{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T128_D768", "num_tokens": 128, "hidden_dim": 768, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.024320999955307343, "p50": 0.025090999997701147, "p90": 0.02569000002949906, "mean": 0.026606800020090304, "iqr": 0.0010690000635804608, "raw_times": [0.03331100015202537, 0.025090999997701147, 0.024320999955307343, 0.02569000002949906, 0.0246209999659186], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03336100007800269, "peak_bytes": 1966080, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T128_D1024", "num_tokens": 128, "hidden_dim": 1024, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.028640999971685233, "p50": 0.02958999994007172, "p90": 0.030561000130546745, "mean": 0.02986059994327661, "iqr": 0.0012610003068402875, "raw_times": [0.029299999823706457, 0.028640999971685233, 0.02958999994007172, 0.0312109998503729, 0.030561000130546745], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03354099999341997, "peak_bytes": 2621440, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T128_D2048", "num_tokens": 128, "hidden_dim": 2048, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.02880000010918593, "p50": 0.030331000061778468, "p90": 0.030401000003621448, "mean": 0.030208600037440192, "iqr": 0.0004209998678561533, "raw_times": [0.02880000010918593, 0.03153099987684982, 0.029980000135765295, 0.030331000061778468, 0.030401000003621448], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03317000005154114, "peak_bytes": 5242880, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T256_D768", "num_tokens": 256, "hidden_dim": 768, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.02921000009337149, "p50": 0.0294310000299447, "p90": 0.029789999871354667, "mean": 0.029938399984530406, "iqr": 0.0004489997991186101, "raw_times": [0.0294310000299447, 0.02921000009337149, 0.03191999985574512, 0.029341000072236056, 0.029789999871354667], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03343000003042107, "peak_bytes": 3932160, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T256_D1024", "num_tokens": 256, "hidden_dim": 1024, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.029799999992974335, "p50": 0.031021000040709623, "p90": 0.031239999998433632, "mean": 0.03210639997632825, "iqr": 0.0009289999525208259, "raw_times": [0.038159999803610845, 0.031021000040709623, 0.029799999992974335, 0.030311000045912806, 0.031239999998433632], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03207100007784902, "peak_bytes": 5242880, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T256_D2048", "num_tokens": 256, "hidden_dim": 2048, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.0278800000614865, "p50": 0.028550999786602915, "p90": 0.029250000125102815, "mean": 0.02903839999817137, "iqr": 0.0010100000054080738, "raw_times": [0.0278800000614865, 0.02824000011969474, 0.028550999786602915, 0.029250000125102815, 0.03127099989796989], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03262000018366962, "peak_bytes": 10485760, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T512_D768", "num_tokens": 512, "hidden_dim": 768, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.02945000005638576, "p50": 0.029881000045861583, "p90": 0.03017099993485317, "mean": 0.03019639998456114, "iqr": 0.0005509998572961194, "raw_times": [0.029881000045861583, 0.03185999980814813, 0.03017099993485317, 0.02945000005638576, 0.02962000007755705], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.031610000178261544, "peak_bytes": 7864320, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T512_D1024", "num_tokens": 512, "hidden_dim": 1024, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.02807000009852345, "p50": 0.028989999918849207, "p90": 0.02929000015683414, "mean": 0.028920200020365883, "iqr": 0.0003590002961573191, "raw_times": [0.029320000066945795, 0.02929000015683414, 0.02807000009852345, 0.028989999918849207, 0.02893099986067682], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.033219999977518455, "peak_bytes": 10485760, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}
{"ts": "2025-10-27T14:46:29Z", "run": "cb61ba9d82ad40cba986f04f71dd51b6", "impl": "hf_kernels_swiglu", "tags": {"family": "hf-kernels", "backend": "cuda"}, "wl": {"name": "cuda_T512_D2048", "num_tokens": 512, "hidden_dim": 2048, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA L40S", "sm": "8.9", "py": "3.11.14", "plat": "Linux-5.10.244-240.970.amzn2.x86_64-x86_64-with-glibc2.35"}, "lat_ms": {"p10": 0.029301000040504732, "p50": 0.03090099994551565, "p90": 0.03149000008306757, "mean": 0.03127060003862425, "iqr": 0.0014889999420120148, "raw_times": [0.029301000040504732, 0.030001000141055556, 0.03149000008306757, 0.03465999998297775, 0.03090099994551565], "has_warnings": false, "reps": 5, "warmup": 2}, "compile_ms": 0.03197000000909611, "peak_bytes": 20971520, "ok": true, "absmax": 0.0, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0, "mae": 0.0, "mse": 0.0, "ref": "swiglu_fp32"}, "err": null}