| { |
| "tunableSpace": { "WORKGROUP_SIZE": [32, 64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "gate-smoke-decode-h64", |
| "preset": "smoke", |
| "provenance": { |
| "notes": "One token: the decode shape, where the whole op is a single tiny dispatch and launch overhead dominates." |
| }, |
| "inputs": { |
| "aT": { |
| "dtype": "float32", |
| "shape": [1, 1, 64], |
| "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 2.0 } |
| }, |
| "dtBiasT": { "dtype": "float32", "shape": [64], "data": { "kind": "linspace", "start": -1.5, "end": 1.0 } }, |
| "decayScaleT": { "dtype": "float32", "shape": [64], "data": { "kind": "linspace", "start": -3.0, "end": -0.2 } }, |
| "bT": { |
| "dtype": "float32", |
| "shape": [1, 1, 64], |
| "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.23, "scale": 2.5 } |
| } |
| }, |
| "outputs": { |
| "decayT": { "dtype": "float32", "shape": [1, 1, 64] }, |
| "betaT": { "dtype": "float32", "shape": [1, 1, 64] } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-smoke-prefill-t512-h32", |
| "preset": "smoke", |
| "provenance": { "notes": "Prefill over 512 tokens on the vectorized path." }, |
| "inputs": { |
| "aT": { |
| "dtype": "float32", |
| "shape": [1, 512, 32], |
| "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 2.0 } |
| }, |
| "dtBiasT": { "dtype": "float32", "shape": [32], "data": { "kind": "linspace", "start": -1.5, "end": 1.0 } }, |
| "decayScaleT": { "dtype": "float32", "shape": [32], "data": { "kind": "linspace", "start": -3.0, "end": -0.2 } }, |
| "bT": { |
| "dtype": "float32", |
| "shape": [1, 512, 32], |
| "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.23, "scale": 2.5 } |
| } |
| }, |
| "outputs": { |
| "decayT": { "dtype": "float32", "shape": [1, 512, 32] }, |
| "betaT": { "dtype": "float32", "shape": [1, 512, 32] } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-smoke-prefill-t256-h30", |
| "preset": "smoke", |
| "provenance": { "notes": "Head count 30 is not a multiple of four, so this is the scalar path at prefill width." }, |
| "inputs": { |
| "aT": { |
| "dtype": "float32", |
| "shape": [1, 256, 30], |
| "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 2.0 } |
| }, |
| "dtBiasT": { "dtype": "float32", "shape": [30], "data": { "kind": "linspace", "start": -1.5, "end": 1.0 } }, |
| "decayScaleT": { "dtype": "float32", "shape": [30], "data": { "kind": "linspace", "start": -3.0, "end": -0.2 } }, |
| "bT": { |
| "dtype": "float32", |
| "shape": [1, 256, 30], |
| "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.23, "scale": 2.5 } |
| } |
| }, |
| "outputs": { |
| "decayT": { "dtype": "float32", "shape": [1, 256, 30] }, |
| "betaT": { "dtype": "float32", "shape": [1, 256, 30] } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-smoke-decay-only-b2-t128-h64", |
| "preset": "smoke", |
| "provenance": { |
| "notes": "No b input: two of the four tensors disappear, which is the bandwidth the optional-pair binding actually saves." |
| }, |
| "inputs": { |
| "aT": { |
| "dtype": "float32", |
| "shape": [2, 128, 64], |
| "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 2.0 } |
| }, |
| "dtBiasT": { "dtype": "float32", "shape": [64], "data": { "kind": "linspace", "start": -1.5, "end": 1.0 } }, |
| "decayScaleT": { "dtype": "float32", "shape": [64], "data": { "kind": "linspace", "start": -3.0, "end": -0.2 } } |
| }, |
| "outputs": { "decayT": { "dtype": "float32", "shape": [2, 128, 64] } }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (2 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-qwen3next-decode-s1-h32", |
| "preset": "model", |
| "provenance": { |
| "notes": "Qwen3-Next class defaults (linear_num_value_heads 32): the decay/beta gate for one decode step." |
| }, |
| "vars": { "batch": 1, "seq": 1, "heads": 32 }, |
| "inputs": { |
| "aT": { "shape": [1, 1, 32], "dtype": "float32", "dist": "normal", "seed": 8400, "scale": 0.5 }, |
| "dtBiasT": { "shape": [32], "dtype": "float32", "dist": "normal", "seed": 8401, "scale": 0.2 }, |
| "decayScaleT": { "shape": [32], "dtype": "float32", "dist": "uniform", "seed": 8402, "min": 0.5, "max": 1.5 }, |
| "bT": { "shape": [1, 1, 32], "dtype": "float32", "dist": "normal", "seed": 8403, "scale": 0.5 } |
| }, |
| "outputs": { |
| "decayT": { "shape": [1, 1, 32], "dtype": "float32" }, |
| "betaT": { "shape": [1, 1, 32], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-qwen3next-prefill-s2048-h32", |
| "preset": "model", |
| "provenance": { "notes": "Qwen3-Next class defaults over a 2048-token prefill chunk." }, |
| "vars": { "batch": 1, "seq": 2048, "heads": 32 }, |
| "inputs": { |
| "aT": { "shape": [1, 2048, 32], "dtype": "float32", "dist": "normal", "seed": 8500, "scale": 0.5 }, |
| "dtBiasT": { "shape": [32], "dtype": "float32", "dist": "normal", "seed": 8501, "scale": 0.2 }, |
| "decayScaleT": { "shape": [32], "dtype": "float32", "dist": "uniform", "seed": 8502, "min": 0.5, "max": 1.5 }, |
| "bT": { "shape": [1, 2048, 32], "dtype": "float32", "dist": "normal", "seed": 8503, "scale": 0.5 } |
| }, |
| "outputs": { |
| "decayT": { "shape": [1, 2048, 32], "dtype": "float32" }, |
| "betaT": { "shape": [1, 2048, 32], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| }, |
| { |
| "name": "gate-qwen3next-prefill-s8192-h32", |
| "preset": "model", |
| "provenance": { |
| "notes": "Qwen3-Next class defaults over an 8192-token prefill chunk, where the gate is purely bandwidth-bound." |
| }, |
| "vars": { "batch": 1, "seq": 8192, "heads": 32 }, |
| "inputs": { |
| "aT": { "shape": [1, 8192, 32], "dtype": "float32", "dist": "normal", "seed": 8600, "scale": 0.5 }, |
| "dtBiasT": { "shape": [32], "dtype": "float32", "dist": "normal", "seed": 8601, "scale": 0.2 }, |
| "decayScaleT": { "shape": [32], "dtype": "float32", "dist": "uniform", "seed": 8602, "min": 0.5, "max": 1.5 }, |
| "bT": { "shape": [1, 8192, 32], "dtype": "float32", "dist": "normal", "seed": 8603, "scale": 0.5 } |
| }, |
| "outputs": { |
| "decayT": { "shape": [1, 8192, 32], "dtype": "float32" }, |
| "betaT": { "shape": [1, 8192, 32], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "4 * (4 * numel(shapes.aT) + 2 * numel(shapes.dtBiasT))" }] |
| } |
| } |
| ] |
| } |
|
|
