diff --git a/book/src/puzzle_04/intro.mojo b/book/src/puzzle_04/intro.mojo
index 6c88703e..8b3aefc9 100644
--- a/book/src/puzzle_04/intro.mojo
+++ b/book/src/puzzle_04/intro.mojo
@@ -17,7 +17,7 @@ fn kernel[
     print(tensor)
 
 
-def main():
+def main() raises:
     ctx = DeviceContext()
 
     a = ctx.enqueue_create_buffer[dtype](HEIGHT * WIDTH)
diff --git a/problems/p01/p01.mojo b/problems/p01/p01.mojo
index 62e80613..21b6ddd2 100644
--- a/problems/p01/p01.mojo
+++ b/problems/p01/p01.mojo
@@ -21,7 +21,7 @@ fn add_10(
 # ANCHOR_END: add_10
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p02/p02.mojo b/problems/p02/p02.mojo
index de5692ae..4365dcb8 100644
--- a/problems/p02/p02.mojo
+++ b/problems/p02/p02.mojo
@@ -22,7 +22,7 @@ fn add(
 # ANCHOR_END: add
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p03/p03.mojo b/problems/p03/p03.mojo
index 12fa7560..c55731e9 100644
--- a/problems/p03/p03.mojo
+++ b/problems/p03/p03.mojo
@@ -22,7 +22,7 @@ fn add_10_guard(
 # ANCHOR_END: add_10_guard
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p04/p04.mojo b/problems/p04/p04.mojo
index 08045d4c..202d3768 100644
--- a/problems/p04/p04.mojo
+++ b/problems/p04/p04.mojo
@@ -23,7 +23,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p04/p04_layout_tensor.mojo b/problems/p04/p04_layout_tensor.mojo
index 1e4a6e7b..543f92a1 100644
--- a/problems/p04/p04_layout_tensor.mojo
+++ b/problems/p04/p04_layout_tensor.mojo
@@ -24,7 +24,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d_layout_tensor
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/problems/p05/p05.mojo b/problems/p05/p05.mojo
index b5d7b644..c6d9a139 100644
--- a/problems/p05/p05.mojo
+++ b/problems/p05/p05.mojo
@@ -22,7 +22,7 @@ fn broadcast_add(
 
 
 # ANCHOR_END: broadcast_add
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p05/p05_layout_tensor.mojo b/problems/p05/p05_layout_tensor.mojo
index d39b9e72..fbbb6f17 100644
--- a/problems/p05/p05_layout_tensor.mojo
+++ b/problems/p05/p05_layout_tensor.mojo
@@ -29,7 +29,7 @@ fn broadcast_add[
 
 
 # ANCHOR_END: broadcast_add_layout_tensor
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/problems/p06/p06.mojo b/problems/p06/p06.mojo
index 03760e63..12dd801b 100644
--- a/problems/p06/p06.mojo
+++ b/problems/p06/p06.mojo
@@ -22,7 +22,7 @@ fn add_10_blocks(
 # ANCHOR_END: add_10_blocks
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p07/p07.mojo b/problems/p07/p07.mojo
index c5071af5..b765756e 100644
--- a/problems/p07/p07.mojo
+++ b/problems/p07/p07.mojo
@@ -23,7 +23,7 @@ fn add_10_blocks_2d(
 # ANCHOR_END: add_10_blocks_2d
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p07/p07_layout_tensor.mojo b/problems/p07/p07_layout_tensor.mojo
index 59990dd3..0397b133 100644
--- a/problems/p07/p07_layout_tensor.mojo
+++ b/problems/p07/p07_layout_tensor.mojo
@@ -28,7 +28,7 @@ fn add_10_blocks_2d[
 # ANCHOR_END: add_10_blocks_2d_layout_tensor
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/problems/p08/p08.mojo b/problems/p08/p08.mojo
index 2e8657bd..33c750c6 100644
--- a/problems/p08/p08.mojo
+++ b/problems/p08/p08.mojo
@@ -38,7 +38,7 @@ fn add_10_shared(
 # ANCHOR_END: add_10_shared
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p08/p08_layout_tensor.mojo b/problems/p08/p08_layout_tensor.mojo
index 90532a42..ab675e2e 100644
--- a/problems/p08/p08_layout_tensor.mojo
+++ b/problems/p08/p08_layout_tensor.mojo
@@ -42,7 +42,7 @@ fn add_10_shared_layout_tensor[
 # ANCHOR_END: add_10_shared_layout_tensor
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p09/p09.mojo b/problems/p09/p09.mojo
index 3460e12b..bc42efc5 100644
--- a/problems/p09/p09.mojo
+++ b/problems/p09/p09.mojo
@@ -90,7 +90,7 @@ fn collaborative_filter(
 # ANCHOR_END: third_crash
 
 
-def main():
+def main() raises:
     if len(argv()) != 2:
         print(
             "Usage: pixi run mojo p09 [--first-case | --second-case |"
diff --git a/problems/p10/p10.mojo b/problems/p10/p10.mojo
index 9ac898ae..f119a298 100644
--- a/problems/p10/p10.mojo
+++ b/problems/p10/p10.mojo
@@ -55,7 +55,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d_no_guard
 
 
-def main():
+def main() raises:
     if len(argv()) != 2:
         print(
             "Expected one command-line argument: '--memory-bug' or"
diff --git a/problems/p11/p11.mojo b/problems/p11/p11.mojo
index 2743a0b8..7595ee88 100644
--- a/problems/p11/p11.mojo
+++ b/problems/p11/p11.mojo
@@ -30,7 +30,7 @@ fn pooling(
 # ANCHOR_END: pooling
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p11/p11_layout_tensor.mojo b/problems/p11/p11_layout_tensor.mojo
index 8f9e6f8e..e394fa9c 100644
--- a/problems/p11/p11_layout_tensor.mojo
+++ b/problems/p11/p11_layout_tensor.mojo
@@ -36,7 +36,7 @@ fn pooling[
 # ANCHOR_END: pooling_layout_tensor
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/problems/p12/p12.mojo b/problems/p12/p12.mojo
index 38a3932e..5ab01850 100644
--- a/problems/p12/p12.mojo
+++ b/problems/p12/p12.mojo
@@ -25,7 +25,7 @@ fn dot_product(
 # ANCHOR_END: dot_product
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](1)
         out.enqueue_fill(0)
diff --git a/problems/p12/p12_layout_tensor.mojo b/problems/p12/p12_layout_tensor.mojo
index d99e0427..d64df5f9 100644
--- a/problems/p12/p12_layout_tensor.mojo
+++ b/problems/p12/p12_layout_tensor.mojo
@@ -31,7 +31,7 @@ fn dot_product[
 # ANCHOR_END: dot_product_layout_tensor
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](1)
         out.enqueue_fill(0)
diff --git a/problems/p13/p13.mojo b/problems/p13/p13.mojo
index 929b109a..16cfced7 100644
--- a/problems/p13/p13.mojo
+++ b/problems/p13/p13.mojo
@@ -56,7 +56,7 @@ fn conv_1d_block_boundary[
 # ANCHOR_END: conv_1d_block_boundary
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         size = SIZE_2 if argv()[1] == "--block-boundary" else SIZE
         conv = CONV_2 if argv()[1] == "--block-boundary" else CONV
diff --git a/problems/p14/p14.mojo b/problems/p14/p14.mojo
index cce933b6..d3c73fab 100644
--- a/problems/p14/p14.mojo
+++ b/problems/p14/p14.mojo
@@ -62,7 +62,7 @@ fn prefix_sum_block_sum_phase[
 # ANCHOR_END: prefix_sum_complete
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         if len(argv()) != 2 or argv()[1] not in [
             "--simple",
diff --git a/problems/p15/p15.mojo b/problems/p15/p15.mojo
index 1f7e2c8b..2371ca15 100644
--- a/problems/p15/p15.mojo
+++ b/problems/p15/p15.mojo
@@ -33,7 +33,7 @@ fn axis_sum[
 # ANCHOR_END: axis_sum
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](BATCH)
         out.enqueue_fill(0)
diff --git a/problems/p16/p16.mojo b/problems/p16/p16.mojo
index 737fc03b..92abe7bc 100644
--- a/problems/p16/p16.mojo
+++ b/problems/p16/p16.mojo
@@ -72,7 +72,7 @@ fn matmul_tiled[
 # ANCHOR_END: matmul_tiled
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         if len(argv()) != 2 or argv()[1] not in [
             "--naive",
diff --git a/problems/p18/test/test_softmax.mojo b/problems/p18/test/test_softmax.mojo
index 5171d550..4eb16fe7 100644
--- a/problems/p18/test/test_softmax.mojo
+++ b/problems/p18/test/test_softmax.mojo
@@ -12,7 +12,7 @@ comptime BLOCK_DIM_X = 1 << log2_ceil(SIZE)
 comptime dtype = DType.float32
 
 
-def test_softmax():
+def test_softmax() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[DType.float32](SIZE)
         out.enqueue_fill(0)
@@ -79,5 +79,5 @@ def test_softmax():
             print("All tests passed 🎉")
 
 
-def main():
+def main() raises:
     test_softmax()
diff --git a/problems/p23/p23.mojo b/problems/p23/p23.mojo
index c188df32..fa4bb30c 100644
--- a/problems/p23/p23.mojo
+++ b/problems/p23/p23.mojo
@@ -311,7 +311,7 @@ fn benchmark_vectorized_parameterized[
     bench_ctx.synchronize()
 
 
-def main():
+def main() raises:
     ctx = DeviceContext()
     out = ctx.enqueue_create_buffer[dtype](SIZE)
     out.enqueue_fill(0)
diff --git a/problems/p24/p24.mojo b/problems/p24/p24.mojo
index acee7a07..8c44e2cb 100644
--- a/problems/p24/p24.mojo
+++ b/problems/p24/p24.mojo
@@ -300,7 +300,7 @@ fn benchmark_traditional_parameterized[
     bench_ctx.synchronize()
 
 
-def main():
+def main() raises:
     if argv()[1] != "--benchmark":
         print("SIZE:", SIZE)
         print("WARP_SIZE:", WARP_SIZE)
diff --git a/problems/p25/p25.mojo b/problems/p25/p25.mojo
index 253ffe25..4ba7162f 100644
--- a/problems/p25/p25.mojo
+++ b/problems/p25/p25.mojo
@@ -133,7 +133,7 @@ fn conditional_broadcast[
 # ANCHOR_END: conditional_broadcast
 
 
-def test_neighbor_difference():
+def test_neighbor_difference() raises:
     with DeviceContext() as ctx:
         # Create test data: [0, 1, 4, 9, 16, 25, ...] (squares)
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -176,7 +176,7 @@ def test_neighbor_difference():
     print("✅ Basic neighbor difference test passed!")
 
 
-def test_moving_average():
+def test_moving_average() raises:
     with DeviceContext() as ctx:
         # Create test data: [1, 2, 4, 7, 11, 16, 22, 29, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE_2)
@@ -239,7 +239,7 @@ def test_moving_average():
     print("✅ Moving average test passed!")
 
 
-def test_broadcast_shuffle_coordination():
+def test_broadcast_shuffle_coordination() raises:
     with DeviceContext() as ctx:
         # Create test data: [2, 4, 6, 8, 1, 3, 5, 7, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -293,7 +293,7 @@ def test_broadcast_shuffle_coordination():
     print("✅ Broadcast + Shuffle coordination test passed!")
 
 
-def test_basic_broadcast():
+def test_basic_broadcast() raises:
     with DeviceContext() as ctx:
         # Create test data: [1, 2, 3, 4, 5, 6, 7, 8, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -338,7 +338,7 @@ def test_basic_broadcast():
     print("✅ Basic broadcast test passed!")
 
 
-def test_conditional_broadcast():
+def test_conditional_broadcast() raises:
     with DeviceContext() as ctx:
         # Create test data: [3, 1, 7, 2, 9, 4, 6, 8, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -398,7 +398,7 @@ def test_conditional_broadcast():
     print("✅ Conditional broadcast test passed!")
 
 
-def main():
+def main() raises:
     print("WARP_SIZE: ", WARP_SIZE)
     if len(argv()) < 1 or len(argv()) > 2:
         print(
diff --git a/problems/p26/p26.mojo b/problems/p26/p26.mojo
index dded54af..c35fee33 100644
--- a/problems/p26/p26.mojo
+++ b/problems/p26/p26.mojo
@@ -156,7 +156,7 @@ fn warp_partition[
 # ANCHOR_END: warp_partition
 
 
-def test_butterfly_pair_swap():
+def test_butterfly_pair_swap() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -201,7 +201,7 @@ def test_butterfly_pair_swap():
     print("✅ Butterfly pair swap test passed!")
 
 
-def test_butterfly_parallel_max():
+def test_butterfly_parallel_max() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -241,7 +241,7 @@ def test_butterfly_parallel_max():
     print("✅ Butterfly parallel max test passed!")
 
 
-def test_butterfly_conditional_max():
+def test_butterfly_conditional_max() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE_2)
         input_buf.enqueue_fill(0)
@@ -301,7 +301,7 @@ def test_butterfly_conditional_max():
     print("✅ Butterfly conditional max test passed!")
 
 
-def test_warp_inclusive_prefix_sum():
+def test_warp_inclusive_prefix_sum() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -342,7 +342,7 @@ def test_warp_inclusive_prefix_sum():
     print("✅ Warp inclusive prefix sum test passed!")
 
 
-def test_warp_partition():
+def test_warp_partition() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -416,7 +416,7 @@ def test_warp_partition():
     print("✅ Warp partition test passed!")
 
 
-def main():
+def main() raises:
     print("WARP_SIZE: ", WARP_SIZE)
     if len(argv()) < 2:
         print(
diff --git a/problems/p27/p27.mojo b/problems/p27/p27.mojo
index 88e35a67..22f8d08f 100644
--- a/problems/p27/p27.mojo
+++ b/problems/p27/p27.mojo
@@ -180,7 +180,7 @@ fn block_normalize_vector[
 # ANCHOR_END: block_normalize
 
 
-def main():
+def main() raises:
     if len(argv()) != 2:
         print(
             "Usage: --traditional-dot-product | --block-sum-dot-product |"
diff --git a/problems/p28/p28.mojo b/problems/p28/p28.mojo
index 0eb59014..cff3ca15 100644
--- a/problems/p28/p28.mojo
+++ b/problems/p28/p28.mojo
@@ -53,7 +53,7 @@ fn async_copy_overlap_convolution[
 # ANCHOR_END: async_copy_overlap_convolution
 
 
-def test_async_copy_overlap_convolution():
+def test_async_copy_overlap_convolution() raises:
     """Test async copy overlap with 1D convolution."""
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](VECTOR_SIZE)
@@ -144,7 +144,7 @@ def test_async_copy_overlap_convolution():
                     print("Async copy overlap convolution test FAILED!")
 
 
-def main():
+def main() raises:
     """Run memory fence tests based on command line arguments."""
     if len(argv()) != 1:
         print("Usage: p25.mojo")
diff --git a/problems/p29/p29.mojo b/problems/p29/p29.mojo
index f25e87dd..44101ee4 100644
--- a/problems/p29/p29.mojo
+++ b/problems/p29/p29.mojo
@@ -196,7 +196,7 @@ fn double_buffered_stencil_computation[
 # ANCHOR_END: double_buffered_stencil
 
 
-def test_multi_stage_pipeline():
+def test_multi_stage_pipeline() raises:
     """Test Puzzle 29A: Multi-Stage Pipeline Coordination."""
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -257,7 +257,7 @@ def test_multi_stage_pipeline():
             print("✅ Multi-stage pipeline coordination test PASSED!")
 
 
-def test_double_buffered_stencil():
+def test_double_buffered_stencil() raises:
     """Test Puzzle 29B: Double-Buffered Stencil Computation."""
     with DeviceContext() as ctx:
         # Test Puzzle 29B: Double-Buffered Stencil Computation
@@ -334,7 +334,7 @@ def test_double_buffered_stencil():
             print("✅ Double-buffered stencil test PASSED!")
 
 
-def main():
+def main() raises:
     """Run GPU synchronization tests based on command line arguments."""
     print("Puzzle 29: GPU Synchronization Primitives")
     print("=" * 50)
diff --git a/problems/p30/p30.mojo b/problems/p30/p30.mojo
index 8f8749c8..ea5cdce9 100644
--- a/problems/p30/p30.mojo
+++ b/problems/p30/p30.mojo
@@ -189,7 +189,7 @@ fn benchmark_kernel3_parameterized[test_size: Int](mut b: Bencher) raises:
     b.iter_custom[kernel3_workflow](bench_ctx)
 
 
-def test_kernel1():
+def test_kernel1() raises:
     """Test kernel 1."""
     print("Testing kernel 1...")
     with DeviceContext() as ctx:
@@ -232,7 +232,7 @@ def test_kernel1():
         print("✅ Kernel 1 test passed")
 
 
-def test_kernel2():
+def test_kernel2() raises:
     """Test kernel 2."""
     print("Testing kernel 2...")
     with DeviceContext() as ctx:
@@ -278,7 +278,7 @@ def test_kernel2():
         print("✅ Kernel 2 test passed,", processed, "elements processed")
 
 
-def test_kernel3():
+def test_kernel3() raises:
     """Test kernel 3."""
     print("Testing kernel 3...")
     with DeviceContext() as ctx:
@@ -321,7 +321,7 @@ def test_kernel3():
         print("✅ Kernel 3 test passed")
 
 
-def main():
+def main() raises:
     """Run the memory access pattern tests."""
     args = argv()
     if len(args) < 2:
diff --git a/problems/p31/p31.mojo b/problems/p31/p31.mojo
index 9ba699a9..3f151327 100644
--- a/problems/p31/p31.mojo
+++ b/problems/p31/p31.mojo
@@ -297,7 +297,7 @@ fn benchmark_balanced_parameterized[test_size: Int](mut b: Bencher) raises:
     b.iter_custom[balanced_workflow](bench_ctx)
 
 
-def test_minimal():
+def test_minimal() raises:
     """Test minimal kernel."""
     print("Testing minimal kernel...")
     with DeviceContext() as ctx:
@@ -340,7 +340,7 @@ def test_minimal():
         print("✅ Minimal kernel test passed")
 
 
-def test_sophisticated():
+def test_sophisticated() raises:
     """Test sophisticated kernel."""
     print("Testing sophisticated kernel...")
     with DeviceContext() as ctx:
@@ -384,7 +384,7 @@ def test_sophisticated():
         print("✅ Sophisticated kernel test passed")
 
 
-def test_balanced():
+def test_balanced() raises:
     """Test balanced kernel."""
     print("Testing balanced kernel...")
     with DeviceContext() as ctx:
@@ -428,7 +428,7 @@ def test_balanced():
         print("✅ Balanced kernel test passed")
 
 
-def main():
+def main() raises:
     """Run the occupancy efficiency mystery tests."""
     args = argv()
     if len(args) < 2:
diff --git a/problems/p33/p33.mojo b/problems/p33/p33.mojo
index 90e9274e..66f20302 100644
--- a/problems/p33/p33.mojo
+++ b/problems/p33/p33.mojo
@@ -241,7 +241,7 @@ fn tensor_core_matrix_multiplication[
 # ANCHOR_END: tensor_core_matrix_multiplication
 
 
-def main():
+def main() raises:
     print("Puzzle 33: Tensor Core Operations")
 
     if len(argv()) < 2:
diff --git a/problems/p34/p34.mojo b/problems/p34/p34.mojo
index 905efdf1..65e9fd2c 100644
--- a/problems/p34/p34.mojo
+++ b/problems/p34/p34.mojo
@@ -115,7 +115,7 @@ fn advanced_cluster_patterns[
 # ANCHOR_END: advanced_cluster_patterns
 
 
-def main():
+def main() raises:
     """Test cluster programming concepts using proper Mojo GPU patterns."""
     if len(argv()) < 2:
         print("Usage: p34.mojo [--coordination | --reduction | --advanced]")
diff --git a/solutions/p01/p01.mojo b/solutions/p01/p01.mojo
index 66b26f70..0369bcd5 100644
--- a/solutions/p01/p01.mojo
+++ b/solutions/p01/p01.mojo
@@ -21,7 +21,7 @@ fn add_10(
 # ANCHOR_END: add_10_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p02/p02.mojo b/solutions/p02/p02.mojo
index 75558b66..6bcf7acc 100644
--- a/solutions/p02/p02.mojo
+++ b/solutions/p02/p02.mojo
@@ -22,7 +22,7 @@ fn add(
 # ANCHOR_END: add_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p03/p03.mojo b/solutions/p03/p03.mojo
index d16a91d1..de825362 100644
--- a/solutions/p03/p03.mojo
+++ b/solutions/p03/p03.mojo
@@ -23,7 +23,7 @@ fn add_10_guard(
 # ANCHOR_END: add_10_guard_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p04/p04.mojo b/solutions/p04/p04.mojo
index 442bce41..dc097b80 100644
--- a/solutions/p04/p04.mojo
+++ b/solutions/p04/p04.mojo
@@ -24,7 +24,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p04/p04_layout_tensor.mojo b/solutions/p04/p04_layout_tensor.mojo
index 4971ccaf..1eaf5d18 100644
--- a/solutions/p04/p04_layout_tensor.mojo
+++ b/solutions/p04/p04_layout_tensor.mojo
@@ -25,7 +25,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/solutions/p05/p05.mojo b/solutions/p05/p05.mojo
index 8944cec5..dacb7ec3 100644
--- a/solutions/p05/p05.mojo
+++ b/solutions/p05/p05.mojo
@@ -25,7 +25,7 @@ fn broadcast_add(
 # ANCHOR_END: broadcast_add_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p05/p05_layout_tensor.mojo b/solutions/p05/p05_layout_tensor.mojo
index 7e12f92c..6a5cbe54 100644
--- a/solutions/p05/p05_layout_tensor.mojo
+++ b/solutions/p05/p05_layout_tensor.mojo
@@ -32,7 +32,7 @@ fn broadcast_add[
 # ANCHOR_END: broadcast_add_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/solutions/p06/p06.mojo b/solutions/p06/p06.mojo
index c7503b18..e0209502 100644
--- a/solutions/p06/p06.mojo
+++ b/solutions/p06/p06.mojo
@@ -23,7 +23,7 @@ fn add_10_blocks(
 # ANCHOR_END: add_10_blocks_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p07/p07.mojo b/solutions/p07/p07.mojo
index 2a30b08b..6e345079 100644
--- a/solutions/p07/p07.mojo
+++ b/solutions/p07/p07.mojo
@@ -24,7 +24,7 @@ fn add_10_blocks_2d(
 # ANCHOR_END: add_10_blocks_2d_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p07/p07_layout_tensor.mojo b/solutions/p07/p07_layout_tensor.mojo
index 96838216..1d633d8a 100644
--- a/solutions/p07/p07_layout_tensor.mojo
+++ b/solutions/p07/p07_layout_tensor.mojo
@@ -29,7 +29,7 @@ fn add_10_blocks_2d[
 # ANCHOR_END: add_10_blocks_2d_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out_buf = ctx.enqueue_create_buffer[dtype](SIZE * SIZE)
         out_buf.enqueue_fill(0)
diff --git a/solutions/p08/p08.mojo b/solutions/p08/p08.mojo
index def0941f..cc877017 100644
--- a/solutions/p08/p08.mojo
+++ b/solutions/p08/p08.mojo
@@ -43,7 +43,7 @@ fn add_10_shared(
 # ANCHOR_END: add_10_shared_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p08/p08_layout_tensor.mojo b/solutions/p08/p08_layout_tensor.mojo
index b9c6fdee..3d075bf4 100644
--- a/solutions/p08/p08_layout_tensor.mojo
+++ b/solutions/p08/p08_layout_tensor.mojo
@@ -47,7 +47,7 @@ fn add_10_shared_layout_tensor[
 # ANCHOR_END: add_10_shared_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p10/p10.mojo b/solutions/p10/p10.mojo
index 936b86a8..13286e18 100644
--- a/solutions/p10/p10.mojo
+++ b/solutions/p10/p10.mojo
@@ -66,7 +66,7 @@ fn add_10_2d(
 # ANCHOR_END: add_10_2d_solution
 
 
-def main():
+def main() raises:
     if len(argv()) != 2:
         print(
             "Expected one command-line argument: '--memory-bug' or"
diff --git a/solutions/p11/p11.mojo b/solutions/p11/p11.mojo
index d40a34ed..69d91460 100644
--- a/solutions/p11/p11.mojo
+++ b/solutions/p11/p11.mojo
@@ -42,7 +42,7 @@ fn pooling(
 # ANCHOR_END: pooling_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p11/p11_layout_tensor.mojo b/solutions/p11/p11_layout_tensor.mojo
index 4f9fafc5..a9e4f929 100644
--- a/solutions/p11/p11_layout_tensor.mojo
+++ b/solutions/p11/p11_layout_tensor.mojo
@@ -53,7 +53,7 @@ fn pooling[
 # ANCHOR_END: pooling_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
         out.enqueue_fill(0)
diff --git a/solutions/p12/p12.mojo b/solutions/p12/p12.mojo
index 91b1a489..b2ab51e8 100644
--- a/solutions/p12/p12.mojo
+++ b/solutions/p12/p12.mojo
@@ -56,7 +56,7 @@ fn dot_product(
 # ANCHOR_END: dot_product_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](1)
         out.enqueue_fill(0)
diff --git a/solutions/p12/p12_layout_tensor.mojo b/solutions/p12/p12_layout_tensor.mojo
index 978ed60e..64f6d079 100644
--- a/solutions/p12/p12_layout_tensor.mojo
+++ b/solutions/p12/p12_layout_tensor.mojo
@@ -55,7 +55,7 @@ fn dot_product[
 # ANCHOR_END: dot_product_layout_tensor_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](1)
         out.enqueue_fill(0)
diff --git a/solutions/p13/p13.mojo b/solutions/p13/p13.mojo
index ca7e4f22..2f1359a4 100644
--- a/solutions/p13/p13.mojo
+++ b/solutions/p13/p13.mojo
@@ -141,7 +141,7 @@ fn conv_1d_block_boundary[
 # ANCHOR_END: conv_1d_block_boundary_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         size = SIZE_2 if argv()[1] == "--block-boundary" else SIZE
         conv = CONV_2 if argv()[1] == "--block-boundary" else CONV
diff --git a/solutions/p14/p14.mojo b/solutions/p14/p14.mojo
index 23d98554..61145ac1 100644
--- a/solutions/p14/p14.mojo
+++ b/solutions/p14/p14.mojo
@@ -152,7 +152,7 @@ fn prefix_sum_block_sum_phase[
 # ANCHOR_END: prefix_sum_complete_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         use_simple = argv()[1] == "--simple"
         size = SIZE if use_simple else SIZE_2
diff --git a/solutions/p15/p15.mojo b/solutions/p15/p15.mojo
index 3bdd09f9..2a2f169f 100644
--- a/solutions/p15/p15.mojo
+++ b/solutions/p15/p15.mojo
@@ -73,7 +73,7 @@ fn axis_sum[
 # ANCHOR_END: axis_sum_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](BATCH)
         out.enqueue_fill(0)
diff --git a/solutions/p16/p16.mojo b/solutions/p16/p16.mojo
index 453c73b9..c519d26f 100644
--- a/solutions/p16/p16.mojo
+++ b/solutions/p16/p16.mojo
@@ -228,7 +228,7 @@ fn matmul_idiomatic_tiled[
 # ANCHOR_END: matmul_idiomatic_tiled_solution
 
 
-def main():
+def main() raises:
     with DeviceContext() as ctx:
         size = (
             SIZE_TILED if argv()[1] == "--idiomatic-tiled"
diff --git a/solutions/p18/test/test_softmax.mojo b/solutions/p18/test/test_softmax.mojo
index 5171d550..4eb16fe7 100644
--- a/solutions/p18/test/test_softmax.mojo
+++ b/solutions/p18/test/test_softmax.mojo
@@ -12,7 +12,7 @@ comptime BLOCK_DIM_X = 1 << log2_ceil(SIZE)
 comptime dtype = DType.float32
 
 
-def test_softmax():
+def test_softmax() raises:
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[DType.float32](SIZE)
         out.enqueue_fill(0)
@@ -79,5 +79,5 @@ def test_softmax():
             print("All tests passed 🎉")
 
 
-def main():
+def main() raises:
     test_softmax()
diff --git a/solutions/p23/p23.mojo b/solutions/p23/p23.mojo
index d8d3a311..c4c9bd25 100644
--- a/solutions/p23/p23.mojo
+++ b/solutions/p23/p23.mojo
@@ -334,7 +334,7 @@ fn benchmark_vectorized_parameterized[
     bench_ctx.synchronize()
 
 
-def main():
+def main() raises:
     ctx = DeviceContext()
     out = ctx.enqueue_create_buffer[dtype](SIZE)
     out.enqueue_fill(0)
diff --git a/solutions/p24/p24.mojo b/solutions/p24/p24.mojo
index 4044ce75..55750a23 100644
--- a/solutions/p24/p24.mojo
+++ b/solutions/p24/p24.mojo
@@ -325,7 +325,7 @@ fn benchmark_traditional_parameterized[
     bench_ctx.synchronize()
 
 
-def main():
+def main() raises:
     if argv()[1] != "--benchmark":
         print("SIZE:", SIZE)
         print("WARP_SIZE:", WARP_SIZE)
diff --git a/solutions/p25/p25.mojo b/solutions/p25/p25.mojo
index 53873da3..ad2ce01e 100644
--- a/solutions/p25/p25.mojo
+++ b/solutions/p25/p25.mojo
@@ -212,7 +212,7 @@ fn conditional_broadcast[
 # ANCHOR_END: conditional_broadcast_solution
 
 
-def test_neighbor_difference():
+def test_neighbor_difference() raises:
     with DeviceContext() as ctx:
         # Create test data: [0, 1, 4, 9, 16, 25, ...] (squares)
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -256,7 +256,7 @@ def test_neighbor_difference():
     print("✅ Basic neighbor difference test passed!")
 
 
-def test_moving_average():
+def test_moving_average() raises:
     with DeviceContext() as ctx:
         # Create test data: [1, 2, 4, 7, 11, 16, 22, 29, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE_2)
@@ -320,7 +320,7 @@ def test_moving_average():
     print("✅ Moving average test passed!")
 
 
-def test_broadcast_shuffle_coordination():
+def test_broadcast_shuffle_coordination() raises:
     with DeviceContext() as ctx:
         # Create test data: [2, 4, 6, 8, 1, 3, 5, 7, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -375,7 +375,7 @@ def test_broadcast_shuffle_coordination():
     print("✅ Broadcast + Shuffle coordination test passed!")
 
 
-def test_basic_broadcast():
+def test_basic_broadcast() raises:
     with DeviceContext() as ctx:
         # Create test data: [1, 2, 3, 4, 5, 6, 7, 8, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -421,7 +421,7 @@ def test_basic_broadcast():
     print("✅ Basic broadcast test passed!")
 
 
-def test_conditional_broadcast():
+def test_conditional_broadcast() raises:
     with DeviceContext() as ctx:
         # Create test data: [3, 1, 7, 2, 9, 4, 6, 8, ...]
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -482,7 +482,7 @@ def test_conditional_broadcast():
     print("✅ Conditional broadcast test passed!")
 
 
-def main():
+def main() raises:
     print("WARP_SIZE: ", WARP_SIZE)
     if len(argv()) != 2:
         print(
diff --git a/solutions/p26/p26.mojo b/solutions/p26/p26.mojo
index 30a57073..cb4eb437 100644
--- a/solutions/p26/p26.mojo
+++ b/solutions/p26/p26.mojo
@@ -222,7 +222,7 @@ fn warp_partition[
 # ANCHOR_END: warp_partition_solution
 
 
-def test_butterfly_pair_swap():
+def test_butterfly_pair_swap() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -267,7 +267,7 @@ def test_butterfly_pair_swap():
     print("✅ Butterfly pair swap test passed!")
 
 
-def test_butterfly_parallel_max():
+def test_butterfly_parallel_max() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -307,7 +307,7 @@ def test_butterfly_parallel_max():
     print("✅ Butterfly parallel max test passed!")
 
 
-def test_butterfly_conditional_max():
+def test_butterfly_conditional_max() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE_2)
         input_buf.enqueue_fill(0)
@@ -367,7 +367,7 @@ def test_butterfly_conditional_max():
     print("✅ Butterfly conditional max test passed!")
 
 
-def test_warp_inclusive_prefix_sum():
+def test_warp_inclusive_prefix_sum() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -409,7 +409,7 @@ def test_warp_inclusive_prefix_sum():
     print("✅ Warp inclusive prefix sum test passed!")
 
 
-def test_warp_partition():
+def test_warp_partition() raises:
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](SIZE)
         input_buf.enqueue_fill(0)
@@ -484,7 +484,7 @@ def test_warp_partition():
     print("✅ Warp partition test passed!")
 
 
-def main():
+def main() raises:
     print("WARP_SIZE: ", WARP_SIZE)
     if len(argv()) != 2:
         print(
diff --git a/solutions/p27/p27.mojo b/solutions/p27/p27.mojo
index fb447c52..39b4622a 100644
--- a/solutions/p27/p27.mojo
+++ b/solutions/p27/p27.mojo
@@ -214,7 +214,7 @@ fn block_normalize_vector[
 # ANCHOR_END: block_normalize_solution
 
 
-def main():
+def main() raises:
     if len(argv()) != 2:
         print(
             "Usage: --traditional-dot-product | --block-sum-dot-product |"
diff --git a/solutions/p28/p28.mojo b/solutions/p28/p28.mojo
index a7419e93..57c80636 100644
--- a/solutions/p28/p28.mojo
+++ b/solutions/p28/p28.mojo
@@ -88,7 +88,7 @@ fn async_copy_overlap_convolution[
 # ANCHOR_END: async_copy_overlap_convolution_solution
 
 
-def test_async_copy_overlap_convolution():
+def test_async_copy_overlap_convolution() raises:
     """Test async copy overlap with 1D convolution."""
     with DeviceContext() as ctx:
         input_buf = ctx.enqueue_create_buffer[dtype](VECTOR_SIZE)
@@ -179,7 +179,7 @@ def test_async_copy_overlap_convolution():
                     print("Async copy overlap convolution test FAILED!")
 
 
-def main():
+def main() raises:
     """Run memory fence tests based on command line arguments."""
     if len(argv()) != 1:
         print("Usage: p25.mojo")
diff --git a/solutions/p29/p29.mojo b/solutions/p29/p29.mojo
index 98b0a275..5dee39fa 100644
--- a/solutions/p29/p29.mojo
+++ b/solutions/p29/p29.mojo
@@ -271,7 +271,7 @@ fn double_buffered_stencil_computation[
 # ANCHOR_END: double_buffered_stencil_solution
 
 
-def test_multi_stage_pipeline():
+def test_multi_stage_pipeline() raises:
     """Test Puzzle 26A: Multi-Stage Pipeline Coordination."""
     with DeviceContext() as ctx:
         out = ctx.enqueue_create_buffer[dtype](SIZE)
@@ -332,7 +332,7 @@ def test_multi_stage_pipeline():
             print("✅ Multi-stage pipeline coordination test PASSED!")
 
 
-def test_double_buffered_stencil():
+def test_double_buffered_stencil() raises:
     """Test Puzzle 26B: Double-Buffered Stencil Computation."""
     with DeviceContext() as ctx:
         # Test Puzzle 26B: Double-Buffered Stencil Computation
@@ -409,7 +409,7 @@ def test_double_buffered_stencil():
             print("✅ Double-buffered stencil test PASSED!")
 
 
-def main():
+def main() raises:
     """Run GPU synchronization tests based on command line arguments."""
     print("Puzzle 26: GPU Synchronization Primitives")
     print("=" * 50)
diff --git a/solutions/p33/p33.mojo b/solutions/p33/p33.mojo
index 8e1adddc..19e033aa 100644
--- a/solutions/p33/p33.mojo
+++ b/solutions/p33/p33.mojo
@@ -256,7 +256,7 @@ fn tensor_core_matrix_multiplication[
 # ANCHOR_END: tensor_core_matrix_multiplication_solution
 
 
-def main():
+def main() raises:
     print("Puzzle 33: Tensor Core Operations")
 
     if len(argv()) < 2:
diff --git a/solutions/p34/p34.mojo b/solutions/p34/p34.mojo
index cad94e4d..31ca775d 100644
--- a/solutions/p34/p34.mojo
+++ b/solutions/p34/p34.mojo
@@ -196,7 +196,7 @@ fn advanced_cluster_patterns[
 # ANCHOR_END: advanced_cluster_patterns_solution
 
 
-def main():
+def main() raises:
     """Test cluster programming concepts using proper Mojo GPU patterns."""
     if len(argv()) < 2:
         print("Usage: p34.mojo [--coordination | --reduction | --advanced]")