diff --git a/flash_attn/impls/artifacts/benchmark/attn.jsonl b/flash_attn/impls/artifacts/benchmark/attn.jsonl index 5bfc57e28bca87740cd188c3cbcd8790a2ef1b44..5acfbcfc5bc3cd28bec9cacd18308543bc7864e5 100644 --- a/flash_attn/impls/artifacts/benchmark/attn.jsonl +++ b/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} diff --git a/flash_attn/impls/artifacts/benchmark_default/attn_default.jsonl b/flash_attn/impls/artifacts/benchmark_default/attn_default.jsonl index c22a735a422739b1d1361b85a03720ea9bc75c22..f0bbce330d32dfe0cbd3c869a8ad5f2aaa045e94 100644 --- a/flash_attn/impls/artifacts/benchmark_default/attn_default.jsonl +++ b/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} diff --git a/flash_attn/impls/artifacts/benchmark_max_autotune/attn_max_autotune.jsonl b/flash_attn/impls/artifacts/benchmark_max_autotune/attn_max_autotune.jsonl index bc641fc5b1ec22e55f344250c86a8d5a9718be99..87baeccdc6ab98529881effd3a62d0dfe5029b5a 100644 --- a/flash_attn/impls/artifacts/benchmark_max_autotune/attn_max_autotune.jsonl +++ b/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} diff --git a/flash_attn/impls/cells/benchmark.py b/flash_attn/impls/cells/benchmark.py index 12c02a967606bf5974e797deac59e0a0af2e284e..f471434e8cb2d02f0e3b081d6133c17ea46bc373 100644 --- a/flash_attn/impls/cells/benchmark.py +++ b/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: diff --git a/flash_attn/impls/compiled_variants.html b/flash_attn/impls/compiled_variants.html index 8614044a8fff911edcd4d7c0e3a71756827e6d80..df7b4a9560dee2b9cc43fb2840e4da8927f7cde9 100644 --- a/flash_attn/impls/compiled_variants.html +++ b/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:
+ + ← back +
light
reset
@@ -3746,7 +3760,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark_default | 46.78s +Cell: benchmark_default | 45.23s | Raw @@ -3830,7 +3844,7 @@ Cell: benchmark_default | 46.78s
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
▶ UV Install Logs
@@ -3906,7 +3920,7 @@ Installed 37 packages in 557ms ▼ output ▶ uv-logs | -Cell: benchmark_max_autotune | 53.65s +Cell: benchmark_max_autotune | 54.06s | Raw @@ -3990,39 +4004,39 @@ Cell: benchmark_max_autotune | 53.65s
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
▶ UV Install Logs
diff --git a/flash_attn/impls/flash_attention.html b/flash_attn/impls/flash_attention.html index cfb5c134292b9f235a89ded8098d37cd808bc251..6c4655b3e27d9fb741de565f4276ebd9e9f5beef 100644 --- a/flash_attn/impls/flash_attention.html +++ b/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:
+ + ← back +
light
reset
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: nv | 0.70s +Cell: nv | 0.67s | Raw @@ -3761,7 +3775,7 @@ Cell: nv | 0.70s
-
Thu Oct 2 18:06:49 2025 +
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 ▼ output ▶ uv-logs | -Cell: benchmark | 36.63s +Cell: benchmark | 35.41s | Raw @@ -3886,7 +3900,7 @@ Cell: benchmark | 36.63s
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
▶ UV Install Logs
diff --git a/flash_attn/impls/hf_kernels_flash_attn.html b/flash_attn/impls/hf_kernels_flash_attn.html index f67683549a9873ee6c5d77297064e02d4d3817b8..eb96379b9cca8d14f8e80d4141cdae4ecd113c07 100644 --- a/flash_attn/impls/hf_kernels_flash_attn.html +++ b/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:
+ + ← back +
light
reset
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark | 39.43s +Cell: benchmark | 38.65s | Raw @@ -3831,40 +3845,40 @@ Cell: benchmark | 39.43s
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
▶ UV Install Logs
Fetching 20 files: 0%| | 0/20 [00:00<?, ?it/s] -Fetching 20 files: 5%|▌ | 1/20 [00:00<00:03, 5.41it/s] -Fetching 20 files: 10%|█ | 2/20 [00:01<00:16, 1.09it/s] -Fetching 20 files: 100%|██████████| 20/20 [00:01<00:00, 12.37it/s]
+Fetching 20 files: 5%|▌ | 1/20 [00:00<00:03, 5.70it/s] +Fetching 20 files: 10%|█ | 2/20 [00:01<00:13, 1.36it/s] +Fetching 20 files: 100%|██████████| 20/20 [00:01<00:00, 15.31it/s]

Artifacts:

attn.jsonl diff --git a/flash_attn/impls/hf_kernels_flash_attn3.html b/flash_attn/impls/hf_kernels_flash_attn3.html index e47cd7efa7b0ba3cfd1c8fae2b977df5d34fc33f..8f2225d6d312b7248216f259f9d91f6f5d1a2bff 100644 --- a/flash_attn/impls/hf_kernels_flash_attn3.html +++ b/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:
+ + ← back +
light
reset
@@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark | 39.41s +Cell: benchmark | 38.16s | Raw @@ -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
@@ -3841,62 +3855,62 @@ hf_kernels_flash_attn3 flux_L512 0.57 True
▶ UV Install Logs
Fetching 4 files: 0%| | 0/4 [00:00<?, ?it/s] -Fetching 4 files: 25%|██▌ | 1/4 [00:00<00:00, 6.35it/s] -Fetching 4 files: 50%|█████ | 2/4 [00:01<00:01, 1.09it/s] -Fetching 4 files: 100%|██████████| 4/4 [00:01<00:00, 2.48it/s]
+Fetching 4 files: 25%|██▌ | 1/4 [00:00<00:00, 5.17it/s] +Fetching 4 files: 50%|█████ | 2/4 [00:01<00:01, 1.22it/s] +Fetching 4 files: 100%|██████████| 4/4 [00:01<00:00, 2.76it/s]

Artifacts:

attn.jsonl diff --git a/flash_attn/impls/index.html b/flash_attn/impls/index.html index 16f13ae51cc53a9df48c5e0d99ab1e6ea0aeb5ce..fbf5a24033c4d68d64ce4f1be0ba13266dd4e89d 100644 --- a/flash_attn/impls/index.html +++ b/flash_attn/impls/index.html @@ -2,22 +2,86 @@ - Directory Index + + Index of /flash_attn/impls +
+ ← back +

Index of /flash_attn/impls

    -
  • ../
  • compiled_variants.html
  • flash_attention.html
  • hf_kernels_flash_attn.html
    • diff --git a/flash_attn/impls/mem_efficient_attention.html b/flash_attn/impls/mem_efficient_attention.html index 2e6d76828b73814a727b5eb63ea4f1ebed867c4a..7213e66ff6d02989f90efc2869171fe1cad361a3 100644 --- a/flash_attn/impls/mem_efficient_attention.html +++ b/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:
    + + ← back +
    light
    reset
    @@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark | 36.09s +Cell: benchmark | 36.80s | Raw @@ -3838,28 +3852,28 @@ torch_mem_eff flux_L512 0.95 True
    ▶ UV Install Logs
    diff --git a/flash_attn/impls/sage_attention.html b/flash_attn/impls/sage_attention.html index f24f704639dc0a5390db0221afb9dba66da9eecb..5a357f2e63a095cefb67f14c37e15b66b39d007a 100644 --- a/flash_attn/impls/sage_attention.html +++ b/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:
    + + ← back +
    light
    reset
    @@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark | 40.08s +Cell: benchmark | 40.58s | Raw @@ -3839,45 +3853,45 @@ Cell: benchmark | 40.08s
    impl wl p50(ms) ok sage_int8_fp16 flux_L128 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd' sage_int8_fp16 flux_L256 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd' sage_int8_fp16 flux_L320 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd' sage_int8_fp16 flux_L384 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd' sage_int8_fp16 flux_L448 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd' sage_int8_fp16 flux_L512 FAIL False - Error: module 'sage_attention_ba12545b014364be' has no attribute 'fwd' + Error: module 'sage_attention_1863f4c92418f0f6' has no attribute 'fwd'
    ▶ UV Install Logs
    Fetching 11 files: 0%| | 0/11 [00:00<?, ?it/s] -Fetching 11 files: 9%|▉ | 1/11 [00:00<00:02, 4.42it/s] -Fetching 11 files: 27%|██▋ | 3/11 [00:00<00:01, 5.95it/s] -Fetching 11 files: 73%|███████▎ | 8/11 [00:00<00:00, 12.27it/s] -Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 14.37it/s]
    +Fetching 11 files: 9%|▉ | 1/11 [00:00<00:01, 5.59it/s] +Fetching 11 files: 73%|███████▎ | 8/11 [00:00<00:00, 12.79it/s] +Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 16.77it/s]

    Artifacts:

    attn.jsonl diff --git a/flash_attn/impls/xformers.html b/flash_attn/impls/xformers.html index 8e523df79ff0dad24891006c04f253b6e26c179e..b4b072cd702fba0ae3c8cc918c12746908e8d2e3 100644 --- a/flash_attn/impls/xformers.html +++ b/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:
    + + ← back +
    light
    reset
    @@ -3745,7 +3759,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: ▼ output ▶ uv-logs | -Cell: benchmark | 40.41s +Cell: benchmark | 42.08s | Raw @@ -3838,38 +3852,38 @@ xformers_meff flux_L512 0.65 True
    ▶ UV Install Logs
    diff --git a/flash_attn/index.html b/flash_attn/index.html index 9f7dc880eb75f8403035bd95857991f89fe04872..eea7df846d9f2d44c6c6e03a5ac30d00cecd90cf 100644 --- a/flash_attn/index.html +++ b/flash_attn/index.html @@ -2,22 +2,86 @@ - Directory Index + + Index of /flash_attn +
    + ← back +

    Index of /flash_attn

    diff --git a/flash_attn/results/artifacts/combine/latency.csv b/flash_attn/results/artifacts/combine/latency.csv index b4d1c6f4b50f7f17437351b83e12a6f6ae66bfcc..3d8227b97b600492490413c0bfcaae4e1e8d53d2 100644 --- a/flash_attn/results/artifacts/combine/latency.csv +++ b/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 diff --git a/flash_attn/results/artifacts/combine/latency.svg b/flash_attn/results/artifacts/combine/latency.svg index f28ef39dd1b578be3abbdfe33b7fa78f5b20327c..242a7df95a0a79e47f8fa9f6f7a99913358dab23 100644 --- a/flash_attn/results/artifacts/combine/latency.svg +++ b/flash_attn/results/artifacts/combine/latency.svg @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:587d477bc9dc161a51c012142295b5a6efa33e48054fc657106ca27ba64b8683 -size 28263 +oid sha256:2c1da56080e7fd1a85c14295083b11d6bac981f6fb3faef98b0753eb2c1676c7 +size 28243 diff --git a/flash_attn/results/combined_results.html b/flash_attn/results/combined_results.html index 7dd6121080ba7791bab4cf718f9a3098bbb6797b..c3231e96afd1eeac602f702cc3b1291530a0a629 100644 --- a/flash_attn/results/combined_results.html +++ b/flash_attn/results/combined_results.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); } @@ -3864,6 +3875,9 @@ g:focus {
    + + ← back +
    light
    reset
    @@ -3906,7 +3920,7 @@ g:focus {

    Flash Attention Benchmarks - Aggregated Results

    This document combines benchmark results from multiple attention implementations using cross-file dependencies.

    -

    Combined Summary and Visualization>

    +

    Combined Summary and Visualization

    @@ -3914,7 +3928,7 @@ using cross-file dependencies.

    - 2025-10-02T18:12:57.416034 + 2025-10-02T20:00:13.145631 image/svg+xml @@ -4024,83 +4038,83 @@ using cross-file dependencies.

    - + - + - 0.4 + 0.4 - + - + - 0.5 + 0.5 - + - + - 0.6 + 0.6 - + - + - 0.7 + 0.7 - + - + - 0.8 + 0.8 - + - + - 0.9 + 0.9 @@ -4108,101 +4122,101 @@ using cross-file dependencies.

         - + - - - - - - + + + + + + - + - - - - - + + + + + - + - - - - - - + + + + + + - + - - - - - - + + + + + + - + - - - - - - + + + + + + - + - - - - - + + + + + - + - - - - - - + + + + + + @@ -4328,10 +4342,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.48577280044555665 -0.47836801409721375 -0.4803520143032074 -0.4827199876308441 +0.49411200881004336 +0.48844799399375916 +0.4936000108718872 +0.4944640100002289 5 83.38 FLASH @@ -4346,10 +4360,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5229184031486511 -0.521727979183197 -0.5228800177574158 -0.5234559774398804 +0.5234112024307251 +0.5224320292472839 +0.5235199928283691 +0.5235840082168579 5 90.62 FLASH @@ -4364,10 +4378,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6515967845916748 -0.6503999829292297 -0.650879979133606 -0.6513599753379822 +0.6527232170104981 +0.6503040194511414 +0.6524800062179565 +0.6545600295066833 5 95.06 FLASH @@ -4382,10 +4396,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6807615995407105 -0.6797440052032471 -0.6808639764785767 -0.6815680265426636 +0.682803213596344 +0.6805760264396667 +0.6828799843788147 +0.6832640171051025 5 99.88 FLASH @@ -4400,10 +4414,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.711027193069458 -0.7058879733085632 -0.7121919989585876 -0.7131519913673401 +0.7075456142425537 +0.7057600021362305 +0.7063360214233398 +0.7070720195770264 5 103.81 FLASH @@ -4418,10 +4432,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7391423940658569 -0.7369279861450195 -0.7383999824523926 -0.7408959865570068 +0.7379711985588073 +0.7368639707565308 +0.7372480034828186 +0.7391039729118347 5 109.12 FLASH @@ -4436,10 +4450,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5875647902488709 -0.5863680243492126 -0.5874559879302979 -0.5876479744911194 +0.5874239921569824 +0.5861759781837463 +0.5873280167579651 +0.5877439975738525 5 83.38 EFFICIENT @@ -4454,10 +4468,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.653657603263855 -0.6485440135002136 -0.6537600159645081 -0.656544029712677 +0.6502719998359681 +0.6490240097045898 +0.649183988571167 +0.6517760157585144 5 90.62 EFFICIENT @@ -4472,10 +4486,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7784512042999268 -0.774944007396698 -0.778656005859375 -0.7801600098609924 +0.7812095880508423 +0.7761600017547607 +0.7803199887275696 +0.7852799892425537 5 95.94 EFFICIENT @@ -4490,10 +4504,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7922943949699401 -0.791263997554779 -0.7924799919128418 -0.7927039861679077 +0.7948480010032654 +0.7911999821662903 +0.7935360074043274 +0.7948480248451233 5 100.0 EFFICIENT @@ -4508,10 +4522,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.848089587688446 -0.8444799780845642 -0.8470079898834229 -0.8499199748039246 +0.8463295936584473 +0.8449919819831848 +0.8459839820861816 +0.8461120128631592 5 103.81 EFFICIENT @@ -4526,10 +4540,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.9523776054382325 -0.95004802942276 -0.9519039988517761 -0.9541119933128357 +0.9538687944412232 +0.9492800235748291 +0.9518399834632874 +0.9581760168075562 5 109.12 EFFICIENT @@ -4544,10 +4558,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.45066879987716674 -0.4474239945411682 -0.44921600818634033 -0.45241600275039673 +0.4515071928501129 +0.44364801049232483 +0.4524799883365631 +0.4557119905948639 5 83.38 memory_efficient @@ -4562,10 +4576,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.47004159688949587 -0.4652479887008667 -0.4705919921398163 -0.4716799855232239 +0.46787199974060056 +0.46489599347114563 +0.4684160053730011 +0.46908798813819885 5 90.62 memory_efficient @@ -4580,10 +4594,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6022783994674683 -0.5987840294837952 -0.6021760106086731 -0.6045759916305542 +0.6001471996307373 +0.596992015838623 +0.5984640121459961 +0.6016640067100525 5 95.06 memory_efficient @@ -4598,10 +4612,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6013055920600892 -0.6000319719314575 -0.600383996963501 -0.6016640067100525 +0.6023231983184815 +0.5997440218925476 +0.6031039953231812 +0.6032639741897583 5 99.88 memory_efficient @@ -4616,10 +4630,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6408192038536071 -0.639136016368866 -0.6404479742050171 -0.6416320204734802 +0.6411136031150818 +0.6381760239601135 +0.6414719820022583 +0.6421440243721008 5 103.81 memory_efficient @@ -4634,10 +4648,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6466111898422241 -0.6447359919548035 -0.6462399959564209 -0.6483839750289917 +0.6594688057899475 +0.6441280245780945 +0.6496639847755432 +0.6527680158615112 5 109.12 memory_efficient @@ -4652,10 +4666,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.527347207069397 -0.5194560289382935 -0.5272960066795349 -0.5312960147857666 +0.5181439876556396 +0.5141760110855103 +0.5175679922103882 +0.5197759866714478 5 83.38 FLASH @@ -4670,10 +4684,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5586367964744567 -0.5560640096664429 -0.5571519732475281 -0.5611839890480042 +0.5579584002494812 +0.5549119710922241 +0.5582720041275024 +0.5598080158233643 5 90.62 FLASH @@ -4688,10 +4702,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6860736012458801 -0.6841920018196106 -0.6860160231590271 -0.6869760155677795 +0.6872959971427918 +0.6853119730949402 +0.687391996383667 +0.6883519887924194 5 95.25 FLASH @@ -4706,10 +4720,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7167360067367554 -0.7152000069618225 -0.7161920070648193 -0.7164160013198853 +0.716153597831726 +0.7128639817237854 +0.7160959839820862 +0.7167680263519287 5 99.88 FLASH @@ -4724,10 +4738,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7423295855522156 +0.7418303966522217 +0.7386879920959473 0.7400959730148315 -0.742143988609314 -0.7431039810180664 +0.7415040135383606 5 103.81 FLASH @@ -4742,10 +4756,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.7743871927261352 -0.7718080282211304 -0.7745919823646545 -0.7748159766197205 +0.7745471954345703 +0.7708160281181335 +0.7740799784660339 +0.7753919959068298 5 109.12 FLASH @@ -4760,10 +4774,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.6489088058471679 -0.6148160099983215 -0.6296960115432739 -0.6522240042686462 +0.6468096017837525 +0.6144000291824341 +0.6245759725570679 +0.6483200192451477 5 67.5 FLASH @@ -4778,10 +4792,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.700761592388153 -0.6615359783172607 -0.6821119785308838 -0.7128959894180298 +0.7060160160064697 +0.6689280271530151 +0.6851199865341187 +0.7184960246086121 5 75.0 FLASH @@ -4796,10 +4810,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.834444797039032 -0.7967039942741394 -0.8164799809455872 -0.8463680148124695 +0.8332608103752136 +0.7953600287437439 +0.8155840039253235 +0.8403519988059998 5 80.38 FLASH @@ -4814,10 +4828,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.8709375977516174 -0.8432319760322571 -0.8498560190200806 -0.8750079870223999 +0.8719295978546142 +0.8470720052719116 +0.849727988243103 +0.8745279908180237 5 82.5 FLASH @@ -4832,10 +4846,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.9069631934165955 -0.8775359988212585 -0.9030719995498657 -0.903872013092041 +0.9034304022789001 +0.8677120208740234 +0.8835520148277283 +0.9034240245819092 5 86.25 FLASH @@ -4850,10 +4864,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.9371584057807922 -0.9145920276641846 -0.9164159893989563 -0.9357439875602722 +0.9387519836425782 +0.9154239892959595 +0.9213759899139404 +0.9359679818153381 5 90.0 FLASH @@ -4868,10 +4882,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.34446719884872434 -0.3438720107078552 -0.3445119857788086 -0.34457600116729736 +0.3455295979976654 +0.34355199337005615 +0.34563198685646057 +0.34643200039863586 5 83.38 flash-attn @@ -4886,10 +4900,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.37571839094161985 -0.37404799461364746 -0.3763839900493622 -0.3766399919986725 +0.3756160080432892 +0.37411201000213623 +0.3752000033855438 +0.3770880103111267 5 90.62 flash-attn @@ -4904,10 +4918,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.4945920050144196 -0.4925439953804016 -0.493120014667511 -0.4938240051269531 +0.4953216016292572 +0.49324798583984375 +0.49433600902557373 +0.49663999676704407 5 95.06 flash-attn @@ -4922,10 +4936,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5139647841453552 -0.5123199820518494 +0.5157055854797363 0.5142719745635986 -0.5147839784622192 +0.516319990158081 +0.516543984413147 5 99.88 flash-attn @@ -4940,10 +4954,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5353855967521668 -0.5339199900627136 -0.5350080132484436 -0.5352320075035095 +0.5356672048568726 +0.5346879959106445 +0.5358080267906189 +0.5361599922180176 5 103.81 flash-attn @@ -4958,10 +4972,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5548800110816956 -0.5538560152053833 -0.5548800230026245 -0.5553280115127563 +0.5587136030197144 +0.5557760000228882 +0.5574079751968384 +0.5581120252609253 5 109.12 flash-attn @@ -4976,10 +4990,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.3617343962192535 -0.36102399230003357 -0.3616960048675537 -0.36211198568344116 +0.3619711995124817 +0.3603839874267578 +0.361952006816864 +0.3624640107154846 5 83.38 flash-attn3 @@ -4994,10 +5008,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.3907967984676361 -0.3885439932346344 -0.39056000113487244 -0.3906239867210388 +0.3912447988986969 +0.3892799913883209 +0.3909760117530823 +0.3922559916973114 5 90.62 flash-attn3 @@ -5012,10 +5026,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5228991985321045 -0.521344006061554 -0.5230720043182373 -0.5232319831848145 +0.5258048176765442 +0.5240640044212341 +0.5248960256576538 +0.5248960256576538 5 95.06 flash-attn3 @@ -5030,10 +5044,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5254656076431274 -0.523904025554657 -0.5249919891357422 -0.526528000831604 +0.5276032090187073 +0.5265600085258484 +0.5277760028839111 +0.5282559990882874 5 99.88 flash-attn3 @@ -5048,10 +5062,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5646592020988465 -0.5627840161323547 -0.565343976020813 -0.565343976020813 +0.5656383991241455 +0.5639039874076843 +0.5657920241355896 +0.5668479800224304 5 103.81 flash-attn3 @@ -5066,10 +5080,10 @@ using cross-file dependencies.

    24 128 bfloat16 -0.5698879957199097 -0.567359983921051 -0.5696640014648438 -0.5698559880256653 +0.5789952039718628 +0.5689600110054016 +0.5698239803314209 +0.5713919997215271 5 109.12 flash-attn3 @@ -5086,7 +5100,7 @@ using cross-file dependencies.

    ▶ output ▶ uv-logs | -Cell: combine | 36.30s +Cell: combine | 36.89s | Raw @@ -5765,31 +5779,31 @@ HF Kernels Flash Attn3 : /repo/flash_attn/impls/.uvnote/cache/65da999faf5 COMBINED BENCHMARK SUMMARY 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 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 -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 torch_flash_compiled_default flux_L448 0.74 True torch_flash_compiled_default flux_L512 0.77 True -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 -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 @@ -5835,28 +5849,28 @@ Total records: 42
    ▶ UV Install Logs

    Artifacts:

    -latency.svg latency.csv -
    - - - - - - - 2025-10-02T18:12:57.416034 - image/svg+xml - - - Matplotlib v3.10.6, https://matplotlib.org/ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - flux_L128 - - - - - - - - - - - - - flux_L256 - - - - - - - - - - - - - flux_L320 - - - - - - - - - - - - - flux_L384 - - - - - - - - - - - - - flux_L448 - - - - - - - - - - - - - flux_L512 - - - - Workload - - - - - - - - - - - - - - - - - 0.4 - - - - - - - - - - - - - 0.5 - - - - - - - - - - - - - 0.6 - - - - - - - - - - - - - 0.7 - - - - - - - - - - - - - 0.8 - - - - - - - - - - - - - 0.9 - - - - Latency P50 (ms) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Attention Implementation Latency - - - - - - - - - - - - - torch_flash_ma - - - - - - - - - torch_mem_eff - - - - - - - - - xformers_meff - - - - - - - - - torch_flash_compiled_default - - - - - - - - - torch_flash_compiled_max_autotune - - - - - - - - - hf_kernels_flash_attn - - - - - - - - - hf_kernels_flash_attn3 - - - - - - - - - - -
    +latency.svg
    @@ -6309,10 +5933,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6327,10 +5951,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6345,10 +5969,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6363,10 +5987,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6381,10 +6005,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6399,10 +6023,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6417,10 +6041,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6435,10 +6059,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6453,10 +6077,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6471,10 +6095,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6489,10 +6113,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6507,10 +6131,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6525,10 +6149,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6543,10 +6167,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6561,10 +6185,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6579,10 +6203,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6597,10 +6221,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6615,10 +6239,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6633,10 +6257,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6651,10 +6275,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6669,10 +6293,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6687,10 +6311,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6705,10 +6329,10 @@ Installed 37 packages in 440ms - + + - - + @@ -6723,10 +6347,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6741,10 +6365,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6759,10 +6383,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6777,10 +6401,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6795,10 +6419,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6813,10 +6437,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6831,10 +6455,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6849,10 +6473,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6867,10 +6491,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6885,10 +6509,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6903,10 +6527,10 @@ Installed 37 packages in 440ms - - + - + + @@ -6921,10 +6545,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6939,10 +6563,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6957,10 +6581,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6975,10 +6599,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -6993,10 +6617,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -7011,10 +6635,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -7029,10 +6653,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -7047,10 +6671,10 @@ Installed 37 packages in 440ms - - - - + + + + @@ -7059,6 +6683,396 @@ Installed 37 packages in 440ms
    24 128 bfloat160.485772800445556650.478368014097213750.48035201430320740.48271998763084410.494112008810043360.488447993993759160.49360001087188720.4944640100002289 5 83.38 FLASH24 128 bfloat160.52291840314865110.5217279791831970.52288001775741580.52345597743988040.52341120243072510.52243202924728390.52351999282836910.5235840082168579 5 90.62 FLASH24 128 bfloat160.65159678459167480.65039998292922970.6508799791336060.65135997533798220.65272321701049810.65030401945114140.65248000621795650.6545600295066833 5 95.06 FLASH24 128 bfloat160.68076159954071050.67974400520324710.68086397647857670.68156802654266360.6828032135963440.68057602643966670.68287998437881470.6832640171051025 5 99.88 FLASH24 128 bfloat160.7110271930694580.70588797330856320.71219199895858760.71315199136734010.70754561424255370.70576000213623050.70633602142333980.7070720195770264 5 103.81 FLASH24 128 bfloat160.73914239406585690.73692798614501950.73839998245239260.74089598655700680.73797119855880730.73686397075653080.73724800348281860.7391039729118347 5 109.12 FLASH24 128 bfloat160.58756479024887090.58636802434921260.58745598793029790.58764797449111940.58742399215698240.58617597818374630.58732801675796510.5877439975738525 5 83.38 EFFICIENT24 128 bfloat160.6536576032638550.64854401350021360.65376001596450810.6565440297126770.65027199983596810.64902400970458980.6491839885711670.6517760157585144 5 90.62 EFFICIENT24 128 bfloat160.77845120429992680.7749440073966980.7786560058593750.78016000986099240.78120958805084230.77616000175476070.78031998872756960.7852799892425537 5 95.94 EFFICIENT24 128 bfloat160.79229439496994010.7912639975547790.79247999191284180.79270398616790770.79484800100326540.79119998216629030.79353600740432740.7948480248451233 5 100.0 EFFICIENT24 128 bfloat160.8480895876884460.84447997808456420.84700798988342290.84991997480392460.84632959365844730.84499198198318480.84598398208618160.8461120128631592 5 103.81 EFFICIENT24 128 bfloat160.95237760543823250.950048029422760.95190399885177610.95411199331283570.95386879444122320.94928002357482910.95183998346328740.9581760168075562 5 109.12 EFFICIENT24 128 bfloat160.450668799877166740.44742399454116820.449216008186340330.452416002750396730.45150719285011290.443648010492324830.45247998833656310.4557119905948639 5 83.38 memory_efficient24 128 bfloat160.470041596889495870.46524798870086670.47059199213981630.47167998552322390.467871999740600560.464895993471145630.46841600537300110.46908798813819885 5 90.62 memory_efficient24 128 bfloat160.60227839946746830.59878402948379520.60217601060867310.60457599163055420.60014719963073730.5969920158386230.59846401214599610.6016640067100525 5 95.06 memory_efficient24 128 bfloat160.60130559206008920.60003197193145750.6003839969635010.60166400671005250.60232319831848150.59974402189254760.60310399532318120.6032639741897583 5 99.88 memory_efficient24 128 bfloat160.64081920385360710.6391360163688660.64044797420501710.64163202047348020.64111360311508180.63817602396011350.64147198200225830.6421440243721008 5 103.81 memory_efficient24 128 bfloat160.64661118984222410.64473599195480350.64623999595642090.64838397502899170.65946880578994750.64412802457809450.64966398477554320.6527680158615112 5 109.12 memory_efficient24 128 bfloat160.5273472070693970.51945602893829350.52729600667953490.53129601478576660.51814398765563960.51417601108551030.51756799221038820.5197759866714478 5 83.38 FLASH24 128 bfloat160.55863679647445670.55606400966644290.55715197324752810.56118398904800420.55795840024948120.55491197109222410.55827200412750240.5598080158233643 5 90.62 FLASH24 128 bfloat160.68607360124588010.68419200181961060.68601602315902710.68697601556777950.68729599714279180.68531197309494020.6873919963836670.6883519887924194 5 95.25 FLASH24 128 bfloat160.71673600673675540.71520000696182250.71619200706481930.71641600131988530.7161535978317260.71286398172378540.71609598398208620.7167680263519287 5 99.88 FLASH24 128 bfloat160.74232958555221560.74183039665222170.7386879920959473 0.74009597301483150.7421439886093140.74310398101806640.7415040135383606 5 103.81 FLASH24 128 bfloat160.77438719272613520.77180802822113040.77459198236465450.77481597661972050.77454719543457030.77081602811813350.77407997846603390.7753919959068298 5 109.12 FLASH24 128 bfloat160.64890880584716790.61481600999832150.62969601154327390.65222400426864620.64680960178375250.61440002918243410.62457597255706790.6483200192451477 5 67.5 FLASH24 128 bfloat160.7007615923881530.66153597831726070.68211197853088380.71289598941802980.70601601600646970.66892802715301510.68511998653411870.7184960246086121 5 75.0 FLASH24 128 bfloat160.8344447970390320.79670399427413940.81647998094558720.84636801481246950.83326081037521360.79536002874374390.81558400392532350.8403519988059998 5 80.38 FLASH24 128 bfloat160.87093759775161740.84323197603225710.84985601902008060.87500798702239990.87192959785461420.84707200527191160.8497279882431030.8745279908180237 5 82.5 FLASH24 128 bfloat160.90696319341659550.87753599882125850.90307199954986570.9038720130920410.90343040227890010.86771202087402340.88355201482772830.9034240245819092 5 86.25 FLASH24 128 bfloat160.93715840578079220.91459202766418460.91641598939895630.93574398756027220.93875198364257820.91542398929595950.92137598991394040.9359679818153381 5 90.0 FLASH24 128 bfloat160.344467198848724340.34387201070785520.34451198577880860.344576001167297360.34552959799766540.343551993370056150.345631986856460570.34643200039863586 5 83.38 flash-attn24 128 bfloat160.375718390941619850.374047994613647460.37638399004936220.37663999199867250.37561600804328920.374112010002136230.37520000338554380.3770880103111267 5 90.62 flash-attn24 128 bfloat160.49459200501441960.49254399538040160.4931200146675110.49382400512695310.49532160162925720.493247985839843750.494336009025573730.49663999676704407 5 95.06 flash-attn24 128 bfloat160.51396478414535520.51231998205184940.5157055854797363 0.51427197456359860.51478397846221920.5163199901580810.516543984413147 5 99.88 flash-attn24 128 bfloat160.53538559675216680.53391999006271360.53500801324844360.53523200750350950.53566720485687260.53468799591064450.53580802679061890.5361599922180176 5 103.81 flash-attn24 128 bfloat160.55488001108169560.55385601520538330.55488002300262450.55532801151275630.55871360301971440.55577600002288820.55740797519683840.5581120252609253 5 109.12 flash-attn24 128 bfloat160.36173439621925350.361023992300033570.36169600486755370.362111985683441160.36197119951248170.36038398742675780.3619520068168640.3624640107154846 5 83.38 flash-attn324 128 bfloat160.39079679846763610.38854399323463440.390560001134872440.39062398672103880.39124479889869690.38927999138832090.39097601175308230.3922559916973114 5 90.62 flash-attn324 128 bfloat160.52289919853210450.5213440060615540.52307200431823730.52323198318481450.52580481767654420.52406400442123410.52489602565765380.5248960256576538 5 95.06 flash-attn324 128 bfloat160.52546560764312740.5239040255546570.52499198913574220.5265280008316040.52760320901870730.52656000852584840.52777600288391110.5282559990882874 5 99.88 flash-attn324 128 bfloat160.56465920209884650.56278401613235470.5653439760208130.5653439760208130.56563839912414550.56390398740768430.56579202413558960.5668479800224304 5 103.81 flash-attn324 128 bfloat160.56988799571990970.5673599839210510.56966400146484380.56985598802566530.57899520397186280.56896001100540160.56982398033142090.5713919997215271 5 109.12 flash-attn3
    +
    + + + + + + + 2025-10-02T20:00:13.145631 + image/svg+xml + + + Matplotlib v3.10.6, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + flux_L128 + + + + + + + + + + + + + flux_L256 + + + + + + + + + + + + + flux_L320 + + + + + + + + + + + + + flux_L384 + + + + + + + + + + + + + flux_L448 + + + + + + + + + + + + + flux_L512 + + + + Workload + + + + + + + + + + + + + + + + + 0.4 + + + + + + + + + + + + + 0.5 + + + + + + + + + + + + + 0.6 + + + + + + + + + + + + + 0.7 + + + + + + + + + + + + + 0.8 + + + + + + + + + + + + + 0.9 + + + + Latency P50 (ms) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Attention Implementation Latency + + + + + + + + + + + + + torch_flash_ma + + + + + + + + + torch_mem_eff + + + + + + + + + xformers_meff + + + + + + + + + torch_flash_compiled_default + + + + + + + + + torch_flash_compiled_max_autotune + + + + + + + + + hf_kernels_flash_attn + + + + + + + + + hf_kernels_flash_attn3 + + + + + + + + + + +
    diff --git a/flash_attn/results/index.html b/flash_attn/results/index.html index 98a15e14d1388566b8c20140be77a606a07d4643..b87b6002f4b781572dbb50f91850e50ee98130ab 100644 --- a/flash_attn/results/index.html +++ b/flash_attn/results/index.html @@ -2,22 +2,86 @@ - Directory Index + + Index of /flash_attn/results +
    + ← back +

    Index of /flash_attn/results

    diff --git a/index.html b/index.html index d04c5bb146362887945a984bd325ef81d54da050..df44040e2dd9e1e4a0fc2d5ee08453d4b9953f11 100644 --- a/index.html +++ b/index.html @@ -2,16 +2,78 @@ - Directory Index + + Index of /