| { |
| "op": "com.microsoft.GroupQueryAttention", |
| "tunableSpace": { |
| "PREFILL_QUERY_TILE": [16, 32], |
| "PREFILL_LANES_PER_QUERY": [4, 8], |
| "CLUSTER_TILE_K_F32": [8, 16, 32], |
| "NO_SG_TILE_K_MAX": [8, 16] |
| }, |
| "cases": [ |
| { |
| "name": "gqa-f32-1x64x64-h4kv2d16", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 64, "kvSeq": 64, "heads": 4, "kvHeads": 2, "headDim": 16 }, |
| "attrs": { "num_heads": 4, "kv_num_heads": 2, "scale": 0.25, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 64, 64], "dtype": "float32", "dist": "normal", "seed": 209, "scale": 0.2 }, |
| "keyT": { "shape": [1, 64, 32], "dtype": "float32", "dist": "normal", "seed": 210, "scale": 0.2 }, |
| "valueT": { "shape": [1, 64, 32], "dtype": "float32", "dist": "normal", "seed": 211, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [63] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [64] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 64, 64], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 2, 64, 16], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 2, 64, 16], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "qwen3-prefill-32h8kv-d128-s512", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 512, "kvSeq": 512, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 512, 4096], "dtype": "float32", "dist": "normal", "seed": 370, "scale": 0.2 }, |
| "keyT": { "shape": [1, 512, 1024], "dtype": "float32", "dist": "normal", "seed": 371, "scale": 0.2 }, |
| "valueT": { "shape": [1, 512, 1024], "dtype": "float32", "dist": "normal", "seed": 372, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 512, 4096], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 8, 512, 128], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 8, 512, 128], "dtype": "float32" } |
| }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "decode-h4kv1-d64-kv512", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 1, "kvSeq": 512, "heads": 4, "kvHeads": 1, "headDim": 64 }, |
| "attrs": { "num_heads": 4, "kv_num_heads": 1, "scale": 0.125, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 1, 256], "dtype": "float32", "dist": "normal", "seed": 376, "scale": 0.2 }, |
| "keyT": { "shape": [1, 512, 64], "dtype": "float32", "dist": "normal", "seed": 377, "scale": 0.2 }, |
| "valueT": { "shape": [1, 512, 64], "dtype": "float32", "dist": "normal", "seed": 378, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 1, 256], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 1, 512, 64], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 1, 512, 64], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(4 * args.kvSeq * args.kvHeads * args.headDim + 2 * args.qSeq * args.heads * args.headDim) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "llama-decode-32h8kv-d128-kv2048", |
| "preset": "model", |
| "vars": { "batch": 1, "qSeq": 1, "kvSeq": 2048, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 1, 4096], "dtype": "float32", "dist": "normal", "seed": 373, "scale": 0.2 }, |
| "keyT": { "shape": [1, 2048, 1024], "dtype": "float32", "dist": "normal", "seed": 374, "scale": 0.2 }, |
| "valueT": { "shape": [1, 2048, 1024], "dtype": "float32", "dist": "normal", "seed": 375, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 1, 4096], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 8, 2048, 128], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 8, 2048, 128], "dtype": "float32" } |
| }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(4 * args.kvSeq * args.kvHeads * args.headDim + 2 * args.qSeq * args.heads * args.headDim) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "llama-decode-32h8kv-d128-kv8192", |
| "preset": "stress", |
| "vars": { "batch": 1, "qSeq": 1, "kvSeq": 8192, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 1, 4096], "dtype": "float32", "dist": "normal", "seed": 373, "scale": 0.2 }, |
| "keyT": { "shape": [1, 8192, 1024], "dtype": "float32", "dist": "normal", "seed": 374, "scale": 0.2 }, |
| "valueT": { "shape": [1, 8192, 1024], "dtype": "float32", "dist": "normal", "seed": 375, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [8191] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [8192] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 1, 4096], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 8, 8192, 128], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 8, 8192, 128], "dtype": "float32" } |
| }, |
| "bench": { |
| "primary": false, |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(4 * args.kvSeq * args.kvHeads * args.headDim + 2 * args.qSeq * args.heads * args.headDim) * 4" |
| } |
| ] |
| }, |
| "provenance": { |
| "notes": "Long-context decode beyond the MAX_SPLITS boundary: ceilDiv(kvSeq, QKV_SPLIT_TILE_K)=64 exceeds the cap of 16. Paired with the headDim-32 control, this case checks the tradeoff between additional split parallelism and merge overhead without assuming that one flat cap is optimal for both geometries." |
| } |
| }, |
| { |
| "name": "qwen3-prefill-32h8kv-d128-s512-f16", |
| "preset": "model", |
| "vars": { "batch": 1, "qSeq": 512, "kvSeq": 512, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 512, 4096], "dtype": "float16", "dist": "normal", "seed": 370, "scale": 0.2 }, |
| "keyT": { "shape": [1, 512, 1024], "dtype": "float16", "dist": "normal", "seed": 371, "scale": 0.2 }, |
| "valueT": { "shape": [1, 512, 1024], "dtype": "float16", "dist": "normal", "seed": 372, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 512, 4096], "dtype": "float16" }, |
| "presentKeyT": { "shape": [1, 8, 512, 128], "dtype": "float16" }, |
| "presentValueT": { "shape": [1, 8, 512, 128], "dtype": "float16" } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "llama-decode-32h8kv-d128-kv2048-f16", |
| "preset": "model", |
| "vars": { "batch": 1, "qSeq": 1, "kvSeq": 2048, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 1, 4096], "dtype": "float16", "dist": "normal", "seed": 373, "scale": 0.2 }, |
| "keyT": { "shape": [1, 2048, 1024], "dtype": "float16", "dist": "normal", "seed": 374, "scale": 0.2 }, |
| "valueT": { "shape": [1, 2048, 1024], "dtype": "float16", "dist": "normal", "seed": 375, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 1, 4096], "dtype": "float16" }, |
| "presentKeyT": { "shape": [1, 8, 2048, 128], "dtype": "float16" }, |
| "presentValueT": { "shape": [1, 8, 2048, 128], "dtype": "float16" } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(4 * args.kvSeq * args.kvHeads * args.headDim + 2 * args.qSeq * args.heads * args.headDim) * 2" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-decode-32h8kv-d128-kv2048", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 2048, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 2048, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 1) * dim(shapes.pastKeyT, 2) * (dim(shapes.queryT, 2) / attrs.num_heads) * attrs.num_heads" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-32h8kv-d128-s64", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 64, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 64, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 64, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [63] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [64] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 64, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 64, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 64, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-32h8kv-d128-s128", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 128, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 128, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 128, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [127] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [128] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 128, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 128, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 128, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-32h8kv-d128-s256", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 256, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 256, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 256, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [255] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [256] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 256, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 256, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 256, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-32h8kv-d128-s512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-rotary-32h8kv-d128-s512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "do_rotary": 1 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "cosCacheT": { |
| "dtype": "float32", |
| "shape": [512, 64], |
| "data": { "kind": "fillFloat32", "scale": 0.5, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "sinCacheT": { |
| "dtype": "float32", |
| "shape": [512, 64], |
| "data": { "kind": "fillFloat32", "scale": 0.5, "sinStep": 0.013, "cosStep": 0.037 } |
| } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-decode-f16-32h8kv-d128-kv2048", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float16", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float16", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float16", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float16", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float16", "shape": [1, 8, 2048, 128] }, |
| "presentValueT": { "dtype": "float16", "shape": [1, 8, 2048, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 1) * dim(shapes.pastKeyT, 2) * (dim(shapes.queryT, 2) / attrs.num_heads) * attrs.num_heads" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-f16-32h8kv-d128-s512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float16", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float16", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float16", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float16", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float16", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float16", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-prefill-rotary-f16-32h8kv-d128-s512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "do_rotary": 1 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float16", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float16", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float16", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float16", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "cosCacheT": { |
| "dtype": "float16", |
| "shape": [512, 64], |
| "data": { "kind": "fillFloat32", "scale": 0.5, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "sinCacheT": { |
| "dtype": "float16", |
| "shape": [512, 64], |
| "data": { "kind": "fillFloat32", "scale": 0.5, "sinStep": 0.013, "cosStep": 0.037 } |
| } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float16", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float16", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float16", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "newkv-past-decode-q1p2048", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.013, "cosStep": 0.017 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.019, "cosStep": 0.023 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2049] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 2049, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 2049, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 1) * dim(shapes.presentKeyT, 2) * (dim(shapes.queryT, 2) / attrs.num_heads) * attrs.num_heads" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "newkv-past-prefill-q512p512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.013, "cosStep": 0.017 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.019, "cosStep": 0.023 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1023] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1024] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 1024, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 1024, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.presentKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "quant-int8-decode-h8kv2-d128-kv2048", |
| "preset": "smoke", |
| "provenance": { |
| "notes": "Buffer-sharing INT8 decode. The bandwidth metric counts the portable physical minimum: WebGPU exposes each logical INT8 cache element through i32 storage, the past cache is preserved into present K/V, and attention reads K/V once per KV head. It intentionally does not credit duplicate cache reads across grouped query heads." |
| }, |
| "attrs": { |
| "num_heads": 8, |
| "kv_num_heads": 2, |
| "kv_cache_bit_width": 8, |
| "k_quant_type": "PER_TENSOR", |
| "v_quant_type": "PER_TENSOR" |
| }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 256], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.013, "cosStep": 0.031 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 256], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.015, "cosStep": 0.033 } |
| }, |
| "pastKeyT": { "dtype": "int8", "shape": [1, 2, 2048, 128], "dist": "uniform", "seed": 880, "min": -8, "max": 8 }, |
| "pastValueT": { |
| "dtype": "int8", |
| "shape": [1, 2, 2048, 128], |
| "dist": "uniform", |
| "seed": 881, |
| "min": -8, |
| "max": 8 |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } }, |
| "kScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } }, |
| "vScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 1024] }, |
| "presentKeyT": { "dtype": "int8", "shape": [1, 2, 2048, 128] }, |
| "presentValueT": { "dtype": "int8", "shape": [1, 2, 2048, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "4 * (6 * numel(shapes.pastKeyT) + 4 * numel(shapes.keyT) + 2 * numel(shapes.queryT))" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "quant-int8-prefill-h32kv8-d128-s512-pathology", |
| "preset": "stress", |
| "provenance": { |
| "notes": "Production-sized prefill for the int8 KV-cache contract, retained as a performance witness for quantized-cache attention." |
| }, |
| "attrs": { |
| "num_heads": 32, |
| "kv_num_heads": 8, |
| "kv_cache_bit_width": 8, |
| "k_quant_type": "PER_TENSOR", |
| "v_quant_type": "PER_TENSOR" |
| }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.013, "cosStep": 0.031 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.015, "cosStep": 0.033 } |
| }, |
| "pastKeyT": { "dtype": "int8", "shape": [1, 8, 512, 128], "dist": "uniform", "seed": 882, "min": -8, "max": 8 }, |
| "pastValueT": { |
| "dtype": "int8", |
| "shape": [1, 8, 512, 128], |
| "dist": "uniform", |
| "seed": 883, |
| "min": -8, |
| "max": 8 |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "kScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } }, |
| "vScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "int8", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "int8", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "quant-int4-prefill-h32kv8-d128-s512-pathology", |
| "preset": "stress", |
| "provenance": { |
| "notes": "Production-sized prefill for the packed int4 KV-cache contract, retained as a performance witness for quantized-cache attention." |
| }, |
| "attrs": { |
| "num_heads": 32, |
| "kv_num_heads": 8, |
| "kv_cache_bit_width": 4, |
| "k_quant_type": "PER_TENSOR", |
| "v_quant_type": "PER_TENSOR" |
| }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.013, "cosStep": 0.031 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 512, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.015, "cosStep": 0.033 } |
| }, |
| "pastKeyT": { "dtype": "uint8", "shape": [1, 8, 512, 64], "dist": "uniform", "seed": 884, "min": 0, "max": 256 }, |
| "pastValueT": { |
| "dtype": "uint8", |
| "shape": [1, 8, 512, 64], |
| "dist": "uniform", |
| "seed": 885, |
| "min": 0, |
| "max": 256 |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "kScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } }, |
| "vScaleT": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.0625] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "uint8", "shape": [1, 8, 512, 64] }, |
| "presentValueT": { "dtype": "uint8", "shape": [1, 8, 512, 64] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "bias-prefill-cliff-h32kv8-d128-s512", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "attentionBiasT": { |
| "dtype": "float32", |
| "shape": [1, 32, 512, 512], |
| "data": { "kind": "fillFloat32", "scale": 0.05, "sinStep": 0.007, "cosStep": 0.009 } |
| } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "hd16-belowfloor-decode-h8kv2-kv2048", |
| "preset": "smoke", |
| "attrs": { "num_heads": 8, "kv_num_heads": 2 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 32], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 32], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 2, 2048, 16], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 2, 2048, 16], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 128] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 2, 2048, 16] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 2, 2048, 16] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * dim(shapes.pastKeyT, 1) * dim(shapes.pastKeyT, 2) * dim(shapes.pastKeyT, 3) * 4 + 2 * dim(shapes.queryT, 2) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sliding-window-decode-h8kv2-d128-kv2048-w256", |
| "preset": "smoke", |
| "attrs": { "num_heads": 8, "kv_num_heads": 2, "local_window_size": 256 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 256], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 256], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 2, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 2, 2048, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2047] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2048] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 1024] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 2, 2048, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 2, 2048, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * dim(shapes.pastKeyT, 1) * dim(shapes.pastKeyT, 2) * dim(shapes.pastKeyT, 3) * 4 + 2 * dim(shapes.queryT, 2) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "scalar-fallback-decode-hd32-h8kv8-kv4096", |
| "preset": "stress", |
| "vars": { "batch": 1, "qSeq": 1, "kvSeq": 4096, "heads": 8, "kvHeads": 8, "headDim": 32 }, |
| "attrs": { "num_heads": 8, "kv_num_heads": 8, "scale": 0.17677669529663687, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 1, 256], "dtype": "float32", "dist": "normal", "seed": 911, "scale": 0.2 }, |
| "keyT": { "shape": [1, 4096, 256], "dtype": "float32", "dist": "normal", "seed": 912, "scale": 0.2 }, |
| "valueT": { "shape": [1, 4096, 256], "dtype": "float32", "dist": "normal", "seed": 913, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [4095] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [4096] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 1, 256], "dtype": "float32", "dist": "empty" }, |
| "presentKeyT": { "shape": [1, 8, 4096, 32], "dtype": "float32", "dist": "empty" }, |
| "presentValueT": { "shape": [1, 8, 4096, 32], "dtype": "float32", "dist": "empty" } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(4 * dim(shapes.keyT, 1) * attrs.kv_num_heads * (dim(shapes.queryT, 2) / attrs.num_heads) + 2 * dim(shapes.queryT, 2)) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "scalar-fallback-prefill-hd32-h8kv8-s512", |
| "preset": "stress", |
| "vars": { "batch": 1, "qSeq": 512, "kvSeq": 512, "heads": 8, "kvHeads": 8, "headDim": 32 }, |
| "attrs": { "num_heads": 8, "kv_num_heads": 8, "scale": 0.17677669529663687, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 512, 256], "dtype": "float32", "dist": "normal", "seed": 901, "scale": 0.2 }, |
| "keyT": { "shape": [1, 512, 256], "dtype": "float32", "dist": "normal", "seed": 902, "scale": 0.2 }, |
| "valueT": { "shape": [1, 512, 256], "dtype": "float32", "dist": "normal", "seed": 903, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 512, 256], "dtype": "float32", "dist": "empty" }, |
| "presentKeyT": { "shape": [1, 8, 512, 32], "dtype": "float32", "dist": "empty" }, |
| "presentValueT": { "shape": [1, 8, 512, 32], "dtype": "float32", "dist": "empty" } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 0) * dim(shapes.queryT, 1) * dim(shapes.keyT, 1) * attrs.num_heads * (dim(shapes.queryT, 2) / attrs.num_heads)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "gqa-f32-q31-kv511-h8-kv2-d64-double-threshold-cliff", |
| "preset": "stress", |
| "vars": { "batch": 1, "qSeq": 31, "kvSeq": 511, "heads": 8, "kvHeads": 2, "headDim": 64 }, |
| "attrs": { "num_heads": 8, "kv_num_heads": 2, "scale": 0.125, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 31, 512], "dtype": "float32", "dist": "normal", "seed": 921, "scale": 0.2 }, |
| "keyT": { "shape": [1, 511, 128], "dtype": "float32", "dist": "normal", "seed": 922, "scale": 0.2 }, |
| "valueT": { "shape": [1, 511, 128], "dtype": "float32", "dist": "normal", "seed": 923, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [510] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 31, 512], "dtype": "float32", "dist": "empty" }, |
| "presentKeyT": { "shape": [1, 2, 511, 64], "dtype": "float32", "dist": "empty" }, |
| "presentValueT": { "shape": [1, 2, 511, 64], "dtype": "float32", "dist": "empty" } |
| }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "gqa-f32-q32-kv512-h8-kv2-d64-threshold-control", |
| "preset": "stress", |
| "vars": { "batch": 1, "qSeq": 32, "kvSeq": 512, "heads": 8, "kvHeads": 2, "headDim": 64 }, |
| "attrs": { "num_heads": 8, "kv_num_heads": 2, "scale": 0.125, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 32, 512], "dtype": "float32", "dist": "normal", "seed": 924, "scale": 0.2 }, |
| "keyT": { "shape": [1, 512, 128], "dtype": "float32", "dist": "normal", "seed": 925, "scale": 0.2 }, |
| "valueT": { "shape": [1, 512, 128], "dtype": "float32", "dist": "normal", "seed": 926, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 32, 512], "dtype": "float32", "dist": "empty" }, |
| "presentKeyT": { "shape": [1, 2, 512, 64], "dtype": "float32", "dist": "empty" }, |
| "presentValueT": { "shape": [1, 2, 512, 64], "dtype": "float32", "dist": "empty" } |
| }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "headsink-prefill-h32kv8-d128-s512-generic-pathology", |
| "preset": "stress", |
| "provenance": { |
| "notes": "Llama-sized GQA prefill with smooth-softmax head sinks, measuring the additional score transformation on a production attention shape." |
| }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "headSinkT": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 927, "scale": 0.2 } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "bias-headsink-prefill-h32kv8-d128-s512-generic-pathology", |
| "preset": "stress", |
| "provenance": { |
| "notes": "The valid additive-bias plus smooth-softmax head-sink combination has a distinct manifest variant but no flash-prefill specialization." |
| }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } }, |
| "attentionBiasT": { |
| "dtype": "float32", |
| "shape": [1, 32, 512, 512], |
| "data": { "kind": "fillFloat32", "scale": 0.05, "sinStep": 0.007, "cosStep": 0.009 } |
| }, |
| "headSinkT": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 928, "scale": 0.2 } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "softcap-prefill-h32kv8-d128-s512-generic-pathology", |
| "preset": "stress", |
| "provenance": { |
| "notes": "Production prefill with model-used score soft-capping, measuring the capped-score semantic on the past-KV contract." |
| }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "softcap": 30 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 512, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "valueT": { "dtype": "float32", "shape": [1, 0, 1024], "data": { "kind": "values", "values": [] } }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 512, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 512, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 512, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 512, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "gflops", |
| "value": "4 * dim(shapes.queryT, 2) * (dim(shapes.queryT, 1) * (dim(shapes.pastKeyT, 2) - dim(shapes.queryT, 1)) + dim(shapes.queryT, 1) * (dim(shapes.queryT, 1) + 1) / 2)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "qkv-prefill-f32-h4kv1-d64-s128", |
| "preset": "model", |
| "provenance": { |
| "notes": "Lower f32 register-topology boundary comparing the Q32-broadcast and subgroup-cluster routes." |
| }, |
| "vars": { "batch": 1, "qSeq": 128, "kvSeq": 128, "heads": 4, "kvHeads": 1, "headDim": 64 }, |
| "attrs": { "num_heads": 4, "kv_num_heads": 1, "scale": 0.125, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 128, 256], "dtype": "float32", "dist": "normal", "seed": 1075, "scale": 0.2 }, |
| "keyT": { "shape": [1, 128, 64], "dtype": "float32", "dist": "normal", "seed": 1076, "scale": 0.2 }, |
| "valueT": { "shape": [1, 128, 64], "dtype": "float32", "dist": "normal", "seed": 1077, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [127] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [128] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 128, 256], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 1, 128, 64], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 1, 128, 64], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "qkv-prefill-f32-h4kv1-d256-s128", |
| "preset": "model", |
| "provenance": { |
| "notes": "Upper f32 register-topology boundary comparing Q32-broadcast, subgroup-cluster, and generic online routes." |
| }, |
| "vars": { "batch": 1, "qSeq": 128, "kvSeq": 128, "heads": 4, "kvHeads": 1, "headDim": 256 }, |
| "attrs": { "num_heads": 4, "kv_num_heads": 1, "scale": 0.0625, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 128, 1024], "dtype": "float32", "dist": "normal", "seed": 1081, "scale": 0.2 }, |
| "keyT": { "shape": [1, 128, 256], "dtype": "float32", "dist": "normal", "seed": 1082, "scale": 0.2 }, |
| "valueT": { "shape": [1, 128, 256], "dtype": "float32", "dist": "normal", "seed": 1083, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [127] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [128] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 128, 1024], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 1, 128, 256], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 1, 128, 256], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "window-decode-32h8kv-d128-cap1024-w1024", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "sliding_window_cache": 1, "local_window_size": 1024 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65535] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65536] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 1024, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 1024, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * 2 * numel(shapes.pastKeyT) * 4 + numel(shapes.queryT) * 4 + numel(shapes.outputT) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "window-decode-32h8kv-d128-cap1024-unfilled", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "sliding_window_cache": 1, "local_window_size": 1024 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [511] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [512] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 1024, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 1024, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * 2 * numel(shapes.pastKeyT) * 4 + numel(shapes.queryT) * 4 + numel(shapes.outputT) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "window-decode-32h8kv-d128-cap256-w256", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "sliding_window_cache": 1, "local_window_size": 256 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 256, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 256, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65535] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65536] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 256, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 256, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * 2 * numel(shapes.pastKeyT) * 4 + numel(shapes.queryT) * 4 + numel(shapes.outputT) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "window-chunk-prefill-32h8kv-d128-cap1024-q128", |
| "preset": "smoke", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "sliding_window_cache": 1, "local_window_size": 1024 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 128, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 128, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 128, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65535] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65536] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 128, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 1024, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 1024, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" }, |
| { |
| "type": "bandwidth", |
| "name": "cache traffic", |
| "value": "2 * 2 * numel(shapes.pastKeyT) * 4 + numel(shapes.queryT) * 4 + numel(shapes.outputT) * 4" |
| } |
| ] |
| }, |
| "vars": { "batch": 1, "qSeq": 128, "kvSeq": 1024, "heads": 32, "headDim": 128 } |
| }, |
| { |
| "name": "window-decode-32h8kv-d128-cap4096-w4096", |
| "preset": "model", |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "sliding_window_cache": 1, "local_window_size": 4096 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 1, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 1, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 4096, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 4096, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65535] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [65536] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 1, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 4096, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 4096, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "2 * 2 * numel(shapes.pastKeyT) * 4 + numel(shapes.queryT) * 4 + numel(shapes.outputT) * 4" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-chunk-append-32h8kv-d128-cap1024-q128", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 128, "kvSeq": 1024, "heads": 32, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 128, 4096], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 128, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 128, 1024], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 8, 1024, 128], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1023] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1024] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 128, 4096] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 8, 1024, 128] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 8, 1024, 128] } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| }, |
| { |
| "name": "sharedkv-chunk-append-8h2kv-d256-cap1024-q128", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 128, "kvSeq": 1024, "heads": 8, "headDim": 256 }, |
| "attrs": { "num_heads": 8, "kv_num_heads": 2 }, |
| "inputs": { |
| "queryT": { |
| "dtype": "float32", |
| "shape": [1, 128, 2048], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.011, "cosStep": 0.029 } |
| }, |
| "keyT": { |
| "dtype": "float32", |
| "shape": [1, 128, 512], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.031, "cosStep": 0.007 } |
| }, |
| "valueT": { |
| "dtype": "float32", |
| "shape": [1, 128, 512], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.037, "cosStep": 0.005 } |
| }, |
| "pastKeyT": { |
| "dtype": "float32", |
| "shape": [1, 2, 1024, 256], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.017, "cosStep": 0.019 } |
| }, |
| "pastValueT": { |
| "dtype": "float32", |
| "shape": [1, 2, 1024, 256], |
| "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.023, "cosStep": 0.013 } |
| }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1023] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1024] } } |
| }, |
| "outputs": { |
| "outputT": { "dtype": "float32", "shape": [1, 128, 2048] }, |
| "presentKeyT": { "dtype": "float32", "shape": [1, 2, 1024, 256] }, |
| "presentValueT": { "dtype": "float32", "shape": [1, 2, 1024, 256] } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| }, |
| "provenance": { |
| "source": "register-geometry gate asymmetry", |
| "notes": "Cached f32 prefill at headDim 256, above the shared-memory cluster's register-geometry boundary. This guards consistent route admission between no-past and cached prefill families." |
| } |
| }, |
| { |
| "name": "qwen3-prefill-32h8kv-d128-s256", |
| "preset": "smoke", |
| "vars": { "batch": 1, "qSeq": 256, "kvSeq": 256, "heads": 32, "kvHeads": 8, "headDim": 128 }, |
| "attrs": { "num_heads": 32, "kv_num_heads": 8, "scale": 0.08838834764831845, "causal": 0 }, |
| "inputs": { |
| "queryT": { "shape": [1, 256, 4096], "dtype": "float32", "dist": "normal", "seed": 960, "scale": 0.2 }, |
| "keyT": { "shape": [1, 256, 1024], "dtype": "float32", "dist": "normal", "seed": 961, "scale": 0.2 }, |
| "valueT": { "shape": [1, 256, 1024], "dtype": "float32", "dist": "normal", "seed": 962, "scale": 0.2 }, |
| "seqlensKT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [255] } }, |
| "totalSequenceLengthT": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [256] } } |
| }, |
| "outputs": { |
| "outputT": { "shape": [1, 256, 4096], "dtype": "float32" }, |
| "presentKeyT": { "shape": [1, 8, 256, 128], "dtype": "float32" }, |
| "presentValueT": { "shape": [1, 8, 256, 128], "dtype": "float32" } |
| }, |
| "bench": { |
| "metrics": [ |
| { "type": "gflops", "value": "4 * args.batch * args.qSeq * args.kvSeq * args.heads * args.headDim" } |
| ] |
| } |
| } |
| ] |
| } |
|
|
