A high-performance inference engine for AI models
naming
Eugene Bokhan17414356dab9
diff --git a/crates/uzu/tests/matmul_correctness_test.rs b/crates/uzu/tests/matmul_correctness_test.rsindex ed9a04c94..07d31c2b9 100644--- a/crates/uzu/tests/matmul_correctness_test.rs+++ b/crates/uzu/tests/matmul_correctness_test.rs@@ -129,7 +129,6 @@ fn generate_test_data(}struct TestCase {- name: &'static str,m: usize,k: usize,n: usize,@@ -139,9 +138,8 @@ struct TestCase {fn test_cases() -> Vec<TestCase> {vec![- // GEMV shapes (m=1)+ // transpose_b = true casesTestCase {- name: "gemv_small",m: 1,k: 2048,n: 2048,@@ -149,7 +147,6 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.01,},TestCase {- name: "gemv_medium",m: 1,k: 4096,n: 4096,@@ -157,16 +154,13 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.01,},TestCase {- name: "gemv_large",m: 1,k: 8192,n: 8192,transpose_b: true,tolerance: 0.02,},- // MLP shapesTestCase {- name: "mlp_up",m: 1,k: 2048,n: 8192,@@ -174,16 +168,27 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.01,},TestCase {- name: "mlp_down",m: 1,k: 8192,n: 2048,transpose_b: true,tolerance: 0.02,},- // Small batch sizesTestCase {- name: "batch_8",+ m: 2,+ k: 4096,+ n: 4096,+ transpose_b: true,+ tolerance: 0.02,+ },+ TestCase {+ m: 4,+ k: 16384,+ n: 4,+ transpose_b: true,+ tolerance: 0.1,+ },+ TestCase {m: 8,k: 2048,n: 2048,@@ -191,7 +196,6 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.02,},TestCase {- name: "batch_16",m: 16,k: 2048,n: 2048,@@ -199,57 +203,55 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.03,},TestCase {- name: "batch_32",- m: 32,+ m: 17,k: 2048,n: 2048,transpose_b: true,tolerance: 0.05,},TestCase {- name: "batch_64",- m: 64,+ m: 32,k: 2048,n: 2048,transpose_b: true,- tolerance: 0.08,+ tolerance: 0.05,},TestCase {- name: "batch_128",- m: 128,+ m: 32,k: 2048,- n: 2048,+ n: 1537,transpose_b: true,tolerance: 0.1,},- // Prefill shapesTestCase {- name: "prefill_256",- m: 256,- k: 2048,+ m: 32,+ k: 1999,n: 2048,transpose_b: true,- tolerance: 0.15,+ tolerance: 0.1,},TestCase {- name: "prefill_512",- m: 512,- k: 2048,- n: 2048,+ m: 33,+ k: 127,+ n: 65,transpose_b: true,- tolerance: 0.2,+ tolerance: 0.1,},TestCase {- name: "prefill_1024",- m: 1024,+ m: 64,+ k: 32,+ n: 64,+ transpose_b: true,+ tolerance: 0.05,+ },+ TestCase {+ m: 64,k: 2048,n: 2048,transpose_b: true,- tolerance: 0.25,+ tolerance: 0.08,},- // Square matricesTestCase {- name: "square_128",m: 128,k: 128,n: 128,@@ -257,32 +259,34 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.05,},TestCase {- name: "square_256",- m: 256,- k: 256,- n: 256,+ m: 128,+ k: 2048,+ n: 1,transpose_b: true,- tolerance: 0.1,+ tolerance: 0.02,},TestCase {- name: "square_512",- m: 512,- k: 512,- n: 512,+ m: 128,+ k: 2048,+ n: 2048,transpose_b: true,tolerance: 0.1,},TestCase {- name: "square_1024",- m: 1024,- k: 1024,- n: 1024,+ m: 256,+ k: 256,+ n: 256,transpose_b: true,tolerance: 0.1,},- // Attention shapesTestCase {- name: "attention_qk",+ m: 256,+ k: 2048,+ n: 2048,+ transpose_b: true,+ tolerance: 0.15,+ },+ TestCase {m: 512,k: 64,n: 512,@@ -290,95 +294,53 @@ fn test_cases() -> Vec<TestCase> {tolerance: 0.1,},TestCase {- name: "attention_av",m: 512,k: 512,- n: 64,- transpose_b: false,+ n: 512,+ transpose_b: true,tolerance: 0.1,},- // Non-transposed B- TestCase {- name: "non_transposed_small",- m: 64,- k: 128,- n: 64,- transpose_b: false,- tolerance: 0.15,- },- TestCase {- name: "non_transposed_medium",- m: 128,- k: 256,- n: 128,- transpose_b: false,- tolerance: 0.15,- },- // Non-aligned sizesTestCase {- name: "unaligned_m",- m: 17,+ m: 512,k: 2048,n: 2048,transpose_b: true,- tolerance: 0.05,+ tolerance: 0.2,},TestCase {- name: "unaligned_n",- m: 32,- k: 2048,- n: 1537,+ m: 1024,+ k: 1024,+ n: 1024,transpose_b: true,tolerance: 0.1,},TestCase {- name: "unaligned_k",- m: 32,- k: 1999,+ m: 1024,+ k: 2048,n: 2048,transpose_b: true,- tolerance: 0.1,- },- TestCase {- name: "unaligned_all",- m: 33,- k: 127,- n: 65,- transpose_b: true,- tolerance: 0.1,+ tolerance: 0.25,},- // Edge cases+ // transpose_b = false casesTestCase {- name: "thin_m",- m: 2,- k: 4096,- n: 4096,- transpose_b: true,- tolerance: 0.02,+ m: 64,+ k: 128,+ n: 64,+ transpose_b: false,+ tolerance: 0.15,},TestCase {- name: "thin_n",m: 128,- k: 2048,- n: 1,- transpose_b: true,- tolerance: 0.02,+ k: 256,+ n: 128,+ transpose_b: false,+ tolerance: 0.15,},TestCase {- name: "small_k",- m: 64,- k: 32,+ m: 512,+ k: 512,n: 64,- transpose_b: true,- tolerance: 0.05,- },- // Large K (tests SplitK path)- TestCase {- name: "large_k_small_mn",- m: 4,- k: 16384,- n: 4,- transpose_b: true,+ transpose_b: false,tolerance: 0.1,},]@@ -450,23 +412,28 @@ fn matmul_correctness_comprehensive() {case.transpose_b,);+ let trans_str = if case.transpose_b {+ "T"+ } else {+ "N"+ };+match compare_results(&metal_result, &reference, case.tolerance) {Ok(()) => {passed += 1;eprintln!(- "✓ {} (m={}, k={}, n={}, transpose_b={})",- case.name, case.m, case.k, case.n, case.transpose_b+ "✓ m={} k={} n={} B={}",+ case.m, case.k, case.n, trans_str);},Err((max_diff, idx, count)) => {failed.push((case, max_diff, idx, count));eprintln!(- "✗ {} (m={}, k={}, n={}, transpose_b={}) - max_diff={:.6} at idx {} ({} elements exceed tolerance {})",- case.name,+ "✗ m={} k={} n={} B={} max_diff={:.6} at idx {} ({} exceed tol {})",case.m,case.k,case.n,- case.transpose_b,+ trans_str,max_diff,idx,count,@@ -481,13 +448,17 @@ fn matmul_correctness_comprehensive() {if !failed.is_empty() {eprintln!("\nFailed tests:");for (case, max_diff, idx, count) in &failed {+ let trans_str = if case.transpose_b {+ "T"+ } else {+ "N"+ };eprintln!(- " {} ({}x{}x{}, transpose_b={}): max_diff={:.6} at idx {}, {} elements exceed tolerance {}",- case.name,+ " m={} k={} n={} B={}: max_diff={:.6} at idx {}, {} exceed tol {}",case.m,case.k,case.n,- case.transpose_b,+ trans_str,max_diff,idx,count,