_strided<T: Clone>(
     .unwrap();
     command_buffer.commit();
     command_buffer.wait_until_completed();
-    output.read_to_vec::<T>(v.len())
+    read_to_vec(&output, v.len())
 }
 
 #[test]
@@ -205,6 +215,25 @@ fn cos_strided_random() {
     );
 }
 
+#[test]
+fn gelu_f16() {
+    let v: Vec<f16> = [-10f32, -1.0, 0., 1., 2., 3., 10.0, 20.0]
+        .iter()
+        .map(|v| f16::from_f32(*v))
+        .collect();
+    let expected: Vec<f32> = vec![-0.0, -0.16, 0.0, 0.84, 1.96, 3.0, 10.0, 20.0];
+    let results = run(&v, unary::contiguous::gelu::HALF);
+    assert_eq!(approx_f16(results, 2), expected);
+}
+
+#[test]
+fn gelu_f32() {
+    let v: Vec<f32> = vec![-10f32, -1.0, 0., 1., 2., 3., 10.0, 20.0];
+    let expected: Vec<f32> = vec![-0.0, -0.159, 0.0, 0.841, 1.955, 2.996, 10.0, 20.0];
+    let results = run(&v, unary::contiguous::gelu::FLOAT);
+    assert_eq!(approx(results, 3), expected);
+}
+
 #[test]
 fn binary_add_f32() {
     let left = vec![1.0f32, 2.0, 3.0];
@@ -221,7 +250,8 @@ fn binary_add_f32() {
 
 fn cast<T: Clone, U: Clone>(v: &[T], name: &'static str) -> Vec<U> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let input = new_buffer(&device, v);
@@ -242,7 +272,7 @@ fn cast<T: Clone, U: Clone>(v: &[T], name: &'static str) -> Vec<U> {
     .unwrap();
     command_buffer.commit();
     command_buffer.wait_until_completed();
-    output.read_to_vec::<U>(v.len())
+    read_to_vec(&output, v.len())
 }
 
 #[test]
@@ -268,7 +298,8 @@ fn cast_u32_f32() {
 
 fn run_affine<T: Clone>(v: &[T], mul: f64, add: f64) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
 
@@ -292,7 +323,7 @@ fn run_affine<T: Clone>(v: &[T], mul: f64, add: f64) -> Vec<T> {
     command_buffer.commit();
     command_buffer.wait_until_completed();
 
-    output.read_to_vec::<T>(v.len())
+    read_to_vec(&output, v.len())
 }
 
 fn run_affine_strided<T: Clone>(
@@ -303,7 +334,8 @@ fn run_affine_strided<T: Clone>(
     add: f64,
 ) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
 
@@ -328,7 +360,7 @@ fn run_affine_strided<T: Clone>(
     command_buffer.wait_until_completed();
 
     let len: usize = shape.iter().product();
-    output.read_to_vec::<T>(len)
+    read_to_vec(&output, len)
 }
 
 #[test]
@@ -431,7 +463,8 @@ fn run_index_select<T: Clone, I: Clone + std::fmt::Debug>(
         _ => unimplemented!(),
     };
 
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     call_index_select(
         &device,
         &command_buffer,
@@ -449,7 +482,7 @@ fn run_index_select<T: Clone, I: Clone + std::fmt::Debug>(
     command_buffer.commit();
     command_buffer.wait_until_completed();
 
-    dst_buffer.read_to_vec::<T>(dst_el)
+    read_to_vec(&dst_buffer, dst_el)
 }
 
 #[test]
@@ -515,7 +548,7 @@ fn index_add() {
     let expected = vec![
         2.0, 3.0, 4.0, 1.0, 1.0, 1.0, 8.0, 9.0, 10.0, 1.0, 1.0, 1.0, 5.0, 6.0, 7.0,
     ];
-    let result = outputs_buffer.read_to_vec::<f32>(right.len());
+    let result: Vec<f32> = read_to_vec(&outputs_buffer, right.len());
     assert_eq!(result, expected);
 }
 
@@ -527,13 +560,14 @@ fn cos_f16() {
         .collect();
     let results = run(&v, unary::contiguous::cos::HALF);
     let expected: Vec<f16> = v.iter().map(|v| f16::from_f32(v.to_f32().cos())).collect();
-    assert_eq!(approx_f16(results, 4), vec![0.5405, -0.4163, -0.9902]);
+    assert_eq!(approx_f16(results, 2), vec![0.54, -0.42, -0.99]);
-    assert_eq!(approx_f16(expected, 4), vec![0.5405, -0.4163, -0.9902]);
+    assert_eq!(approx_f16(expected, 2), vec![0.54, -0.42, -0.99]);
 }
 
 fn run_reduce<T: Clone>(v: &[T], out_length: usize, name: &'static str) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let input = new_buffer(&device, v);
@@ -555,12 +589,13 @@ fn run_reduce<T: Clone>(v: &[T], out_length: usize, name: &'static str) -> Vec<T
     command_buffer.commit();
     command_buffer.wait_until_completed();
 
-    output.read_to_vec::<T>(out_length)
+    read_to_vec(&output, out_length)
 }
 
 fn run_softmax<T: Clone + std::fmt::Debug>(v: &[T], last_dim: usize, name: &'static str) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let input = new_buffer(&device, v);
@@ -579,7 +614,7 @@ fn run_softmax<T: Clone + std::fmt::Debug>(v: &[T], last_dim: usize, name: &'sta
     command_buffer.commit();
     command_buffer.wait_until_completed();
 
-    output.read_to_vec::<T>(v.len())
+    read_to_vec(&output, v.len())
 }
 
 #[test]
@@ -610,6 +645,24 @@ fn softmax() {
         vec![0.0043, 0.0116, 0.0315, 0.0858, 0.2331, 0.6337]
     );
 
+    let last_dim = 4096;
+    let n = 200;
+    let mut v = vec![0.0; n * last_dim];
+    for i in 0..n {
+        v[i * last_dim] = 20.0;
+    }
+    let results = run_softmax(&v, last_dim, "softmax_float");
+    let results = approx(results, 4);
+    println!("{results:?}");
+    assert_eq!(
+        results.iter().map(|&s| s.round() as usize).sum::<usize>(),
+        n
+    );
+    assert_eq!(results[0], 1.0);
+    assert_eq!(results[1], 0.0);
+    assert_eq!(results[last_dim], 1.0);
+    assert_eq!(results[2 * last_dim], 1.0);
+
     let v = vec![0.0f32, 1.0, 2.0, 3.0, 4.0, 5.0];
     let last_dim = 6;
     let results = run_softmax(&v, last_dim, "softmax_float");
@@ -660,7 +713,8 @@ fn run_where_cond<I: Clone, T: Clone>(
     name: &'static str,
 ) -> Vec<T> {
     let device = device();
-    let kernels = Kernels::new();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
     let command_queue = device.new_command_queue();
     let command_buffer = command_queue.new_command_buffer();
     let options = MTLResourceOptions::StorageModeManaged;
@@ -701,7 +755,7 @@ fn run_where_cond<I: Clone, T: Clone>(
     command_buffer.commit();
     command_buffer.wait_until_completed();
 
-    output.read_to_vec::<T>(length)
+    read_to_vec(&output, length)
 }
 
 #[test]
@@ -725,3 +779,93 @@ fn where_cond() {
     );
     assert_eq!(approx(results, 4), vec![-1.0f32, 2.0, -3.0, -4.0, 5.0, 6.0]);
 }
+
+fn run_gemm<T: Clone>(
+    (b, m, n, k): (usize, usize, usize, usize),
+    lhs: &[T],
+    lhs_stride: Vec<usize>,
+    lhs_offset: usize,
+    rhs: &[T],
+    rhs_stride: Vec<usize>,
+    rhs_offset: usize,
+) -> Vec<T> {
+    let device = device();
+    let fence = device.new_fence();
+    let kernels = Kernels::new(fence);
+    let command_queue = device.new_command_queue();
+    let command_buffer = command_queue.new_command_buffer();
+    let options = MTLResourceOptions::StorageModeManaged;
+
+    let lhs = device.new_buffer_with_data(
+        lhs.as_ptr() as *const core::ffi::c_void,
+        std::mem::size_of_val(lhs) as u64,
+        options,
+    );
+    let rhs = device.new_buffer_with_data(
+        rhs.as_ptr() as *const core::ffi::c_void,
+        std::mem::size_of_val(rhs) as u64,
+        options,
+    );
+    let length = b * m * n;
+    let output = device.new_buffer((length * core::mem::size_of::<T>()) as u64, options);
+    call_gemm(
+        &device,
+        command_buffer,
+        &kernels,
+        "sgemm",
+        (b, m, n, k),
+        &lhs_stride,
+        lhs_offset,
+        &lhs,
+        &rhs_stride,
+        rhs_offset,
+        &rhs,
+        &output,
+    )
+    .unwrap();
+    command_buffer.commit();
+    command_buffer.wait_until_completed();
+
+    read_to_vec(&output, length)
+}
+
+#[test]
+fn gemm() {
+    let (b, m, n, k) = (1, 2, 4, 3);
+    let lhs_stride = vec![m * k, k, 1];
+    let lhs: Vec<f32> = (0..b * m * k).map(|f| f as f32).collect();
+    let rhs_stride = vec![n * k, n, 1];
+    let rhs: Vec<f32> = (0..b * n * k).map(|f| f as f32).collect();
+    let results = run_gemm((b, m, n, k), &lhs, lhs_stride, 0, &rhs, rhs_stride, 0);
+    assert_eq!(
+        approx(results, 4),
+        vec![20.0, 23.0, 26.0, 29.0, 56.0, 68.0, 80.0, 92.0]
+    );
+
+    let (b, m, n, k) = (2, 2, 4, 3);
+    let lhs_stride = vec![m * k, k, 1];
+    let lhs: Vec<f32> = (0..b * m * k).map(|f| f as f32).collect();
+    let rhs_stride = vec![n * k, n, 1];
+    let rhs: Vec<f32> = (0..b * n * k).map(|f| f as f32).collect();
+    let results = run_gemm((b, m, n, k), &lhs, lhs_stride, 0, &rhs, rhs_stride, 0);
+    assert_eq!(
+        approx(results, 4),
+        vec![
+            20.0, 23.0, 26.0, 29.0, 56.0, 68.0, 80.0, 92.0, 344.0, 365.0, 386.0, 407.0, 488.0,
+            518.0, 548.0, 578.0
+        ]
+    );
+
+    // OFFSET
+    let (b, m, n, k) = (2, 2, 4, 3);
+    let lhs_stride = vec![m * k, k, 1];
+    let lhs: Vec<f32> = (0..b * m * k).map(|f| f as f32).collect();
+    let rhs_stride = vec![n * k, n, 1];
+    let rhs: Vec<f32> = (0..b * n * k).map(|f| f as f32).collect();
+    // Manually set batch_size=1 and offset 12 elements * 4 the number of bytes for f32
+    let results = run_gemm((1, m, n, k), &lhs, lhs_stride, 0, &rhs, rhs_stride, 12 * 4);
+    assert_eq!(
+        approx(results, 4),
+        vec![56.0, 59.0, 62.0, 65.0, 200.0, 212.0, 224.0, 236.0]
+    );
+}

candle-metal-kernels/src/unary.metal

@@ -43,8 +43,13 @@ template <typename T> METAL_FUNC T erf(T in){
     return T(sign*y);
 }
 template <typename T> METAL_FUNC T id(T in) { return in; }
-template <typename T> METAL_FUNC T gelu_erf(T x){ return T(x * (1 + erf(x * M_SQRT1_2_F)) / 2); }
+template <typename T> METAL_FUNC T gelu_erf(T x) {
+    return T(x * (1 + erf(x * M_SQRT1_2_F)) / 2);
+}
 template <typename T> METAL_FUNC T gelu(T x) {
+    if (x > 5) {
+        return x;
+    }
     T x_sq = x * x;
     T x_cube = x_sq * x;
     T alpha = x + static_cast<T>(0.044715) * x_cube;
@@ -64,7 +69,7 @@ kernel void FN_NAME( \
     if (thread_position_in_grid >= dim) { \
         return; \
     } \
-    output[thread_position_in_grid] = TYPENAME(FN(input[thread_position_in_grid])); \
+    output[thread_position_in_grid] = TYPENAME(FN(float(input[thread_position_in_grid]))); \
 }\
 kernel void FN_NAME_STRIDED( \
     constant size_t &dim, \
@@ -78,7 +83,7 @@ kernel void FN_NAME_STRIDED( \
     if (thread_position_in_grid >= dim) { \
         return; \
     } \
-    output[thread_position_in_grid] = TYPENAME(FN(input[get_strided_index(thread_position_in_grid, num_dims, dims, strides)])); \
+    output[thread_position_in_grid] = TYPENAME(FN(float(input[get_strided_index(thread_position_in_grid, num_dims, dims, strides)]))); \
 }
 
 #define UNARY_OP(NAME) \
@@ -102,6 +107,7 @@ UNARY_OP(floor)
 UNARY_OP(round)
 UNARY_OP(gelu_erf)
 UNARY_OP(erf)
+UNARY_OP(tanh)
 UNARY(id, float, copy_float, copy_float_strided)
 UNARY(id, half, copy_half, copy_half_strided)
 UNARY(id, uint8_t, copy_u8, copy_u8_strided)
@@ -121,6 +127,7 @@ BFLOAT_UNARY_OP(floor)
 BFLOAT_UNARY_OP(round)
 BFLOAT_UNARY_OP(gelu_erf)
 BFLOAT_UNARY_OP(erf)
+BFLOAT_UNARY_OP(tanh)
 
 UNARY(id, bfloat, copy_bfloat, copy_bfloat_strided)
 #endif

candle-nn/Cargo.toml

@@ -19,6 +19,7 @@ num-traits = { workspace = true }
 rayon = { workspace = true }
 safetensors = { workspace = true }
 serde = { workspace = true }
+metal = { workspace = true, optional = true }
 candle-metal-kernels = { path = "../candle-metal-kernels", version = "0.3.0", optional = true }
 
 [dev-dependencies]
@@ -30,4 +31,4 @@ default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate"]
 cuda = ["candle/cuda"]
 mkl = ["dep:intel-mkl-src", "candle/mkl"]
-metal = ["candle/metal", "dep:candle-metal-kernels"]
+metal = ["candle/metal", "dep:candle-metal-kernels", "dep:metal"]

candle-nn/src/ops.rs

@@ -220,13 +220,13 @@ impl candle::CustomOp1 for SoftmaxLastDim {
         };
 
         let n = layout.stride().len();
-        if !(layout.stride()[n - 1] == 1 && layout.start_offset() == 0) {
+        if !(layout.is_contiguous() && layout.stride()[n - 1] == 1 && layout.start_offset() == 0) {
             candle::bail!("Non contiguous softmax-last-dim is not implemented");
         }
 
         let last_dim = layout.dims()[layout.shape().rank() - 1];
         let elem_count = layout.shape().elem_count();
-        let mut output = device.new_buffer(elem_count, storage.dtype());
+        let mut output = device.new_buffer(elem_count, storage.dtype(), "softmax");
         candle_metal_kernels::call_last_softmax(
             device.metal_device(),
             &command_buffer,

candle-transformers/Cargo.toml

@@ -31,3 +31,4 @@ accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate"]
 cuda = ["candle/cuda", "candle-nn/cuda"]
 flash-attn = ["cuda", "dep:candle-flash-attn"]
 mkl = ["dep:intel-mkl-src", "candle/mkl", "candle-nn/mkl"]
+metal = ["candle/metal", "candle-nn/metal"]

candle-transformers/src/models/mixformer.rs

@@ -142,10 +142,9 @@ impl RotaryEmbedding {
             .to_dtype(DType::F32)?
             .reshape((max_seq_len, 1))?;
         let freqs = t.matmul(&inv_freq)?;
-        Ok(Self {
+        let sin = freqs.sin()?;
-            sin: freqs.sin()?,
+        let cos = freqs.cos()?;
-            cos: freqs.cos()?,
+        Ok(Self { sin, cos })
-        })
     }
 
     fn apply_rotary_emb_qkv(
@@ -408,3 +407,38 @@ impl MixFormerSequentialForCausalLM {
         self.blocks.iter_mut().for_each(|b| b.clear_kv_cache())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    #[test]
+    fn test_rotary() {
+        let dev = Device::new_metal(0).unwrap();
+        for i in 0..10000 {
+            let dim = 8;
+            let max_seq_len = 12;
+            let inv_freq: Vec<_> = (0..dim)
+                .step_by(2)
+                .map(|i| 1f32 / 10000f32.powf(i as f32 / dim as f32))
+                .collect();
+            let inv_freq_len = inv_freq.len();
+            let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), &dev).unwrap();
+            let t = Tensor::arange(0u32, max_seq_len as u32, &dev)
+                .unwrap()
+                .to_dtype(DType::F32)
+                .unwrap()
+                .reshape((max_seq_len, 1))
+                .unwrap();
+            let x: f32 = t.i((1, 0)).unwrap().to_scalar().unwrap();
+            assert_eq!(x, 1.0);
+            let x: f32 = inv_freq.i((0, 1)).unwrap().to_scalar().unwrap();
+            assert_eq!(x, 0.1);
+            let freqs = t.matmul(&inv_freq).unwrap();
+            let x: f32 = freqs.i((1, 1)).unwrap().to_scalar().unwrap();
+            assert_eq!(x, 0.1);
+            let sin = freqs.sin().unwrap().contiguous().unwrap();
+            let x: f32 = sin.i((1, 1)).unwrap().to_scalar().unwrap();
+            assert_eq!(x, 0.099833414);
+        }
+    }
+}