{ "benchmark_source_sha256": "c0876934ab9f5967591fc11ed48791063dae9f13c4d5bd6e0691a485da7b1324", "cases": [ { "bytes_read_and_written": 131072, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 32, 1, 1 ], "tile_shape": [ 32 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 4, 8 ], "output_shape": [ 1024, 8, 4 ], "seed": 1073, "tile_shape": [ 4, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00028675, "scatter_seconds": 0.000278167, "speedup": 1.0308555651820668 } } }, { "bytes_read_and_written": 262144, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 32, 1, 1 ], "tile_shape": [ 32 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 4, 8 ], "output_shape": [ 1024, 8, 4 ], "seed": 1073, "tile_shape": [ 4, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00024816650000000003, "scatter_seconds": 0.000230083, "speedup": 1.078595550301413 } } }, { "bytes_read_and_written": 2097152, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 32, 1, 1 ], "tile_shape": [ 32 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 4, 8 ], "output_shape": [ 16384, 8, 4 ], "seed": 16433, "tile_shape": [ 4, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00022120900000000002, "scatter_seconds": 0.000219334, "speedup": 1.0085486062352393 } } }, { "bytes_read_and_written": 4194304, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 32, 1, 1 ], "tile_shape": [ 32 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 4, 8 ], "output_shape": [ 16384, 8, 4 ], "seed": 16433, "tile_shape": [ 4, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00022989600000000002, "scatter_seconds": 0.00022910400000000002, "speedup": 1.0034569453174105 } } }, { "bytes_read_and_written": 262144, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 64, 1, 1 ], "tile_shape": [ 64 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 8, 8 ], "output_shape": [ 1024, 8, 8 ], "seed": 1105, "tile_shape": [ 8, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.000206292, "scatter_seconds": 0.000202125, "speedup": 1.0206159554730985 } } }, { "bytes_read_and_written": 524288, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 64, 1, 1 ], "tile_shape": [ 64 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 8, 8 ], "output_shape": [ 1024, 8, 8 ], "seed": 1105, "tile_shape": [ 8, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.0002030415, "scatter_seconds": 0.0002034995, "speedup": 0.9977493802196075 } } }, { "bytes_read_and_written": 4194304, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 64, 1, 1 ], "tile_shape": [ 64 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 8, 8 ], "output_shape": [ 16384, 8, 8 ], "seed": 16465, "tile_shape": [ 8, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00023181250000000003, "scatter_seconds": 0.000229542, "speedup": 1.0098914359899278 } } }, { "bytes_read_and_written": 8388608, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 64, 1, 1 ], "tile_shape": [ 64 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 8, 8 ], "output_shape": [ 16384, 8, 8 ], "seed": 16465, "tile_shape": [ 8, 8 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.000248083, "scatter_seconds": 0.000253708, "speedup": 0.977828842606461 } } }, { "bytes_read_and_written": 524288, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 128, 1, 1 ], "tile_shape": [ 128 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 8, 16 ], "output_shape": [ 1024, 16, 8 ], "seed": 1169, "tile_shape": [ 8, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.000197979, "scatter_seconds": 0.00020337500000000002, "speedup": 0.9734677320221266 } } }, { "bytes_read_and_written": 1048576, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 128, 1, 1 ], "tile_shape": [ 128 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 8, 16 ], "output_shape": [ 1024, 16, 8 ], "seed": 1169, "tile_shape": [ 8, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00020383300000000003, "scatter_seconds": 0.00020583300000000002, "speedup": 0.9902833850743078 } } }, { "bytes_read_and_written": 8388608, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 128, 1, 1 ], "tile_shape": [ 128 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 8, 16 ], "output_shape": [ 16384, 16, 8 ], "seed": 16529, "tile_shape": [ 8, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00024391600000000003, "scatter_seconds": 0.000254208, "speedup": 0.959513469284995 } } }, { "bytes_read_and_written": 16777216, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 128, 1, 1 ], "tile_shape": [ 128 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 8, 16 ], "output_shape": [ 16384, 16, 8 ], "seed": 16529, "tile_shape": [ 8, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.000325708, "scatter_seconds": 0.000329208, "speedup": 0.989368423610605 } } }, { "bytes_read_and_written": 1048576, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 256, 1, 1 ], "tile_shape": [ 256 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 16, 16 ], "output_shape": [ 1024, 16, 16 ], "seed": 1297, "tile_shape": [ 16, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00020164600000000001, "scatter_seconds": 0.0002098125, "speedup": 0.9610771522192434 } } }, { "bytes_read_and_written": 2097152, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 256, 1, 1 ], "tile_shape": [ 256 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 1024, 16, 16 ], "output_shape": [ 1024, 16, 16 ], "seed": 1297, "tile_shape": [ 16, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.000217667, "scatter_seconds": 0.000220167, "speedup": 0.9886449831264449 } } }, { "bytes_read_and_written": 16777216, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float16", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 256, 1, 1 ], "tile_shape": [ 256 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 16, 16 ], "output_shape": [ 16384, 16, 16 ], "seed": 16657, "tile_shape": [ 16, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00032283400000000003, "scatter_seconds": 0.000331833, "speedup": 0.9728809370978776 } } }, { "bytes_read_and_written": 33554432, "correctness": { "bitwise_exact": true, "passed": true, "reference": "numpy_contiguous_batched_transpose", "signed_zero_bits_checked": true, "validated_before_timing": true }, "destination_layout": null, "dtype": "float32", "execution_reports": { "ownership": null, "scatter": { "allocations": [], "double_buffered": false, "epilogue_regions": 0, "loops": [], "notes": [ "Layout widths describe emitted Metal operations, not final machine instructions.", "Masked tails may use scalar accesses even when aligned interiors are vectorized.", "No software double buffering materialized." ], "passes": [ "split_elementwise_loops", "vectorize_elementwise", "fold_constants" ], "plan": { "backend": "elementwise", "cooperative_layout": "scalar lanes", "double_buffer": null, "outer_bounds_proven": false, "reasons": [ "Lane geometry follows the traced arange/BLOCK contract." ], "simdgroup_grid": null, "staging": "device", "threadgroup_size": [ 256, 1, 1 ], "tile_shape": [ 256 ], "vector_width": null }, "vectorized_loads": 0 } }, "input_shape": [ 16384, 16, 16 ], "output_shape": [ 16384, 16, 16 ], "seed": 16657, "tile_shape": [ 16, 16 ], "wall_pairs": { "scatter_vs_mlx": { "mlx_seconds": 0.00046972900000000006, "scatter_seconds": 0.0004694165, "speedup": 1.0006657201014453 } } } ], "configuration": { "batches": [ 1024, 16384 ], "dtypes": [ "float16", "float32" ], "metile_root": "/tmp/metile-layout-baseline.Jdz5Ug", "output_json": "benchmarks/results/m5-thread-layouts-baseline.json", "rep_ms": 200.0, "scatter_only": true, "seed": 17, "shapes": [ "4x8", "8x8", "8x16", "16x16" ], "warmup_ms": 50.0 }, "environment": {}, "hardware": { "chip": "Apple M5", "machine": "arm64", "memory": "32 GB", "model_name": "Mac17,2", "processor": "arm", "release": "25.2.0", "system": "Darwin" }, "measurement": { "clock": "perf_counter_ns", "communication": "compiler_selected_identity_simd_shuffle_or_synchronous_threadgroup", "input_allocation_and_compilation_excluded": true, "metric": "synchronized_wall_seconds", "mlx_output": "mx.contiguous_materialized_transpose_not_a_strided_view", "order": "alternating_AB_BA_within_each_pair", "output_allocation": "meTile_preallocated_MLX_framework_managed", "rep_ms": 200.0, "statistic": "middle_80_percent_median", "synchronization": "meTile_queue_and_mlx_default_stream_after_each_call", "timed_scope": "prepared_meTile_dispatch_vs_compiled_MLX_call_and_eval", "warmup_ms": 50.0 }, "metile_root": "/private/tmp/metile-layout-baseline.Jdz5Ug", "precision_comparison": { "baseline_weights": "matching_per_case_float16_or_float32_storage", "bitwise_exact": true, "class": "same_storage_precision", "operation": "pure_reordering_without_arithmetic_or_quantization", "optimized_weights": "matching_per_case_float16_or_float32_storage", "same_weight_representation": true, "validation": "numpy_bitwise_reference_and_mlx_contiguous_materialization" }, "recorded_at": "2026-10-02T16:36:59.238203+00:00", "schema_version": 1, "software": { "metal_compiler": "Apple metal version 32023.864 (metalfe-32023.864)\nTarget: air64-apple-darwin25.2.0\nThread model: posix\nInstalledDir: /private/var/run/com.apple.security.cryptexd/mnt/com.apple.MobileAsset.MetalToolchain-v17.3.7003.10.GzPOjD/Metal.xctoolchain/usr/metal/current/bin", "metile": "0.0.1", "mlx": "0.32.0", "numpy": "2.5.3", "python": "3.12.12" }, "source_sha256": { "metile/__init__.py": "e467b57b1fe0ed9bdbe3a55ad400f4618e696e88b1bc28f65d2c7e8bb18df2ff", "metile/backends/__init__.py": "a69e602909b06e96197769e4bd8c20feff1f19c0ad3179b9024fc0dd491306fb", "metile/backends/mlx.py": "badc9c3f0e2e4354128ab4ebb4980612c7272098e3290e190139ee82a3b22dc0", "metile/backends/mlx_affine.py": "69dac0c8e776d83ebf01ab30375f5c750b609428fc43df4381ba88d10c80fc2a", "metile/backends/mlx_attention.py": "53edf481eb0ea14780840948ec7ecbb20ba176381468ee68a691394205cd7f31", "metile/backends/mlx_block_scaled.py": "558f2c934768fdcb9a00e9710ab4d642a4d933a903eeb7ce7dc042df4553e76a", "metile/backends/mlx_compressed_down.py": "b31a94c27475448975c010dd59bc7b2c1889576b61b536cf0b9af084b5c1c047", "metile/backends/mlx_dense.py": "feea78dad21a53cb56f32e1c16b340f30d018cd8906f813dae4eb2a3a3fd769e", "metile/backends/mlx_dense_residual.py": "b27771e2735f712fbe94ddf85fb70702f6b5aa592ab0d3d47232d1433e2f7bc5", "metile/backends/mlx_dense_swiglu.py": "88341838b024abcaec77709403e506862f9488bc767a452671e1e4df1e0a73a2", "metile/backends/mlx_graph.py": "03963e7902645c8db23102279031313b7228dd46166a2d9f41fcc09f3b2b1256", "metile/backends/mlx_quantized/__init__.py": "2ad98e4382336c9e2ec8e251b8d4db5e14cca02db2658bec8c4a915fa453d27d", "metile/backends/mlx_quantized/common.py": "211f3797aa8df823da6f12919c908f98eb6d0a5f1a9bb6833de4261a7009d5b9", "metile/backends/mlx_quantized/dispatch.py": "37ad5d4cb08acbfc8b6eb34dd072a83953585c3780159a65ac40e8b792957cb2", "metile/backends/mlx_quantized/qmv.py": "996ee75a0b9113f8e12d162a5e5d8da929e96a4de4a96d635df62336c61dcf49", "metile/backends/mlx_quantized/residual.py": "be5d001d4abecb6faf6fcea0074001063258812b644d39eba11a1900de785592", "metile/backends/mlx_quantized/swiglu.py": "88f8250c94227d948ce1d7752d1c8b60f89eb3db8fab9a41e81ec49ce1229915", "metile/codegen/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "metile/codegen/msl_emitter/__init__.py": "722808c4b4d2b0582820fd499c93eca09f595b91ae598bec69ad4577880716ec", "metile/codegen/msl_emitter/block_scaled.py": "3c7cd7f2b895624a47e11e7b72748ea66efecec187e48e391f014f210e15ea54", "metile/codegen/msl_emitter/common.py": "b76f97778198b33e4e8039b7b8656d8948e7d3b6c3bdaebd0c8d803d9a0120ce", "metile/codegen/msl_emitter/elementwise.py": "2f54db22cc9b5dc2ecf0e4415ecb5415914ef9f5901828bb0091a8424b3c3bd4", "metile/codegen/msl_emitter/gemm.py": "11b24ca4a0b980a71e19944f6534652a23bfe2785aa06c44fa51b1bbd32dd8d3", "metile/codegen/msl_emitter/nax.py": "53ac752554c39fa80d89c22fd368ef07149f3ad7240ce380aba375507d406f24", "metile/codegen/msl_emitter/simdgroup.py": "61be7105bd75f60a473677634fff1836b5934804e24e2a93c740d386426ae7fc", "metile/compile.py": "5102f426f21d8c4e43ff8ef490db87b735dd7da686f2b047a831872d52613747", "metile/compiler/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "metile/compiler/affine_quantized.py": "6eee87860d983fb33f0f347717b924efdcd8c4bd213c3c2c328df4196f10bce1", "metile/compiler/agx_schedule.py": "27dc8832ce99d916a367bf659074272ef6fdee3dfba5aa9dc99ebf191c55d2b8", "metile/compiler/algo_discovery.py": "cec98735866080e102c623f119f7b27b1e0086b1404d59518d54925ad2c392c7", "metile/compiler/block_scaled.py": "251ac93ad2c0fa977dee2424b42f676d233e3c4cc9a82eb4d5e7add582b3eed0", "metile/compiler/dense.py": "d3555ee5f04697fb748139f1deeacd5b5cba9a4681c73fe0d163c7647b760493", "metile/compiler/epilogue.py": "2a882239dc276f3ed03a856a4b1e9129d4b46ff7579717eff931e6a6f7358831", "metile/compiler/execution_report.py": "02717ba9cbd765437f18ef8c4242ad6e828c5dcb642c646b142d37d4c445a8de", "metile/compiler/gemm_configs.py": "78f21e12b0a1b379dd90e0bc59a0e31a8c3fab9b7adf5f8760a2d28091c1cacf", "metile/compiler/graph_fusion.py": "477f56a99d9ac1c85b18c544269bf86a633792e2f4495c6f0fff7b80750e25a3", "metile/compiler/lowering/__init__.py": "9226cfdd2e825aadd7ace8502e9d3a108d95a44d0a395532db44fc286b8614d3", "metile/compiler/lowering/common.py": "e26a45a35dd6db9fb24e6dd43a682292ff1f3de760bb63f4c4e0a31ec7f13cb3", "metile/compiler/lowering/elementwise.py": "f7686032d741c26bae517c504ee6c79d76de8cf641926669d96cf095feb906f0", "metile/compiler/lowering/gemm.py": "ac42bc12856c769b7ea74afc073020f3fa48cb0e4eb87a4b7c82935fe4cd49df", "metile/compiler/max_flow.py": "aa55ce3e55d18410274e4a23a08c851d7bdfc6cd71b36fdc861b2138dd86e9db", "metile/compiler/options.py": "f303563d31f1aed532f31adcde8870d2cb107770769066395dd359d5040dc757", "metile/compiler/passes.py": "a9df3d3f88897ae06617601615f2156a355659d8120f07853e03e1c274611ef4", "metile/compiler/planning.py": "f715dc5482bdea23169c03744f5bafed4db9fe6444f385d6d10a613b87bd224b", "metile/compiler/reduction_algebra.py": "1721e995a6d28b625e5f8bee6b12d1b15121ae9433cce5226a165b34a814c6b1", "metile/compiler/schedule_expr.py": "5a1dedb24b86fdac712cf7dd5a05a67c67f4520192108d11859e1dcd4dfd918c", "metile/compiler/schedule_search.py": "c9ab6f30e4491285b68b3085e8d6ed3e89aab91bdcadb3bacc103757ee90d79b", "metile/compiler/schedules.py": "6db4d77a4fc4d8c6be5fca88563e05ecf04a46a05c267ce3e56cd80f2e04c094", "metile/compiler/scheduling.py": "ce51105ff2983dbac66d5f8936dd08a389945992aa80443c34c4eaf7ff3c0894", "metile/frontend/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "metile/frontend/autotune.py": "8764e7b43be4457d2c337bdd7f867d0373eed720c7327aef482380458669ba1b", "metile/frontend/kernel.py": "5ade85cc2a9dc49dabddf74b53d802c6d0f9c79b25fa59adb80804a9b4b4f6d2", "metile/frontend/tensor.py": "4329b14aad92de8356a244bb7736004549a6e6c5670cfe0c0ae47addc33009b2", "metile/frontend/tracing.py": "16d907faf48f8838c937ce3bb56b70b35aea952d10ac377ea3a7214f18babad3", "metile/integrations/__init__.py": "652dd5312ee9458f627f4cae0cb57bcffb4b186f5aca432c36ca89c6d2a8506f", "metile/integrations/mlx_lm/__init__.py": "f8de1d946a2107bcdee92059e1cf5b4efc6dec78296ba513a8186438386c2001", "metile/integrations/mlx_lm/_state.py": "172cbafb9be16eea09609c34a8a37309c0e710fed946c286d9ee2eb9c8355826", "metile/integrations/mlx_lm/apply.py": "fde44e04e3be0f99ca1d11a3ba7cb2abf99c84f841e8c18daacdc5aa0264eea9", "metile/integrations/mlx_lm/autotune.py": "627396bc354abd8471af54c122657fdb704981bbbf06688eb4d7b50dc02e0a9f", "metile/integrations/mlx_lm/compressed.py": "b717da12554b983203cfc87db88155b6e73cea4c64027c49f7ad748d7ba07097", "metile/integrations/mlx_lm/core.py": "5697e0fa1e0abac92d044314e3e6e0c2f2aa724e9e783cf8749967a5ad970b10", "metile/integrations/mlx_lm/modules.py": "f5ff7536efbb643e3e501fccd46809194ae8cc174f36ef04bef86116141a7c74", "metile/integrations/mlx_lm/patching.py": "6b2b6448b171d88245bbcc13c668cb16539a4472eb26dbc082382c7ece70ed37", "metile/integrations/mlx_lm/plan_model.py": "71be57f1b9bfc66e4767beed48a96d14b8fec156d8eb37a9b4a78ba9d5b29797", "metile/integrations/mlx_lm/tuning.py": "a9fd5665fc5508d0d03f78392b05e01247310b8a130fb82842fb8ed048d3e843", "metile/ir/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "metile/ir/graph_ir.py": "df6baf153562e58c2352fd99dcb77ad447d9e1433316353e295d71c6bd6002fc", "metile/ir/layout.py": "25adadd723df708c0ac8902ebc19273a7dfe397c148a0ecab7f8c6064d106c28", "metile/ir/metal_ir.py": "7116661a2cba4bfe5b07e2cff53a16efdc48463bd70d3df3f1da3c49554c26d9", "metile/ir/printer.py": "a962f96fd27481e70e7fae08656fe2dbc694205430902bc32116ed43d1548d6f", "metile/ir/tile_ir.py": "bc9babb628fe8ad8b5ee88a5f642f7952e872ef46477c16331fc2d6f6a3ca776", "metile/ir/types.py": "4217c9c6d3928e074f945ff5681d35eebb711acf7622ba7a18cfbdd1439ff20d", "metile/kernels/__init__.py": "30cfa7e525e2c5f0813971e526ffa30d7faf4c31fadaf067187cd704246e51b0", "metile/kernels/add_rmsnorm.py": "b042a33d0e81e060820d745ccfc5da86f4c7ee99c275eca6e73f89a03dba5b8f", "metile/kernels/affine_qmv.py": "760b329265388e2a5986a1265cdc73903d73cd7c9db4b833b5814b7572f4d317", "metile/kernels/attention.py": "463c03631f9b3255d4182fa9136122336128d0a33792426442b3f477e5d380bc", "metile/kernels/fft.py": "987fe73b9bb1a50c802ec83472f043a8a043730526e18dd46141d50904c0e591", "metile/kernels/gemm.py": "f5b353fc181e84ab0eb6d275e87d6a026a415d4b23ce367ce685f4e59b76d97e", "metile/kernels/layernorm.py": "6253a3aa5a43f668b414320cfc003e6135a27f75d9880497410cf341a39b98c6", "metile/kernels/mlp.py": "dea5cfcf3c8192c913de05f06bf11c350953bc1d80330a46b61f25b92b89c689", "metile/kernels/reduce.py": "e67c3e690b1a91dd32c0016bff62262e14d12d7e2dc882876cddf6b4c1278bf3", "metile/kernels/rmsnorm.py": "34a8856bbc5c945c47e561ef3e49f8b92771672496d041d83dd2b44cf60dea88", "metile/kernels/simdgroup_specialized_elementwise.py": "bd4b70d07c06e73f8fd7e53489f6178fa08c8d16320c3d7265077cfc95d1fad0", "metile/kernels/softmax.py": "0ac234200855dd91befb15ffdc8fce56e110a7a4650d1f9eafb5f3a06347619c", "metile/runtime/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "metile/runtime/address_space.py": "730103157cae2459201ddc1e3aa45736c7d068e77a8743f4186d9ce94337506b", "metile/runtime/attention.py": "35bb663fff776606ce7f45f5ff982808c378d72199b76485f7fdf49ad128c8e1", "metile/runtime/block_scaled.py": "32ae21160bdd6b0e9c6ed8f7d9b4845c8e6238ae3e99db8724188e5bc99b751b", "metile/runtime/buffer.py": "20ed8aed26a26f1057c5faff49b4fc7b326c0d7568bed072cba205b6611f4eeb", "metile/runtime/cache.py": "0025c8cf10416b9c5e9e1614d200596ae86a694c75d4a8f1e5d2fdacc8c94da3", "metile/runtime/metal_device.py": "b0190c367ba4cd1a13cfbddeb7223a4cc5c8be625adbbd2c90fa4335b06cd4ff", "metile/target/__init__.py": "41c7c5614012905213769bdb3f40d31f2a82715e415f28e59585061327261e34", "metile/target/agx.py": "c113f2c1071902316342b3f8127d2bf021b7907d1b6ae1ff4ca0007980e4d1d5", "metile/target/agx_isa.py": "25b12612f5976f5ed1234c06cad47506a445eb98c8987657983ba751bde0634d", "metile/tuning/__init__.py": "bcd0afdc50bd4f7acfb2cc716ffab5c0d63fc48a7a471534a68c5c45e95963c8", "metile/tuning/tournament.py": "a624be50382365230616c8694ee230a9b135de29325bbb27dcbde2b5cceccefc" } }