{ "op": "com.microsoft.SkipSimplifiedLayerNormalization", "cases": [ { "name": "skip-rmsnorm-f32-256x128", "preset": "smoke", "vars": { "rows": 256, "hidden": 128 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 215, "scale": 0.2 }, "skipT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 216, "scale": 0.2 }, "gammaT": { "shape": [128], "dtype": "float32", "dist": "uniform", "seed": 217, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [256, 128], "dtype": "float32" }, "residualT": { "shape": [256, 128], "dtype": "float32" } }, "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "kimi-linear-decode-f32-1x2304", "preset": "model", "provenance": { "notes": "Original Kimi Linear hidden_size=2304 decode residual-plus-RMSNorm shape; occurs twice per decoder layer." }, "vars": { "rows": 1, "hidden": 2304 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [1, 2304], "dtype": "float32", "dist": "normal", "seed": 1201, "scale": 0.2 }, "skipT": { "shape": [1, 2304], "dtype": "float32", "dist": "normal", "seed": 1202, "scale": 0.2 }, "gammaT": { "shape": [2304], "dtype": "float32", "dist": "uniform", "seed": 1203, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [1, 2304], "dtype": "float32" }, "residualT": { "shape": [1, 2304], "dtype": "float32" } }, "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "kimi-linear-prefill-f32-64x2304", "preset": "model", "provenance": { "notes": "Original Kimi Linear hidden_size=2304 at the 64-token recurrent chunk boundary; this is a future vectorized-prefill target because the current graph replays T=1." }, "vars": { "rows": 64, "hidden": 2304 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [64, 2304], "dtype": "float32", "dist": "normal", "seed": 1211, "scale": 0.2 }, "skipT": { "shape": [64, 2304], "dtype": "float32", "dist": "normal", "seed": 1212, "scale": 0.2 }, "gammaT": { "shape": [2304], "dtype": "float32", "dist": "uniform", "seed": 1213, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [64, 2304], "dtype": "float32" }, "residualT": { "shape": [64, 2304], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "kimi-linear-prefill-f32-512x2304", "preset": "model", "provenance": { "notes": "Original Kimi Linear hidden_size=2304 representative future vectorized-prefill shape; the current graph replays T=1." }, "vars": { "rows": 512, "hidden": 2304 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [512, 2304], "dtype": "float32", "dist": "normal", "seed": 1221, "scale": 0.2 }, "skipT": { "shape": [512, 2304], "dtype": "float32", "dist": "normal", "seed": 1222, "scale": 0.2 }, "gammaT": { "shape": [2304], "dtype": "float32", "dist": "uniform", "seed": 1223, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [512, 2304], "dtype": "float32" }, "residualT": { "shape": [512, 2304], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "kimi-linear-prefill-f32-1024x2304", "preset": "model", "provenance": { "notes": "Original Kimi Linear hidden_size=2304 representative future vectorized-prefill shape; the current graph replays T=1." }, "vars": { "rows": 1024, "hidden": 2304 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [1024, 2304], "dtype": "float32", "dist": "normal", "seed": 1231, "scale": 0.2 }, "skipT": { "shape": [1024, 2304], "dtype": "float32", "dist": "normal", "seed": 1232, "scale": 0.2 }, "gammaT": { "shape": [2304], "dtype": "float32", "dist": "uniform", "seed": 1233, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [1024, 2304], "dtype": "float32" }, "residualT": { "shape": [1024, 2304], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-bias-256x128", "preset": "smoke", "vars": { "rows": 256, "hidden": 128 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 216, "scale": 0.2 }, "skipT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 217, "scale": 0.2 }, "gammaT": { "shape": [128], "dtype": "float32", "dist": "uniform", "seed": 218, "scale": 0.1, "offset": 1 }, "biasT": { "shape": [128], "dtype": "float32", "dist": "normal", "seed": 219, "scale": 0.05 } }, "outputs": { "outputT": { "shape": [256, 128], "dtype": "float32" }, "residualT": { "shape": [256, 128], "dtype": "float32" } }, "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + 2 * args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-output-256x128", "preset": "smoke", "vars": { "rows": 256, "hidden": 128 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 224, "scale": 0.2 }, "skipT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 225, "scale": 0.2 }, "gammaT": { "shape": [128], "dtype": "float32", "dist": "uniform", "seed": 226, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [256, 128], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (3 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-bias-output-256x128", "preset": "smoke", "vars": { "rows": 256, "hidden": 128 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 227, "scale": 0.2 }, "skipT": { "shape": [256, 128], "dtype": "float32", "dist": "normal", "seed": 228, "scale": 0.2 }, "gammaT": { "shape": [128], "dtype": "float32", "dist": "uniform", "seed": 229, "scale": 0.1, "offset": 1 }, "biasT": { "shape": [128], "dtype": "float32", "dist": "normal", "seed": 230, "scale": 0.05 } }, "outputs": { "outputT": { "shape": [256, 128], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (3 * args.rows * args.hidden + 2 * args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-4096x768", "preset": "model", "vars": { "rows": 4096, "hidden": 768 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 319, "scale": 0.2 }, "skipT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 320, "scale": 0.2 }, "gammaT": { "shape": [768], "dtype": "float32", "dist": "uniform", "seed": 321, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [4096, 768], "dtype": "float32" }, "residualT": { "shape": [4096, 768], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-8192x4096", "preset": "model", "vars": { "rows": 8192, "hidden": 4096 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [8192, 4096], "dtype": "float32", "dist": "normal", "seed": 419, "scale": 0.2 }, "skipT": { "shape": [8192, 4096], "dtype": "float32", "dist": "normal", "seed": 420, "scale": 0.2 }, "gammaT": { "shape": [4096], "dtype": "float32", "dist": "uniform", "seed": 421, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [8192, 4096], "dtype": "float32" }, "residualT": { "shape": [8192, 4096], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-output-4096x768", "preset": "model", "vars": { "rows": 4096, "hidden": 768 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 324, "scale": 0.2 }, "skipT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 325, "scale": 0.2 }, "gammaT": { "shape": [768], "dtype": "float32", "dist": "uniform", "seed": 326, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [4096, 768], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (3 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-bias-4096x768", "preset": "model", "vars": { "rows": 4096, "hidden": 768 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 320, "scale": 0.2 }, "skipT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 321, "scale": 0.2 }, "gammaT": { "shape": [768], "dtype": "float32", "dist": "uniform", "seed": 322, "scale": 0.1, "offset": 1 }, "biasT": { "shape": [768], "dtype": "float32", "dist": "normal", "seed": 323, "scale": 0.05 } }, "outputs": { "outputT": { "shape": [4096, 768], "dtype": "float32" }, "residualT": { "shape": [4096, 768], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + 2 * args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-bias-output-4096x768", "preset": "model", "vars": { "rows": 4096, "hidden": 768 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 327, "scale": 0.2 }, "skipT": { "shape": [4096, 768], "dtype": "float32", "dist": "normal", "seed": 328, "scale": 0.2 }, "gammaT": { "shape": [768], "dtype": "float32", "dist": "uniform", "seed": 329, "scale": 0.1, "offset": 1 }, "biasT": { "shape": [768], "dtype": "float32", "dist": "normal", "seed": 330, "scale": 0.05 } }, "outputs": { "outputT": { "shape": [4096, 768], "dtype": "float32" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (3 * args.rows * args.hidden + 2 * args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-scalar-fallback-8192x4095", "preset": "stress", "vars": { "rows": 8192, "hidden": 4095 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [8192, 4095], "dtype": "float32", "dist": "normal", "seed": 621, "scale": 0.2 }, "skipT": { "shape": [8192, 4095], "dtype": "float32", "dist": "normal", "seed": 622, "scale": 0.2 }, "gammaT": { "shape": [4095], "dtype": "float32", "dist": "uniform", "seed": 623, "scale": 0.1, "offset": 1 } }, "outputs": { "outputT": { "shape": [8192, 4095], "dtype": "float32", "dist": "empty" }, "residualT": { "shape": [8192, 4095], "dtype": "float32", "dist": "empty" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + args.hidden)" }] } }, { "name": "skip-rmsnorm-f32-bias-scalar-fallback-8192x2049", "preset": "stress", "vars": { "rows": 8192, "hidden": 2049 }, "attrs": { "epsilon": 0.00001 }, "inputs": { "inputT": { "shape": [8192, 2049], "dtype": "float32", "dist": "normal", "seed": 631, "scale": 0.2 }, "skipT": { "shape": [8192, 2049], "dtype": "float32", "dist": "normal", "seed": 632, "scale": 0.2 }, "gammaT": { "shape": [2049], "dtype": "float32", "dist": "uniform", "seed": 633, "scale": 0.1, "offset": 1 }, "biasT": { "shape": [2049], "dtype": "float32", "dist": "normal", "seed": 634, "scale": 0.05 } }, "outputs": { "outputT": { "shape": [8192, 2049], "dtype": "float32", "dist": "empty" }, "residualT": { "shape": [8192, 2049], "dtype": "float32", "dist": "empty" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (4 * args.rows * args.hidden + 2 * args.hidden)" }] } } ] }