A high-performance inference engine for AI models
Weaver speedup(10%) + cleanup (#658)
CCYeh780d4901cbdb
diff --git a/crates/backend-uzu/src/backends/cpu/kernel/attention/ancestor_attention.rs b/crates/backend-uzu/src/backends/cpu/kernel/attention/ancestor_attention.rsindex 0522ecf47..b8d5fc7f2 100644--- a/crates/backend-uzu/src/backends/cpu/kernel/attention/ancestor_attention.rs+++ b/crates/backend-uzu/src/backends/cpu/kernel/attention/ancestor_attention.rs@@ -7,10 +7,11 @@ use super::attention_single_pass::attention_single_pass;#[variants(HEAD_DIM, 128)]fn ancestor_attention<const HEAD_DIM: u32>(prefix_kv: *const bf16,- node_kv: *const bf16,+ node_kv: *mut bf16,current_qkv: *const bf16,ancestor_indices: *const u32,ancestor_counts: *const u32,+ node_indices: *const u32,output: *mut bf16,rows: u32,prefix_length: u32,@@ -91,6 +92,19 @@ fn ancestor_attention<const HEAD_DIM: u32>(false,false,);+ if node_capacity > 0 {+ let node = (*node_indices.add(row) as usize).min(last_node);+ std::ptr::copy_nonoverlapping(+ current_row.add(KEY_COMPONENT * model_dim),+ node_kv.add(node * model_dim),+ model_dim,+ );+ std::ptr::copy_nonoverlapping(+ current_row.add(VALUE_COMPONENT * model_dim),+ node_kv.add(node_capacity * model_dim + node * model_dim),+ model_dim,+ );+ }}}}diff --git a/crates/backend-uzu/src/backends/cpu/kernel/weaver/mod.rs b/crates/backend-uzu/src/backends/cpu/kernel/weaver/mod.rsindex f7c44e340..a199938d1 100644--- a/crates/backend-uzu/src/backends/cpu/kernel/weaver/mod.rs+++ b/crates/backend-uzu/src/backends/cpu/kernel/weaver/mod.rs@@ -1,4 +1,3 @@-pub mod weaver_frontier_scatter;+pub mod weaver_frontier_insert_children;pub mod weaver_frontier_select;-pub mod weaver_node_cache_write;pub mod weaver_top_children;diff --git a/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_scatter.rs b/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_insert_children.rssimilarity index 97%rename from crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_scatter.rsrename to crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_insert_children.rsindex c8f94c7f3..71f6647c8 100644--- a/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_scatter.rs+++ b/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_frontier_insert_children.rs@@ -13,8 +13,8 @@ fn top_k_score_key(score: f32) -> u32 {}}-#[kernel(WeaverFrontierScatter)]-pub fn weaver_frontier_scatter(+#[kernel(WeaverFrontierInsertChildren)]+pub fn weaver_frontier_insert_children(packed_tree: *const u32,node_metadata: *const u32,node_valid: *const u32,diff --git a/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_node_cache_write.rs b/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_node_cache_write.rsdeleted file mode 100644index 1f4c177d6..000000000--- a/crates/backend-uzu/src/backends/cpu/kernel/weaver/weaver_node_cache_write.rs+++ /dev/null@@ -1,34 +0,0 @@-use half::bf16;-use num_traits::Float;-use proc_macros::kernel;--use crate::array::ArrayElement;--#[kernel(WeaverNodeCacheWrite)]-#[variants(T, f32, bf16)]-pub fn weaver_node_cache_write<T: ArrayElement + Float>(- current_qkv: *const T,- node_kv: *mut T,- node_indices: *const u32,- model_dim: u32,- node_capacity: u32,- total: u32,-) {- if node_capacity == 0 {- return;- }- let model_dim = model_dim as usize;- let qkv_width = 3 * model_dim;- let last_node = (node_capacity - 1) as usize;- unsafe {- for position in 0..total as usize {- let row = position / (2 * model_dim);- let offset = position - row * 2 * model_dim;- let component = offset / model_dim;- let component_offset = offset % model_dim;- let node = (*node_indices.add(row) as usize).min(last_node);- *node_kv.add(component * node_capacity as usize * model_dim + node * model_dim + component_offset) =- *current_qkv.add(row * qkv_width + (component + 1) * model_dim + component_offset);- }- }-}diff --git a/crates/backend-uzu/src/backends/metal/kernel/attention/ancestor_attention.metal b/crates/backend-uzu/src/backends/metal/kernel/attention/ancestor_attention.metalindex e615d580f..519d6ec6d 100644--- a/crates/backend-uzu/src/backends/metal/kernel/attention/ancestor_attention.metal+++ b/crates/backend-uzu/src/backends/metal/kernel/attention/ancestor_attention.metal@@ -53,10 +53,11 @@ template <uint HEAD_DIM>VARIANTS(HEAD_DIM, 128)PUBLIC KERNEL(AncestorAttention)(const device bfloat* prefix_kv,- const device bfloat* node_kv,+ device bfloat* node_kv,const device bfloat* current_qkv,const device uint* ancestor_indices,const device uint* ancestor_counts,+ const device uint* node_indices,device bfloat* output,constant uint& rows,constant uint& prefix_length,@@ -86,7 +87,7 @@ PUBLIC KERNEL(AncestorAttention)(const uint prefix_value_offset = prefix_length * model_vectors + head_offset;const uint node_value_offset = node_capacity * model_vectors + head_offset;const device bfloat4* prefix_vectors = (const device bfloat4*)prefix_kv;- const device bfloat4* node_vectors = (const device bfloat4*)node_kv;+ device bfloat4* node_vectors = (device bfloat4*)node_kv;const device bfloat4* current_row = (const device bfloat4*)current_qkv + row * qkv_vectors;const float4 query = float4(current_row[head_offset]) * scale;@@ -127,5 +128,12 @@ PUBLIC KERNEL(AncestorAttention)(return current_row;});+ if (node_capacity > 0u) {+ const uint node = min(node_indices[row], last_node);++ node_vectors[node * model_vectors + head_offset] = current_row[current_key_offset];+ node_vectors[node_value_offset + node * model_vectors] = current_row[current_value_offset];+ }+((device bfloat4*)output)[(row * num_heads + head) * vectors_per_head + lane] = bfloat4(values / sum);}diff --git a/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_scatter.metal b/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_insert_children.metalsimilarity index 97%rename from crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_scatter.metalrename to crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_insert_children.metalindex ee2f95c33..d79cf432e 100644--- a/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_scatter.metal+++ b/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_frontier_insert_children.metal@@ -5,7 +5,7 @@using namespace metal;-PUBLIC KERNEL(WeaverFrontierScatter)(+PUBLIC KERNEL(WeaverFrontierInsertChildren)(const device uint* packed_tree,const device uint* node_metadata,const device uint* node_valid,diff --git a/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_node_cache_write.metal b/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_node_cache_write.metaldeleted file mode 100644index a9b6c6bcf..000000000--- a/crates/backend-uzu/src/backends/metal/kernel/weaver/weaver_node_cache_write.metal+++ /dev/null@@ -1,29 +0,0 @@-#include <metal_stdlib>-#include "../common/dsl.h"--// Scatters the key/value halves of the current step's packed QKV rows into their-// node-cache slots. Kept out of AncestorAttention so that the node arena stays-// read-only for the whole attention dispatch.-template <typename T>-VARIANTS(T, float, bfloat)-PUBLIC KERNEL(WeaverNodeCacheWrite)(- const device T* current_qkv,- device T* node_kv,- const device uint* node_indices,- constant uint& model_dim,- constant uint& node_capacity,- constant uint& total,- const uint position AXIS(total, 256)-) {- if (node_capacity == 0) {- return;- }- const uint row = position / (2 * model_dim);- const uint offset = position - row * 2 * model_dim;- const uint component = offset / model_dim;- const uint component_offset = offset % model_dim;- const uint qkv_width = 3 * model_dim;- const uint node = min(node_indices[row], node_capacity - 1u);- node_kv[component * node_capacity * model_dim + node * model_dim + component_offset] =- current_qkv[row * qkv_width + (component + 1) * model_dim + component_offset];-}diff --git a/crates/backend-uzu/src/encodable_block/weaver.rs b/crates/backend-uzu/src/encodable_block/weaver.rsindex 4ee117304..b0f928c55 100644--- a/crates/backend-uzu/src/encodable_block/weaver.rs+++ b/crates/backend-uzu/src/encodable_block/weaver.rs@@ -3,14 +3,13 @@ use thiserror::Error;use crate::{array::size_for_shape,backends::common::{- Allocation, AllocationType, AsBufferRangeMut, Backend, Context, Encoder, Kernels,+ Allocation, AsBufferRangeMut, Backend, Encoder, Kernels,gpu_types::weaver::{CANDIDATES_MAX, FRONTIER_MAX_SLOTS, FRONTIER_NO_WINNER, FrontierIdx, MetadataIdx, TreeIdx,},kernel::{ActivationKernel, AncestorAttentionKernel, AttentionPrepareKernel, TensorAddBiasKernel,- WeaverFrontierScatterKernel, WeaverFrontierSelectKernel, WeaverNodeCacheWriteKernel,- WeaverTopChildrenKernel,+ WeaverFrontierInsertChildrenKernel, WeaverFrontierSelectKernel, WeaverTopChildrenKernel,},},config::{normalization::NormalizationConfig, weaver::WeaverConfig},@@ -120,7 +119,7 @@ pub struct Weaver<B: Backend> {node_position_add: <B::Kernels as Kernels>::TensorAddBiasKernel,top_children: <B::Kernels as Kernels>::WeaverTopChildrenKernel,frontier_select: <B::Kernels as Kernels>::WeaverFrontierSelectKernel,- frontier_scatter: <B::Kernels as Kernels>::WeaverFrontierScatterKernel,+ frontier_insert_children: <B::Kernels as Kernels>::WeaverFrontierInsertChildrenKernel,model_dim: usize,target_model_dim: usize,max_depth: usize,@@ -133,7 +132,6 @@ struct WeaverLayer<B: Backend> {attention_prepare: <B::Kernels as Kernels>::AttentionPrepareKernel,prefix_attention: AttentionCores<B>,ancestor_attention: <B::Kernels as Kernels>::AncestorAttentionKernel,- node_cache_write: <B::Kernels as Kernels>::WeaverNodeCacheWriteKernel,out_projection: Box<dyn Linear<B>>,// MLP@@ -275,8 +273,8 @@ impl<B: Backend> Weaver<B> {<B::Kernels as Kernels>::WeaverTopChildrenKernel::new(context).map_err(WeaverNewError::Backend)?;let frontier_select =<B::Kernels as Kernels>::WeaverFrontierSelectKernel::new(context).map_err(WeaverNewError::Backend)?;- let frontier_scatter =- <B::Kernels as Kernels>::WeaverFrontierScatterKernel::new(context).map_err(WeaverNewError::Backend)?;+ let frontier_insert_children = <B::Kernels as Kernels>::WeaverFrontierInsertChildrenKernel::new(context)+ .map_err(WeaverNewError::Backend)?;Ok(Self {token_embedding_norm,token_embedding_projection,@@ -290,7 +288,7 @@ impl<B: Backend> Weaver<B> {node_position_add,top_children,frontier_select,- frontier_scatter,+ frontier_insert_children,model_dim: config.model_dim,target_model_dim: config.target_model_dim,max_depth: config.max_depth,@@ -301,7 +299,6 @@ impl<B: Backend> Weaver<B> {&self,inputs: WeaverInputs<'_, B>,shape: TreeShape,- context: &B::Context,encoder: &mut Encoder<B>,) -> Result<EncodedWeaverTree<B>, WeaverEncodeError<B>> {let tree_slot_count = shape.budget + 1;@@ -323,7 +320,7 @@ impl<B: Backend> Weaver<B> {.encode_prefix(inputs.target_hidden, inputs.draft_hidden, lookahead_count, encoder).map_err(WeaverEncodeError::Backend)?;let mut node_cache =- self.create_node_expansion_kv_cache(tree_slot_count, context).map_err(WeaverEncodeError::Backend)?;+ self.create_node_expansion_kv_cache(tree_slot_count, encoder).map_err(WeaverEncodeError::Backend)?;let mut tree_init = vec![0u32; TreeIdx::COUNT * tree_slot_count];for slot in 0..tree_slot_count {@@ -395,6 +392,9 @@ impl<B: Backend> Weaver<B> {encoder,);}+ if batch_start_slot + batch_node_count == tree_slot_count {+ break; // packed-tree slots filled+ }let (batch_candidate_ids, batch_candidate_logits) = if batch_start_slot == 0 {(inputs.candidates.ids, inputs.candidates.logits)} else {@@ -413,7 +413,7 @@ impl<B: Backend> Weaver<B> {candidates_per_node: inputs.candidates.candidates_per_depth,depth_seeds: &depth_seeds,};- let selected_children = self.encode_node_expansion(+ let top_children = self.encode_node_expansion(&prefix_cache,&nodes,&candidates,@@ -422,12 +422,12 @@ impl<B: Backend> Weaver<B> {inputs.target_embedding,encoder,)?;- self.frontier_scatter.encode(+ self.frontier_insert_children.encode(&packed_tree,&node_metadata,&node_valid,- &selected_children.token_ids,- &selected_children.logprobs,+ &top_children.token_ids,+ &top_children.logprobs,&mut frontier,frontier_capacity as u32,tree_slot_count as u32,@@ -497,13 +497,12 @@ impl<B: Backend> Weaver<B> {fn create_node_expansion_kv_cache(&self,capacity: usize,- context: &B::Context,+ encoder: &mut Encoder<B>,) -> Result<NodeExpansionKvCache<B>, B::Error> {assert!(capacity > 0, "Weaver node capacity must be positive");let kv_size = size_for_shape(&[2, capacity, self.model_dim], DATA_TYPE);- let layers = (0..self.layers.len())- .map(|_| context.create_allocation(kv_size, AllocationType::Global))- .collect::<Result<Box<[_]>, _>>()?;+ let layers =+ (0..self.layers.len()).map(|_| encoder.allocate_scratch(kv_size)).collect::<Result<Box<[_]>, _>>()?;Ok(NodeExpansionKvCache {layers,capacity: capacity as u32,@@ -711,8 +710,6 @@ impl<B: Backend> WeaverLayer<B> {let ancestor_attention =<B::Kernels as Kernels>::AncestorAttentionKernel::new(context, head_dim as u32, num_heads as u32).map_err(WeaverNewError::Backend)?;- let node_cache_write = <B::Kernels as Kernels>::WeaverNodeCacheWriteKernel::new(context, DATA_TYPE)- .map_err(WeaverNewError::Backend)?;Ok(Self {qkv_projection,out_projection,@@ -724,7 +721,6 @@ impl<B: Backend> WeaverLayer<B> {down_projection,activation,ancestor_attention,- node_cache_write,attention_scale,model_dim,hidden_dim,@@ -780,9 +776,7 @@ impl<B: Backend> WeaverLayer<B> {let mut queries =encoder.allocate_scratch(size_for_shape(&[self.num_heads, token_count, self.head_dim], DATA_TYPE))?;let kv_plane_bytes = size_for_shape(&[token_count, self.model_dim], DATA_TYPE);- let mut kv_cache = encoder- .context()- .create_allocation(size_for_shape(&[2, token_count, self.model_dim], DATA_TYPE), AllocationType::Global)?;+ let mut kv_cache = encoder.allocate_scratch(size_for_shape(&[2, token_count, self.model_dim], DATA_TYPE))?;let (keys, values) = kv_cache.as_buffer_range_mut().split_at(kv_plane_bytes);self.attention_prepare.encode(&qkv,@@ -826,10 +820,11 @@ impl<B: Backend> WeaverLayer<B> {encoder.allocate_scratch(size_for_shape(&[nodes.count, self.model_dim], DATA_TYPE))?;self.ancestor_attention.encode(prefix_kv,- &*node_kv_cache,+ node_kv_cache,¤t_qkv,nodes.ancestor_indices,ancestor_counts,+ tree_slot_indices,&mut attention_output,nodes.count as u32,prefix_length as u32,@@ -838,17 +833,6 @@ impl<B: Backend> WeaverLayer<B> {self.attention_scale,encoder,);- // Separate dispatch so the node arena is read-only above; the encoder- // serializes it after the attention that reads the ancestor slots.- self.node_cache_write.encode(- ¤t_qkv,- node_kv_cache,- tree_slot_indices,- self.model_dim as u32,- node_capacity,- (nodes.count * 2 * self.model_dim) as u32,- encoder,- );self.encode_post_attention(attention_output, residual_state, nodes.count, encoder)}diff --git a/crates/backend-uzu/src/speculators/dflash_speculator.rs b/crates/backend-uzu/src/speculators/dflash_speculator.rsindex 3f9a9d8a3..6ec23c0ad 100644--- a/crates/backend-uzu/src/speculators/dflash_speculator.rs+++ b/crates/backend-uzu/src/speculators/dflash_speculator.rs@@ -186,7 +186,6 @@ impl<B: Backend> DFlashSpeculator<B> {frontier_width: options.frontier_width,children_per_node: options.children_per_node,},- &self.context,&mut encoder,)?;let completed = encoder.end_encoding().submit().wait_until_completed().map_err(DFlashTreeError::Backend)?;diff --git a/crates/backend-uzu/tests/unit/backends/common/kernel/attention/ancestor_attention_test.rs b/crates/backend-uzu/tests/unit/backends/common/kernel/attention/ancestor_attention_test.rsindex 04356b975..d07a35363 100644--- a/crates/backend-uzu/tests/unit/backends/common/kernel/attention/ancestor_attention_test.rs+++ b/crates/backend-uzu/tests/unit/backends/common/kernel/attention/ancestor_attention_test.rs@@ -28,6 +28,7 @@ struct Runner<B: Backend> {current_qkv: Allocation<B>,ancestor_indices: Allocation<B>,ancestor_counts: Allocation<B>,+ node_indices: Allocation<B>,output: Allocation<B>,rows: u32,prefix_length: u32,@@ -52,6 +53,7 @@ impl<B: Backend> Runner<B> {.collect::<Vec<_>>()};let first_output_node = nodes - rows;+ let node_indices = (first_output_node..nodes).map(|node| node as u32).collect::<Vec<_>>();let ancestor_counts = (0..rows).map(|row| (row % (ancestor_stride + 1)) as u32).collect::<Vec<_>>();let mut ancestor_indices = vec![0; rows * ancestor_stride];for row in 0..rows {@@ -71,6 +73,7 @@ impl<B: Backend> Runner<B> {current_qkv: alloc_allocation_with_data::<B, bf16>(&context, &values(rows * qkv_width, 29)),ancestor_indices: alloc_allocation_with_data::<B, u32>(&context, &ancestor_indices),ancestor_counts: alloc_allocation_with_data::<B, u32>(&context, &ancestor_counts),+ node_indices: alloc_allocation_with_data::<B, u32>(&context, &node_indices),output: alloc_allocation::<B, bf16>(&context, rows * NUM_HEADS as usize * HEAD_DIM),context,kernel,@@ -89,10 +92,11 @@ impl<B: Backend> Runner<B> {for _ in 0..repetitions {self.kernel.encode(&self.prefix_kv,- &self.node_kv,+ &mut self.node_kv,&self.current_qkv,&self.ancestor_indices,&self.ancestor_counts,+ &self.node_indices,&mut self.output,self.rows,self.prefix_length,@@ -109,18 +113,17 @@ impl<B: Backend> Runner<B> {#[uzu_test]fn ancestor_attention_matches_cpu() {let mut cpu = Runner::<Cpu>::new(4, 5, 3, 16);- let initial_nodes = allocation_to_vec::<Cpu, bf16>(&cpu.node_kv);cpu.encode(1);let expected_output = allocation_to_vec::<Cpu, bf16>(&cpu.output);-- assert_eq!(allocation_to_vec::<Cpu, bf16>(&cpu.node_kv), initial_nodes);+ let expected_node_kv = allocation_to_vec::<Cpu, bf16>(&cpu.node_kv);let mut metal = Runner::<Metal>::new(4, 5, 3, 16);metal.encode(1);let actual_output = allocation_to_vec::<Metal, bf16>(&metal.output);+ let actual_node_kv = allocation_to_vec::<Metal, bf16>(&metal.node_kv);assert_eq_float(&expected_output, &actual_output, 0.02, "AncestorAttention output");- assert_eq!(allocation_to_vec::<Metal, bf16>(&metal.node_kv), initial_nodes);+ assert_eq!(actual_node_kv, expected_node_kv);}#[uzu_test]diff --git a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/mod.rs b/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/mod.rsindex 52d3b29e3..2fb0afeb7 100644--- a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/mod.rs+++ b/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/mod.rs@@ -1,3 +1,2 @@mod weaver_frontier_test;-mod weaver_node_cache_write_test;mod weaver_top_children_test;diff --git a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_frontier_test.rs b/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_frontier_test.rsindex 949292fd5..bce06ae78 100644--- a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_frontier_test.rs+++ b/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_frontier_test.rs@@ -5,7 +5,7 @@ use crate::{common::{Backend, Encoder, Kernels,gpu_types::weaver::{FrontierIdx, MetadataIdx, TreeIdx},- kernel::{WeaverFrontierScatterKernel, WeaverFrontierSelectKernel},+ kernel::{WeaverFrontierInsertChildrenKernel, WeaverFrontierSelectKernel},},cpu::Cpu,},@@ -77,7 +77,7 @@ fn select<B: Backend>() -> Vec<u32> {.collect()}-fn scatter<B: Backend>() -> Vec<u32> {+fn insert_children<B: Backend>() -> Vec<u32> {let context = create_context::<B>();let mut tree = vec![0; TreeIdx::COUNT * 4];tree[TreeIdx::PathLogprobBits as usize * 4..(TreeIdx::PathLogprobBits as usize + 1) * 4]@@ -91,7 +91,7 @@ fn scatter<B: Backend>() -> Vec<u32> {let ids = alloc_allocation_with_data::<B, u32>(&context, &(10..19).collect::<Vec<_>>());let scores = alloc_allocation_with_data::<B, f32>(&context, &[-0.1, -0.2, -0.3, 8.0, 8.0, 8.0, 0.1, 0.2, 0.3]);let mut frontier = alloc_allocation_with_data::<B, u32>(&context, &[42; FrontierIdx::COUNT * 16]);- let kernel = <B::Kernels as Kernels>::WeaverFrontierScatterKernel::new(&context).unwrap();+ let kernel = <B::Kernels as Kernels>::WeaverFrontierInsertChildrenKernel::new(&context).unwrap();let mut encoder = Encoder::new(context.as_ref()).unwrap();kernel.encode(&tree, &metadata, &valid, &ids, &scores, &mut frontier, 16, 4, 3, 3, &mut encoder);encoder.end_encoding().submit().wait_until_completed().unwrap();@@ -102,6 +102,6 @@ fn scatter<B: Backend>() -> Vec<u32> {fn weaver_frontier_kernels_match_cpu() {for_each_non_cpu_backend!(|B| {assert_eq!(select::<B>(), select::<Cpu>());- assert_eq!(scatter::<B>(), scatter::<Cpu>());+ assert_eq!(insert_children::<B>(), insert_children::<Cpu>());});}diff --git a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_node_cache_write_test.rs b/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_node_cache_write_test.rsdeleted file mode 100644index bd1260d26..000000000--- a/crates/backend-uzu/tests/unit/backends/common/kernel/weaver/weaver_node_cache_write_test.rs+++ /dev/null@@ -1,98 +0,0 @@-#![cfg(backend = "metal")]--use half::bf16;-use proc_macros::uzu_test;--use crate::{- backends::{- common::{Allocation, Backend, Encoder, Kernels, kernel::WeaverNodeCacheWriteKernel},- cpu::Cpu,- metal::Metal,- },- data_type::DataType,- tests::helpers::{alloc_allocation_with_data, allocation_to_vec, create_context},-};--const MODEL_DIM: usize = 64;-const ROWS: usize = 4;-const NODES: usize = 9;--struct Runner<B: Backend> {- context: std::sync::Arc<B::Context>,- kernel: <B::Kernels as Kernels>::WeaverNodeCacheWriteKernel,- current_qkv: Allocation<B>,- node_kv: Allocation<B>,- node_indices: Allocation<B>,-}--fn values(- length: usize,- offset: usize,-) -> Vec<bf16> {- (0..length).map(|index| bf16::from_f32((((index + offset) * 17 % 251) as f32 - 125.0) / 128.0)).collect()-}--/// Slots are deliberately non-contiguous and out of order so the scatter is-/// exercised rather than a coincidentally-contiguous copy.-const NODE_INDICES: [u32; ROWS] = [7, 1, 8, 4];--impl<B: Backend> Runner<B> {- fn new() -> Self {- let qkv_width = 3 * MODEL_DIM;- let kv_width = 2 * MODEL_DIM;- let context = create_context::<B>();- let kernel =- <B::Kernels as Kernels>::WeaverNodeCacheWriteKernel::new(context.as_ref(), DataType::BF16).unwrap();- Self {- current_qkv: alloc_allocation_with_data::<B, bf16>(&context, &values(ROWS * qkv_width, 29)),- node_kv: alloc_allocation_with_data::<B, bf16>(&context, &values(NODES * kv_width, 11)),- node_indices: alloc_allocation_with_data::<B, u32>(&context, &NODE_INDICES),- context,- kernel,- }- }-- fn run(&mut self) -> Vec<bf16> {- let mut encoder = Encoder::new(self.context.as_ref()).unwrap();- self.kernel.encode(- &self.current_qkv,- &mut self.node_kv,- &self.node_indices,- MODEL_DIM as u32,- NODES as u32,- (ROWS * 2 * MODEL_DIM) as u32,- &mut encoder,- );- encoder.end_encoding().submit().wait_until_completed().unwrap();- allocation_to_vec::<B, bf16>(&self.node_kv)- }-}--#[uzu_test]-fn weaver_node_cache_write_matches_cpu() {- let qkv_width = 3 * MODEL_DIM;-- let mut cpu = Runner::<Cpu>::new();- let initial_nodes = allocation_to_vec::<Cpu, bf16>(&cpu.node_kv);- let current_qkv = allocation_to_vec::<Cpu, bf16>(&cpu.current_qkv);- let expected = cpu.run();-- for component in 0..2 {- for node in 0..NODES {- let destination = component * NODES * MODEL_DIM + node * MODEL_DIM;- match NODE_INDICES.iter().position(|&index| index as usize == node) {- Some(row) => {- let source = row * qkv_width + (component + 1) * MODEL_DIM;- assert_eq!(expected[destination..destination + MODEL_DIM], current_qkv[source..source + MODEL_DIM],);- },- None => assert_eq!(- expected[destination..destination + MODEL_DIM],- initial_nodes[destination..destination + MODEL_DIM],- ),- }- }- }-- let mut metal = Runner::<Metal>::new();- assert_eq!(metal.run(), expected);-}