Skip to content

armgit / uzu

Public Imported from GitHub
0

A high-performance inference engine for AI models

Custom test harness (#533)

Aleksandr Golokoz643faf72bf28
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 36e72261d..68b03ccbf 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -121,7 +121,7 @@ jobs:
- name: Run tests
if: ${{ matrix.backend != 'cpu' && matrix.tracing && matrix.grammar || matrix.backend == 'cpu' && !matrix.tracing && !matrix.grammar && matrix.target == 'wasm32-wasip1-threads' }}
run: |
- cargo +stable nextest run --profile=ci --no-default-features $FEATURE_FLAGS \
+ cargo +nightly nextest run --profile=ci --no-default-features $FEATURE_FLAGS \
${{ matrix.target == 'wasm32-wasip1-threads' && '-p backend-uzu backends::common::kernel::tests --lib' || '' }}
validate-bindings:
diff --git a/crates/backend-uzu/src/inference/chat/bridging.rs b/crates/backend-uzu/src/inference/chat/bridging.rs
index 54457482d..e756babdc 100644
--- a/crates/backend-uzu/src/inference/chat/bridging.rs
+++ b/crates/backend-uzu/src/inference/chat/bridging.rs
@@ -204,6 +204,7 @@ fn build_stats(stats: &Stats) -> ShojiStats {
#[cfg(test)]
mod tests {
+ use proc_macros::uzu_test;
use shoji::types::session::chat::ChatMessage;
use super::*;
@@ -216,12 +217,12 @@ mod tests {
ChatMessage::user().with_text("hi".to_string()).with_reasoning_effort(reasoning_effort)
}
- #[test]
+ #[uzu_test]
fn enable_thinking_defaults_on_without_effort() {
assert!(enable_thinking_for(vec![ChatMessage::user().with_text("hi".to_string())]));
}
- #[test]
+ #[uzu_test]
fn enable_thinking_follows_latest_turn() {
let reenabled = vec![
user(ReasoningEffort::Disabled),
diff --git a/crates/backend-uzu/src/lib.rs b/crates/backend-uzu/src/lib.rs
index f4507b352..f7ea4d243 100644
--- a/crates/backend-uzu/src/lib.rs
+++ b/crates/backend-uzu/src/lib.rs
@@ -1,3 +1,6 @@
+#![cfg_attr(test, feature(custom_test_frameworks, test))]
+#![cfg_attr(test, test_runner(crate::harness::uzu_harness))]
+
// needed for tests to resolve `backend_uzu::` imports
#[cfg(test)]
extern crate self as backend_uzu;
@@ -34,3 +37,7 @@ pub mod _benchmarks {
trie::{TrieCreationConfig, TrieNode},
};
}
+
+#[cfg(test)]
+#[path = "../tests/harness/mod.rs"]
+mod harness;
diff --git a/crates/backend-uzu/src/session/config/tts_run_config.rs b/crates/backend-uzu/src/session/config/tts_run_config.rs
index d8c7780c6..e76a2081d 100644
--- a/crates/backend-uzu/src/session/config/tts_run_config.rs
+++ b/crates/backend-uzu/src/session/config/tts_run_config.rs
@@ -109,9 +109,11 @@ impl Default for TtsRunConfig {
#[cfg(test)]
mod tests {
+ use proc_macros::uzu_test;
+
use super::{TtsChunkPolicy, TtsNonStreamingMode, TtsRunConfig};
- #[test]
+ #[uzu_test]
fn default_run_config_matches_production_defaults() {
let config = TtsRunConfig::default();
assert!(config.streaming_enabled);
@@ -128,7 +130,7 @@ mod tests {
assert!((config.chunk_hysteresis_fraction - 0.25).abs() < f64::EPSILON);
}
- #[test]
+ #[uzu_test]
fn fixed_chunk_constructor_sets_bounds() {
let config = TtsRunConfig::fixed_chunk_frames(64);
assert_eq!(config.chunk_policy, TtsChunkPolicy::Fixed);
@@ -137,7 +139,7 @@ mod tests {
assert_eq!(config.max_chunk_frames, 64);
}
- #[test]
+ #[uzu_test]
fn run_config_validation_rejects_invalid_bounds() {
let invalid = TtsRunConfig {
min_chunk_frames: 32,
@@ -147,7 +149,7 @@ mod tests {
assert!(invalid.validate().is_err());
}
- #[test]
+ #[uzu_test]
fn run_config_validation_rejects_zero_initial_chunk() {
let invalid = TtsRunConfig {
initial_chunk_frames: 0,
@@ -156,7 +158,7 @@ mod tests {
assert!(invalid.validate().is_err());
}
- #[test]
+ #[uzu_test]
fn run_config_validation_rejects_zero_semantic_cap() {
let invalid = TtsRunConfig {
max_semantic_frames: 0,
@@ -165,7 +167,7 @@ mod tests {
assert!(invalid.validate().is_err());
}
- #[test]
+ #[uzu_test]
fn run_config_validation_rejects_zero_workspace_frames() {
let invalid = TtsRunConfig {
max_stream_workspace_frames: 0,
@@ -174,7 +176,7 @@ mod tests {
assert!(invalid.validate().is_err());
}
- #[test]
+ #[uzu_test]
fn run_config_validation_rejects_workspace_smaller_than_max_chunk() {
let invalid = TtsRunConfig {
max_chunk_frames: 256,
diff --git a/crates/backend-uzu/src/session/helpers/output_parser.rs b/crates/backend-uzu/src/session/helpers/output_parser.rs
index 8a36682f6..a035b2c63 100644
--- a/crates/backend-uzu/src/session/helpers/output_parser.rs
+++ b/crates/backend-uzu/src/session/helpers/output_parser.rs
@@ -76,11 +76,13 @@ impl OutputParser {
#[cfg(test)]
mod tests {
+ use proc_macros::uzu_test;
+
use super::*;
const QWEN_REGEX: &str = r"(?s)(?:<think>)?(?P<chain_of_thought>.*?)(?:</think>\s*(?P<response>.*))?\Z";
- #[test]
+ #[uzu_test]
fn thinking_enabled_splits_on_close_tag() {
let parser = OutputParser::new(Some(QWEN_REGEX.to_string())).unwrap();
let parsed = parser.parse("Let me compute.</think>2 + 2 equals 4.".to_string(), true).parsed;
@@ -88,7 +90,7 @@ mod tests {
assert_eq!(parsed.response.as_deref(), Some("2 + 2 equals 4."));
}
- #[test]
+ #[uzu_test]
fn thinking_disabled_keeps_answer_as_response() {
let parser = OutputParser::new(Some(QWEN_REGEX.to_string())).unwrap();
let parsed = parser.parse("2 + 2 equals 4.".to_string(), false).parsed;
@@ -96,7 +98,7 @@ mod tests {
assert_eq!(parsed.response.as_deref(), Some("2 + 2 equals 4."));
}
- #[test]
+ #[uzu_test]
fn thinking_disabled_strips_emitted_reasoning_tags() {
let parser = OutputParser::new(Some(QWEN_REGEX.to_string())).unwrap();
let parsed = parser.parse("<think>wait, reconsider</think>2 + 2 equals 4.".to_string(), false).parsed;
@@ -104,7 +106,7 @@ mod tests {
assert_eq!(parsed.response.as_deref(), Some("2 + 2 equals 4."));
}
- #[test]
+ #[uzu_test]
fn thinking_enabled_without_close_tag_is_still_reasoning() {
let parser = OutputParser::new(Some(QWEN_REGEX.to_string())).unwrap();
let parsed = parser.parse("Let me compute".to_string(), true).parsed;
diff --git a/crates/backend-uzu/tests/harness/mod.rs b/crates/backend-uzu/tests/harness/mod.rs
new file mode 100644
index 000000000..4cccf7a44
--- /dev/null
+++ b/crates/backend-uzu/tests/harness/mod.rs
@@ -0,0 +1,18 @@
+#![cfg(test)]
+
+mod uzu_test;
+
+pub extern crate test;
+
+pub use uzu_test::UzuTest;
+
+pub fn uzu_harness(tests: &[&UzuTest]) {
+ let tests = tests
+ .iter()
+ .filter_map(|test| match test {
+ UzuTest::Test(test) => Some(*test),
+ UzuTest::Bench(_) => None,
+ })
+ .collect::<Vec<_>>();
+ test::test_main_static(&tests)
+}
diff --git a/crates/backend-uzu/tests/harness/uzu_test.rs b/crates/backend-uzu/tests/harness/uzu_test.rs
new file mode 100644
index 000000000..abfcb1036
--- /dev/null
+++ b/crates/backend-uzu/tests/harness/uzu_test.rs
@@ -0,0 +1,7 @@
+extern crate test;
+
+pub enum UzuTest {
+ #[allow(dead_code)]
+ Bench(&'static dyn Fn()),
+ Test(&'static test::TestDescAndFn),
+}
diff --git a/crates/backend-uzu/tests/integration_entry.rs b/crates/backend-uzu/tests/integration_entry.rs
index 6d3bbe604..0b9b7d9f6 100644
--- a/crates/backend-uzu/tests/integration_entry.rs
+++ b/crates/backend-uzu/tests/integration_entry.rs
@@ -1 +1,7 @@
+#![feature(custom_test_frameworks, test)]
+#![test_runner(crate::harness::uzu_harness)]
+
+#[path = "harness/mod.rs"]
+mod harness;
+
mod integration;
diff --git a/crates/backend-uzu/tests/performance_entry.rs b/crates/backend-uzu/tests/performance_entry.rs
index f0d6f6ba7..e291f028d 100644
--- a/crates/backend-uzu/tests/performance_entry.rs
+++ b/crates/backend-uzu/tests/performance_entry.rs
@@ -1,2 +1,8 @@
+#![feature(custom_test_frameworks, test)]
+#![test_runner(crate::harness::uzu_harness)]
+
+#[path = "harness/mod.rs"]
+mod harness;
+
mod common;
mod performance;
diff --git a/crates/backend-uzu/unit/audio/types_test.rs b/crates/backend-uzu/unit/audio/types_test.rs
index 89814d78e..3cc69dbe0 100644
--- a/crates/backend-uzu/unit/audio/types_test.rs
+++ b/crates/backend-uzu/unit/audio/types_test.rs
@@ -1,6 +1,8 @@
+use proc_macros::uzu_test;
+
use super::{AudioError, AudioTokenGrid};
-#[test]
+#[uzu_test]
fn token_grid_get_reads_expected_slots() {
let grid = AudioTokenGrid::new(
vec![
@@ -21,7 +23,7 @@ fn token_grid_get_reads_expected_slots() {
assert_eq!(grid.get(1, 1, 1), 7);
}
-#[test]
+#[uzu_test]
fn invalid_grid_shape_is_rejected() {
let error = AudioTokenGrid::new(vec![0, 1, 2].into_boxed_slice(), 1, 2, 2, vec![2].into_boxed_slice())
.expect_err("shape mismatch should fail");
diff --git a/crates/backend-uzu/unit/backends/common/allocator/allocator.rs b/crates/backend-uzu/unit/backends/common/allocator/allocator.rs
index bc4351f06..a9d00f340 100644
--- a/crates/backend-uzu/unit/backends/common/allocator/allocator.rs
+++ b/crates/backend-uzu/unit/backends/common/allocator/allocator.rs
@@ -6,6 +6,7 @@ use std::{
time::Instant,
};
+use proc_macros::uzu_test;
use serde::{Deserialize, Serialize};
use crate::backends::{
@@ -48,7 +49,7 @@ enum TraceEventType {
},
}
-#[test]
+#[uzu_test]
#[ignore]
fn bench_allocator_generation_trace() {
let file = BufReader::with_capacity(1 << 20, File::open("/tmp/allocator_trace.jsonl").unwrap());
diff --git a/crates/backend-uzu/unit/backends/common/hazard_tracker_test.rs b/crates/backend-uzu/unit/backends/common/hazard_tracker_test.rs
index b3d51d89b..8a794cda4 100644
--- a/crates/backend-uzu/unit/backends/common/hazard_tracker_test.rs
+++ b/crates/backend-uzu/unit/backends/common/hazard_tracker_test.rs
@@ -1,6 +1,8 @@
+use proc_macros::uzu_test;
+
use super::*;
-#[test]
+#[uzu_test]
fn test_hazard_tracker_simple_compute_sequence() {
let mut ht = HazardTracker::new();
diff --git a/crates/backend-uzu/unit/backends/common/kernel/matmul/dispatch_paths_test.rs b/crates/backend-uzu/unit/backends/common/kernel/matmul/dispatch_paths_test.rs
index 09439d4dd..daffa67c5 100644
--- a/crates/backend-uzu/unit/backends/common/kernel/matmul/dispatch_paths_test.rs
+++ b/crates/backend-uzu/unit/backends/common/kernel/matmul/dispatch_paths_test.rs
@@ -4,6 +4,7 @@ use std::fmt::{Debug, Display};
use half::bf16;
use num_traits::Float;
+use proc_macros::uzu_test;
use rstest::rstest;
use crate::{
@@ -63,6 +64,7 @@ fn run_matrix<T: ArrayElement + Float + Debug + Display>(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::base(1.0, false)]
#[case::ab_scale(0.5, false)]
#[case::accumulate(1.0, true)]
@@ -74,12 +76,12 @@ fn matches_cpu_reference_bf16(
run_matrix::<bf16>(|shape| Case::new(shape).with_ab_scale(ab_scale).with_accumulate(accumulate), 1.0);
}
-#[test]
+#[uzu_test]
fn matches_cpu_reference_f32() {
run_matrix::<f32>(Case::new, 0.01);
}
-#[test]
+#[uzu_test]
fn b_transpose_false_bf16() {
run_matrix::<bf16>(
|shape| Case {
@@ -95,7 +97,7 @@ fn rht_shapes() -> impl Iterator<Item = crate::common::matmul::Shape> {
[Shape::new(8, 128, 64), Shape::new(64, 128, 128), Shape::new(128, 2048, 256), Shape::new(33, 128, 64)].into_iter()
}
-#[test]
+#[uzu_test]
fn rht_parity_bf16() {
let context = MetalContext::new().expect("Metal context");
let mut kernel = <<Metal as Backend>::Kernels as Kernels>::MatmulKernel::new(
@@ -113,7 +115,7 @@ fn rht_parity_bf16() {
}
}
-#[test]
+#[uzu_test]
fn bias_parity_bf16() {
use crate::common::matmul::Shape;
let context = MetalContext::new().expect("Metal context");
@@ -133,7 +135,7 @@ fn bias_parity_bf16() {
}
}
-#[test]
+#[uzu_test]
fn gemv_fp_partial_output_block_bf16() {
use crate::common::matmul::Shape;
let context = MetalContext::new().expect("Metal context");
@@ -152,7 +154,7 @@ fn gemv_fp_partial_output_block_bf16() {
}
}
-#[test]
+#[uzu_test]
fn gemv_fp_output_transforms_bf16() {
use crate::common::matmul::Shape;
let context = MetalContext::new().expect("Metal context");
diff --git a/crates/backend-uzu/unit/backends/common/kernel/matmul/gemv_test.rs b/crates/backend-uzu/unit/backends/common/kernel/matmul/gemv_test.rs
index d080c73f5..b65d1bf59 100644
--- a/crates/backend-uzu/unit/backends/common/kernel/matmul/gemv_test.rs
+++ b/crates/backend-uzu/unit/backends/common/kernel/matmul/gemv_test.rs
@@ -2,6 +2,7 @@ use std::fmt::{Debug, Display};
use half::bf16;
use num_traits::Float;
+use proc_macros::uzu_test;
use rstest::rstest;
use crate::{
@@ -106,6 +107,7 @@ fn test<T: ArrayElement + Float + Debug + Display>(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::m1(1, 128, 64)]
#[case::batched(4, 128, 64)]
#[case::max_batch(8, 128, 64)]
@@ -122,6 +124,7 @@ fn gemv_bf16(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::m1(1, 128, 64)]
#[case::batched(4, 128, 64)]
#[case::max_batch(8, 128, 64)]
diff --git a/crates/backend-uzu/unit/backends/common/kernel/matmul/quant_dispatch_test.rs b/crates/backend-uzu/unit/backends/common/kernel/matmul/quant_dispatch_test.rs
index fba4d7f8b..2524b0a1c 100644
--- a/crates/backend-uzu/unit/backends/common/kernel/matmul/quant_dispatch_test.rs
+++ b/crates/backend-uzu/unit/backends/common/kernel/matmul/quant_dispatch_test.rs
@@ -91,6 +91,7 @@ fn run_parity<T: ArrayElement + Float + Debug + Display>(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::gs16_4bit_zp_prefill(64, 256, 64, 16, 4, QuantizationMethod::ScaleZeroPoint)]
#[case::gs32_4bit_mlx_prefill(64, 256, 64, 32, 4, QuantizationMethod::ScaleBias)]
#[case::gs64_4bit_mlx_prefill(64, 256, 64, 64, 4, QuantizationMethod::ScaleBias)]
@@ -112,6 +113,7 @@ fn parity_bf16(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::gs32_8bit_mlx_prefill(64, 256, 64, 32, 8, QuantizationMethod::ScaleBias)]
#[case::gs64_8bit_zp_prefill(64, 256, 64, 64, 8, QuantizationMethod::ScaleZeroPoint)]
#[case::gs128_8bit_zp_prefill(64, 256, 64, 128, 8, QuantizationMethod::ScaleZeroPoint)]
@@ -154,6 +156,7 @@ fn run_parity_gemv<T: ArrayElement + Float + Debug + Display>(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::m1_gs32_4bit_mlx(1, 256, 64, 32, 4, QuantizationMethod::ScaleBias)]
#[case::m1_gs64_4bit_mlx(1, 256, 64, 64, 4, QuantizationMethod::ScaleBias)]
#[case::m1_gs128_4bit_mlx(1, 256, 64, 128, 4, QuantizationMethod::ScaleBias)]
@@ -261,6 +264,7 @@ fn parity_bf16_gemv_qmv_fused_scale_bias() {
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::gs64_4bit(1, 96, 64, 64, 4, QuantizationMethod::ScaleBias)]
#[case::gs64_4bit_zp(2, 96, 64, 64, 4, QuantizationMethod::ScaleZeroPoint)]
#[case::gs128_8bit(1, 192, 64, 128, 8, QuantizationMethod::ScaleBias)]
@@ -278,6 +282,7 @@ fn parity_gemv_partial_group_bf16(
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::n12_gs32_4bit(1, 256, 12, 32, 4, QuantizationMethod::ScaleBias)]
#[case::n20_gs64_4bit_zp(2, 256, 20, 64, 4, QuantizationMethod::ScaleZeroPoint)]
#[case::n36_gs32_8bit(1, 256, 36, 32, 8, QuantizationMethod::ScaleBias)]
@@ -497,6 +502,7 @@ fn quant_gemm_accumulate_returns_unsupported_dop() {
}
#[rstest]
+#[test_attr(uzu_test)]
#[case::gs32_4bit_mlx(128, 256, 64, 32, 4, QuantizationMethod::ScaleBias)]
#[case::gs64_4bit_mlx(128, 256, 64, 64, 4, QuantizationMethod::ScaleBias)]
#[case::gs128_4bit_mlx(128, 256, 64, 128, 4, QuantizationMethod::ScaleBias)]
diff --git a/crates/backend-uzu/unit/backends/metal/sparse/sparse_buffer_test.rs b/crates/backend-uzu/unit/backends/metal/sparse/sparse_buffer_test.rs
index e35cb489d..f16167f9e 100644
--- a/crates/backend-uzu/unit/backends/metal/sparse/sparse_buffer_test.rs
+++ b/crates/backend-uzu/unit/backends/metal/sparse/sparse_buffer_test.rs
@@ -1,5 +1,7 @@
use std::ops::Range;
+use proc_macros::uzu_test;
+
use crate::{
backends::{
common::{Backend, SparseBuffer},
@@ -26,7 +28,7 @@ fn pages_for_heaps(
(heap_offset * pages_per_heap)..((heap_offset + heap_count) * pages_per_heap)
}
-#[test]
+#[uzu_test]
fn test_mapping_succeeds() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
@@ -36,7 +38,7 @@ fn test_mapping_succeeds() {
sparse_buffer.map(ctx.as_ref(), &pages).expect("Failed to map sparse buffer");
}
-#[test]
+#[uzu_test]
fn test_unmapping_succeeds() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
@@ -47,7 +49,7 @@ fn test_unmapping_succeeds() {
sparse_buffer.unmap(ctx.as_ref(), &pages).expect("Failed to unmap sparse buffer");
}
-#[test]
+#[uzu_test]
fn test_partial_unmapping_succeeds() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
@@ -59,7 +61,7 @@ fn test_partial_unmapping_succeeds() {
sparse_buffer.unmap(ctx.as_ref(), &unmapped).expect("Failed to unmap sparse buffer");
}
-#[test]
+#[uzu_test]
fn test_mapping_uses_minimum_heaps() {
let ctx = create_context::<Metal>();
let heap_count = 4;
@@ -74,7 +76,7 @@ fn test_mapping_uses_minimum_heaps() {
assert_eq!(heaps.heaps_count(), expected, "mapping should allocate the minimum number of heaps");
}
-#[test]
+#[uzu_test]
fn test_remapping_reuses_freed_pages() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 6);
@@ -96,7 +98,7 @@ fn test_remapping_reuses_freed_pages() {
);
}
-#[test]
+#[uzu_test]
fn test_full_unmap_then_remap_reuses_heaps() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
@@ -116,7 +118,7 @@ fn test_full_unmap_then_remap_reuses_heaps() {
);
}
-#[test]
+#[uzu_test]
fn test_remapping_same_pages_is_noop() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
@@ -135,7 +137,7 @@ fn test_remapping_same_pages_is_noop() {
);
}
-#[test]
+#[uzu_test]
fn test_mapping_multiple_gaps_reserves_pages_between_gaps() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -152,7 +154,7 @@ fn test_mapping_multiple_gaps_reserves_pages_between_gaps() {
);
}
-#[test]
+#[uzu_test]
fn test_drop_releases_pool_heaps() {
// Regression: a mapped buffer dropped without an explicit unmap must
// still release its heap pages back to the shared pool.
@@ -173,7 +175,7 @@ fn test_drop_releases_pool_heaps() {
);
}
-#[test]
+#[uzu_test]
fn test_drop_does_not_disturb_other_buffer_mappings() {
// Dropping one buffer must not unmap heap pages held by another buffer
// sharing the same pool.
@@ -194,7 +196,7 @@ fn test_drop_does_not_disturb_other_buffer_mappings() {
assert_eq!(current_heaps, baseline, "dropping a transient buffer must leave keeper buffer's mappings intact",);
}
-#[test]
+#[uzu_test]
fn test_sequential_mappings_are_compact() {
let ctx = create_context::<Metal>();
let capacity = buffer_capacity(ctx.as_ref(), 4);
diff --git a/crates/backend-uzu/unit/backends/metal/sparse/sparse_heap_pool_test.rs b/crates/backend-uzu/unit/backends/metal/sparse/sparse_heap_pool_test.rs
index a6d4c794c..4b548c75d 100644
--- a/crates/backend-uzu/unit/backends/metal/sparse/sparse_heap_pool_test.rs
+++ b/crates/backend-uzu/unit/backends/metal/sparse/sparse_heap_pool_test.rs
@@ -1,3 +1,5 @@
+use proc_macros::uzu_test;
+
use crate::{
backends::{
common::SparseBuffer,
@@ -6,13 +8,13 @@ use crate::{
common::helpers::{create_context, sparse_buffer_create},
};
-#[test]
+#[uzu_test]
fn test_new_pool_is_empty() {
let ctx = create_context::<Metal>();
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 0);
}
-#[test]
+#[uzu_test]
fn test_map_empty_range_allocates_no_heaps() {
let ctx = create_context::<Metal>();
let cap = ctx.sparse_heap_pool().heap_capacity_bytes();
@@ -23,7 +25,7 @@ fn test_map_empty_range_allocates_no_heaps() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 0);
}
-#[test]
+#[uzu_test]
fn test_map_single_heap_allocates_one_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -35,7 +37,7 @@ fn test_map_single_heap_allocates_one_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_map_multi_heap_allocates_minimum_heaps() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -47,7 +49,7 @@ fn test_map_multi_heap_allocates_minimum_heaps() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 4);
}
-#[test]
+#[uzu_test]
fn test_map_partial_heap_rounds_up_heap_count() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -59,7 +61,7 @@ fn test_map_partial_heap_rounds_up_heap_count() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 2);
}
-#[test]
+#[uzu_test]
fn test_sequential_mappings_pack_into_one_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -73,7 +75,7 @@ fn test_sequential_mappings_pack_into_one_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_map_overflows_existing_heap_into_new_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -87,7 +89,7 @@ fn test_map_overflows_existing_heap_into_new_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 2);
}
-#[test]
+#[uzu_test]
fn test_two_buffers_share_one_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -102,7 +104,7 @@ fn test_two_buffers_share_one_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_full_unmap_removes_heaps() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -116,7 +118,7 @@ fn test_full_unmap_removes_heaps() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 0);
}
-#[test]
+#[uzu_test]
fn test_partial_unmap_keeps_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -129,7 +131,7 @@ fn test_partial_unmap_keeps_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_unmap_with_other_buffer_keeps_mappings() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -144,7 +146,7 @@ fn test_unmap_with_other_buffer_keeps_mappings() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_unmap_one_buffer_keeps_heap_for_other() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -160,7 +162,7 @@ fn test_unmap_one_buffer_keeps_heap_for_other() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_remap_after_unmap_does_not_grow_pool() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -176,7 +178,7 @@ fn test_remap_after_unmap_does_not_grow_pool() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), initial);
}
-#[test]
+#[uzu_test]
fn test_remap_into_freed_gap_reuses_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -191,7 +193,7 @@ fn test_remap_into_freed_gap_reuses_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_sequential_partial_unmaps_release_full_heap() {
// Regression: after a partial unmap splits a rangemap entry, the surviving
// entry must still report the correct buffer↔heap correspondence so that a
@@ -209,7 +211,7 @@ fn test_sequential_partial_unmaps_release_full_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 0);
}
-#[test]
+#[uzu_test]
fn test_three_buffers_share_one_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -226,7 +228,7 @@ fn test_three_buffers_share_one_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_two_buffers_exactly_fill_one_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -241,7 +243,7 @@ fn test_two_buffers_exactly_fill_one_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_third_buffer_spills_when_heap_filled_by_others() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -258,7 +260,7 @@ fn test_third_buffer_spills_when_heap_filled_by_others() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 2);
}
-#[test]
+#[uzu_test]
fn test_unmap_buffer_frees_pages_for_another_in_same_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -276,7 +278,7 @@ fn test_unmap_buffer_frees_pages_for_another_in_same_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_unmap_middle_of_three_buffers_keeps_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -294,7 +296,7 @@ fn test_unmap_middle_of_three_buffers_keeps_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_unmap_all_shared_buffers_releases_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -314,7 +316,7 @@ fn test_unmap_all_shared_buffers_releases_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 0);
}
-#[test]
+#[uzu_test]
fn test_second_buffer_spans_existing_and_new_heap() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -329,7 +331,7 @@ fn test_second_buffer_spans_existing_and_new_heap() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 2);
}
-#[test]
+#[uzu_test]
fn test_unmap_spanning_buffer_keeps_heap_with_other_buffer() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -345,7 +347,7 @@ fn test_unmap_spanning_buffer_keeps_heap_with_other_buffer() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), 1);
}
-#[test]
+#[uzu_test]
fn test_remap_two_buffers_after_unmap_does_not_grow_pool() {
let ctx = create_context::<Metal>();
let pages_per_heap = ctx.sparse_heap_pool().heap_capacity_pages();
@@ -365,7 +367,7 @@ fn test_remap_two_buffers_after_unmap_does_not_grow_pool() {
assert_eq!(ctx.sparse_heap_pool().heaps_count(), initial);
}
-#[test]
+#[uzu_test]
fn test_heap_capacity_pages_matches_byte_capacity() {
let ctx = create_context::<Metal>();
let pool = ctx.sparse_heap_pool();
diff --git a/crates/backend-uzu/unit/encodable_block/attention_gemm_test.rs b/crates/backend-uzu/unit/encodable_block/attention_gemm_test.rs
index dd97c47ee..b9575780d 100644
--- a/crates/backend-uzu/unit/encodable_block/attention_gemm_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/attention_gemm_test.rs
@@ -2,6 +2,7 @@ use std::fmt::{Debug, Display};
use half::{bf16, f16};
use num_traits::Float;
+use proc_macros::uzu_test;
use super::{AttentionGemmArguments, AttentionGemmBlock};
use crate::{
@@ -194,81 +195,81 @@ fn test_unaligned<T: ArrayElement + Float + Debug + Display>() {
}
// Basic tests
-#[test]
+#[uzu_test]
fn test_basic_f32() {
test_basic::<f32>();
}
-#[test]
+#[uzu_test]
fn test_basic_f16() {
test_basic::<f16>();
}
-#[test]
+#[uzu_test]
fn test_basic_bf16() {
test_basic::<bf16>();
}
// Causal tests
-#[test]
+#[uzu_test]
fn test_causal_f32() {
test_causal::<f32>();
}
-#[test]
+#[uzu_test]
fn test_causal_f16() {
test_causal::<f16>();
}
-#[test]
+#[uzu_test]
fn test_causal_bf16() {
test_causal::<bf16>();
}
// GQA tests
-#[test]
+#[uzu_test]
fn test_gqa_f32() {
test_gqa::<f32>();
}
-#[test]
+#[uzu_test]
fn test_gqa_f16() {
test_gqa::<f16>();
}
-#[test]
+#[uzu_test]
fn test_gqa_bf16() {
test_gqa::<bf16>();
}
// Head dim 128
-#[test]
+#[uzu_test]
fn test_head_dim_128_f32() {
test_head_dim::<f32>(128);
}
-#[test]
+#[uzu_test]
fn test_head_dim_128_f16() {
test_head_dim::<f16>(128);
}
-#[test]
+#[uzu_test]
fn test_head_dim_128_bf16() {
test_head_dim::<bf16>(128);
}
// Unaligned tests
-#[test]
+#[uzu_test]
fn test_unaligned_f32() {
test_unaligned::<f32>();
}
-#[test]
+#[uzu_test]
fn test_unaligned_f16() {
test_unaligned::<f16>();
}
-#[test]
+#[uzu_test]
fn test_unaligned_bf16() {
test_unaligned::<bf16>();
}
diff --git a/crates/backend-uzu/unit/encodable_block/attention_test.rs b/crates/backend-uzu/unit/encodable_block/attention_test.rs
index fcc747468..e652055cf 100644
--- a/crates/backend-uzu/unit/encodable_block/attention_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/attention_test.rs
@@ -1,6 +1,7 @@
#![cfg(metal_backend)]
use ndarray::{Array4, s};
+use proc_macros::uzu_test;
use test_tag::tag;
use super::{AttentionGemmArguments, AttentionGemmBlock};
@@ -374,7 +375,7 @@ fn compare_results(
Ok(())
}
-#[test]
+#[uzu_test]
fn test_single_pass_attention_basic() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -394,7 +395,7 @@ fn test_single_pass_attention_basic() {
compare_results(&kernel_output, &reference_output, 1e-2, "Single-pass attention").unwrap();
}
-#[test]
+#[uzu_test]
fn test_gemm_attention_basic() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -415,7 +416,7 @@ fn test_gemm_attention_basic() {
compare_results(&kernel_output, &reference_output, 1e-2, "Gemm attention").unwrap();
}
-#[test]
+#[uzu_test]
fn test_gemm_attention_f32_head_dim_128() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -436,7 +437,7 @@ fn test_gemm_attention_f32_head_dim_128() {
compare_results(&kernel_output, &reference_output, 1e-2, "Gemm attention f32 head_dim=128").unwrap();
}
-#[test]
+#[uzu_test]
fn test_matrix_attention_matches_vector_and_cpu_seq256() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -472,7 +473,7 @@ fn test_matrix_attention_matches_vector_and_cpu_seq256() {
);
}
-#[test]
+#[uzu_test]
fn test_single_pass_attention_with_sinks() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -495,7 +496,7 @@ fn test_single_pass_attention_with_sinks() {
compare_results(&kernel_output, &reference_output, 1e-2, "Single-pass attention with sinks").unwrap();
}
-#[test]
+#[uzu_test]
fn test_single_pass_attention_with_sinks_long_sequence() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -519,7 +520,7 @@ fn test_single_pass_attention_with_sinks_long_sequence() {
.unwrap();
}
-#[test]
+#[uzu_test]
fn test_single_pass_attention_gqa() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -608,7 +609,7 @@ fn run_two_pass_attention(
Ok(kernel_output)
}
-#[test]
+#[uzu_test]
fn test_two_pass_attention() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -632,7 +633,7 @@ fn test_two_pass_attention() {
compare_results(&kernel_output, &reference_output, 1e-2, "Two-pass attention").unwrap();
}
-#[test]
+#[uzu_test]
fn test_two_pass_attention_gqa() {
let context = <Metal as Backend>::Context::new().expect("Failed to create <Metal as Backend>::Context");
@@ -657,7 +658,7 @@ fn test_two_pass_attention_gqa() {
}
#[tag(heavy)]
-#[test]
+#[uzu_test]
fn perf_two_pass_attention() {
use std::time::Instant;
diff --git a/crates/backend-uzu/unit/encodable_block/kv_cache_update_test.rs b/crates/backend-uzu/unit/encodable_block/kv_cache_update_test.rs
index 948b816fd..90568f4cd 100644
--- a/crates/backend-uzu/unit/encodable_block/kv_cache_update_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/kv_cache_update_test.rs
@@ -1,6 +1,7 @@
#![cfg(metal_backend)]
use ndarray::{Array, Array3, s};
+use proc_macros::uzu_test;
use super::*;
use crate::{
@@ -107,7 +108,7 @@ fn test_random_pattern<B: Backend>(context: &B::Context) {
assert_eq!(value_result, expected_values);
}
-#[test]
+#[uzu_test]
fn test_kv_cache_update_kernel() {
let metal_context = match <Metal as Backend>::Context::new() {
Ok(ctx) => ctx,
@@ -120,7 +121,7 @@ fn test_kv_cache_update_kernel() {
test_random_pattern::<Metal>(&metal_context);
}
-#[test]
+#[uzu_test]
fn test_direct_swaps() {
let sources = [0, 2, 4];
let destinations = [1, 3, 5];
diff --git a/crates/backend-uzu/unit/encodable_block/mamba_mixer/ssd_prefill_test.rs b/crates/backend-uzu/unit/encodable_block/mamba_mixer/ssd_prefill_test.rs
index c4a301759..91ff51944 100644
--- a/crates/backend-uzu/unit/encodable_block/mamba_mixer/ssd_prefill_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/mamba_mixer/ssd_prefill_test.rs
@@ -1,3 +1,5 @@
+use proc_macros::uzu_test;
+
use super::{SSDPrefillArguments, SSDPrefillBlock, SSDPrefillMode};
use crate::{
backends::common::{Backend, Context, Encoder, Kernels, gpu_types::ActivationType, kernel::Conv1dScanKernel},
@@ -387,35 +389,35 @@ fn conv1d_scan_deterministic_internal<B: Backend>() {
}
}
-#[test]
+#[uzu_test]
fn ssd_prefill_sequential_is_deterministic() {
for_each_non_cpu_backend!(|B| {
assert_deterministic_for_mode::<B>(SSDPrefillMode::Sequential);
});
}
-#[test]
+#[uzu_test]
fn ssd_prefill_single_pass_is_deterministic() {
for_each_non_cpu_backend!(|B| {
assert_deterministic_for_mode::<B>(SSDPrefillMode::SinglePass);
});
}
-#[test]
+#[uzu_test]
fn ssd_prefill_sequential_matches_cpu_reference() {
for_each_non_cpu_backend!(|B| {
assert_matches_cpu_reference::<B>(SSDPrefillMode::Sequential);
});
}
-#[test]
+#[uzu_test]
fn ssd_prefill_single_pass_matches_cpu_reference() {
for_each_non_cpu_backend!(|B| {
assert_matches_cpu_reference::<B>(SSDPrefillMode::SinglePass);
});
}
-#[test]
+#[uzu_test]
fn conv1d_scan_is_deterministic() {
for_each_non_cpu_backend!(|B| {
conv1d_scan_deterministic_internal::<B>();
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_block_e2e_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_block_e2e_test.rs
index 4fc9c9797..507f83bf7 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_block_e2e_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_block_e2e_test.rs
@@ -1,4 +1,5 @@
use half::bf16;
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use super::{MoeExpertsTwoPassArguments, MoeExpertsTwoPassPrefillBlock, MoeGather};
@@ -626,7 +627,7 @@ fn run_moe_parity_test_internal<B: Backend>(
}
// Test 1: Minimal (K=1, small dims, no clipping, alpha=1.0)
-#[test]
+#[uzu_test]
fn test_moe_minimal() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -648,7 +649,7 @@ fn test_moe_minimal() {
}
// Test 2: Multi-expert (K=2, tests finalize weighted sum)
-#[test]
+#[uzu_test]
fn test_moe_multi_expert() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -670,7 +671,7 @@ fn test_moe_multi_expert() {
}
// Test 3: Large d_model only (K=1 to isolate from finalize complexity)
-#[test]
+#[uzu_test]
fn test_moe_large_d_model_only() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -692,7 +693,7 @@ fn test_moe_large_d_model_only() {
}
// Test 4a: d_ff=32 (exactly 1 full chunk - baseline for multi-chunk)
-#[test]
+#[uzu_test]
fn test_1_full_chunk() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -714,7 +715,7 @@ fn test_1_full_chunk() {
}
// Test 4b: d_ff=48 (1.5 chunks - tests partial second chunk)
-#[test]
+#[uzu_test]
fn test_1_5_chunks() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -736,7 +737,7 @@ fn test_1_5_chunks() {
}
// Test 4c: d_ff=64 (exactly 2 full chunks)
-#[test]
+#[uzu_test]
fn test_2_full_chunks() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -758,7 +759,7 @@ fn test_2_full_chunks() {
}
// Test 5: GELU activation (gating_code=0)
-#[test]
+#[uzu_test]
fn test_gelu_activation() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -780,7 +781,7 @@ fn test_gelu_activation() {
}
// Test 6: SiLU activation (gating_code=1)
-#[test]
+#[uzu_test]
fn test_silu_activation() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -802,7 +803,7 @@ fn test_silu_activation() {
}
// Test 7: GEGLU activation (gating_code=3)
-#[test]
+#[uzu_test]
fn test_geglu_activation() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -824,7 +825,7 @@ fn test_geglu_activation() {
}
// Test 8: Bucket stress - multiple tokens, larger E, stress scatter/gather
-#[test]
+#[uzu_test]
fn test_bucket_stress() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -846,7 +847,7 @@ fn test_bucket_stress() {
}
// Test 9a: K=4 small-scale to verify multi-expert accumulation
-#[test]
+#[uzu_test]
fn test_small_scale() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -868,7 +869,7 @@ fn test_small_scale() {
}
// Test 9b: d_model=128 (4 k0 chunks, 2 n-tiles)
-#[test]
+#[uzu_test]
fn test_model_128() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -890,7 +891,7 @@ fn test_model_128() {
}
// Test 9c: d_model=192 (6 k0 chunks, 3 n-tiles - test threshold)
-#[test]
+#[uzu_test]
fn test_model_192() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -912,7 +913,7 @@ fn test_model_192() {
}
// Test 9d: d_model=256 (8 k0 chunks, 4 n-tiles)
-#[test]
+#[uzu_test]
fn test_model_256() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -934,7 +935,7 @@ fn test_model_256() {
}
// Test 10a: Isolate large d_ff with small d_model (FF accumulation test)
-#[test]
+#[uzu_test]
fn test_ff_accumulation() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -956,7 +957,7 @@ fn test_ff_accumulation() {
}
// Test 10b: d_model=1024 (16 n-tiles)
-#[test]
+#[uzu_test]
fn test_model_1024() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -979,7 +980,7 @@ fn test_model_1024() {
// Test 10c: Verify layout with d_model=96, d_ff=96 (3 chunks each)
// This tests non-power-of-2 to catch stride bugs
-#[test]
+#[uzu_test]
fn test_stride() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -1001,7 +1002,7 @@ fn test_stride() {
}
// Test 10d: d_model=1536
-#[test]
+#[uzu_test]
fn test_model_1536() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -1023,7 +1024,7 @@ fn test_model_1536() {
}
// Test 10e: 2 n-tiles
-#[test]
+#[uzu_test]
fn test_2n_tiles() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -1045,7 +1046,7 @@ fn test_2n_tiles() {
}
// Test 10g: 8 n-tiles - find threshold
-#[test]
+#[uzu_test]
fn test_8n_tiles() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -1067,7 +1068,7 @@ fn test_8n_tiles() {
}
// Test 11b: Production scale with clipping but K=1
-#[test]
+#[uzu_test]
#[ignore]
fn test_prod_scale_clipping() {
for_each_non_cpu_backend!(|B| {
@@ -1090,7 +1091,7 @@ fn test_prod_scale_clipping() {
}
// Test 11c: PRODUCTION SCALE T=1 decode (triggers GEMV v2)
-#[test]
+#[uzu_test]
#[ignore]
fn test_prod_scale_decode() {
for_each_non_cpu_backend!(|B| {
@@ -1113,7 +1114,7 @@ fn test_prod_scale_decode() {
}
// Test 11d: FULL PRODUCTION SCALE with K=2, T=4 (uses tiled MMA)
-#[test]
+#[uzu_test]
#[ignore]
fn test_prod_scale_tile_mma() {
for_each_non_cpu_backend!(|B| {
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_experts_perf_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_experts_perf_test.rs
index 4700a271f..a9d14fc2c 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_experts_perf_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_experts_perf_test.rs
@@ -1,6 +1,7 @@
use std::time::Instant;
use half::bf16;
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use super::{MoeExpertsTwoPassArguments, MoeExpertsTwoPassDecodeBlock, MoeExpertsTwoPassPrefillBlock};
@@ -263,7 +264,7 @@ fn run_two_pass_prefill_case<B: Backend>(
eprintln!(" → Throughput: {:.1} µs/token (mean / sum_k)", (mean / sum_k as f64) * 1000.0);
}
-#[test]
+#[uzu_test]
#[ignore]
fn test_two_pass_decode_speed() {
for_each_non_cpu_backend!(|B| {
@@ -280,7 +281,7 @@ fn test_two_pass_decode_speed() {
});
}
-#[test]
+#[uzu_test]
#[ignore]
fn test_two_pass_prefill_speed() {
for_each_non_cpu_backend!(|B| {
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_experts_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_experts_test.rs
index ceb05586b..ddb947d15 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_experts_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_experts_test.rs
@@ -7,6 +7,7 @@
//! - Numerical correctness against CPU reference
use half::bf16;
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use super::{
@@ -279,7 +280,7 @@ fn cpu_moe_reference(
y
}
-#[test]
+#[uzu_test]
fn test_two_pass_decode_correctness() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -443,7 +444,7 @@ fn test_two_pass_decode_correctness() {
});
}
-#[test]
+#[uzu_test]
fn test_two_pass_decode_multi_token() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -527,7 +528,7 @@ fn test_two_pass_decode_multi_token() {
});
}
-#[test]
+#[uzu_test]
fn test_two_pass_prefill_correctness() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -608,7 +609,7 @@ fn test_two_pass_prefill_correctness() {
});
}
-#[test]
+#[uzu_test]
fn test_tile_infrastructure() {
// Test that tile counts/scan/map are computed correctly
// This is already tested in moe_tiles_test.rs, but we verify here
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_gather_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_gather_test.rs
index c3cff890e..d415dc7d0 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_gather_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_gather_test.rs
@@ -1,5 +1,6 @@
use half::bf16;
use num_traits::Float;
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use super::MoeGather;
@@ -75,22 +76,22 @@ fn test_gather_internal(
})
}
-#[test]
+#[uzu_test]
fn test_gather_single_token() {
test_gather_internal(1, 2, 64);
}
-#[test]
+#[uzu_test]
fn test_gather_small_batch() {
test_gather_internal(4, 8, 128);
}
-#[test]
+#[uzu_test]
fn test_gather_medium_batch() {
test_gather_internal(16, 32, 256)
}
-#[test]
+#[uzu_test]
fn test_gather_large_batch() {
test_gather_internal(64, 128, 512)
}
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_perf_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_perf_test.rs
index 2e15ba0c9..ef3ddd272 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_perf_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_perf_test.rs
@@ -1,4 +1,5 @@
use half::bf16;
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use super::{MoeExpertsTwoPassArguments, MoeExpertsTwoPassDecodeBlock, MoeGather};
@@ -18,7 +19,7 @@ use crate::{
};
// Test E2E MoE performance with timing breakdown (decode mode, T=1)
-#[test]
+#[uzu_test]
#[ignore]
fn test_moe_e2e_decode_perf() {
eprintln!("\n=== End-to-End MoE Performance (DECODE, T=1) ===");
@@ -73,7 +74,7 @@ fn test_moe_e2e_decode_perf() {
}
// Test E2E MoE performance with timing breakdown (prefill mode, T>1)
-#[test]
+#[uzu_test]
#[ignore]
fn test_moe_e2e_prefill_perf() {
eprintln!("\n=== End-to-End MoE Performance (PREFILL, T>1) ===");
@@ -134,7 +135,7 @@ fn test_moe_e2e_prefill_perf() {
}
// Test complete MoE pipeline timing breakdown (decode mode, T=1)
-#[test]
+#[uzu_test]
#[ignore]
fn test_moe_pipeline_breakdown_decode() {
for_each_non_cpu_backend!(|B| {
diff --git a/crates/backend-uzu/unit/encodable_block/moe/moe_tiles_test.rs b/crates/backend-uzu/unit/encodable_block/moe/moe_tiles_test.rs
index 9ec58b682..dd21c370f 100644
--- a/crates/backend-uzu/unit/encodable_block/moe/moe_tiles_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/moe/moe_tiles_test.rs
@@ -1,3 +1,4 @@
+use proc_macros::uzu_test;
use rand::{RngExt, SeedableRng, rngs::StdRng};
use crate::{
@@ -11,7 +12,7 @@ use crate::{
encodable_block::mlp::moe::tests::{cpu_tile_counts, cpu_tile_scan},
};
-#[test]
+#[uzu_test]
fn test_tile_counts_correctness() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -54,7 +55,7 @@ fn test_tile_counts_correctness() {
});
}
-#[test]
+#[uzu_test]
fn test_tile_scan_correctness() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
@@ -97,7 +98,7 @@ fn test_tile_scan_correctness() {
});
}
-#[test]
+#[uzu_test]
fn test_tile_edge_cases() {
for_each_non_cpu_backend!(|B| {
let ctx = create_context::<B>();
diff --git a/crates/backend-uzu/unit/encodable_block/sampling_test.rs b/crates/backend-uzu/unit/encodable_block/sampling_test.rs
index 4bc4e0af8..06c5abba8 100644
--- a/crates/backend-uzu/unit/encodable_block/sampling_test.rs
+++ b/crates/backend-uzu/unit/encodable_block/sampling_test.rs
@@ -6,6 +6,7 @@ use backend_uzu::{
dispatch_dtype,
};
use num_traits::{Float, NumCast};
+use proc_macros::uzu_test;
use proptest::prelude::*;
use rand::{RngExt, SeedableRng, rngs::SmallRng};
use rand_distr::Normal;
@@ -167,7 +168,7 @@ fn sampling_case() -> impl Strategy<Value = SamplingCase> {
)
}
-#[test]
+#[uzu_test]
fn test_sampling_prop() {
let contextes = TestContextes::new();
diff --git a/crates/backend-uzu/unit/forward_pass/cache_layers_test.rs b/crates/backend-uzu/unit/forward_pass/cache_layers_test.rs
index 8b11e717e..26fa2607e 100644
--- a/crates/backend-uzu/unit/forward_pass/cache_layers_test.rs
+++ b/crates/backend-uzu/unit/forward_pass/cache_layers_test.rs
@@ -2,6 +2,8 @@
use std::{mem::size_of, ops::Range};
+use proc_macros::uzu_test;
+
use super::{CacheEntryIndex, CacheLayer, CacheLayers, LayerCacheBinding};
use crate::{
array::ArrayContextExt,
@@ -102,7 +104,7 @@ fn make_cache_layers(
}
}
-#[test]
+#[uzu_test]
fn test_cache_layers_copy_from_clones_only_live_window_rows() {
let Some(context) = <Metal as Backend>::Context::new().ok() else {
return;
diff --git a/crates/backend-uzu/unit/forward_pass/kv_cache_state_test.rs b/crates/backend-uzu/unit/forward_pass/kv_cache_state_test.rs
index 5f6da027c..ca0298488 100644
--- a/crates/backend-uzu/unit/forward_pass/kv_cache_state_test.rs
+++ b/crates/backend-uzu/unit/forward_pass/kv_cache_state_test.rs
@@ -1,5 +1,7 @@
#![cfg(metal_backend)]
+use proc_macros::uzu_test;
+
use crate::{
backends::{
common::{Backend, Context, Encoder},
@@ -211,7 +213,7 @@ fn run_scenario<B: Backend>(
);
}
-#[test]
+#[uzu_test]
fn kv_cache_state_scenarios() {
let Some(context) = <Metal as Backend>::Context::new().ok() else {
return;
@@ -291,7 +293,7 @@ fn kv_cache_state_scenarios() {
}
}
-#[test]
+#[uzu_test]
fn kv_cache_slice_apply_contiguous_window() {
let Some(context) = <Metal as Backend>::Context::new().ok() else {
return;
@@ -328,7 +330,7 @@ fn kv_cache_slice_apply_contiguous_window() {
assert_eq!(values_after[0..4], initial_values[0..4], "values restored for contiguous slice");
}
-#[test]
+#[uzu_test]
fn kv_cache_slice_apply_wrap_window() {
let Some(context) = <Metal as Backend>::Context::new().ok() else {
return;
@@ -365,7 +367,7 @@ fn kv_cache_slice_apply_wrap_window() {
assert_eq!(values_after[0..4], initial_values[0..4], "values restored for wrapped slice");
}
-#[test]
+#[uzu_test]
fn kv_cache_slice_apply_full_restores_metadata() {
let Some(context) = <Metal as Backend>::Context::new().ok() else {
return;
diff --git a/crates/backend-uzu/unit/trie_test.rs b/crates/backend-uzu/unit/trie_test.rs
index 0c61fccb5..0da4f21b0 100644
--- a/crates/backend-uzu/unit/trie_test.rs
+++ b/crates/backend-uzu/unit/trie_test.rs
@@ -1,5 +1,7 @@
use std::collections::HashMap;
+use proc_macros::uzu_test;
+
use crate::{
language_model::rng::PRng,
speculators::{empty_speculator::EmptySpeculator, speculator::Speculator},
@@ -45,14 +47,14 @@ fn verify_sprout(
assert_eq!(flat_trie.token_seeds().collect::<Vec<u64>>(), vec![expected_seed]);
}
-#[test]
+#[uzu_test]
fn test_trie_manual_sprout() {
let trie_root = TrieNode::new(0, None, 0);
verify_sprout(&trie_root, 0);
}
-#[test]
+#[uzu_test]
fn test_trie_from_speculator_sprout() {
let rng = PRng::new(42);
@@ -106,7 +108,7 @@ fn verify_stick(
}
}
-#[test]
+#[uzu_test]
fn test_trie_manual_stick() {
let rng = PRng::new(0);
let mut trie_root = TrieNode::new(0, None, rng.derive(0));
@@ -120,7 +122,7 @@ fn test_trie_manual_stick() {
verify_stick(&trie_root, &rng);
}
-#[test]
+#[uzu_test]
fn test_trie_from_speculator_stick() {
let rng = PRng::new(42);
@@ -176,7 +178,7 @@ fn verify_bush(
}
}
-#[test]
+#[uzu_test]
fn test_trie_manual_bush() {
let rng = PRng::new(0);
let mut trie_root = TrieNode::new(0, None, rng.derive(0));
@@ -191,7 +193,7 @@ fn test_trie_manual_bush() {
verify_bush(&trie_root, &rng);
}
-#[test]
+#[uzu_test]
fn test_trie_from_speculator_bush() {
let rng = PRng::new(0);
@@ -250,7 +252,7 @@ fn verify_tree(
}
}
-#[test]
+#[uzu_test]
fn test_trie_manual_tree() {
let rng = PRng::new(0);
let mut trie_root = TrieNode::new(0, None, rng.derive(0));
diff --git a/crates/backend-uzu/unit/utils/model_size_test.rs b/crates/backend-uzu/unit/utils/model_size_test.rs
index dd730e752..32b3b64a8 100644
--- a/crates/backend-uzu/unit/utils/model_size_test.rs
+++ b/crates/backend-uzu/unit/utils/model_size_test.rs
@@ -1,5 +1,7 @@
use std::path::PathBuf;
+use proc_macros::uzu_test;
+
use super::*;
use crate::utils::VERSION;
pub const MODEL_DIR_NAME: &str = "Llama-3.2-1B-Instruct";
@@ -18,7 +20,7 @@ pub fn get_test_model_path() -> PathBuf {
path
}
-#[test]
+#[uzu_test]
fn test_model_size_from_path() {
let model_path = get_test_model_path();
diff --git a/crates/proc-macros/src/uzu_test.rs b/crates/proc-macros/src/uzu_test.rs
index 7267a5918..65a8342e8 100644
--- a/crates/proc-macros/src/uzu_test.rs
+++ b/crates/proc-macros/src/uzu_test.rs
@@ -1,16 +1,59 @@
use proc_macro::TokenStream;
-use quote::quote;
-use syn::{ItemFn, parse_macro_input};
+use quote::{format_ident, quote};
+use syn::{ItemFn, Meta, parse_macro_input};
pub fn uzu_test(
_args: TokenStream,
input: TokenStream,
) -> TokenStream {
- let func = parse_macro_input!(input as ItemFn);
+ let mut func = parse_macro_input!(input as ItemFn);
+
+ let ignore_attr =
+ func.attrs.iter().position(|attr| attr.path().is_ident("ignore")).map(|index| func.attrs.remove(index));
+ let (ignore, ignore_message) = match ignore_attr.as_ref().map(|attr| &attr.meta) {
+ None => (quote! { false }, quote! { ::core::option::Option::None }),
+ Some(Meta::Path(_)) => (quote! { true }, quote! { ::core::option::Option::None }),
+ Some(Meta::NameValue(name_value)) => {
+ let message = &name_value.value;
+ (quote! { true }, quote! { ::core::option::Option::Some(#message) })
+ },
+ Some(meta) => {
+ return syn::Error::new_spanned(meta, "unsupported #[ignore] form").to_compile_error().into();
+ },
+ };
+
+ let name = &func.sig.ident;
+ let const_name = format_ident!("__UZU_TEST_CASE_{}", name);
quote! {
- #[test]
#func
+
+ #[test_case]
+ #[allow(non_upper_case_globals)]
+ const #const_name: crate::harness::UzuTest =
+ crate::harness::UzuTest::Test(&crate::harness::test::TestDescAndFn {
+ desc: crate::harness::test::TestDesc {
+ name: crate::harness::test::StaticTestName(concat!(
+ module_path!(),
+ "::",
+ stringify!(#name),
+ )),
+ ignore: #ignore,
+ ignore_message: #ignore_message,
+ source_file: file!(),
+ start_line: line!() as usize,
+ start_col: column!() as usize,
+ end_line: line!() as usize,
+ end_col: column!() as usize,
+ should_panic: crate::harness::test::ShouldPanic::No,
+ compile_fail: false,
+ no_run: false,
+ test_type: crate::harness::test::TestType::Unknown,
+ },
+ testfn: crate::harness::test::StaticTestFn(|| {
+ crate::harness::test::assert_test_result(#name())
+ }),
+ });
}
.into()
}