Upload folder using huggingface_hub

flash_attn/impls/artifacts/benchmark/attn.jsonl

@@ -1,6 +1,6 @@
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.4474239945411682, "p50": 0.44921600818634033, "p90": 0.45241600275039673, "mean": 0.45066879987716674, "reps": 5, "warmup": 2}, "compile_ms": 1.7530560493469238, "peak_bytes": 87425024, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.4652479887008667, "p50": 0.4705919921398163, "p90": 0.4716799855232239, "mean": 0.47004159688949587, "reps": 5, "warmup": 2}, "compile_ms": 0.36032000184059143, "peak_bytes": 95027200, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5987840294837952, "p50": 0.6021760106086731, "p90": 0.6045759916305542, "mean": 0.6022783994674683, "reps": 5, "warmup": 2}, "compile_ms": 0.4950079917907715, "peak_bytes": 99680256, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.905726432800293e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6000319719314575, "p50": 0.600383996963501, "p90": 0.6016640067100525, "mean": 0.6013055920600892, "reps": 5, "warmup": 2}, "compile_ms": 0.49647998809814453, "peak_bytes": 104726528, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003604888916015625, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.639136016368866, "p50": 0.6404479742050171, "p90": 0.6416320204734802, "mean": 0.6408192038536071, "reps": 5, "warmup": 2}, "compile_ms": 0.530239999294281, "peak_bytes": 108855296, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003566741943359375, "mse": 2.86102294921875e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:12:20Z", "run": "a2d84c2da2864a0ead6e0da36e5784e9", "impl": "xformers_meff", "tags": {"family": "xformers", "backend": "memory_efficient", "compile": "none"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6447359919548035, "p50": 0.6462399959564209, "p90": 0.6483839750289917, "mean": 0.6466111898422241, "reps": 5, "warmup": 2}, "compile_ms": 0.5342720150947571, "peak_bytes": 114425856, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8759241104125977e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.3603839874267578, "p50": 0.361952006816864, "p90": 0.3624640107154846, "mean": 0.3619711995124817, "reps": 5, "warmup": 2}, "compile_ms": 1.5701119899749756, "peak_bytes": 87425024, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.3892799913883209, "p50": 0.3909760117530823, "p90": 0.3922559916973114, "mean": 0.3912447988986969, "reps": 5, "warmup": 2}, "compile_ms": 0.35811200737953186, "peak_bytes": 95027200, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5240640044212341, "p50": 0.5248960256576538, "p90": 0.5248960256576538, "mean": 0.5258048176765442, "reps": 5, "warmup": 2}, "compile_ms": 0.4891839921474457, "peak_bytes": 99680256, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.905726432800293e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5265600085258484, "p50": 0.5277760028839111, "p90": 0.5282559990882874, "mean": 0.5276032090187073, "reps": 5, "warmup": 2}, "compile_ms": 0.4968000054359436, "peak_bytes": 104726528, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003604888916015625, "mse": 2.8908252716064453e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5639039874076843, "p50": 0.5657920241355896, "p90": 0.5668479800224304, "mean": 0.5656383991241455, "reps": 5, "warmup": 2}, "compile_ms": 0.5312319993972778, "peak_bytes": 108855296, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003566741943359375, "mse": 2.86102294921875e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:59:35Z", "run": "8bc1bbc1e0504355abbb1f58e69828d3", "impl": "hf_kernels_flash_attn3", "tags": {"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5689600110054016, "p50": 0.5698239803314209, "p90": 0.5713919997215271, "mean": 0.5789952039718628, "reps": 5, "warmup": 2}, "compile_ms": 0.5350080132484436, "peak_bytes": 114425856, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00035858154296875, "mse": 2.8759241104125977e-06, "ref": "sdpa_math_fp32"}, "err": null}

flash_attn/impls/artifacts/benchmark_default/attn_default.jsonl

@@ -1,6 +1,6 @@
-{"ts": "2025-10-02T18:08:46Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5194560289382935, "p50": 0.5272960066795349, "p90": 0.5312960147857666, "mean": 0.527347207069397, "reps": 5, "warmup": 2}, "compile_ms": 3354.235107421875, "peak_bytes": 87425024, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.000339508056640625, "mse": 2.726912498474121e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:08:47Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5560640096664429, "p50": 0.5571519732475281, "p90": 0.5611839890480042, "mean": 0.5586367964744567, "reps": 5, "warmup": 2}, "compile_ms": 471.23529052734375, "peak_bytes": 95027200, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003414154052734375, "mse": 2.7418136596679688e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:08:47Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6841920018196106, "p50": 0.6860160231590271, "p90": 0.6869760155677795, "mean": 0.6860736012458801, "reps": 5, "warmup": 2}, "compile_ms": 468.1533508300781, "peak_bytes": 99876864, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:08:48Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7152000069618225, "p50": 0.7161920070648193, "p90": 0.7164160013198853, "mean": 0.7167360067367554, "reps": 5, "warmup": 2}, "compile_ms": 465.7891540527344, "peak_bytes": 104726528, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:08:48Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7400959730148315, "p50": 0.742143988609314, "p90": 0.7431039810180664, "mean": 0.7423295855522156, "reps": 5, "warmup": 2}, "compile_ms": 468.6272888183594, "peak_bytes": 108855296, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:08:49Z", "run": "3ea490632d4f4be2a19f477a48f12fc5", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7718080282211304, "p50": 0.7745919823646545, "p90": 0.7748159766197205, "mean": 0.7743871927261352, "reps": 5, "warmup": 2}, "compile_ms": 475.9334716796875, "peak_bytes": 114425856, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003452301025390625, "mse": 2.771615982055664e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:18Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5141760110855103, "p50": 0.5175679922103882, "p90": 0.5197759866714478, "mean": 0.5181439876556396, "reps": 5, "warmup": 2}, "compile_ms": 3084.621826171875, "peak_bytes": 87425024, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.000339508056640625, "mse": 2.726912498474121e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:19Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.5549119710922241, "p50": 0.5582720041275024, "p90": 0.5598080158233643, "mean": 0.5579584002494812, "reps": 5, "warmup": 2}, "compile_ms": 270.21795654296875, "peak_bytes": 95027200, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003414154052734375, "mse": 2.7418136596679688e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:19Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6853119730949402, "p50": 0.687391996383667, "p90": 0.6883519887924194, "mean": 0.6872959971427918, "reps": 5, "warmup": 2}, "compile_ms": 269.78741455078125, "peak_bytes": 99876864, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:19Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7128639817237854, "p50": 0.7160959839820862, "p90": 0.7167680263519287, "mean": 0.716153597831726, "reps": 5, "warmup": 2}, "compile_ms": 269.8607177734375, "peak_bytes": 104726528, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:19Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7386879920959473, "p50": 0.7400959730148315, "p90": 0.7415040135383606, "mean": 0.7418303966522217, "reps": 5, "warmup": 2}, "compile_ms": 269.20501708984375, "peak_bytes": 108855296, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:58:20Z", "run": "9ebc449a917f4f2196503654e5549239", "impl": "torch_flash_compiled_default", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "default"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7708160281181335, "p50": 0.7740799784660339, "p90": 0.7753919959068298, "mean": 0.7745471954345703, "reps": 5, "warmup": 2}, "compile_ms": 270.93829345703125, "peak_bytes": 114425856, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003452301025390625, "mse": 2.771615982055664e-06, "ref": "sdpa_math_fp32"}, "err": null}

flash_attn/impls/artifacts/benchmark_max_autotune/attn_max_autotune.jsonl

@@ -1,6 +1,6 @@
-{"ts": "2025-10-02T18:09:34Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6148160099983215, "p50": 0.6296960115432739, "p90": 0.6522240042686462, "mean": 0.6489088058471679, "reps": 5, "warmup": 2}, "compile_ms": 4649.109375, "peak_bytes": 70779904, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.000339508056640625, "mse": 2.726912498474121e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:09:35Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6615359783172607, "p50": 0.6821119785308838, "p90": 0.7128959894180298, "mean": 0.700761592388153, "reps": 5, "warmup": 2}, "compile_ms": 1487.6849365234375, "peak_bytes": 78644224, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003414154052734375, "mse": 2.7418136596679688e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:09:37Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7967039942741394, "p50": 0.8164799809455872, "p90": 0.8463680148124695, "mean": 0.834444797039032, "reps": 5, "warmup": 2}, "compile_ms": 1492.66748046875, "peak_bytes": 84280320, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:09:39Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.8432319760322571, "p50": 0.8498560190200806, "p90": 0.8750079870223999, "mean": 0.8709375977516174, "reps": 5, "warmup": 2}, "compile_ms": 1477.6558837890625, "peak_bytes": 86508544, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:09:41Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.8775359988212585, "p50": 0.9030719995498657, "p90": 0.903872013092041, "mean": 0.9069631934165955, "reps": 5, "warmup": 2}, "compile_ms": 1919.1016845703125, "peak_bytes": 90440704, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
-{"ts": "2025-10-02T18:09:43Z", "run": "02313c2372ed4884add9b1c644335af9", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.9145920276641846, "p50": 0.9164159893989563, "p90": 0.9357439875602722, "mean": 0.9371584057807922, "reps": 5, "warmup": 2}, "compile_ms": 1487.1219482421875, "peak_bytes": 94372864, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003452301025390625, "mse": 2.771615982055664e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:25Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L128", "batch": 1, "seq_len": 1152, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6144000291824341, "p50": 0.6245759725570679, "p90": 0.6483200192451477, "mean": 0.6468096017837525, "reps": 5, "warmup": 2}, "compile_ms": 4407.3388671875, "peak_bytes": 70779904, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.000339508056640625, "mse": 2.726912498474121e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:27Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L256", "batch": 1, "seq_len": 1280, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.6689280271530151, "p50": 0.6851199865341187, "p90": 0.7184960246086121, "mean": 0.7060160160064697, "reps": 5, "warmup": 2}, "compile_ms": 1686.2735595703125, "peak_bytes": 78644224, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003414154052734375, "mse": 2.7418136596679688e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:29Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L320", "batch": 1, "seq_len": 1344, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.7953600287437439, "p50": 0.8155840039253235, "p90": 0.8403519988059998, "mean": 0.8332608103752136, "reps": 5, "warmup": 2}, "compile_ms": 1462.938232421875, "peak_bytes": 84280320, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:31Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L384", "batch": 1, "seq_len": 1408, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.8470720052719116, "p50": 0.849727988243103, "p90": 0.8745279908180237, "mean": 0.8719295978546142, "reps": 5, "warmup": 2}, "compile_ms": 1689.3455810546875, "peak_bytes": 86508544, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:33Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L448", "batch": 1, "seq_len": 1472, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.8677120208740234, "p50": 0.8835520148277283, "p90": 0.9034240245819092, "mean": 0.9034304022789001, "reps": 5, "warmup": 2}, "compile_ms": 1693.035888671875, "peak_bytes": 90440704, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.00034332275390625, "mse": 2.7567148208618164e-06, "ref": "sdpa_math_fp32"}, "err": null}
+{"ts": "2025-10-02T19:57:34Z", "run": "edb73be653834cdf8524ee78b403db7f", "impl": "torch_flash_compiled_max_autotune", "tags": {"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"}, "wl": {"name": "flux_L512", "batch": 1, "seq_len": 1536, "heads": 24, "head_dim": 128, "dtype": "bfloat16", "device": "cuda", "seed": 0}, "env": {"torch": "2.8.0+cu128", "cuda": "12.8", "gpu": "NVIDIA A10G", "sm": "8.6", "py": "3.11.13", "plat": "Linux-6.12.40-64.114.amzn2023.x86_64-x86_64-with-glibc2.36"}, "lat_ms": {"p10": 0.9154239892959595, "p50": 0.9213759899139404, "p90": 0.9359679818153381, "mean": 0.9387519836425782, "reps": 5, "warmup": 2}, "compile_ms": 1689.36279296875, "peak_bytes": 94372864, "ok": true, "absmax": 0.0625, "corr": {"ok": true, "rtol": 0.02, "atol": 0.02, "absmax": 0.0625, "mae": 0.0003452301025390625, "mse": 2.771615982055664e-06, "ref": "sdpa_math_fp32"}, "err": null}

flash_attn/impls/cells/benchmark.py

@@ -4,7 +4,7 @@
 #     "numpy",
 #     "torch",
 #     "kernels-benchmark-tools",
-#     "xformers",
+#     "kernels",
 # ]
 #
 # [tool.uv.sources]
@@ -14,32 +14,35 @@ import torch
 import sys
 import os
 import kernels_benchmark_tools as kbt
-import xformers.ops as xops
+from kernels import get_kernel
 
+hf_kernels_flash_attn3 = get_kernel("kernels-community/flash-attn3")
 
-def xformers_attention(q, k, v):
-    """xFormers memory efficient attention"""
-    # xFormers expects [batch, seq_len, heads, head_dim]
-    return xops.memory_efficient_attention(q, k, v)
+
+def hf_flash_attention3(query, key, value):
+    return hf_kernels_flash_attn3.flash_attn_func(query, key, value, causal=False)[0]
 
 
 kbt.add(
-    "xformers_meff",
-    xformers_attention,
-    tags={"family": "xformers", "backend": "memory_efficient", "compile": "none"},
+    "hf_kernels_flash_attn3",
+    hf_flash_attention3,
+    tags={"family": "hf-kernels", "backend": "flash-attn3", "compile": "none"},
 )
 
 if __name__ == "__main__":
     device = "cuda" if torch.cuda.is_available() else "cpu"
-    dtype = "float32" if device == "cpu" else "bfloat16"
+
+    if device == "cpu":
+        print("HF Kernels Flash Attention 3 requires CUDA - skipping benchmark")
+        sys.exit(0)
+
+    dtype = "bfloat16"
 
     # Flux-like workloads
-    base = 1024 if device == "cuda" else 512
-    flux_sizes = (
-        [128, 256, 320, 384, 448, 512] if device == "cuda" else [64, 128, 192, 256]
-    )
-    heads = 24 if device == "cuda" else 8
-    head_dim = 128 if device == "cuda" else 64
+    base = 1024
+    flux_sizes = [128, 256, 320, 384, 448, 512]
+    heads = 24
+    head_dim = 128
 
     wl = []
     for L in flux_sizes:

flash_attn/impls/compiled_variants.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3746,7 +3760,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark_default')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark_default" onclick="toggleUvLogsFromHeader('benchmark_default')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark_default | 46.78s
+Cell: benchmark_default | 45.23s
  | <button class="run-btn" onclick="runCell('benchmark_default')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark_default')">Copy</button>
 <a href="cells/benchmark_default.py" target="_blank" class="raw-btn">Raw</a>
@@ -3830,7 +3844,7 @@ Cell: benchmark_default | 46.78s
 </div>
 <div id="output-benchmark_default" class="cell-output">
 <div class="cell-stdout">impl                     wl                  p50(ms)  ok
-torch_flash_compiled_default flux_L128              0.53  True
+torch_flash_compiled_default flux_L128              0.52  True
 torch_flash_compiled_default flux_L256              0.56  True
 torch_flash_compiled_default flux_L320              0.69  True
 torch_flash_compiled_default flux_L384              0.72  True
@@ -3841,41 +3855,41 @@ torch_flash_compiled_default flux_L512              0.77  True
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading matplotlib (8.3MiB)
+Downloading pillow (6.3MiB)
+Downloading numpy (16.2MiB)
 Downloading networkx (1.9MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading pillow (6.3MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading fonttools (4.7MiB)
+Downloading kiwisolver (1.4MiB)
 Downloading sympy (6.0MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading torch (846.9MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading matplotlib (8.3MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading torch (846.9MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading fonttools (4.7MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading kiwisolver (1.4MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading triton (148.3MiB)
-Downloading numpy (16.2MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
  Downloading kiwisolver
  Downloading setuptools
- Downloading networkx
  Downloading fonttools
+ Downloading networkx
  Downloading pillow
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cuda-cupti-cu12
  Downloading matplotlib
- Downloading sympy
  Downloading numpy
+ Downloading sympy
  Downloading nvidia-nvjitlink-cu12
  Downloading nvidia-curand-cu12
  Downloading nvidia-cuda-nvrtc-cu12
@@ -3888,7 +3902,7 @@ Downloading numpy (16.2MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 557ms
+Installed 37 packages in 551ms
 </div>
 </div>
 <div class="cell-artifacts">
@@ -3906,7 +3920,7 @@ Installed 37 packages in 557ms
 <span onclick="toggleOutput('benchmark_max_autotune')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark_max_autotune" onclick="toggleUvLogsFromHeader('benchmark_max_autotune')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark_max_autotune | 53.65s
+Cell: benchmark_max_autotune | 54.06s
  | <button class="run-btn" onclick="runCell('benchmark_max_autotune')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark_max_autotune')">Copy</button>
 <a href="cells/benchmark_max_autotune.py" target="_blank" class="raw-btn">Raw</a>
@@ -3990,39 +4004,39 @@ Cell: benchmark_max_autotune | 53.65s
 </div>
 <div id="output-benchmark_max_autotune" class="cell-output">
 <div class="cell-stdout">impl                     wl                  p50(ms)  ok
-torch_flash_compiled_max_autotune flux_L128              0.63  True
-torch_flash_compiled_max_autotune flux_L256              0.68  True
+torch_flash_compiled_max_autotune flux_L128              0.62  True
+torch_flash_compiled_max_autotune flux_L256              0.69  True
 torch_flash_compiled_max_autotune flux_L320              0.82  True
 torch_flash_compiled_max_autotune flux_L384              0.85  True
-torch_flash_compiled_max_autotune flux_L448              0.90  True
+torch_flash_compiled_max_autotune flux_L448              0.88  True
 torch_flash_compiled_max_autotune flux_L512              0.92  True
 </div>
 <div class="uv-install-logs" id="uv-logs-benchmark_max_autotune">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading sympy (6.0MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading networkx (1.9MiB)
 Downloading matplotlib (8.3MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading triton (148.3MiB)
-Downloading networkx (1.9MiB)
-Downloading fonttools (4.7MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
 Downloading torch (846.9MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading kiwisolver (1.4MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
 Downloading numpy (16.2MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading sympy (6.0MiB)
+Downloading kiwisolver (1.4MiB)
 Downloading pillow (6.3MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading fonttools (4.7MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
@@ -4042,13 +4056,13 @@ Downloading nvidia-curand-cu12 (60.7MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 525ms
+Installed 37 packages in 513ms
 </div>
 </div>
 <div class="cell-artifacts">

flash_attn/impls/flash_attention.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('nv')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-nv" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: nv | 0.70s
+Cell: nv | 0.67s
  | <button class="run-btn" onclick="runCell('nv')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('nv')">Copy</button>
 <a href="cells/nv.py" target="_blank" class="raw-btn">Raw</a>
@@ -3761,7 +3775,7 @@ Cell: nv | 0.70s
 </div>
 </div>
 <div id="output-nv" class="cell-output">
-<div class="cell-stdout">Thu Oct  2 18:06:49 2025       
+<div class="cell-stdout">Thu Oct  2 19:58:23 2025       
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 570.172.08             Driver Version: 570.172.08     CUDA Version: 12.8     |
 |-----------------------------------------+------------------------+----------------------+
@@ -3770,19 +3784,19 @@ Cell: nv | 0.70s
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA A10G                    On  |   00000000:00:1B.0 Off |                    0 |
-|  0%   26C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   37C    P0             92W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   1  NVIDIA A10G                    On  |   00000000:00:1C.0 Off |                    0 |
-|  0%   26C    P8             26W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   29C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   2  NVIDIA A10G                    On  |   00000000:00:1D.0 Off |                    0 |
-|  0%   26C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   29C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   3  NVIDIA A10G                    On  |   00000000:00:1E.0 Off |                    0 |
-|  0%   27C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   30C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 
@@ -3806,7 +3820,7 @@ Cell: nv | 0.70s
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 36.63s
+Cell: benchmark | 35.41s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3886,7 +3900,7 @@ Cell: benchmark | 36.63s
 </div>
 <div id="output-benchmark" class="cell-output">
 <div class="cell-stdout">impl                     wl                  p50(ms)  ok
-torch_flash_ma           flux_L128              0.48  True
+torch_flash_ma           flux_L128              0.49  True
 torch_flash_ma           flux_L256              0.52  True
 torch_flash_ma           flux_L320              0.65  True
 torch_flash_ma           flux_L384              0.68  True
@@ -3897,35 +3911,35 @@ torch_flash_ma           flux_L512              0.74  True
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading triton (148.3MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading matplotlib (8.3MiB)
+Downloading sympy (6.0MiB)
+Downloading fonttools (4.7MiB)
 Downloading networkx (1.9MiB)
-Downloading kiwisolver (1.4MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading sympy (6.0MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading torch (846.9MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading pillow (6.3MiB)
-Downloading numpy (16.2MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading setuptools (1.1MiB)
-Downloading matplotlib (8.3MiB)
-Downloading triton (148.3MiB)
-Downloading fonttools (4.7MiB)
-Downloading torch (846.9MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading kiwisolver (1.4MiB)
 Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading pillow (6.3MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
  Downloading kiwisolver
  Downloading setuptools
- Downloading networkx
  Downloading fonttools
+ Downloading networkx
  Downloading pillow
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cuda-cupti-cu12
@@ -3941,10 +3955,10 @@ Downloading nvidia-cudnn-cu12 (674.0MiB)
  Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
- Downloading nvidia-cudnn-cu12
  Downloading nvidia-cublas-cu12
+ Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 548ms
+Installed 37 packages in 491ms
 </div>
 </div>
 <div class="cell-artifacts">

flash_attn/impls/hf_kernels_flash_attn.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 39.43s
+Cell: benchmark | 38.65s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3831,40 +3845,40 @@ Cell: benchmark | 39.43s
 </div>
 <div id="output-benchmark" class="cell-output">
 <div class="cell-stdout">impl                     wl                  p50(ms)  ok
-hf_kernels_flash_attn    flux_L128              0.34  True
+hf_kernels_flash_attn    flux_L128              0.35  True
 hf_kernels_flash_attn    flux_L256              0.38  True
 hf_kernels_flash_attn    flux_L320              0.49  True
-hf_kernels_flash_attn    flux_L384              0.51  True
+hf_kernels_flash_attn    flux_L384              0.52  True
 hf_kernels_flash_attn    flux_L448              0.54  True
-hf_kernels_flash_attn    flux_L512              0.55  True
+hf_kernels_flash_attn    flux_L512              0.56  True
 </div>
 <div class="uv-install-logs" id="uv-logs-benchmark">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
+Downloading kiwisolver (1.4MiB)
 Downloading sympy (6.0MiB)
-Downloading matplotlib (8.3MiB)
+Downloading hf-xet (3.0MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
 Downloading networkx (1.9MiB)
-Downloading torch (846.9MiB)
-Downloading setuptools (1.1MiB)
+Downloading pillow (6.3MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading matplotlib (8.3MiB)
 Downloading triton (148.3MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading fonttools (4.7MiB)
 Downloading numpy (16.2MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading kiwisolver (1.4MiB)
-Downloading hf-xet (3.0MiB)
-Downloading pillow (6.3MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading torch (846.9MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading fonttools (4.7MiB)
+Downloading setuptools (1.1MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
@@ -3875,29 +3889,29 @@ Downloading nvidia-cufft-cu12 (184.2MiB)
  Downloading fonttools
  Downloading pillow
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
- Downloading matplotlib
  Downloading nvidia-cuda-cupti-cu12
- Downloading numpy
+ Downloading matplotlib
  Downloading sympy
+ Downloading numpy
  Downloading nvidia-nvjitlink-cu12
  Downloading nvidia-curand-cu12
  Downloading nvidia-cuda-nvrtc-cu12
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparse-cu12
  Downloading nvidia-cusparselt-cu12
+ Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 47 packages in 552ms
+Installed 47 packages in 527ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 20 files:   0%|          | 0/20 [00:00&lt;?, ?it/s]
-Fetching 20 files:   5%|▌         | 1/20 [00:00&lt;00:03,  5.41it/s]
-Fetching 20 files:  10%|█         | 2/20 [00:01&lt;00:16,  1.09it/s]
-Fetching 20 files: 100%|██████████| 20/20 [00:01&lt;00:00, 12.37it/s]</div>
+Fetching 20 files:   5%|▌         | 1/20 [00:00&lt;00:03,  5.70it/s]
+Fetching 20 files:  10%|█         | 2/20 [00:01&lt;00:13,  1.36it/s]
+Fetching 20 files: 100%|██████████| 20/20 [00:01&lt;00:00, 15.31it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/attn.jsonl" class="artifact" target="_blank">attn.jsonl</a>

flash_attn/impls/hf_kernels_flash_attn3.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 39.41s
+Cell: benchmark | 38.16s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3833,7 +3847,7 @@ Cell: benchmark | 39.41s
 hf_kernels_flash_attn3   flux_L128              0.36  True
 hf_kernels_flash_attn3   flux_L256              0.39  True
 hf_kernels_flash_attn3   flux_L320              0.52  True
-hf_kernels_flash_attn3   flux_L384              0.52  True
+hf_kernels_flash_attn3   flux_L384              0.53  True
 hf_kernels_flash_attn3   flux_L448              0.57  True
 hf_kernels_flash_attn3   flux_L512              0.57  True
 </div>
@@ -3841,62 +3855,62 @@ hf_kernels_flash_attn3   flux_L512              0.57  True
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading setuptools (1.1MiB)
+Downloading networkx (1.9MiB)
+Downloading kiwisolver (1.4MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading pillow (6.3MiB)
-Downloading numpy (16.2MiB)
+Downloading matplotlib (8.3MiB)
+Downloading setuptools (1.1MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading networkx (1.9MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading sympy (6.0MiB)
 Downloading hf-xet (3.0MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading torch (846.9MiB)
-Downloading triton (148.3MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading pillow (6.3MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading fonttools (4.7MiB)
-Downloading kiwisolver (1.4MiB)
-Downloading matplotlib (8.3MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading sympy (6.0MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading triton (148.3MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
  Downloading kiwisolver
  Downloading hf-xet
  Downloading setuptools
- Downloading networkx
  Downloading fonttools
+ Downloading networkx
  Downloading pillow
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cuda-cupti-cu12
  Downloading matplotlib
- Downloading sympy
  Downloading numpy
+ Downloading sympy
  Downloading nvidia-nvjitlink-cu12
  Downloading nvidia-curand-cu12
  Downloading nvidia-cuda-nvrtc-cu12
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparse-cu12
  Downloading nvidia-cusparselt-cu12
+ Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 47 packages in 529ms
+Installed 47 packages in 565ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 4 files:   0%|          | 0/4 [00:00&lt;?, ?it/s]
-Fetching 4 files:  25%|██▌       | 1/4 [00:00&lt;00:00,  6.35it/s]
-Fetching 4 files:  50%|█████     | 2/4 [00:01&lt;00:01,  1.09it/s]
-Fetching 4 files: 100%|██████████| 4/4 [00:01&lt;00:00,  2.48it/s]</div>
+Fetching 4 files:  25%|██▌       | 1/4 [00:00&lt;00:00,  5.17it/s]
+Fetching 4 files:  50%|█████     | 2/4 [00:01&lt;00:01,  1.22it/s]
+Fetching 4 files: 100%|██████████| 4/4 [00:01&lt;00:00,  2.76it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/attn.jsonl" class="artifact" target="_blank">attn.jsonl</a>

flash_attn/impls/index.html

@@ -2,22 +2,86 @@
 <html>
 <head>
   <meta charset='UTF-8'>
-  <title>Directory Index</title>
+  <meta name='viewport' content='width=device-width, initial-scale=1.0'>
+  <title>Index of /flash_attn/impls</title>
   <style>
-    body { font-family: monospace; margin: 20px; }
-    h1 { font-size: 1.5em; }
-    ul { list-style-type: none; padding-left: 20px; }
-    li { margin: 5px 0; }
-    .dir { font-weight: bold; }
-    .file { color: #0066cc; }
-    a { text-decoration: none; }
-    a:hover { text-decoration: underline; }
+    :root {
+      --bg-primary: #0a0a0a;
+      --bg-secondary: #121212;
+      --bg-tertiary: #181818;
+      --text-primary: #e0e0e0;
+      --text-secondary: #888888;
+      --text-link: #64b5f6;
+      --border-primary: #2a2a2a;
+    }
+    body {
+      font-family: system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial, sans-serif;
+      background: var(--bg-primary);
+      color: var(--text-primary);
+      margin: 0;
+      padding: 16px;
+      max-width: 900px;
+      margin: 0 auto;
+    }
+    .controls {
+      display: flex;
+      justify-content: flex-end;
+      margin-bottom: 1rem;
+    }
+    .back-button {
+      background: var(--bg-secondary);
+      border: 1px solid var(--border-primary);
+      padding: 8px 12px;
+      border-radius: 4px;
+      color: var(--text-secondary);
+      cursor: pointer;
+      font-size: 0.9rem;
+      text-decoration: none;
+      display: inline-block;
+    }
+    .back-button:hover {
+      color: var(--text-primary);
+      background: var(--bg-tertiary);
+    }
+    h1 {
+      font-size: 1.5em;
+      margin: 1rem 0;
+      color: var(--text-primary);
+      border-bottom: 1px solid var(--border-primary);
+      padding-bottom: 0.5rem;
+    }
+    ul {
+      list-style-type: none;
+      padding: 0;
+    }
+    li {
+      margin: 0;
+      border-bottom: 1px solid var(--border-primary);
+    }
+    li:last-child {
+      border-bottom: none;
+    }
+    a {
+      display: block;
+      padding: 0.75rem 0.5rem;
+      text-decoration: none;
+      color: var(--text-link);
+      transition: background 0.2s ease;
+    }
+    a:hover {
+      background: var(--bg-secondary);
+    }
+    .dir {
+      font-weight: 500;
+    }
   </style>
 </head>
 <body>
+  <div class='controls'>
+    <a href='../index.html' class='back-button'>← back</a>
+  </div>
   <h1>Index of /flash_attn/impls</h1>
   <ul>
-    <li><a href='../index.html' class='dir'>../</a></li>
     <li><a href='compiled_variants.html' class='file'>compiled_variants.html</a></li>
     <li><a href='flash_attention.html' class='file'>flash_attention.html</a></li>
     <li><a href='hf_kernels_flash_attn.html' class='file'>hf_kernels_flash_attn.html</a></li>

flash_attn/impls/mem_efficient_attention.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 36.09s
+Cell: benchmark | 36.80s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3838,28 +3852,28 @@ torch_mem_eff            flux_L512              0.95  True
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading numpy (16.2MiB)
 Downloading sympy (6.0MiB)
-Downloading setuptools (1.1MiB)
+Downloading networkx (1.9MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading kiwisolver (1.4MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading torch (846.9MiB)
-Downloading matplotlib (8.3MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading pillow (6.3MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading networkx (1.9MiB)
-Downloading numpy (16.2MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading triton (148.3MiB)
+Downloading setuptools (1.1MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading pillow (6.3MiB)
 Downloading fonttools (4.7MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading triton (148.3MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading torch (846.9MiB)
+Downloading matplotlib (8.3MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
@@ -3885,7 +3899,7 @@ Downloading triton (148.3MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 447ms
+Installed 37 packages in 448ms
 </div>
 </div>
 <div class="cell-artifacts">

flash_attn/impls/sage_attention.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 40.08s
+Cell: benchmark | 40.58s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3839,45 +3853,45 @@ Cell: benchmark | 40.08s
 <div id="output-benchmark" class="cell-output">
 <div class="cell-stdout">impl                     wl                  p50(ms)  ok
 sage_int8_fp16           flux_L128             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 sage_int8_fp16           flux_L256             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 sage_int8_fp16           flux_L320             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 sage_int8_fp16           flux_L384             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 sage_int8_fp16           flux_L448             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 sage_int8_fp16           flux_L512             FAIL  False
-  Error: module &#x27;sage_attention_ba12545b014364be&#x27; has no attribute &#x27;fwd&#x27;
+  Error: module &#x27;sage_attention_1863f4c92418f0f6&#x27; has no attribute &#x27;fwd&#x27;
 </div>
 <div class="uv-install-logs" id="uv-logs-benchmark">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading sympy (6.0MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading numpy (16.2MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading networkx (1.9MiB)
-Downloading hf-xet (3.0MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading pillow (6.3MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading triton (148.3MiB)
+Downloading pillow (6.3MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading hf-xet (3.0MiB)
+Downloading networkx (1.9MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading kiwisolver (1.4MiB)
+Downloading matplotlib (8.3MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading fonttools (4.7MiB)
 Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading matplotlib (8.3MiB)
+Downloading triton (148.3MiB)
+Downloading sympy (6.0MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading kiwisolver (1.4MiB)
-Downloading fonttools (4.7MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading torch (846.9MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
@@ -3888,8 +3902,8 @@ Downloading torch (846.9MiB)
  Downloading networkx
  Downloading pillow
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
- Downloading nvidia-cuda-cupti-cu12
  Downloading matplotlib
+ Downloading nvidia-cuda-cupti-cu12
  Downloading numpy
  Downloading sympy
  Downloading nvidia-nvjitlink-cu12
@@ -3898,20 +3912,19 @@ Downloading torch (846.9MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 48 packages in 531ms
+Installed 48 packages in 591ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 11 files:   0%|          | 0/11 [00:00&lt;?, ?it/s]
-Fetching 11 files:   9%|▉         | 1/11 [00:00&lt;00:02,  4.42it/s]
-Fetching 11 files:  27%|██▋       | 3/11 [00:00&lt;00:01,  5.95it/s]
-Fetching 11 files:  73%|███████▎  | 8/11 [00:00&lt;00:00, 12.27it/s]
-Fetching 11 files: 100%|██████████| 11/11 [00:00&lt;00:00, 14.37it/s]</div>
+Fetching 11 files:   9%|▉         | 1/11 [00:00&lt;00:01,  5.59it/s]
+Fetching 11 files:  73%|███████▎  | 8/11 [00:00&lt;00:00, 12.79it/s]
+Fetching 11 files: 100%|██████████| 11/11 [00:00&lt;00:00, 16.77it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/attn.jsonl" class="artifact" target="_blank">attn.jsonl</a>

flash_attn/impls/xformers.html

@@ -101,10 +101,12 @@
         :root[data-ui="monocolor"] a { color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button,
         :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle,
+        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-button:hover,
         :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
+        :root[data-ui="monocolor"] .reset-toggle:hover,
+        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
         :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
         :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
         :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
@@ -214,7 +216,8 @@
         /* Keep default control styling when widgets are enabled, even in minimal UI */
         :root[data-ui="none"][data-widgets="on"] .menu-button,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
+        :root[data-ui="none"][data-widgets="on"] .back-button {
             background: #f6f6f6;
             border: 1px solid #cccccc;
             color: #222222;
@@ -244,7 +247,8 @@
         :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
         :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
         :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover {
+        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
             background: #ededed;
             border-color: #bbbbbb;
             color: #000000;
@@ -302,7 +306,8 @@
         }
         
         .theme-toggle,
-        .reset-toggle {
+        .reset-toggle,
+        .back-button {
             background: var(--bg-secondary);
             border: 1px solid var(--border-primary);
             padding: 8px 12px;
@@ -313,9 +318,15 @@
             font-size: 0.9rem;
             user-select: none;
         }
-        
+
+        .back-button {
+            text-decoration: none;
+            display: inline-block;
+        }
+
         .theme-toggle:hover,
-        .reset-toggle:hover {
+        .reset-toggle:hover,
+        .back-button:hover {
             color: var(--text-primary);
             background: var(--bg-tertiary);
         }
@@ -3697,6 +3708,9 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <body>
     <div class="controls">
         <div class="controls-buttons">
+            
+            <a href="index.html" class="back-button">← back</a>
+            
             <div class="theme-toggle" onclick="toggleTheme()">light</div>
             <div class="reset-toggle" onclick="resetLayout()">reset</div>
             <div class="menu-button" onclick="toggleMenu()">
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 40.41s
+Cell: benchmark | 42.08s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -3838,38 +3852,38 @@ xformers_meff            flux_L512              0.65  True
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
    Updating https://github.com/drbh/kernels-benchmark-tools.git (main)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading pillow (6.3MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading networkx (1.9MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading numpy (16.2MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading networkx (1.9MiB)
-Downloading kiwisolver (1.4MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading torch (846.9MiB)
 Downloading matplotlib (8.3MiB)
-Downloading triton (148.3MiB)
-Downloading sympy (6.0MiB)
 Downloading fonttools (4.7MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading kiwisolver (1.4MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading xformers (111.8MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading pillow (6.3MiB)
+Downloading sympy (6.0MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading triton (148.3MiB)
     Updated https://github.com/drbh/kernels-benchmark-tools.git (f457279bca6573cd2fa54a74e67118f5e6b7a31c)
    Building kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
  Downloading nvidia-cufile-cu12
  Downloading kiwisolver
  Downloading setuptools
- Downloading fonttools
  Downloading networkx
- Downloading pillow
+ Downloading fonttools
       Built kernels-benchmark-tools @ git+https://github.com/drbh/kernels-benchmark-tools.git@f457279bca6573cd2fa54a74e67118f5e6b7a31c
+ Downloading pillow
  Downloading nvidia-cuda-cupti-cu12
  Downloading matplotlib
  Downloading sympy
@@ -3884,10 +3898,10 @@ Downloading xformers (111.8MiB)
  Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
- Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
+ Downloading nvidia-cublas-cu12
  Downloading torch
-Installed 38 packages in 452ms
+Installed 38 packages in 541ms
 </div>
 </div>
 <div class="cell-artifacts">

flash_attn/index.html

@@ -2,22 +2,86 @@
 <html>
 <head>
   <meta charset='UTF-8'>
-  <title>Directory Index</title>
+  <meta name='viewport' content='width=device-width, initial-scale=1.0'>
+  <title>Index of /flash_attn</title>
   <style>
-    body { font-family: monospace; margin: 20px; }
-    h1 { font-size: 1.5em; }
-    ul { list-style-type: none; padding-left: 20px; }
-    li { margin: 5px 0; }
-    .dir { font-weight: bold; }
-    .file { color: #0066cc; }
-    a { text-decoration: none; }
-    a:hover { text-decoration: underline; }
+    :root {
+      --bg-primary: #0a0a0a;
+      --bg-secondary: #121212;
+      --bg-tertiary: #181818;
+      --text-primary: #e0e0e0;
+      --text-secondary: #888888;
+      --text-link: #64b5f6;
+      --border-primary: #2a2a2a;
+    }
+    body {
+      font-family: system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial, sans-serif;
+      background: var(--bg-primary);
+      color: var(--text-primary);
+      margin: 0;
+      padding: 16px;
+      max-width: 900px;
+      margin: 0 auto;
+    }
+    .controls {
+      display: flex;
+      justify-content: flex-end;
+      margin-bottom: 1rem;
+    }
+    .back-button {
+      background: var(--bg-secondary);
+      border: 1px solid var(--border-primary);
+      padding: 8px 12px;
+      border-radius: 4px;
+      color: var(--text-secondary);
+      cursor: pointer;
+      font-size: 0.9rem;
+      text-decoration: none;
+      display: inline-block;
+    }
+    .back-button:hover {
+      color: var(--text-primary);
+      background: var(--bg-tertiary);
+    }
+    h1 {
+      font-size: 1.5em;
+      margin: 1rem 0;
+      color: var(--text-primary);
+      border-bottom: 1px solid var(--border-primary);
+      padding-bottom: 0.5rem;
+    }
+    ul {
+      list-style-type: none;
+      padding: 0;
+    }
+    li {
+      margin: 0;
+      border-bottom: 1px solid var(--border-primary);
+    }
+    li:last-child {
+      border-bottom: none;
+    }
+    a {
+      display: block;
+      padding: 0.75rem 0.5rem;
+      text-decoration: none;
+      color: var(--text-link);
+      transition: background 0.2s ease;
+    }
+    a:hover {
+      background: var(--bg-secondary);
+    }
+    .dir {
+      font-weight: 500;
+    }
   </style>
 </head>
 <body>
+  <div class='controls'>
+    <a href='../index.html' class='back-button'>← back</a>
+  </div>
   <h1>Index of /flash_attn</h1>
   <ul>
-    <li><a href='../index.html' class='dir'>../</a></li>
     <li><a href='impls/index.html' class='dir'>impls/</a></li>
     <li><a href='results/index.html' class='dir'>results/</a></li>
   </ul>

flash_attn/results/artifacts/combine/latency.csv

@@ -1,43 +1,43 @@
 Implementation,Impl ID,Workload,Batch,Seq Length,Heads,Head Dim,Dtype,Mean (ms),P10 (ms),P50 (ms),P90 (ms),Reps,Peak Mem (MB),Backend,Family
-Flash (PyTorch SDPA),torch_flash_ma,flux_L128,1,1152,24,128,bfloat16,0.48577280044555665,0.47836801409721375,0.4803520143032074,0.4827199876308441,5,83.38,FLASH,torch-sdpa
-Flash (PyTorch SDPA),torch_flash_ma,flux_L256,1,1280,24,128,bfloat16,0.5229184031486511,0.521727979183197,0.5228800177574158,0.5234559774398804,5,90.62,FLASH,torch-sdpa
-Flash (PyTorch SDPA),torch_flash_ma,flux_L320,1,1344,24,128,bfloat16,0.6515967845916748,0.6503999829292297,0.650879979133606,0.6513599753379822,5,95.06,FLASH,torch-sdpa
-Flash (PyTorch SDPA),torch_flash_ma,flux_L384,1,1408,24,128,bfloat16,0.6807615995407105,0.6797440052032471,0.6808639764785767,0.6815680265426636,5,99.88,FLASH,torch-sdpa
-Flash (PyTorch SDPA),torch_flash_ma,flux_L448,1,1472,24,128,bfloat16,0.711027193069458,0.7058879733085632,0.7121919989585876,0.7131519913673401,5,103.81,FLASH,torch-sdpa
-Flash (PyTorch SDPA),torch_flash_ma,flux_L512,1,1536,24,128,bfloat16,0.7391423940658569,0.7369279861450195,0.7383999824523926,0.7408959865570068,5,109.12,FLASH,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L128,1,1152,24,128,bfloat16,0.5875647902488709,0.5863680243492126,0.5874559879302979,0.5876479744911194,5,83.38,EFFICIENT,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L256,1,1280,24,128,bfloat16,0.653657603263855,0.6485440135002136,0.6537600159645081,0.656544029712677,5,90.62,EFFICIENT,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L320,1,1344,24,128,bfloat16,0.7784512042999268,0.774944007396698,0.778656005859375,0.7801600098609924,5,95.94,EFFICIENT,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L384,1,1408,24,128,bfloat16,0.7922943949699401,0.791263997554779,0.7924799919128418,0.7927039861679077,5,100.0,EFFICIENT,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L448,1,1472,24,128,bfloat16,0.848089587688446,0.8444799780845642,0.8470079898834229,0.8499199748039246,5,103.81,EFFICIENT,torch-sdpa
-MemEff (PyTorch SDPA),torch_mem_eff,flux_L512,1,1536,24,128,bfloat16,0.9523776054382325,0.95004802942276,0.9519039988517761,0.9541119933128357,5,109.12,EFFICIENT,torch-sdpa
-xFormers,xformers_meff,flux_L128,1,1152,24,128,bfloat16,0.45066879987716674,0.4474239945411682,0.44921600818634033,0.45241600275039673,5,83.38,memory_efficient,xformers
-xFormers,xformers_meff,flux_L256,1,1280,24,128,bfloat16,0.47004159688949587,0.4652479887008667,0.4705919921398163,0.4716799855232239,5,90.62,memory_efficient,xformers
-xFormers,xformers_meff,flux_L320,1,1344,24,128,bfloat16,0.6022783994674683,0.5987840294837952,0.6021760106086731,0.6045759916305542,5,95.06,memory_efficient,xformers
-xFormers,xformers_meff,flux_L384,1,1408,24,128,bfloat16,0.6013055920600892,0.6000319719314575,0.600383996963501,0.6016640067100525,5,99.88,memory_efficient,xformers
-xFormers,xformers_meff,flux_L448,1,1472,24,128,bfloat16,0.6408192038536071,0.639136016368866,0.6404479742050171,0.6416320204734802,5,103.81,memory_efficient,xformers
-xFormers,xformers_meff,flux_L512,1,1536,24,128,bfloat16,0.6466111898422241,0.6447359919548035,0.6462399959564209,0.6483839750289917,5,109.12,memory_efficient,xformers
-Compiled (default),torch_flash_compiled_default,flux_L128,1,1152,24,128,bfloat16,0.527347207069397,0.5194560289382935,0.5272960066795349,0.5312960147857666,5,83.38,FLASH,torch-sdpa
-Compiled (default),torch_flash_compiled_default,flux_L256,1,1280,24,128,bfloat16,0.5586367964744567,0.5560640096664429,0.5571519732475281,0.5611839890480042,5,90.62,FLASH,torch-sdpa
-Compiled (default),torch_flash_compiled_default,flux_L320,1,1344,24,128,bfloat16,0.6860736012458801,0.6841920018196106,0.6860160231590271,0.6869760155677795,5,95.25,FLASH,torch-sdpa
-Compiled (default),torch_flash_compiled_default,flux_L384,1,1408,24,128,bfloat16,0.7167360067367554,0.7152000069618225,0.7161920070648193,0.7164160013198853,5,99.88,FLASH,torch-sdpa
-Compiled (default),torch_flash_compiled_default,flux_L448,1,1472,24,128,bfloat16,0.7423295855522156,0.7400959730148315,0.742143988609314,0.7431039810180664,5,103.81,FLASH,torch-sdpa
-Compiled (default),torch_flash_compiled_default,flux_L512,1,1536,24,128,bfloat16,0.7743871927261352,0.7718080282211304,0.7745919823646545,0.7748159766197205,5,109.12,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L128,1,1152,24,128,bfloat16,0.6489088058471679,0.6148160099983215,0.6296960115432739,0.6522240042686462,5,67.5,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L256,1,1280,24,128,bfloat16,0.700761592388153,0.6615359783172607,0.6821119785308838,0.7128959894180298,5,75.0,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L320,1,1344,24,128,bfloat16,0.834444797039032,0.7967039942741394,0.8164799809455872,0.8463680148124695,5,80.38,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L384,1,1408,24,128,bfloat16,0.8709375977516174,0.8432319760322571,0.8498560190200806,0.8750079870223999,5,82.5,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L448,1,1472,24,128,bfloat16,0.9069631934165955,0.8775359988212585,0.9030719995498657,0.903872013092041,5,86.25,FLASH,torch-sdpa
-Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L512,1,1536,24,128,bfloat16,0.9371584057807922,0.9145920276641846,0.9164159893989563,0.9357439875602722,5,90.0,FLASH,torch-sdpa
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L128,1,1152,24,128,bfloat16,0.34446719884872434,0.3438720107078552,0.3445119857788086,0.34457600116729736,5,83.38,flash-attn,hf-kernels
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L256,1,1280,24,128,bfloat16,0.37571839094161985,0.37404799461364746,0.3763839900493622,0.3766399919986725,5,90.62,flash-attn,hf-kernels
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L320,1,1344,24,128,bfloat16,0.4945920050144196,0.4925439953804016,0.493120014667511,0.4938240051269531,5,95.06,flash-attn,hf-kernels
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L384,1,1408,24,128,bfloat16,0.5139647841453552,0.5123199820518494,0.5142719745635986,0.5147839784622192,5,99.88,flash-attn,hf-kernels
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L448,1,1472,24,128,bfloat16,0.5353855967521668,0.5339199900627136,0.5350080132484436,0.5352320075035095,5,103.81,flash-attn,hf-kernels
-HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L512,1,1536,24,128,bfloat16,0.5548800110816956,0.5538560152053833,0.5548800230026245,0.5553280115127563,5,109.12,flash-attn,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L128,1,1152,24,128,bfloat16,0.3617343962192535,0.36102399230003357,0.3616960048675537,0.36211198568344116,5,83.38,flash-attn3,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L256,1,1280,24,128,bfloat16,0.3907967984676361,0.3885439932346344,0.39056000113487244,0.3906239867210388,5,90.62,flash-attn3,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L320,1,1344,24,128,bfloat16,0.5228991985321045,0.521344006061554,0.5230720043182373,0.5232319831848145,5,95.06,flash-attn3,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L384,1,1408,24,128,bfloat16,0.5254656076431274,0.523904025554657,0.5249919891357422,0.526528000831604,5,99.88,flash-attn3,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L448,1,1472,24,128,bfloat16,0.5646592020988465,0.5627840161323547,0.565343976020813,0.565343976020813,5,103.81,flash-attn3,hf-kernels
-HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L512,1,1536,24,128,bfloat16,0.5698879957199097,0.567359983921051,0.5696640014648438,0.5698559880256653,5,109.12,flash-attn3,hf-kernels
+Flash (PyTorch SDPA),torch_flash_ma,flux_L128,1,1152,24,128,bfloat16,0.49411200881004336,0.48844799399375916,0.4936000108718872,0.4944640100002289,5,83.38,FLASH,torch-sdpa
+Flash (PyTorch SDPA),torch_flash_ma,flux_L256,1,1280,24,128,bfloat16,0.5234112024307251,0.5224320292472839,0.5235199928283691,0.5235840082168579,5,90.62,FLASH,torch-sdpa
+Flash (PyTorch SDPA),torch_flash_ma,flux_L320,1,1344,24,128,bfloat16,0.6527232170104981,0.6503040194511414,0.6524800062179565,0.6545600295066833,5,95.06,FLASH,torch-sdpa
+Flash (PyTorch SDPA),torch_flash_ma,flux_L384,1,1408,24,128,bfloat16,0.682803213596344,0.6805760264396667,0.6828799843788147,0.6832640171051025,5,99.88,FLASH,torch-sdpa
+Flash (PyTorch SDPA),torch_flash_ma,flux_L448,1,1472,24,128,bfloat16,0.7075456142425537,0.7057600021362305,0.7063360214233398,0.7070720195770264,5,103.81,FLASH,torch-sdpa
+Flash (PyTorch SDPA),torch_flash_ma,flux_L512,1,1536,24,128,bfloat16,0.7379711985588073,0.7368639707565308,0.7372480034828186,0.7391039729118347,5,109.12,FLASH,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L128,1,1152,24,128,bfloat16,0.5874239921569824,0.5861759781837463,0.5873280167579651,0.5877439975738525,5,83.38,EFFICIENT,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L256,1,1280,24,128,bfloat16,0.6502719998359681,0.6490240097045898,0.649183988571167,0.6517760157585144,5,90.62,EFFICIENT,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L320,1,1344,24,128,bfloat16,0.7812095880508423,0.7761600017547607,0.7803199887275696,0.7852799892425537,5,95.94,EFFICIENT,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L384,1,1408,24,128,bfloat16,0.7948480010032654,0.7911999821662903,0.7935360074043274,0.7948480248451233,5,100.0,EFFICIENT,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L448,1,1472,24,128,bfloat16,0.8463295936584473,0.8449919819831848,0.8459839820861816,0.8461120128631592,5,103.81,EFFICIENT,torch-sdpa
+MemEff (PyTorch SDPA),torch_mem_eff,flux_L512,1,1536,24,128,bfloat16,0.9538687944412232,0.9492800235748291,0.9518399834632874,0.9581760168075562,5,109.12,EFFICIENT,torch-sdpa
+xFormers,xformers_meff,flux_L128,1,1152,24,128,bfloat16,0.4515071928501129,0.44364801049232483,0.4524799883365631,0.4557119905948639,5,83.38,memory_efficient,xformers
+xFormers,xformers_meff,flux_L256,1,1280,24,128,bfloat16,0.46787199974060056,0.46489599347114563,0.4684160053730011,0.46908798813819885,5,90.62,memory_efficient,xformers
+xFormers,xformers_meff,flux_L320,1,1344,24,128,bfloat16,0.6001471996307373,0.596992015838623,0.5984640121459961,0.6016640067100525,5,95.06,memory_efficient,xformers
+xFormers,xformers_meff,flux_L384,1,1408,24,128,bfloat16,0.6023231983184815,0.5997440218925476,0.6031039953231812,0.6032639741897583,5,99.88,memory_efficient,xformers
+xFormers,xformers_meff,flux_L448,1,1472,24,128,bfloat16,0.6411136031150818,0.6381760239601135,0.6414719820022583,0.6421440243721008,5,103.81,memory_efficient,xformers
+xFormers,xformers_meff,flux_L512,1,1536,24,128,bfloat16,0.6594688057899475,0.6441280245780945,0.6496639847755432,0.6527680158615112,5,109.12,memory_efficient,xformers
+Compiled (default),torch_flash_compiled_default,flux_L128,1,1152,24,128,bfloat16,0.5181439876556396,0.5141760110855103,0.5175679922103882,0.5197759866714478,5,83.38,FLASH,torch-sdpa
+Compiled (default),torch_flash_compiled_default,flux_L256,1,1280,24,128,bfloat16,0.5579584002494812,0.5549119710922241,0.5582720041275024,0.5598080158233643,5,90.62,FLASH,torch-sdpa
+Compiled (default),torch_flash_compiled_default,flux_L320,1,1344,24,128,bfloat16,0.6872959971427918,0.6853119730949402,0.687391996383667,0.6883519887924194,5,95.25,FLASH,torch-sdpa
+Compiled (default),torch_flash_compiled_default,flux_L384,1,1408,24,128,bfloat16,0.716153597831726,0.7128639817237854,0.7160959839820862,0.7167680263519287,5,99.88,FLASH,torch-sdpa
+Compiled (default),torch_flash_compiled_default,flux_L448,1,1472,24,128,bfloat16,0.7418303966522217,0.7386879920959473,0.7400959730148315,0.7415040135383606,5,103.81,FLASH,torch-sdpa
+Compiled (default),torch_flash_compiled_default,flux_L512,1,1536,24,128,bfloat16,0.7745471954345703,0.7708160281181335,0.7740799784660339,0.7753919959068298,5,109.12,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L128,1,1152,24,128,bfloat16,0.6468096017837525,0.6144000291824341,0.6245759725570679,0.6483200192451477,5,67.5,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L256,1,1280,24,128,bfloat16,0.7060160160064697,0.6689280271530151,0.6851199865341187,0.7184960246086121,5,75.0,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L320,1,1344,24,128,bfloat16,0.8332608103752136,0.7953600287437439,0.8155840039253235,0.8403519988059998,5,80.38,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L384,1,1408,24,128,bfloat16,0.8719295978546142,0.8470720052719116,0.849727988243103,0.8745279908180237,5,82.5,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L448,1,1472,24,128,bfloat16,0.9034304022789001,0.8677120208740234,0.8835520148277283,0.9034240245819092,5,86.25,FLASH,torch-sdpa
+Compiled (max-autotune),torch_flash_compiled_max_autotune,flux_L512,1,1536,24,128,bfloat16,0.9387519836425782,0.9154239892959595,0.9213759899139404,0.9359679818153381,5,90.0,FLASH,torch-sdpa
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L128,1,1152,24,128,bfloat16,0.3455295979976654,0.34355199337005615,0.34563198685646057,0.34643200039863586,5,83.38,flash-attn,hf-kernels
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L256,1,1280,24,128,bfloat16,0.3756160080432892,0.37411201000213623,0.3752000033855438,0.3770880103111267,5,90.62,flash-attn,hf-kernels
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L320,1,1344,24,128,bfloat16,0.4953216016292572,0.49324798583984375,0.49433600902557373,0.49663999676704407,5,95.06,flash-attn,hf-kernels
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L384,1,1408,24,128,bfloat16,0.5157055854797363,0.5142719745635986,0.516319990158081,0.516543984413147,5,99.88,flash-attn,hf-kernels
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L448,1,1472,24,128,bfloat16,0.5356672048568726,0.5346879959106445,0.5358080267906189,0.5361599922180176,5,103.81,flash-attn,hf-kernels
+HF Kernels Flash Attn,hf_kernels_flash_attn,flux_L512,1,1536,24,128,bfloat16,0.5587136030197144,0.5557760000228882,0.5574079751968384,0.5581120252609253,5,109.12,flash-attn,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L128,1,1152,24,128,bfloat16,0.3619711995124817,0.3603839874267578,0.361952006816864,0.3624640107154846,5,83.38,flash-attn3,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L256,1,1280,24,128,bfloat16,0.3912447988986969,0.3892799913883209,0.3909760117530823,0.3922559916973114,5,90.62,flash-attn3,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L320,1,1344,24,128,bfloat16,0.5258048176765442,0.5240640044212341,0.5248960256576538,0.5248960256576538,5,95.06,flash-attn3,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L384,1,1408,24,128,bfloat16,0.5276032090187073,0.5265600085258484,0.5277760028839111,0.5282559990882874,5,99.88,flash-attn3,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L448,1,1472,24,128,bfloat16,0.5656383991241455,0.5639039874076843,0.5657920241355896,0.5668479800224304,5,103.81,flash-attn3,hf-kernels
+HF Kernels Flash Attn3,hf_kernels_flash_attn3,flux_L512,1,1536,24,128,bfloat16,0.5789952039718628,0.5689600110054016,0.5698239803314209,0.5713919997215271,5,109.12,flash-attn3,hf-kernels

flash_attn/results/index.html

@@ -2,22 +2,86 @@
 <html>
 <head>
   <meta charset='UTF-8'>
-  <title>Directory Index</title>
+  <meta name='viewport' content='width=device-width, initial-scale=1.0'>
+  <title>Index of /flash_attn/results</title>
   <style>
-    body { font-family: monospace; margin: 20px; }
-    h1 { font-size: 1.5em; }
-    ul { list-style-type: none; padding-left: 20px; }
-    li { margin: 5px 0; }
-    .dir { font-weight: bold; }
-    .file { color: #0066cc; }
-    a { text-decoration: none; }
-    a:hover { text-decoration: underline; }
+    :root {
+      --bg-primary: #0a0a0a;
+      --bg-secondary: #121212;
+      --bg-tertiary: #181818;
+      --text-primary: #e0e0e0;
+      --text-secondary: #888888;
+      --text-link: #64b5f6;
+      --border-primary: #2a2a2a;
+    }
+    body {
+      font-family: system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial, sans-serif;
+      background: var(--bg-primary);
+      color: var(--text-primary);
+      margin: 0;
+      padding: 16px;
+      max-width: 900px;
+      margin: 0 auto;
+    }
+    .controls {
+      display: flex;
+      justify-content: flex-end;
+      margin-bottom: 1rem;
+    }
+    .back-button {
+      background: var(--bg-secondary);
+      border: 1px solid var(--border-primary);
+      padding: 8px 12px;
+      border-radius: 4px;
+      color: var(--text-secondary);
+      cursor: pointer;
+      font-size: 0.9rem;
+      text-decoration: none;
+      display: inline-block;
+    }
+    .back-button:hover {
+      color: var(--text-primary);
+      background: var(--bg-tertiary);
+    }
+    h1 {
+      font-size: 1.5em;
+      margin: 1rem 0;
+      color: var(--text-primary);
+      border-bottom: 1px solid var(--border-primary);
+      padding-bottom: 0.5rem;
+    }
+    ul {
+      list-style-type: none;
+      padding: 0;
+    }
+    li {
+      margin: 0;
+      border-bottom: 1px solid var(--border-primary);
+    }
+    li:last-child {
+      border-bottom: none;
+    }
+    a {
+      display: block;
+      padding: 0.75rem 0.5rem;
+      text-decoration: none;
+      color: var(--text-link);
+      transition: background 0.2s ease;
+    }
+    a:hover {
+      background: var(--bg-secondary);
+    }
+    .dir {
+      font-weight: 500;
+    }
   </style>
 </head>
 <body>
+  <div class='controls'>
+    <a href='../index.html' class='back-button'>← back</a>
+  </div>
   <h1>Index of /flash_attn/results</h1>
   <ul>
-    <li><a href='../index.html' class='dir'>../</a></li>
     <li><a href='combined_results.html' class='file'>combined_results.html</a></li>
   </ul>
 </body>

index.html

@@ -2,16 +2,78 @@
 <html>
 <head>
   <meta charset='UTF-8'>
-  <title>Directory Index</title>
+  <meta name='viewport' content='width=device-width, initial-scale=1.0'>
+  <title>Index of /</title>
   <style>
-    body { font-family: monospace; margin: 20px; }
-    h1 { font-size: 1.5em; }
-    ul { list-style-type: none; padding-left: 20px; }
-    li { margin: 5px 0; }
-    .dir { font-weight: bold; }
-    .file { color: #0066cc; }
-    a { text-decoration: none; }
-    a:hover { text-decoration: underline; }
+    :root {
+      --bg-primary: #0a0a0a;
+      --bg-secondary: #121212;
+      --bg-tertiary: #181818;
+      --text-primary: #e0e0e0;
+      --text-secondary: #888888;
+      --text-link: #64b5f6;
+      --border-primary: #2a2a2a;
+    }
+    body {
+      font-family: system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial, sans-serif;
+      background: var(--bg-primary);
+      color: var(--text-primary);
+      margin: 0;
+      padding: 16px;
+      max-width: 900px;
+      margin: 0 auto;
+    }
+    .controls {
+      display: flex;
+      justify-content: flex-end;
+      margin-bottom: 1rem;
+    }
+    .back-button {
+      background: var(--bg-secondary);
+      border: 1px solid var(--border-primary);
+      padding: 8px 12px;
+      border-radius: 4px;
+      color: var(--text-secondary);
+      cursor: pointer;
+      font-size: 0.9rem;
+      text-decoration: none;
+      display: inline-block;
+    }
+    .back-button:hover {
+      color: var(--text-primary);
+      background: var(--bg-tertiary);
+    }
+    h1 {
+      font-size: 1.5em;
+      margin: 1rem 0;
+      color: var(--text-primary);
+      border-bottom: 1px solid var(--border-primary);
+      padding-bottom: 0.5rem;
+    }
+    ul {
+      list-style-type: none;
+      padding: 0;
+    }
+    li {
+      margin: 0;
+      border-bottom: 1px solid var(--border-primary);
+    }
+    li:last-child {
+      border-bottom: none;
+    }
+    a {
+      display: block;
+      padding: 0.75rem 0.5rem;
+      text-decoration: none;
+      color: var(--text-link);
+      transition: background 0.2s ease;
+    }
+    a:hover {
+      background: var(--bg-secondary);
+    }
+    .dir {
+      font-weight: 500;
+    }
   </style>
 </head>
 <body>