Rotary Position Embeddings Benchmarks - Aggregated Results

This document combines benchmark results from multiple Rotary Position Embeddings implementations.

Combined Summary and Visualization

▶ code ▼ output ▶ uv-logs | Cell: combine | 4.37s | Raw

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "numpy",
#     "torch==2.8.0",
#     "kernels-benchmark-tools",
#     "matplotlib",
# ]
#
# [tool.uv.sources]
# kernels-benchmark-tools = { path = "../../../../../tools", editable = true }
# ///
from kernels_benchmark_tools.core.visuals import generate_combined_results

# Map display names to uvnote environment variables
cache_env_map = {
    "HF Kernels Rotary": "UVNOTE_FILE_HF_KERNELS_ROTARY_BENCHMARK",
    "PyTorch Rotary": "UVNOTE_FILE_TORCH_ROTARY_BENCHMARK",
}

# Generate combined results with visualization
generate_combined_results(
    cache_env_map=cache_env_map,
    output_filename="rotary.jsonl",
    svg_filename="latency.svg"
)

======================================================================
LOADING BENCHMARK DATA
======================================================================
✓ HF Kernels Rotary             : /__w/kernels-benchmarks/kernels-benchmarks/benches/rotary/impls/.uvnote/cache/3884170bda871392d403d55c822a8b7de8970f81c4733ae7630938c3bf0db88a
✓ PyTorch Rotary                : /__w/kernels-benchmarks/kernels-benchmarks/benches/rotary/impls/.uvnote/cache/abf801d6445dfa81a8dd7b2e6257930c39c18160a9b97a739858c3b244e16cc5

  ✓ Found HF Kernels Rotary
     Path: /__w/kernels-benchmarks/kernels-benchmarks/benches/rotary/impls/.uvnote/cache/3884170bda871392d403d55c822a8b7de8970f81c4733ae7630938c3bf0db88a/rotary.jsonl
  ✓ Found PyTorch Rotary
     Path: /__w/kernels-benchmarks/kernels-benchmarks/benches/rotary/impls/.uvnote/cache/abf801d6445dfa81a8dd7b2e6257930c39c18160a9b97a739858c3b244e16cc5/rotary.jsonl

======================================================================
Summary: 2 found, 0 skipped, 0 missing
======================================================================

COMBINED BENCHMARK SUMMARY

impl                     wl                  p50(ms)  ok
hf_kernels_rotary        cuda_B1_S128_H32_D128_R64     0.09  True
hf_kernels_rotary        cuda_B1_S128_H32_D64_R32     0.09  True
hf_kernels_rotary        cuda_B1_S128_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B1_S128_H8_D64_R32     0.08  True
hf_kernels_rotary        cuda_B1_S2048_H32_D128_R64     0.26  True
hf_kernels_rotary        cuda_B1_S2048_H32_D64_R32     0.09  True
hf_kernels_rotary        cuda_B1_S2048_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B1_S2048_H8_D64_R32     0.09  True
hf_kernels_rotary        cuda_B1_S512_H32_D128_R64     0.09  True
hf_kernels_rotary        cuda_B1_S512_H32_D64_R32     0.09  True
hf_kernels_rotary        cuda_B1_S512_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B1_S512_H8_D64_R32     0.09  True
hf_kernels_rotary        cuda_B2_S128_H32_D128_R64     0.09  True
hf_kernels_rotary        cuda_B2_S128_H32_D64_R32     0.09  True
hf_kernels_rotary        cuda_B2_S128_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B2_S128_H8_D64_R32     0.09  True
hf_kernels_rotary        cuda_B2_S2048_H32_D128_R64     0.85  True
hf_kernels_rotary        cuda_B2_S2048_H32_D64_R32     0.26  True
hf_kernels_rotary        cuda_B2_S2048_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B2_S2048_H8_D64_R32     0.09  True
hf_kernels_rotary        cuda_B2_S512_H32_D128_R64     0.09  True
hf_kernels_rotary        cuda_B2_S512_H32_D64_R32     0.09  True
hf_kernels_rotary        cuda_B2_S512_H8_D128_R64     0.09  True
hf_kernels_rotary        cuda_B2_S512_H8_D64_R32     0.09  True
torch_eager              cuda_B1_S128_H32_D128_R64     0.22  True
torch_eager              cuda_B1_S128_H32_D64_R32     0.23  True
torch_eager              cuda_B1_S128_H8_D128_R64     0.23  True
torch_eager              cuda_B1_S128_H8_D64_R32     0.17  True
torch_eager              cuda_B1_S2048_H32_D128_R64     0.23  True
torch_eager              cuda_B1_S2048_H32_D64_R32     0.22  True
torch_eager              cuda_B1_S2048_H8_D128_R64     0.22  True
torch_eager              cuda_B1_S2048_H8_D64_R32     0.22  True
torch_eager              cuda_B1_S512_H32_D128_R64     0.22  True
torch_eager              cuda_B1_S512_H32_D64_R32     0.22  True
torch_eager              cuda_B1_S512_H8_D128_R64     0.22  True
torch_eager              cuda_B1_S512_H8_D64_R32     0.22  True
torch_eager              cuda_B2_S128_H32_D128_R64     0.22  True
torch_eager              cuda_B2_S128_H32_D64_R32     0.22  True
torch_eager              cuda_B2_S128_H8_D128_R64     0.22  True
torch_eager              cuda_B2_S128_H8_D64_R32     0.22  True
torch_eager              cuda_B2_S2048_H32_D128_R64     0.64  True
torch_eager              cuda_B2_S2048_H32_D64_R32     0.23  True
torch_eager              cuda_B2_S2048_H8_D128_R64     0.23  True
torch_eager              cuda_B2_S2048_H8_D64_R32     0.22  True
torch_eager              cuda_B2_S512_H32_D128_R64     0.22  True
torch_eager              cuda_B2_S512_H32_D64_R32     0.23  True
torch_eager              cuda_B2_S512_H8_D128_R64     0.22  True
torch_eager              cuda_B2_S512_H8_D64_R32     0.22  True

GENERATING COMBINED VISUALIZATION

Loaded 48 records
✓ Visualization saved as latency.svg
Saved latency.png
✓ Visualization saved as latency.svg
✓ SVG visualization ready!

ANALYSIS COMPLETE
Total implementations analyzed: 2

Implementations included:
  ✓ HF Kernels Rotary
  ✓ PyTorch Rotary

▶ UV Install Logs

Artifacts:

latency.svg