+
+▼ code
+▼ output
+ ▶ uv-logs
+ |
+Cell: benchmark | 37.93s
+ |
+
+Raw
+
+
+
+
+
+
+
+
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+# "numpy",
+# "torch",
+# "kernels-benchmark-tools",
+# "kernels",
+# ]
+#
+# [tool.uv.sources]
+# kernels-benchmark-tools = { git = "https://github.com/drbh/kernels-benchmark-tools.git", branch = "main" }
+# ///
+import torch
+import sys
+import os
+import kernels_benchmark_tools as kbt
+from kernels import get_kernel
+
+hf_kernels_flash_attn = get_kernel("kernels-community/flash-attn")
+
+
+def hf_flash_attention(query, key, value):
+ """HuggingFace Kernels Flash Attention"""
+ return hf_kernels_flash_attn.fwd(query, key, value, is_causal=False)[0]
+
+
+kbt.add(
+ "hf_kernels_flash_attn",
+ hf_flash_attention,
+ tags={"family": "hf-kernels", "backend": "flash-attn", "compile": "none"},
+)
+
+if __name__ == "__main__":
+ device = "cuda" if torch.cuda.is_available() else "cpu"
+
+ if device == "cpu":
+ print("HF Kernels Flash Attention requires CUDA - skipping benchmark")
+ sys.exit(0)
+
+ dtype = "bfloat16"
+
+ # Flux-like workloads
+ base = 1024
+ flux_sizes = [128, 256, 320, 384, 448, 512]
+ heads = 24
+ head_dim = 128
+
+ wl = []
+ for L in flux_sizes:
+ wl.append(
+ {
+ "name": f"flux_L{L}",
+ "batch": 1,
+ "seq_len": base + L,
+ "heads": heads,
+ "head_dim": head_dim,
+ "dtype": dtype,
+ "device": device,
+ "seed": 0,
+ }
+ )
+
+ kbt.run(
+ wl,
+ jsonl="attn.jsonl",
+ reps=5,
+ warmup=2,
+ gen=kbt.attn.gen_qkv,
+ ref=kbt.attn.ref_math,
+ cmp=kbt.attn.cmp_allclose,
+ )
+ kbt.summarize(["attn.jsonl"])
+
+
+impl wl p50(ms) ok
+hf_kernels_flash_attn flux_L128 0.35 True
+hf_kernels_flash_attn flux_L256 0.38 True
+hf_kernels_flash_attn flux_L320 0.50 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 0.56 True
+
+
+
+▶ UV Install Logs
+
+Fetching 20 files: 0%| | 0/20 [00:00<?, ?it/s]
+Fetching 20 files: 5%|▌ | 1/20 [00:00<00:03, 6.33it/s]
+Fetching 20 files: 10%|█ | 2/20 [00:01<00:10, 1.75it/s]
+Fetching 20 files: 100%|██████████| 20/20 [00:01<00:00, 19.65it/s]
+
+
+