A high-performance inference engine for AI models
Remove KernelDataType (#181)
uuuvna117bfd42684
diff --git a/crates/uzu/build/metal/bindgen.rs b/crates/uzu/build/metal/bindgen.rsindex 951e164b0..89dba2a53 100644--- a/crates/uzu/build/metal/bindgen.rs+++ b/crates/uzu/build/metal/bindgen.rs@@ -34,7 +34,7 @@ pub fn bindgen((variant_names.iter().map(|name| quote! { #[allow(non_snake_case)] #name: crate::DataType }).collect(),- quote! { &format!(#kernel_format, #entry_name #(, KernelDataType::from(#variant_names).function_name_suffix())*) },+ quote! { &format!(#kernel_format, #entry_name #(, #variant_names.metal_type())*) },)} else {(Vec::new(), quote! { #entry_name })diff --git a/crates/uzu/build/metal/compiler.rs b/crates/uzu/build/metal/compiler.rsindex 793eafac9..e32620491 100644--- a/crates/uzu/build/metal/compiler.rs+++ b/crates/uzu/build/metal/compiler.rs@@ -286,7 +286,7 @@ impl MetalCompiler {use crate::backends::metal::{ComputeEncoderSetValue,FunctionConstantValuesSetValue,- KernelDataType,+ data_type::MetalDataTypeExt,MTLContext,MTLError,MTLFunctionConstantValues,diff --git a/crates/uzu/src/backends/metal/data_type.rs b/crates/uzu/src/backends/metal/data_type.rsnew file mode 100644index 000000000..cf91d58c7--- /dev/null+++ b/crates/uzu/src/backends/metal/data_type.rs@@ -0,0 +1,16 @@+use crate::DataType;++pub trait MetalDataTypeExt {+ fn metal_type(&self) -> &'static str;+}++impl MetalDataTypeExt for DataType {+ fn metal_type(&self) -> &'static str {+ match self {+ DataType::F16 => "half",+ DataType::BF16 => "bfloat",+ DataType::F32 => "float",+ _ => panic!("Unsupported data type: {0:?}", self),+ }+ }+}diff --git a/crates/uzu/src/backends/metal/encodable_block/attention.rs b/crates/uzu/src/backends/metal/encodable_block/attention.rsindex 31eb40078..b37c35821 100644--- a/crates/uzu/src/backends/metal/encodable_block/attention.rs+++ b/crates/uzu/src/backends/metal/encodable_block/attention.rs@@ -2,9 +2,9 @@use super::{EncodableBlock, Metal};use crate::{+ DataType,backends::metal::{- KernelDataType, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, ProtocolObject,- Retained,+ MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, ProtocolObject, Retained,kernel::attention::{AttentionError, AttentionGemmArguments, AttentionKernel, AttentionKernelVariant,AttentionSinglePassArguments, AttentionTwoPassArguments, KVCacheUpdateArguments,@@ -42,7 +42,7 @@ pub struct Attention {impl Attention {pub fn new(context: &MTLContext,- data_type: KernelDataType,+ data_type: DataType,layer_index: usize,attention_scale: Option<f32>,has_sinks: bool,diff --git a/crates/uzu/src/backends/metal/encodable_block/classifier_layer.rs b/crates/uzu/src/backends/metal/encodable_block/classifier_layer.rsindex d3eddbcdf..d548afdb6 100644--- a/crates/uzu/src/backends/metal/encodable_block/classifier_layer.rs+++ b/crates/uzu/src/backends/metal/encodable_block/classifier_layer.rs@@ -9,7 +9,7 @@ use crate::{DataType,backends::metal::{MTLCommandBuffer, MTLComputeCommandEncoder, MTLContext, Metal, ProtocolObject, Retained,- compilation_parameters::CompilationConfig, kernel::KernelDataType,+ compilation_parameters::CompilationConfig,},config::TransformerLayerConfig,encodable_block::EncodingParameters,@@ -56,7 +56,6 @@ impl ClassifierLayer {let attention_config =layer_config.mixer_config.as_attention().expect("Classifier layers must use attention");let intermediate_data_type: DataType = attention_config.qkv_projection_config.activation_precision().into();- let kernel_data_type: KernelDataType = intermediate_data_type.into();let copy_main_to_shortcut_mixer: Box<dyn EncodableBlock<Metal>> = Box::new(TensorCopy::<Metal>::new(@@ -208,7 +207,7 @@ impl ClassifierLayer {let attention: Box<dyn EncodableBlock<Metal>> = Box::new(Attention::new(ctx,- kernel_data_type,+ intermediate_data_type,layer_index,attention_scale,attention_config.has_sinks,diff --git a/crates/uzu/src/backends/metal/encodable_block/layer/executables.rs b/crates/uzu/src/backends/metal/encodable_block/layer/executables.rsindex a8c1a96fc..8955a3296 100644--- a/crates/uzu/src/backends/metal/encodable_block/layer/executables.rs+++ b/crates/uzu/src/backends/metal/encodable_block/layer/executables.rs@@ -17,7 +17,7 @@ use crate::{DataType, DecoderLayerConfig,backends::metal::{MTLCommandBuffer, MTLComputeCommandEncoder, MTLContext, Metal, ProtocolObject, Retained,- compilation_parameters::CompilationConfig, kernel::KernelDataType,+ compilation_parameters::CompilationConfig,},config::{DecoderLayerType, MixerConfig},encodable_block::EncodingParameters,@@ -62,8 +62,6 @@ impl LayerExecutables {MixerConfig::Mamba(mamba) => mamba.in_projection_config.activation_precision().into(),MixerConfig::ShortConv(short_conv) => short_conv.in_projection_config.activation_precision().into(),};- let kernel_data_type: KernelDataType = intermediate_data_type.into();-let copy_main_to_shortcut: Box<dyn EncodableBlock<Metal>> = Box::new(TensorCopy::<Metal>::new(ctx,@@ -136,7 +134,7 @@ impl LayerExecutables {let attention = Box::new(Attention::new(ctx,- kernel_data_type,+ intermediate_data_type,layer_index,attention_scale,attention_config.has_sinks,diff --git a/crates/uzu/src/backends/metal/encodable_block/mamba_mixer.rs b/crates/uzu/src/backends/metal/encodable_block/mamba_mixer.rsindex 0472d34d1..578ab69e8 100644--- a/crates/uzu/src/backends/metal/encodable_block/mamba_mixer.rs+++ b/crates/uzu/src/backends/metal/encodable_block/mamba_mixer.rs@@ -8,8 +8,8 @@ use crate::{backends::{common::kernel::{SSDUpdateKernel, SplitInProjKernel},metal::{- KernelDataType, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, MetalArray,- ProtocolObject, Retained,+ MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, MetalArray, ProtocolObject,+ Retained,compilation_parameters::CompilationConfig,kernel::{dsl::{SSDUpdateMetalKernel, SplitInProjMetalKernel},@@ -65,7 +65,6 @@ impl MambaMixer {let conv_tree = resolve_subtree(&split_tree, &["conv", "conv1d"]);let data_type: DataType = mamba_config.in_projection_config.activation_precision().into();- let kernel_data_type: KernelDataType = data_type.into();let in_projection = transformer_layer::linear_block(&mamba_config.in_projection_config,@@ -100,14 +99,12 @@ impl MambaMixer {let gate_bias = split_tree.leaf("gate_bias").unwrap().clone();let skip_connection_weight = split_tree.leaf("skip_connection_weight").unwrap().clone();- let split_inproj = SplitInProjMetalKernel::new(mtl_context, kernel_data_type.into())- .expect("Failed to create split in-projection kernel");- let conv_scan = Conv1dScanKernel::new(mtl_context, kernel_data_type, &mamba_config.activation)+ let split_inproj =+ SplitInProjMetalKernel::new(mtl_context, data_type).expect("Failed to create split in-projection kernel");+ let conv_scan = Conv1dScanKernel::new(mtl_context, data_type, &mamba_config.activation).expect("Failed to create conv scan kernel");- let ssd_prefill =- SSDPrefillKernels::new(mtl_context, kernel_data_type).expect("Failed to create SSD prefill kernel");- let ssd_update = SSDUpdateMetalKernel::new(mtl_context, kernel_data_type.into())- .expect("Failed to create SSD decode kernel");+ let ssd_prefill = SSDPrefillKernels::new(mtl_context, data_type).expect("Failed to create SSD prefill kernel");+ let ssd_update = SSDUpdateMetalKernel::new(mtl_context, data_type).expect("Failed to create SSD decode kernel");let prefill_mode = resolve_prefill_mode_from_env();Self {diff --git a/crates/uzu/src/backends/metal/encodable_block/moe_block/moe_block.rs b/crates/uzu/src/backends/metal/encodable_block/moe_block/moe_block.rsindex 9baeff4ff..8580b9226 100644--- a/crates/uzu/src/backends/metal/encodable_block/moe_block/moe_block.rs+++ b/crates/uzu/src/backends/metal/encodable_block/moe_block/moe_block.rs@@ -11,8 +11,8 @@ use crate::{backends::{common::kernel::{MoeCountsOffsetsFusedKernel, MoeFinalizeKernel},metal::{- KernelDataType, MTLBlitCommandEncoderExt, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder,- MTLComputeCommandEncoder, MTLContext, MetalArray, ProtocolObject, Retained,+ MTLBlitCommandEncoderExt, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder,+ MTLContext, MetalArray, ProtocolObject, Retained,kernel::{MoeGatherKernels,dsl::{MoeCountsOffsetsFusedMetalKernel, MoeFinalizeMetalKernel},@@ -38,7 +38,7 @@ enum RouterBlock {pub struct MoeBlock {router: RouterBlock,- router_data_type: KernelDataType,+ router_data_type: DataType,router_renorm: bool,router_topk_kernel: MoeRouterTopKKernel,counts_offsets_kernel: MoeCountsOffsetsFusedMetalKernel,@@ -50,7 +50,7 @@ pub struct MoeBlock {moe_config: MixtureOfExpertsConfig,model_dim: usize,hidden_dim: usize,- data_type: KernelDataType,+ data_type: DataType,shared_weights: SharedMoeWeights,}@@ -62,11 +62,9 @@ impl MoeBlock {hidden_dim: usize,parameter_tree: &ParameterTree<MTLContext>,) -> Result<Self, crate::backends::metal::MTLError> {- let activation_data_type: DataType = moe_config.expert_config.linear_config.activation_precision().into();- let data_type: KernelDataType = activation_data_type.into();+ let data_type: DataType = moe_config.expert_config.linear_config.activation_precision().into();let router_data_type: DataType = moe_config.router_config.activation_precision().into();- let router_kernel_data_type: KernelDataType = router_data_type.into();let router_renorm = matches!(moe_config.routing_function, RoutingFunctionConfig::SoftmaxRouting);@@ -121,7 +119,7 @@ impl MoeBlock {let experts_two_pass_prefill_kernel = MoeExpertsTwoPassPrefillKernel::new(context).map_err(|e| {crate::backends::metal::MTLError::Generic(format!("Experts two-pass prefill kernel error: {:?}", e))})?;- let finalize_kernel = MoeFinalizeMetalKernel::new(context, data_type.into())+ let finalize_kernel = MoeFinalizeMetalKernel::new(context, data_type).map_err(|e| crate::backends::metal::MTLError::Generic(format!("Finalize kernel error: {:?}", e)))?;let experts_tree = parameter_tree@@ -165,7 +163,7 @@ impl MoeBlock {Ok(Self {router,- router_data_type: router_kernel_data_type,+ router_data_type,router_renorm,router_topk_kernel,counts_offsets_kernel,@@ -260,11 +258,6 @@ impl EncodableBlock<Metal> for MoeBlock {let k_tile = 128;// Clear internal MoE buffers- let dtype_size = match self.data_type {- KernelDataType::BFloat16 | KernelDataType::Float16 => 2,- KernelDataType::Float32 => 4,- };-let blit_encoder = root.new_blit_command_encoder().expect("Failed to create blit command encoder");if suffix_length > 0 && k > 0 {@@ -281,19 +274,19 @@ impl EncodableBlock<Metal> for MoeBlock {}// Clear hidden buffer- let hidden_bytes = suffix_length * k * self.hidden_dim * dtype_size;+ let hidden_bytes = suffix_length * k * self.hidden_dim * self.data_type.size_in_bytes();if hidden_bytes > 0 {blit_encoder.fill_buffer_range_value(&hidden_buf, 0..hidden_bytes, 0);}// Clear y_partial buffer- let y_partial_bytes = suffix_length * k * self.model_dim * dtype_size;+ let y_partial_bytes = suffix_length * k * self.model_dim * self.data_type.size_in_bytes();if y_partial_bytes > 0 {blit_encoder.fill_buffer_range_value(&y_partial_buf, 0..y_partial_bytes, 0);}// Clear x_perm buffer- let x_perm_bytes = suffix_length * k * self.model_dim * dtype_size;+ let x_perm_bytes = suffix_length * k * self.model_dim * self.data_type.size_in_bytes();if x_perm_bytes > 0 {blit_encoder.fill_buffer_range_value(&x_perm_buf, 0..x_perm_bytes, 0);}diff --git a/crates/uzu/src/backends/metal/encodable_block/short_conv_mixer.rs b/crates/uzu/src/backends/metal/encodable_block/short_conv_mixer.rsindex c5e484e51..5dca379e9 100644--- a/crates/uzu/src/backends/metal/encodable_block/short_conv_mixer.rs+++ b/crates/uzu/src/backends/metal/encodable_block/short_conv_mixer.rs@@ -6,8 +6,8 @@ use crate::{backends::{common::Context,metal::{- KernelDataType, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, MetalArray,- ProtocolObject, Retained,+ MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLContext, MetalArray, ProtocolObject,+ Retained,compilation_parameters::CompilationConfig,kernel::short_conv::{ShortConvDecodeArguments, ShortConvKernel, ShortConvPackArguments, ShortConvPrefillArguments,@@ -63,7 +63,6 @@ impl ShortConvMixer {let conv_tree = resolve_subtree(&mixer_tree, &["conv"]);let data_type: DataType = short_conv_config.in_projection_config.activation_precision().into();- let kernel_data_type: KernelDataType = data_type.into();let in_projection = transformer_layer::linear_block(&short_conv_config.in_projection_config,@@ -97,7 +96,7 @@ impl ShortConvMixer {};let short_conv_kernel =- ShortConvKernel::new(mtl_context, kernel_data_type).expect("Failed to create short conv kernel");+ ShortConvKernel::new(mtl_context, data_type).expect("Failed to create short conv kernel");Self {layer_index,diff --git a/crates/uzu/src/backends/metal/kernel/attention/mod.rs b/crates/uzu/src/backends/metal/kernel/attention/mod.rsindex 8a91a5f3c..51a48d9d4 100644--- a/crates/uzu/src/backends/metal/kernel/attention/mod.rs+++ b/crates/uzu/src/backends/metal/kernel/attention/mod.rs@@ -3,11 +3,15 @@ use std::{collections::HashMap, mem::size_of, ptr::NonNull};use objc2::rc::Retained;use thiserror::Error;-use crate::backends::{- common::gpu_types::{AttnMaskParams, AttnParams},- metal::{- FunctionConstantValuesSetValue, KernelDataType, MTLBuffer, MTLComputeCommandEncoder, MTLComputePipelineState,- MTLContext, MTLError, MTLFunctionConstantValues, MTLSize, ProtocolObject,+use crate::{+ DataType,+ backends::{+ common::gpu_types::{AttnMaskParams, AttnParams},+ metal::{+ FunctionConstantValuesSetValue, MTLBuffer, MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext,+ MTLError, MTLFunctionConstantValues, MTLSize, ProtocolObject, data_type::MetalDataTypeExt,+ kernel::moe::dtype_suffix,+ },},};@@ -27,7 +31,7 @@ pub struct AttentionKernelPipelines {}pub struct AttentionKernel {- data_type: KernelDataType,+ data_type: DataType,pipelines: AttentionKernelPipelines,}@@ -145,9 +149,9 @@ fn make_function_constants(impl AttentionKernel {pub fn new(context: &MTLContext,- data_type: KernelDataType,+ data_type: DataType,) -> Result<Self, AttentionError> {- let data_suffix = data_type.function_name_suffix();+ let data_suffix = data_type.metal_type();let supported_head_dims = [64, 128, 256];let mut single_pass = HashMap::new();@@ -579,11 +583,7 @@ impl AttentionKernel {// Kernel name matches gemm_attention.metal instantiations:// attention_gemm_{f16|bf16|f32}_{head_dim}_bk{bk}- let type_name = match self.data_type {- KernelDataType::Float16 => "f16",- KernelDataType::BFloat16 => "bf16",- KernelDataType::Float32 => "f32",- };+ let type_name = dtype_suffix(self.data_type);let function_name = format!("attention_gemm_{}_{}_bk{}", type_name, args.head_dim, bk);diff --git a/crates/uzu/src/backends/metal/kernel/data_type.rs b/crates/uzu/src/backends/metal/kernel/data_type.rsdeleted file mode 100644index 1237a4185..000000000--- a/crates/uzu/src/backends/metal/kernel/data_type.rs+++ /dev/null@@ -1,39 +0,0 @@-use crate::DataType;--#[derive(Debug, Clone, Copy, PartialEq, Eq)]-pub enum KernelDataType {- Float32,- Float16,- BFloat16,-}--impl KernelDataType {- pub fn function_name_suffix(&self) -> &'static str {- match self {- KernelDataType::Float32 => "float",- KernelDataType::Float16 => "half",- KernelDataType::BFloat16 => "bfloat",- }- }-}--impl From<KernelDataType> for DataType {- fn from(val: KernelDataType) -> Self {- match val {- KernelDataType::BFloat16 => DataType::BF16,- KernelDataType::Float16 => DataType::F16,- KernelDataType::Float32 => DataType::F32,- }- }-}--impl From<DataType> for KernelDataType {- fn from(dtype: DataType) -> Self {- match dtype {- DataType::BF16 => KernelDataType::BFloat16,- DataType::F16 => KernelDataType::Float16,- DataType::F32 => KernelDataType::Float32,- _ => panic!("Unsupported data type: {0:?}", dtype),- }- }-}diff --git a/crates/uzu/src/backends/metal/kernel/mod.rs b/crates/uzu/src/backends/metal/kernel/mod.rsindex 202e9b4d7..f46e41e24 100644--- a/crates/uzu/src/backends/metal/kernel/mod.rs+++ b/crates/uzu/src/backends/metal/kernel/mod.rs@@ -1,5 +1,4 @@pub mod attention;-mod data_type;use crate::backends::common::kernel::matmul::MatmulKernels;pub mod dsl {include!(concat!(env!("OUT_DIR"), "/dsl.rs"));@@ -22,7 +21,6 @@ pub use attention::{AttentionError, AttentionKernel, AttentionKernelVariant, AttentionSinglePassArguments, AttentionTwoPassArguments,KVCacheUpdateArguments,};-pub use data_type::KernelDataType;pub use matmul::{MatmulArguments, MatmulKernel};pub use moe::{MoeBlockBasesArguments, MoeExpertsArguments, MoeExpertsError, MoeExpertsTwoPassArguments,diff --git a/crates/uzu/src/backends/metal/kernel/moe/experts.rs b/crates/uzu/src/backends/metal/kernel/moe/experts.rsindex 7de2161bf..fc5d21e19 100644--- a/crates/uzu/src/backends/metal/kernel/moe/experts.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/experts.rs@@ -1,13 +1,16 @@use std::{mem::size_of, ptr::NonNull};use super::{dtype_index, dtype_suffix};-use crate::backends::metal::{- FunctionConstantValuesSetValue, KernelDataType, MTLBlitCommandEncoderExt, MTLBuffer, MTLCommandBuffer,- MTLCommandEncoder, MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext, MTLError,- MTLFunctionConstantValues, MTLSize, ProtocolObject, Retained,- kernel::moe::tiles_map::{- MoeTileCountsArguments, MoeTileDispatchArguments, MoeTileMapBuildArguments, MoeTileMapKernels,- MoeTileScanArguments,+use crate::{+ DataType,+ backends::metal::{+ FunctionConstantValuesSetValue, MTLBlitCommandEncoderExt, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder,+ MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext, MTLError, MTLFunctionConstantValues, MTLSize,+ ProtocolObject, Retained,+ kernel::moe::tiles_map::{+ MoeTileCountsArguments, MoeTileDispatchArguments, MoeTileMapBuildArguments, MoeTileMapKernels,+ MoeTileScanArguments,+ },},};@@ -50,7 +53,7 @@ pub struct MoeExpertsArguments<'a> {pub up_clip_min: f32,pub up_clip_max: f32,pub silu_alpha: f32,- pub data_type: KernelDataType,+ pub data_type: DataType,}pub struct MoeExpertsTwoPassPrefillKernel {@@ -86,12 +89,12 @@ pub struct MoeExpertsTwoPassArguments<'a> {pub up_clip_min: f32,pub up_clip_max: f32,pub silu_alpha: f32,- pub data_type: KernelDataType,+ pub data_type: DataType,}impl MoeExpertsTwoPassPrefillKernel {pub fn new(ctx: &MTLContext) -> Result<Self, MoeExpertsError> {- let dtypes = [KernelDataType::Float16, KernelDataType::BFloat16, KernelDataType::Float32];+ let dtypes = [DataType::F16, DataType::BF16, DataType::F32];let mut pass_a_indirect = vec![Vec::with_capacity(dtypes.len()); 4];for gate in 0u32..4u32 {for dtype in &dtypes {@@ -130,11 +133,7 @@ impl MoeExpertsTwoPassPrefillKernel {}let gate_idx = (args.gating_code.min(3)) as usize;let dtype_idx = dtype_index(args.data_type);- let dtype_size = match args.data_type {- KernelDataType::BFloat16 | KernelDataType::Float16 => 2,- KernelDataType::Float32 => 4,- };- let hidden_bytes = args.total_rows * args.d_ff * dtype_size;+ let hidden_bytes = args.total_rows * args.d_ff * args.data_type.size_in_bytes();let blit_encoder = command_buffer.new_blit_command_encoder().expect("Failed to create blit command encoder");blit_encoder.fill_buffer_range_value(args.hidden_buffer, 0..hidden_bytes, 0);blit_encoder.end_encoding();diff --git a/crates/uzu/src/backends/metal/kernel/moe/experts_single.rs b/crates/uzu/src/backends/metal/kernel/moe/experts_single.rsindex 5ebb1ec6b..586833efd 100644--- a/crates/uzu/src/backends/metal/kernel/moe/experts_single.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/experts_single.rs@@ -1,15 +1,18 @@use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder};use objc2::{__framework_prelude::ProtocolObject, Message, rc::Retained};-use crate::backends::{- common::kernel::{MoeExpertsDecodeSinglePassAKernel as _, MoeExpertsDecodeSinglePassBKernel as _},- metal::{- KernelDataType, MTLContext, MTLError,- kernel::dsl::{MoeExpertsDecodeSinglePassAMetalKernel, MoeExpertsDecodeSinglePassBMetalKernel},+use crate::{+ DataType,+ backends::{+ common::kernel::{MoeExpertsDecodeSinglePassAKernel as _, MoeExpertsDecodeSinglePassBKernel as _},+ metal::{+ MTLContext, MTLError,+ kernel::dsl::{MoeExpertsDecodeSinglePassAMetalKernel, MoeExpertsDecodeSinglePassBMetalKernel},+ },},};-static DTYPES: [KernelDataType; 3] = [KernelDataType::Float16, KernelDataType::BFloat16, KernelDataType::Float32];+static DTYPES: [DataType; 3] = [DataType::F16, DataType::BF16, DataType::F32];/// Arguments for single-token MoE decode (T=1 optimized path)#[derive(Debug)]@@ -51,7 +54,7 @@ pub struct MoeExpertsSingleDecodeArguments<'a> {/// Up clipping maxpub up_clip_max: f32,/// Data type- pub data_type: KernelDataType,+ pub data_type: DataType,}pub struct MoeExpertsSingleDecodeKernels {diff --git a/crates/uzu/src/backends/metal/kernel/moe/gather.rs b/crates/uzu/src/backends/metal/kernel/moe/gather.rsindex 977e7d674..eebffe8d9 100644--- a/crates/uzu/src/backends/metal/kernel/moe/gather.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/gather.rs@@ -5,8 +5,7 @@ use crate::{backends::{common::kernel::{MoeGatherXPerm1DKernel, MoeGatherXPerm2DKernel},metal::{- KernelDataType, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLContext, MTLError, ProtocolObject,- Retained,+ MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLContext, MTLError, ProtocolObject, Retained,kernel::dsl::{MoeGatherXPerm1DMetalKernel, MoeGatherXPerm2DMetalKernel},},},@@ -41,12 +40,12 @@ impl MoeGatherKernels {pub fn encode(&self,command_buffer: &Retained<ProtocolObject<dyn MTLCommandBuffer>>,- dtype: KernelDataType,+ dtype: DataType,args: &MoeGatherArguments,) {let encoder = command_buffer.new_compute_command_encoder().expect("Failed to create compute command encoder");match dtype {- KernelDataType::Float32 => self.f32.encode(+ DataType::F32 => self.f32.encode(&args.x_buffer.retain(),&args.bucketed_ids_buffer.retain(),&args.x_perm_buffer.retain(),@@ -56,7 +55,7 @@ impl MoeGatherKernels {args.k as u32,&encoder,),- KernelDataType::Float16 => self.f16.encode(+ DataType::F16 => self.f16.encode(&args.x_buffer.retain(),&args.bucketed_ids_buffer.retain(),&args.x_perm_buffer.retain(),@@ -66,7 +65,7 @@ impl MoeGatherKernels {args.k as u32,&encoder,),- KernelDataType::BFloat16 => self.bf16.encode(+ DataType::BF16 => self.bf16.encode(&args.x_buffer.retain(),&args.bucketed_ids_buffer.retain(),&args.x_perm_buffer.retain(),@@ -76,6 +75,7 @@ impl MoeGatherKernels {args.k as u32,&encoder,),+ _ => panic!("Unsupported data type: {:?}", dtype),}encoder.end_encoding();}diff --git a/crates/uzu/src/backends/metal/kernel/moe/mod.rs b/crates/uzu/src/backends/metal/kernel/moe/mod.rsindex 5981a3a92..d868ace3a 100644--- a/crates/uzu/src/backends/metal/kernel/moe/mod.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/mod.rs@@ -1,4 +1,4 @@-use crate::backends::metal::KernelDataType;+use crate::DataType;// Submodulesmod experts;@@ -25,18 +25,20 @@ pub use tiles_pass_a::{};// Common utility functions-pub(crate) fn dtype_suffix(dtype: KernelDataType) -> &'static str {+pub(crate) fn dtype_suffix(dtype: DataType) -> &'static str {match dtype {- KernelDataType::Float16 => "f16",- KernelDataType::BFloat16 => "bf16",- KernelDataType::Float32 => "f32",+ DataType::F16 => "f16",+ DataType::BF16 => "bf16",+ DataType::F32 => "f32",+ _ => panic!("Unsupported data type: {:?}", dtype),}}-pub(crate) fn dtype_index(dtype: KernelDataType) -> usize {+pub(crate) fn dtype_index(dtype: DataType) -> usize {match dtype {- KernelDataType::Float16 => 0,- KernelDataType::BFloat16 => 1,- KernelDataType::Float32 => 2,+ DataType::F16 => 0,+ DataType::BF16 => 1,+ DataType::F32 => 2,+ _ => panic!("Unsupported data type: {:?}", dtype),}}diff --git a/crates/uzu/src/backends/metal/kernel/moe/router_topk.rs b/crates/uzu/src/backends/metal/kernel/moe/router_topk.rsindex 8c30769fd..adfa009aa 100644--- a/crates/uzu/src/backends/metal/kernel/moe/router_topk.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/router_topk.rs@@ -1,6 +1,9 @@-use crate::backends::metal::{- KernelDataType, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLComputePipelineState,- MTLContext, MTLError, MTLSize, ProtocolObject, Retained, metal_extensions::ComputeEncoderSetValue,+use crate::{+ DataType,+ backends::metal::{+ MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext,+ MTLError, MTLSize, ProtocolObject, Retained, metal_extensions::ComputeEncoderSetValue,+ },};const THREADS_PER_THREADGROUP: usize = 256;@@ -55,7 +58,7 @@ impl MoeRouterTopKKernel {pub fn encode(&self,command_buffer: &Retained<ProtocolObject<dyn MTLCommandBuffer>>,- dtype: KernelDataType,+ dtype: DataType,args: MoeRouterTopKArguments,) -> Result<(), MoeRouterTopKError> {if args.t == 0 || args.e == 0 || args.k == 0 {@@ -81,9 +84,10 @@ impl MoeRouterTopKKernel {let compute_encoder =command_buffer.new_compute_command_encoder().expect("Failed to create compute command encoder");match dtype {- KernelDataType::Float16 => compute_encoder.set_compute_pipeline_state(&self.pipeline_f16),- KernelDataType::Float32 => compute_encoder.set_compute_pipeline_state(&self.pipeline_f32),- KernelDataType::BFloat16 => compute_encoder.set_compute_pipeline_state(&self.pipeline_bf16),+ DataType::F16 => compute_encoder.set_compute_pipeline_state(&self.pipeline_f16),+ DataType::F32 => compute_encoder.set_compute_pipeline_state(&self.pipeline_f32),+ DataType::BF16 => compute_encoder.set_compute_pipeline_state(&self.pipeline_bf16),+ _ => panic!("Unsupported data type: {:?}", dtype),}compute_encoder.set_buffer(Some(args.input_buffer), 0, 0);diff --git a/crates/uzu/src/backends/metal/kernel/moe/scatter.rs b/crates/uzu/src/backends/metal/kernel/moe/scatter.rsindex bc55f413a..9de76dcf6 100644--- a/crates/uzu/src/backends/metal/kernel/moe/scatter.rs+++ b/crates/uzu/src/backends/metal/kernel/moe/scatter.rs@@ -1,6 +1,9 @@-use crate::backends::metal::{- KernelDataType, MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLComputePipelineState,- MTLContext, MTLSize, ProtocolObject, Retained, metal_extensions::ComputeEncoderSetValue,+use crate::{+ DataType,+ backends::metal::{+ MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext,+ MTLSize, ProtocolObject, Retained, metal_extensions::ComputeEncoderSetValue,+ },};// ---- Scatter Buckets Kernels ----@@ -110,21 +113,22 @@ impl MoeScatterKernels {&self,command_buffer: &Retained<ProtocolObject<dyn MTLCommandBuffer>>,args: MoeScatterArguments,- dtype: KernelDataType,+ dtype: DataType,) -> Result<(), MoeScatterError> {let compute_encoder =command_buffer.new_compute_command_encoder().expect("Failed to create compute command encoder");// Select pipeline based on dtypematch dtype {- KernelDataType::Float16 => {+ DataType::F16 => {compute_encoder.set_compute_pipeline_state(&self.pipeline_scatter_f16);},- KernelDataType::Float32 => {+ DataType::F32 => {compute_encoder.set_compute_pipeline_state(&self.pipeline_scatter_f32);},- KernelDataType::BFloat16 => {+ DataType::BF16 => {compute_encoder.set_compute_pipeline_state(&self.pipeline_scatter_bf16);},+ _ => panic!("Unsupported data type: {:?}", dtype),}compute_encoder.set_buffer(Some(args.topk_ids_buffer), 0, 0);compute_encoder.set_buffer(Some(args.topk_probs_buffer), 0, 1);@@ -157,14 +161,15 @@ impl MoeScatterKernels {&self,command_buffer: &Retained<ProtocolObject<dyn MTLCommandBuffer>>,args: MoeScatterWithMapArguments,- dtype: KernelDataType,+ dtype: DataType,) -> Result<(), MoeScatterError> {let compute_encoder =command_buffer.new_compute_command_encoder().expect("Failed to create compute command encoder");let pipeline = match dtype {- KernelDataType::Float16 => &self.pipeline_scatter_map_f16,- KernelDataType::Float32 => &self.pipeline_scatter_map_f32,- KernelDataType::BFloat16 => &self.pipeline_scatter_map_bf16,+ DataType::F16 => &self.pipeline_scatter_map_f16,+ DataType::F32 => &self.pipeline_scatter_map_f32,+ DataType::BF16 => &self.pipeline_scatter_map_bf16,+ _ => panic!("Unsupported data type: {:?}", dtype),};compute_encoder.set_compute_pipeline_state(pipeline);let base = &args.base;diff --git a/crates/uzu/src/backends/metal/kernel/short_conv/mod.rs b/crates/uzu/src/backends/metal/kernel/short_conv/mod.rsindex fb63af9c1..e0ee7dc4e 100644--- a/crates/uzu/src/backends/metal/kernel/short_conv/mod.rs+++ b/crates/uzu/src/backends/metal/kernel/short_conv/mod.rs@@ -1,6 +1,10 @@-use crate::backends::metal::{- ComputeEncoderSetValue, FunctionConstantValuesSetValue, KernelDataType, MTLBuffer, MTLComputeCommandEncoder,- MTLComputePipelineState, MTLContext, MTLError, MTLFunctionConstantValues, MTLSize, ProtocolObject, Retained,+use crate::{+ DataType,+ backends::metal::{+ ComputeEncoderSetValue, FunctionConstantValuesSetValue, MTLBuffer, MTLComputeCommandEncoder,+ MTLComputePipelineState, MTLContext, MTLError, MTLFunctionConstantValues, MTLSize, ProtocolObject, Retained,+ data_type::MetalDataTypeExt,+ },};#[derive(Debug, thiserror::Error)]@@ -9,10 +13,6 @@ pub enum ShortConvKernelError {MetalError(#[from] MTLError),}-fn fn_suffix(dt: KernelDataType) -> &'static str {- dt.function_name_suffix()-}-fn make_function_constants(has_bias: bool) -> Retained<MTLFunctionConstantValues> {let function_constants = MTLFunctionConstantValues::new();function_constants.set_value(&has_bias, 0);@@ -95,12 +95,12 @@ pub struct ShortConvTrieArguments<'a> {impl ShortConvKernel {pub fn new(context: &MTLContext,- data_type: KernelDataType,+ data_type: DataType,) -> Result<Self, ShortConvKernelError> {- let pack_name = format!("short_conv_pack_kernel_{}", fn_suffix(data_type));- let prefill_name = format!("short_conv_prefill_kernel_{}", fn_suffix(data_type));- let decode_name = format!("short_conv_decode_kernel_{}", fn_suffix(data_type));- let trie_name = format!("short_conv_trie_kernel_{}", fn_suffix(data_type));+ let pack_name = format!("short_conv_pack_kernel_{}", data_type.metal_type());+ let prefill_name = format!("short_conv_prefill_kernel_{}", data_type.metal_type());+ let decode_name = format!("short_conv_decode_kernel_{}", data_type.metal_type());+ let trie_name = format!("short_conv_trie_kernel_{}", data_type.metal_type());let pack_pipeline =context.compute_pipeline_state(&pack_name, None).map_err(ShortConvKernelError::MetalError)?;diff --git a/crates/uzu/src/backends/metal/kernel/ssm/conv1d_scan.rs b/crates/uzu/src/backends/metal/kernel/ssm/conv1d_scan.rsindex ff0427c0a..6056644ba 100644--- a/crates/uzu/src/backends/metal/kernel/ssm/conv1d_scan.rs+++ b/crates/uzu/src/backends/metal/kernel/ssm/conv1d_scan.rs@@ -1,8 +1,9 @@-use super::{SSMKernelError, fn_suffix};+use super::SSMKernelError;use crate::{+ DataType,backends::metal::{- FunctionConstantValuesSetValue, KernelDataType, MTLBuffer, MTLComputeCommandEncoder, MTLComputePipelineState,- MTLContext, MTLFunctionConstantValues, MTLSize, ProtocolObject, Retained,+ FunctionConstantValuesSetValue, MTLBuffer, MTLComputeCommandEncoder, MTLComputePipelineState, MTLContext,+ MTLFunctionConstantValues, MTLSize, ProtocolObject, Retained, data_type::MetalDataTypeExt,metal_extensions::ComputeEncoderSetValue,},config::Activation,@@ -93,12 +94,12 @@ pub struct Conv1dDecodeArguments<'a> {impl Conv1dScanKernel {pub fn new(context: &MTLContext,- data_type: KernelDataType,+ data_type: DataType,activation: &Activation,) -> Result<Self, SSMKernelError> {- let fn_name = format!("conv1d_scan_kernel_{}", fn_suffix(data_type));- let pack_name = format!("conv1d_pack_prefix_kernel_{}", fn_suffix(data_type));- let decode_name = format!("conv1d_decode_kernel_{}", fn_suffix(data_type));+ let fn_name = format!("conv1d_scan_kernel_{}", data_type.metal_type());+ let pack_name = format!("conv1d_pack_prefix_kernel_{}", data_type.metal_type());+ let decode_name = format!("conv1d_decode_kernel_{}", data_type.metal_type());let activation_id = activation_to_int(activation);let pipeline_no_bias = {let function_constants = make_function_constants(activation, false);diff --git a/crates/uzu/src/backends/metal/kernel/ssm/mod.rs b/crates/uzu/src/backends/metal/kernel/ssm/mod.rsindex e21c5cad0..c64f8af63 100644--- a/crates/uzu/src/backends/metal/kernel/ssm/mod.rs+++ b/crates/uzu/src/backends/metal/kernel/ssm/mod.rs@@ -1,4 +1,4 @@-use crate::backends::metal::{KernelDataType, MTLError};+use crate::backends::metal::MTLError;#[derive(Debug, thiserror::Error)]pub enum SSMKernelError {@@ -8,10 +8,6 @@ pub enum SSMKernelError {FunctionNotFound(String),}-fn fn_suffix(dt: KernelDataType) -> &'static str {- dt.function_name_suffix()-}-pub mod conv1d_scan;pub mod ssd_prefill;diff --git a/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.rs b/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.rsindex 6fc06c806..ae7993b92 100644--- a/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.rs+++ b/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.rs@@ -1,10 +1,13 @@use objc2::Message;-use crate::backends::{- common::kernel::{SSDPrefill64Kernel as _, SSDPrefillKernel as _, SSDPrefillSequentialKernel as _},- metal::{- KernelDataType, MTLBuffer, MTLComputeCommandEncoder, MTLContext, MTLError, ProtocolObject,- kernel::dsl::{SSDPrefill64MetalKernel, SSDPrefillMetalKernel, SSDPrefillSequentialMetalKernel},+use crate::{+ DataType,+ backends::{+ common::kernel::{SSDPrefill64Kernel as _, SSDPrefillKernel as _, SSDPrefillSequentialKernel as _},+ metal::{+ MTLBuffer, MTLComputeCommandEncoder, MTLContext, MTLError, ProtocolObject,+ kernel::dsl::{SSDPrefill64MetalKernel, SSDPrefillMetalKernel, SSDPrefillSequentialMetalKernel},+ },},};@@ -43,11 +46,11 @@ pub struct SSDPrefillKernels {impl SSDPrefillKernels {pub fn new(context: &MTLContext,- data_type: KernelDataType,+ data_type: DataType,) -> Result<Self, MTLError> {- let single = SSDPrefillMetalKernel::new(context, data_type.into())?;- let single_64 = SSDPrefill64MetalKernel::new(context, data_type.into())?;- let sequential = SSDPrefillSequentialMetalKernel::new(context, data_type.into())?;+ let single = SSDPrefillMetalKernel::new(context, data_type)?;+ let single_64 = SSDPrefill64MetalKernel::new(context, data_type)?;+ let sequential = SSDPrefillSequentialMetalKernel::new(context, data_type)?;Ok(Self {single,single_64,diff --git a/crates/uzu/src/backends/metal/language_model_generator_context.rs b/crates/uzu/src/backends/metal/language_model_generator_context.rsindex 73b1fe021..6fcdafb97 100644--- a/crates/uzu/src/backends/metal/language_model_generator_context.rs+++ b/crates/uzu/src/backends/metal/language_model_generator_context.rs@@ -6,7 +6,7 @@ use std::{rc::Rc,};-use super::{Decoder, KVCacheUpdate, KernelDataType, MTLContext, Metal, compilation_parameters::CompilationConfig};+use super::{Decoder, KVCacheUpdate, MTLContext, Metal, compilation_parameters::CompilationConfig};use crate::{DataType,backends::{@@ -206,7 +206,6 @@ impl LanguageModelGeneratorContext {)));let intermediate_data_type: DataType = decoder_config.output_norm_config.scale_precision.into();- let kernel_data_type: KernelDataType = intermediate_data_type.into();let kv_cache_update = Box::new(KVCacheUpdate::new(context.as_ref(), intermediate_data_type, max_prefix_length).map_err(|_| Error::UnableToCreateMetalContext)?,@@ -224,7 +223,7 @@ impl LanguageModelGeneratorContext {// Create mask update kernel if model has attention layerslet mask_update = if decoder_config.has_attention_layers() {Some(- MaskUpdateMetalKernel::new(&context, kernel_data_type.into())+ MaskUpdateMetalKernel::new(&context, intermediate_data_type).map_err(|_| Error::UnableToCreateMetalContext)?,)} else {diff --git a/crates/uzu/src/backends/metal/mod.rs b/crates/uzu/src/backends/metal/mod.rsindex 93c1c5306..5aed4dd6f 100644--- a/crates/uzu/src/backends/metal/mod.rs+++ b/crates/uzu/src/backends/metal/mod.rs@@ -6,6 +6,7 @@ mod command_buffer;pub mod compilation_parameters;mod context;mod copy_encoder;+mod data_type;pub mod encodable_block;pub mod error;pub mod kernel;@@ -17,7 +18,7 @@ pub use classifier_context::ClassifierContext;pub use context::{DeviceArchitecture, DeviceClass, DeviceGeneration, MTLContext};pub use encodable_block::Decoder;pub use error::MTLError;-pub use kernel::{KernelDataType, dsl::MetalKernels};+pub use kernel::dsl::MetalKernels;pub use language_model_generator_context::LanguageModelGeneratorContext;pub use metal::prelude::*;pub use metal_extensions::{ComputeEncoderSetValue, FunctionConstantValuesSetValue};diff --git a/crates/uzu/tests/attention_kernel_test.rs b/crates/uzu/tests/attention_kernel_test.rsindex d8383a721..c2a0cfb02 100644--- a/crates/uzu/tests/attention_kernel_test.rs+++ b/crates/uzu/tests/attention_kernel_test.rs@@ -7,15 +7,18 @@ use std::mem::size_of;use bytemuck;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue, MTLDeviceExt, MTLResourceOptions};use ndarray::{Array3, Array4, s};-use uzu::backends::{- common::Context,- metal::{- KernelDataType, MTLContext, ProtocolObject, Retained,- kernel::attention::{- AttentionGemmArguments, AttentionKernel, AttentionKernelVariant, AttentionSinglePassArguments,- AttentionTwoPassArguments,+use uzu::{+ DataType,+ backends::{+ common::Context,+ metal::{+ MTLContext, ProtocolObject, Retained,+ kernel::attention::{+ AttentionGemmArguments, AttentionKernel, AttentionKernelVariant, AttentionSinglePassArguments,+ AttentionTwoPassArguments,+ },+ metal_extensions::CommandBufferTimingExt,},- metal_extensions::CommandBufferTimingExt,},};@@ -553,7 +556,7 @@ fn test_single_pass_attention_basic() {println!("Reference output range (no mask): [{}, {}]", ref_min, ref_max);println!("Reference sample values: {:?}", &reference_output.slice(s![0, 0, 0, 0..4]).to_vec());- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -602,7 +605,7 @@ fn test_gemm_attention_basic() {let (queries, keys, values, _mask) = create_test_data(batch_size, num_heads, num_kv_heads, seq_len, head_dim, 123);- let kernel = AttentionKernel::new(&context, KernelDataType::Float32).expect("Failed to create AttentionKernel");+ let kernel = AttentionKernel::new(&context, DataType::F32).expect("Failed to create AttentionKernel");let reference_output = reference_attention(&queries, &keys, &values, None, None, scale);@@ -630,7 +633,7 @@ fn test_gemm_attention_f32_head_dim_128() {let (queries, keys, values, _mask) = create_test_data(batch_size, num_heads, num_kv_heads, seq_len, head_dim, 456);- let kernel = AttentionKernel::new(&context, KernelDataType::Float32).expect("Failed to create AttentionKernel");+ let kernel = AttentionKernel::new(&context, DataType::F32).expect("Failed to create AttentionKernel");let reference_output = reference_attention(&queries, &keys, &values, None, None, scale);@@ -659,7 +662,7 @@ fn test_matrix_attention_matches_vector_and_cpu_seq256() {let (queries, keys, values, mask) = create_test_data(batch_size, num_heads, num_kv_heads, seq_len, head_dim, 2026);- let kernel = AttentionKernel::new(&context, KernelDataType::Float32).expect("Failed to create AttentionKernel");+ let kernel = AttentionKernel::new(&context, DataType::F32).expect("Failed to create AttentionKernel");// CPU reference (uses the same additive mask)let reference_output = reference_attention(&queries, &keys, &values, Some(&mask), None, scale);@@ -714,7 +717,7 @@ fn test_single_pass_attention_with_mask() {let (queries, keys, values, mask) = create_test_data(batch_size, num_heads, num_kv_heads, seq_len, head_dim, 42);- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -766,7 +769,7 @@ fn test_single_pass_attention_with_sinks() {let sinks: Vec<f32> = (0..num_heads).map(|h| (h as f32 - (num_heads as f32 / 2.0)) * 0.25).collect();println!("Using sinks: {:?}", sinks);- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -810,7 +813,7 @@ fn test_single_pass_attention_with_sinks_long_sequence() {let sinks: Vec<f32> = (0..num_heads).map(|h| (h as f32 * 0.1) - 0.15).collect();println!("Long sequence sinks: {:?}", sinks);- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -856,7 +859,7 @@ fn test_single_pass_attention_gqa() {let (queries, keys, values, _mask) = create_test_data(batch_size, num_heads, num_kv_heads, seq_len, head_dim, 42);- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -982,7 +985,7 @@ fn run_two_pass_attention(fn test_two_pass_attention() {let context = MTLContext::new().expect("Failed to create MTLContext");- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -1028,7 +1031,7 @@ fn test_two_pass_attention() {fn test_two_pass_attention_gqa() {let context = MTLContext::new().expect("Failed to create MTLContext");- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);@@ -1076,7 +1079,7 @@ fn perf_two_pass_attention() {let context = MTLContext::new().expect("Failed to create MTLContext");- let kernel = match AttentionKernel::new(&context, KernelDataType::Float32) {+ let kernel = match AttentionKernel::new(&context, DataType::F32) {Ok(k) => k,Err(e) => {panic!("Failed to create AttentionKernel: {:?}", e);diff --git a/crates/uzu/tests/moe/counts_offsets_fused_test.rs b/crates/uzu/tests/moe/counts_offsets_fused_test.rsindex 342ef513c..66bd80d69 100644--- a/crates/uzu/tests/moe/counts_offsets_fused_test.rs+++ b/crates/uzu/tests/moe/counts_offsets_fused_test.rs@@ -3,14 +3,16 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::{- common::kernel::MoeCountsOffsetsFusedKernel,- metal::{- MTLContext, ProtocolObject, Retained,- kernel::{- KernelDataType,- dsl::MoeCountsOffsetsFusedMetalKernel,- moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+use uzu::{+ DataType,+ backends::{+ common::kernel::MoeCountsOffsetsFusedKernel,+ metal::{+ MTLContext, ProtocolObject, Retained,+ kernel::{+ dsl::MoeCountsOffsetsFusedMetalKernel,+ moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+ },},},};@@ -89,7 +91,7 @@ fn gen_topk_ids_from_logits(k,renorm: true,};- router_topk.encode(&cb, KernelDataType::BFloat16, args).expect("encode router_topk");+ router_topk.encode(&cb, DataType::BF16, args).expect("encode router_topk");cb.commit();cb.wait_until_completed();diff --git a/crates/uzu/tests/moe/moe_block_e2e_test.rs b/crates/uzu/tests/moe/moe_block_e2e_test.rsindex cc5df1800..175fca611 100644--- a/crates/uzu/tests/moe/moe_block_e2e_test.rs+++ b/crates/uzu/tests/moe/moe_block_e2e_test.rs@@ -1,16 +1,19 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::{- common::kernel::{MoeCountsOffsetsFusedKernel, MoeFinalizeKernel},- metal::{- MTLContext,- kernel::{- KernelDataType, MoeBlockBasesArguments, MoeScatterKernels, MoeScatterWithMapArguments,- dsl::{MoeCountsOffsetsFusedMetalKernel, MoeFinalizeMetalKernel},- moe::{- MoeExpertsTwoPassArguments, MoeExpertsTwoPassPrefillKernel, MoeGatherArguments, MoeGatherKernels,- MoeRouterTopKArguments, MoeRouterTopKKernel,+use uzu::{+ DataType,+ backends::{+ common::kernel::{MoeCountsOffsetsFusedKernel, MoeFinalizeKernel},+ metal::{+ MTLContext,+ kernel::{+ MoeBlockBasesArguments, MoeScatterKernels, MoeScatterWithMapArguments,+ dsl::{MoeCountsOffsetsFusedMetalKernel, MoeFinalizeMetalKernel},+ moe::{+ MoeExpertsTwoPassArguments, MoeExpertsTwoPassPrefillKernel, MoeGatherArguments, MoeGatherKernels,+ MoeRouterTopKArguments, MoeRouterTopKKernel,+ },},},},@@ -351,7 +354,7 @@ fn run_moe_parity_test_internal(router_topk.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,MoeRouterTopKArguments {input_buffer: &x_buf,weight_buffer: &router_w_buf,@@ -407,14 +410,14 @@ fn run_moe_parity_test_internal(},tok2row_buffer: &tok2row_buf,},- KernelDataType::BFloat16,+ DataType::BF16,).expect("scatter");let gather = MoeGatherKernels::new(&ctx).expect("gather");gather.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,&MoeGatherArguments {x_buffer: &x_buf,bucketed_ids_buffer: &bucketed_ids_buf,@@ -462,12 +465,12 @@ fn run_moe_parity_test_internal(up_clip_min: up_clip.0,up_clip_max: up_clip.1,silu_alpha,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},).expect("experts encode");- let finalize = MoeFinalizeMetalKernel::new(&ctx, KernelDataType::BFloat16.into()).expect("finalize");+ let finalize = MoeFinalizeMetalKernel::new(&ctx, DataType::BF16).expect("finalize");let encoder = cb.new_compute_command_encoder().expect("encoder");finalize.encode(&tok2row_buf,diff --git a/crates/uzu/tests/moe/moe_experts_perf_test.rs b/crates/uzu/tests/moe/moe_experts_perf_test.rsindex 1a684516f..1178cbf0a 100644--- a/crates/uzu/tests/moe/moe_experts_perf_test.rs+++ b/crates/uzu/tests/moe/moe_experts_perf_test.rs@@ -3,11 +3,14 @@ use std::time::Instant;use half::bf16;use metal::{MTLCommandBuffer, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::metal::{- KernelDataType, MTLContext,- kernel::moe::{- MoeExpertsSingleDecodeArguments, MoeExpertsSingleDecodeKernels, MoeExpertsTwoPassArguments,- MoeExpertsTwoPassDecodeKernels, MoeExpertsTwoPassPrefillKernel,+use uzu::{+ DataType,+ backends::metal::{+ MTLContext,+ kernel::moe::{+ MoeExpertsSingleDecodeArguments, MoeExpertsSingleDecodeKernels, MoeExpertsTwoPassArguments,+ MoeExpertsTwoPassDecodeKernels, MoeExpertsTwoPassPrefillKernel,+ },},};@@ -125,7 +128,7 @@ fn run_decode_case(up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha: 1.702,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,};for _ in 0..warmup {@@ -258,7 +261,7 @@ fn run_two_pass_prefill_case(up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha: 1.702,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,};for _ in 0..warmup {@@ -367,7 +370,7 @@ fn run_fused_single_token_case(gate_clip_max: f32::INFINITY,up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,};for _ in 0..warmup {@@ -530,7 +533,7 @@ fn run_indirect_decode_timed(up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha: 1.0,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,};for _ in 0..warmup {@@ -611,7 +614,7 @@ fn run_fused_decode_timed(gate_clip_max: f32::INFINITY,up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,};for _ in 0..warmup {diff --git a/crates/uzu/tests/moe/moe_experts_test.rs b/crates/uzu/tests/moe/moe_experts_test.rsindex a7a0e3892..663d0ed4b 100644--- a/crates/uzu/tests/moe/moe_experts_test.rs+++ b/crates/uzu/tests/moe/moe_experts_test.rs@@ -11,9 +11,9 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::metal::{- KernelDataType,- kernel::moe::{+use uzu::{+ DataType,+ backends::metal::kernel::moe::{MoeExpertsSingleDecodeKernels, MoeExpertsTwoPassArguments, MoeExpertsTwoPassDecodeKernels,MoeExpertsTwoPassPrefillKernel,},@@ -446,7 +446,7 @@ fn test_two_pass_decode_correctness() {up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},);@@ -586,7 +586,7 @@ fn test_two_pass_decode_multi_token() {up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},);cb.commit();@@ -689,7 +689,7 @@ fn test_two_pass_prefill_correctness() {up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,silu_alpha,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},).expect("encode");@@ -815,7 +815,7 @@ fn test_fused_single_token_decode() {gate_clip_max: f32::INFINITY,up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},);@@ -938,7 +938,7 @@ fn test_fused_single_token_k4() {gate_clip_max: f32::INFINITY,up_clip_min: f32::NEG_INFINITY,up_clip_max: f32::INFINITY,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},);cb.commit();diff --git a/crates/uzu/tests/moe/moe_finalize_test.rs b/crates/uzu/tests/moe/moe_finalize_test.rsindex 225fa4565..c74e621e6 100644--- a/crates/uzu/tests/moe/moe_finalize_test.rs+++ b/crates/uzu/tests/moe/moe_finalize_test.rs@@ -3,9 +3,9 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::{- common::kernel::MoeFinalizeKernel,- metal::{KernelDataType, kernel::dsl::MoeFinalizeMetalKernel},+use uzu::{+ DataType,+ backends::{common::kernel::MoeFinalizeKernel, metal::kernel::dsl::MoeFinalizeMetalKernel},};use super::test_utils::{alloc_buffer, alloc_buffer_with_data, assert_bf16_close, create_ctx};@@ -85,7 +85,7 @@ fn test_finalize_correctness() {let y_out_buf = alloc_buffer::<bf16>(&ctx, t * d_model);// Execute finalize kernel- let finalize = MoeFinalizeMetalKernel::new(&ctx, KernelDataType::BFloat16.into()).expect("finalize kernel");+ let finalize = MoeFinalizeMetalKernel::new(&ctx, DataType::BF16).expect("finalize kernel");let cb = ctx.command_queue.command_buffer().expect("Failed to create command buffer");let encoder = cb.new_compute_command_encoder().expect("encoder");finalize.encode(diff --git a/crates/uzu/tests/moe/moe_gather_test.rs b/crates/uzu/tests/moe/moe_gather_test.rsindex a568bc3ee..e9e1bc743 100644--- a/crates/uzu/tests/moe/moe_gather_test.rs+++ b/crates/uzu/tests/moe/moe_gather_test.rs@@ -3,9 +3,9 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::metal::kernel::{- KernelDataType,- moe::{MoeGatherArguments, MoeGatherKernels},+use uzu::{+ DataType,+ backends::metal::kernel::moe::{MoeGatherArguments, MoeGatherKernels},};use super::test_utils::{alloc_buffer, alloc_buffer_with_data, assert_bf16_close, create_ctx};@@ -78,7 +78,7 @@ fn test_gather_correctness() {let cb = ctx.command_queue.command_buffer().expect("Failed to create command buffer");gather.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,&MoeGatherArguments {x_buffer: &x_buf,bucketed_ids_buffer: &ids_buf,diff --git a/crates/uzu/tests/moe/moe_perf_test.rs b/crates/uzu/tests/moe/moe_perf_test.rsindex 6766e7eda..5759b538e 100644--- a/crates/uzu/tests/moe/moe_perf_test.rs+++ b/crates/uzu/tests/moe/moe_perf_test.rs@@ -3,11 +3,11 @@ use std::time::Instant;use half::bf16;use metal::{MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::{- common::kernel::{MoeCountsOffsetsFusedKernel, MoeFinalizeKernel},- metal::{- KernelDataType,- kernel::{+use uzu::{+ DataType,+ backends::{+ common::kernel::{MoeCountsOffsetsFusedKernel, MoeFinalizeKernel},+ metal::kernel::{dsl::{MoeCountsOffsetsFusedMetalKernel, MoeFinalizeMetalKernel},moe::{MoeExpertsTwoPassArguments, MoeExpertsTwoPassDecodeKernels, MoeGatherArguments, MoeGatherKernels,@@ -123,7 +123,7 @@ fn test_moe_e2e_decode_perf() {router_topk.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,MoeRouterTopKArguments {input_buffer: &x_buf,weight_buffer: &router_w_buf,@@ -186,7 +186,7 @@ fn test_moe_e2e_prefill_perf() {router_topk.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,MoeRouterTopKArguments {input_buffer: &x_buf,weight_buffer: &router_w_buf,@@ -301,7 +301,7 @@ fn test_moe_pipeline_breakdown_decode() {let scatter_kernel = MoeScatterKernels::new(&ctx).expect("scatter");let gather_kernel = MoeGatherKernels::new(&ctx).expect("gather");let experts_kernel = MoeExpertsTwoPassDecodeKernels::new(&ctx).expect("experts two-pass decode");- let finalize_kernel = MoeFinalizeMetalKernel::new(&ctx, KernelDataType::BFloat16.into()).expect("finalize");+ let finalize_kernel = MoeFinalizeMetalKernel::new(&ctx, DataType::BF16).expect("finalize");let router_topk_fused_kernel = MoeRouterTopKKernel::new(&ctx).expect("router+topk fused");// Testing: Router + TopK + Counts+Offsets (FUSED)@@ -310,7 +310,7 @@ fn test_moe_pipeline_breakdown_decode() {router_topk_fused_kernel.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,MoeRouterTopKArguments {input_buffer: &x_buf,weight_buffer: &router_w_buf,@@ -382,7 +382,7 @@ fn test_moe_pipeline_breakdown_decode() {},tok2row_buffer: &tok2row_buf,},- KernelDataType::BFloat16,+ DataType::BF16,).expect("scatter");cb.commit();@@ -393,7 +393,7 @@ fn test_moe_pipeline_breakdown_decode() {let cb = ctx.command_queue.command_buffer().expect("Failed to create command buffer");gather_kernel.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,&MoeGatherArguments {x_buffer: &x_buf,bucketed_ids_buffer: &bucketed_ids_buf,@@ -438,7 +438,7 @@ fn test_moe_pipeline_breakdown_decode() {up_clip_min: -19.0,up_clip_max: 21.0,silu_alpha: 1.702,- data_type: KernelDataType::BFloat16,+ data_type: DataType::BF16,},);cb.commit();diff --git a/crates/uzu/tests/moe/moe_router_topk_test.rs b/crates/uzu/tests/moe/moe_router_topk_test.rsindex 7e5a11ead..55f0841f7 100644--- a/crates/uzu/tests/moe/moe_router_topk_test.rs+++ b/crates/uzu/tests/moe/moe_router_topk_test.rs@@ -3,9 +3,12 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::metal::{- KernelDataType, MTLContext,- kernel::moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+use uzu::{+ DataType,+ backends::metal::{+ MTLContext,+ kernel::moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+ },};use super::test_utils::{alloc_buffer, alloc_buffer_with_data, create_ctx};@@ -167,7 +170,7 @@ fn run_router_topk_once(k,renorm,};- kernel.encode(&cb, KernelDataType::BFloat16, args).expect("encode fused router+topk");+ kernel.encode(&cb, DataType::BF16, args).expect("encode fused router+topk");cb.commit();cb.wait_until_completed();diff --git a/crates/uzu/tests/moe/moe_scatter_test.rs b/crates/uzu/tests/moe/moe_scatter_test.rsindex c68f7fb68..857f0a781 100644--- a/crates/uzu/tests/moe/moe_scatter_test.rs+++ b/crates/uzu/tests/moe/moe_scatter_test.rs@@ -3,12 +3,15 @@use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue};use rand::{RngExt, SeedableRng, rngs::StdRng};-use uzu::backends::{- common::kernel::MoeCountsOffsetsFusedKernel,- metal::kernel::{- KernelDataType, MoeBlockBasesArguments, MoeScatterArguments, MoeScatterKernels,- dsl::MoeCountsOffsetsFusedMetalKernel,- moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+use uzu::{+ DataType,+ backends::{+ common::kernel::MoeCountsOffsetsFusedKernel,+ metal::kernel::{+ MoeBlockBasesArguments, MoeScatterArguments, MoeScatterKernels,+ dsl::MoeCountsOffsetsFusedMetalKernel,+ moe::{MoeRouterTopKArguments, MoeRouterTopKKernel},+ },},};@@ -82,7 +85,7 @@ fn test_scatter_buckets_parity() {router_topk.encode(&cb,- KernelDataType::BFloat16,+ DataType::BF16,MoeRouterTopKArguments {input_buffer: &input_buf,weight_buffer: &weight_buf,@@ -175,7 +178,7 @@ fn test_scatter_buckets_parity() {num_blocks,num_tiles,},- KernelDataType::BFloat16,+ DataType::BF16,).expect("encode scatter");cb.commit();diff --git a/crates/uzu/tests/sampling_test.rs b/crates/uzu/tests/sampling_test.rsindex 14a27695b..a186cb98c 100644--- a/crates/uzu/tests/sampling_test.rs+++ b/crates/uzu/tests/sampling_test.rs@@ -14,7 +14,7 @@ use uzu::{},},metal::{- KernelDataType, MTLContext, Metal,+ MTLContext, Metal,kernel::dsl::{GumbelMetalKernel, MinPMetalKernel, TemperatureMetalKernel, TopKMetalKernel, TopPMetalKernel,},@@ -1012,8 +1012,7 @@ fn test_temperature_gpu_cpu_match() {const RTOL: f32 = 1e-6;const ATOL: f32 = 1e-6;- let kernel = TemperatureMetalKernel::new(&context, KernelDataType::Float32.into())- .expect("Failed to create temperature kernel");+ let kernel = TemperatureMetalKernel::new(&context, DataType::F32).expect("Failed to create temperature kernel");let logits: Vec<f32> = (0..BATCH * VOCAB).map(|i| ((i * 37 % 1000) as f32 - 500.0) * 0.01).collect();@@ -1074,7 +1073,7 @@ fn test_topk_gpu_cpu_match() {const VOCAB: usize = 1024;const TOPK: u32 = 16;- let kernel = TopKMetalKernel::new(&context, KernelDataType::Float32.into()).expect("Failed to create topk kernel");+ let kernel = TopKMetalKernel::new(&context, DataType::F32).expect("Failed to create topk kernel");let mut rng = StdRng::seed_from_u64(42);let mut logits = vec![0.0f32; BATCH * VOCAB];@@ -1138,7 +1137,7 @@ fn test_topp_gpu_cpu_match() {const VOCAB: usize = 1024;const TOPP: f32 = 0.9;- let kernel = TopPMetalKernel::new(&context, KernelDataType::Float32.into()).expect("Failed to create topp kernel");+ let kernel = TopPMetalKernel::new(&context, DataType::F32).expect("Failed to create topp kernel");let mut rng = StdRng::seed_from_u64(42);let mut logits = vec![0.0f32; BATCH * VOCAB];@@ -1207,7 +1206,7 @@ fn test_minp_gpu_cpu_match() {const VOCAB: usize = 1024;const MINP: f32 = 0.1;- let kernel = MinPMetalKernel::new(&context, KernelDataType::Float32.into()).expect("Failed to create minp kernel");+ let kernel = MinPMetalKernel::new(&context, DataType::F32).expect("Failed to create minp kernel");let mut rng = StdRng::seed_from_u64(42);let mut logits = vec![0.0f32; BATCH * VOCAB];@@ -1391,8 +1390,7 @@ fn test_gumbel_gpu_cpu_match() {const RTOL: f32 = 0.01;const ATOL: f32 = 1e-6;- let kernel =- GumbelMetalKernel::new(&context, KernelDataType::Float32.into()).expect("Failed to create gumbel kernel");+ let kernel = GumbelMetalKernel::new(&context, DataType::F32).expect("Failed to create gumbel kernel");let logits = vec![0.0f32; BATCH * VOCAB];let seeds: Vec<u64> = (0_u64..BATCH as u64).collect();diff --git a/crates/uzu/tests/ssd_prefill_kernel_test.rs b/crates/uzu/tests/ssd_prefill_kernel_test.rsindex 75d36df52..a4bd00717 100644--- a/crates/uzu/tests/ssd_prefill_kernel_test.rs+++ b/crates/uzu/tests/ssd_prefill_kernel_test.rs@@ -7,10 +7,11 @@ use metal::{MTLBlitCommandEncoder, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue, MTLDeviceExt, MTLResourceOptions,};use uzu::{+ DataType,backends::{common::Context,metal::{- KernelDataType, MTLBuffer, MTLContext, ProtocolObject,+ MTLBuffer, MTLContext, ProtocolObject,kernel::ssm::{SSDPrefillArguments, SSDPrefillKernels, SSDPrefillMode,conv1d_scan::{Conv1dScanArguments, Conv1dScanKernel},@@ -362,7 +363,7 @@ fn assert_deterministic_for_mode(mode: SSDPrefillMode) {eprintln!("Skipping SSD prefill determinism test: no Metal device");return;};- let kernel = SSDPrefillKernels::new(&ctx, KernelDataType::Float32).unwrap();+ let kernel = SSDPrefillKernels::new(&ctx, DataType::F32).unwrap();let fixture = SSDPrefillFixture::new();let (y_a, state_a, _) = run_prefill_kernel_mode(&ctx, &kernel, &fixture, mode);@@ -377,7 +378,7 @@ fn assert_matches_cpu_reference(mode: SSDPrefillMode) {eprintln!("Skipping SSD prefill reference test: no Metal device");return;};- let kernel = SSDPrefillKernels::new(&ctx, KernelDataType::Float32).unwrap();+ let kernel = SSDPrefillKernels::new(&ctx, DataType::F32).unwrap();let fixture = SSDPrefillFixture::new();let (y_ref, state_ref) = ssd_prefill_cpu_reference(@@ -464,7 +465,7 @@ fn conv1d_scan_is_deterministic() {return;};let activation = Activation::Identity;- let kernel = Conv1dScanKernel::new(&ctx, KernelDataType::Float32, &activation).unwrap();+ let kernel = Conv1dScanKernel::new(&ctx, DataType::F32, &activation).unwrap();let suffix_len = 192usize;let channels = 8usize;diff --git a/crates/uzu/tests/ssm_update_test.rs b/crates/uzu/tests/ssm_update_test.rsindex 67fa85419..115fed290 100644--- a/crates/uzu/tests/ssm_update_test.rs+++ b/crates/uzu/tests/ssm_update_test.rs@@ -3,9 +3,12 @@use bytemuck;use half::bf16;use metal::{MTLBuffer, MTLCommandBuffer, MTLCommandEncoder, MTLCommandQueue, MTLDeviceExt, MTLResourceOptions};-use uzu::backends::{- common::{Context, kernel::SSDUpdateKernel},- metal::{KernelDataType, MTLContext, kernel::dsl::SSDUpdateMetalKernel},+use uzu::{+ DataType,+ backends::{+ common::{Context, kernel::SSDUpdateKernel},+ metal::{MTLContext, kernel::dsl::SSDUpdateMetalKernel},+ },};#[allow(dead_code)]@@ -235,7 +238,7 @@ fn ssd_update_with_z_bf16() {.new_buffer(bsz * h * dh * n * std::mem::size_of::<bf16>(), MTLResourceOptions::STORAGE_MODE_SHARED).expect("Failed to create buffer");- let kernel = SSDUpdateMetalKernel::new(&ctx, KernelDataType::BFloat16.into()).unwrap();+ let kernel = SSDUpdateMetalKernel::new(&ctx, DataType::BF16).unwrap();let cb_ref = ctx.command_queue.command_buffer().expect("Failed to create command buffer");let cb = cb_ref.to_owned();let enc = cb.new_compute_command_encoder().expect("Failed to create compute encoder");