Dataset Viewer
Auto-converted to Parquet Duplicate
name
stringlengths
16
64
description
stringlengths
69
172
hf_id
stringclasses
1 value
axes
stringlengths
79
1.08k
inputs
stringclasses
10 values
outputs
stringclasses
9 values
reference
stringlengths
70
7.23k
custom_inputs_entrypoint
null
workloads
stringlengths
1.1k
41.7k
001_fused_add_rmsnorm_h2048
Fused Add + RMSNorm with hidden_size=2048 for Qwen3-30B-A3B. Epsilon is fixed at 1e-6.
{"batch_size": {"type": "var"}, "hidden_size": {"type": "const", "value": 2048}}
{"hidden_states": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "residual": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "weight": {"shape": ["hidden_size"], "dtype": "bfloat16"}}
{"output": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}}
import torch @torch.no_grad() def run(hidden_states, residual, weight): _, hidden_size = hidden_states.shape # Check constants assert hidden_size == 2048 EPS = 1e-6 x = hidden_states.to(torch.float32) + residual.to(torch.float32) inv_rms = torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + EPS...
null
[{"uuid": "906efdc6-aa3b-4b4f-a291-3a2be509d354", "axes": {"batch_size": 6}, "inputs": {"hidden_states": {"type": "random"}, "residual": {"type": "random"}, "weight": {"type": "random"}}}, {"uuid": "1526086b-0c06-41a1-b7b9-47b5cfc830d7", "axes": {"batch_size": 1}, "inputs": {"hidden_states": {"type": "random"}, "residu...
002_fused_add_rmsnorm_h4096
Fused Add + RMSNorm with hidden_size=4096 for Llama-3.1-8B. Epsilon is fixed at 1e-5.
{"batch_size": {"type": "var"}, "hidden_size": {"type": "const", "value": 4096}}
{"hidden_states": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "residual": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "weight": {"shape": ["hidden_size"], "dtype": "bfloat16"}}
{"output": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}}
import torch @torch.no_grad() def run(hidden_states, residual, weight): _, hidden_size = hidden_states.shape # Check constants assert hidden_size == 4096 EPS = 1e-5 x = hidden_states.to(torch.float32) + residual.to(torch.float32) inv_rms = torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + EPS...
null
[{"uuid": "9f5150a0-744c-4699-a970-f15b0ea14a38", "axes": {"batch_size": 7}, "inputs": {"hidden_states": {"type": "random"}, "residual": {"type": "random"}, "weight": {"type": "random"}}}, {"uuid": "146cc899-3cd6-4713-a85f-0cca97509a48", "axes": {"batch_size": 1}, "inputs": {"hidden_states": {"type": "random"}, "residu...
003_fused_add_rmsnorm_h7168
Fused Add + RMSNorm with hidden_size=7168 for DeepSeek-V3/R1. Epsilon is fixed at 1e-6.
{"batch_size": {"type": "var"}, "hidden_size": {"type": "const", "value": 7168}}
{"hidden_states": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "residual": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}, "weight": {"shape": ["hidden_size"], "dtype": "bfloat16"}}
{"output": {"shape": ["batch_size", "hidden_size"], "dtype": "bfloat16"}}
import torch @torch.no_grad() def run(hidden_states, residual, weight): _, hidden_size = hidden_states.shape # Check constants assert hidden_size == 7168 EPS = 1e-6 x = hidden_states.to(torch.float32) + residual.to(torch.float32) inv_rms = torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + EPS...
null
[{"uuid": "c4831914-583d-4dbf-a185-e246bf426098", "axes": {"batch_size": 7}, "inputs": {"hidden_states": {"type": "random"}, "residual": {"type": "random"}, "weight": {"type": "random"}}}, {"uuid": "d8b7a476-6877-429c-821d-35c779c56087", "axes": {"batch_size": 1}, "inputs": {"hidden_states": {"type": "random"}, "residu...
004_gemm_n128_k2048
General matrix multiply (GEMM) C = A @ B.T. Captured from Qwen 3 30B A3B moe.gate.
{"M": {"type": "var"}, "N": {"type": "const", "value": 128}, "K": {"type": "const", "value": 2048}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "6ba7c7de-dc5a-48d2-8ada-1382feb5ceac", "axes": {"M": 6}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "91e43228-c0a7-457e-9424-1d3461621362", "axes": {"M": 1}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "4fa7da89-3488-4343-a154-591cbbf584eb", "axes": {"...
005_gemm_n256_k7168
General matrix multiply (GEMM) C = A @ B.T. Captured from DeepSeek-V3 moe.gate.
{"M": {"type": "var"}, "N": {"type": "const", "value": 256}, "K": {"type": "const", "value": 7168}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "b01eaa0e-4579-4d3a-a2d5-4c32360db254", "axes": {"M": 1}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "5f8e8baf-eee5-4d68-b72b-685d15714e0a", "axes": {"M": 32}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "bb5b51a5-34df-4a8c-9f07-2d3eb3d07627", "axes": {...
006_gemm_n2048_k4096
General matrix multiply (GEMM) C = A @ B.T. Captured from Qwen 3 30B A3B attn.o_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 2048}, "K": {"type": "const", "value": 4096}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "bcc2cc41b7464e54b0c4d11e65da96d8", "axes": {"M": 6}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "ff70a7abf26848aba6424c7571f777d4", "axes": {"M": 1}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "f4513ae887ae4d42946dcf6630671c22", "axes": {"M": 34}, "in...
007_gemm_n4096_k4096
General matrix multiply (GEMM) C = A @ B.T. Captured from Llama 3.1 8B attn.o_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 4096}, "K": {"type": "const", "value": 4096}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "280860e6-08f0-427c-b7c5-9cffcfab1a10", "axes": {"M": 256}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "2e90109a-282e-484f-b94b-61f49e72fde2", "axes": {"M": 248}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "e7c939ae-2083-4b6f-a51a-8c76ffd08926", "axes"...
008_gemm_n4096_k14336
General matrix multiply (GEMM) C = A @ B.T. Captured from Llama 3.1 8B mlp.down_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 4096}, "K": {"type": "const", "value": 14336}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "501f81c6-97cd-477c-b091-5068a52c385d", "axes": {"M": 256}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "9e5d631f-b19a-454c-86d5-522d389afaa9", "axes": {"M": 248}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "6e497f43-f8e1-4eff-a888-3413c107b918", "axes"...
009_gemm_n5120_k2048
General matrix multiply (GEMM) C = A @ B.T. Captured from Qwen 3 30B A3B attn.qkv_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 5120}, "K": {"type": "const", "value": 2048}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "b0a78c88-11ee-4f4e-a818-184c13c6f2e5", "axes": {"M": 6}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "107f3ae3-6c78-43c6-acd3-5e080208f536", "axes": {"M": 1}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "22298d10-a6b5-4f90-9130-11335cb6b6ec", "axes": {"...
010_gemm_n6144_k4096
General matrix multiply (GEMM) C = A @ B.T. Captured from Llama 3.1 8B attn.qkv_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 6144}, "K": {"type": "const", "value": 4096}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "000bed54-6e96-4655-a8ef-61dd810c6e1b", "axes": {"M": 256}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "342e6dfa-076a-499f-8921-054f503100b7", "axes": {"M": 248}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "a73527f6-7701-43ac-8032-077e7461926c", "axes"...
011_gemm_n28672_k4096
General matrix multiply (GEMM) C = A @ B.T. Captured from Llama 3.1 8B mlp.gate_up_proj.
{"M": {"type": "var"}, "N": {"type": "const", "value": 28672}, "K": {"type": "const", "value": 4096}}
{"A": {"shape": ["M", "K"], "dtype": "float16"}, "B": {"shape": ["N", "K"], "dtype": "float16"}}
{"C": {"shape": ["M", "N"], "dtype": "float16"}}
import torch def run(A, B): C = torch.matmul(A, B.T) return C
null
[{"uuid": "181e4a1e-6dec-4bca-8a21-703c35d0762a", "axes": {"M": 256}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "4723ac5c-205a-454d-9318-2ea368b0a57d", "axes": {"M": 248}, "inputs": {"A": {"type": "random"}, "B": {"type": "random"}}}, {"uuid": "014b776b-dc2b-4468-b5ae-aa0d7d1491aa", "axes"...
012_gqa_paged_decode_h32_kv4_d128_ps1
Batched Grouped Query Attention decode with a paged KV cache. Captured from Qwen3-30B-A3B.
{"batch_size": {"type": "var", "description": "Total number of query tokens."}, "num_qo_heads": {"type": "const", "value": 32, "description": null}, "num_kv_heads": {"type": "const", "value": 4, "description": null}, "head_dim": {"type": "const", "value": 128, "description": null}, "num_pages": {"type": "var", "descrip...
{"q": {"shape": ["batch_size", "num_qo_heads", "head_dim"], "dtype": "bfloat16", "description": null}, "k_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16", "description": null}, "v_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16",...
{"output": {"shape": ["batch_size", "num_qo_heads", "head_dim"], "dtype": "bfloat16", "description": null}, "lse": {"shape": ["batch_size", "num_qo_heads"], "dtype": "float32", "description": "The 2-based log-sum-exp of attention logits."}}
import torch import math @torch.no_grad() def run(q, k_cache, v_cache, kv_indptr, kv_indices, sm_scale): batch_size, num_qo_heads, head_dim = q.shape _, page_size, num_kv_heads, _ = k_cache.shape len_indptr = kv_indptr.shape[0] num_kv_indices = kv_indices.shape[0] # Check constants assert num...
null
[{"uuid": "0c2489b2-f878-428b-b1bd-d0c6d4c39338", "axes": {"batch_size": 1, "num_pages": 8, "len_indptr": 2, "num_kv_indices": 7}, "inputs": {"q": {"type": "random"}, "k_cache": {"type": "random"}, "v_cache": {"type": "random"}, "kv_indptr": {"type": "safetensors", "path": "data/flashinfer-trace/blob/workloads/gqa_page...
013_gqa_paged_decode_h32_kv8_d128_ps1
Batched Grouped Query Attention decode with a paged KV cache. Captured from Llama-3.1-8B.
{"batch_size": {"type": "var", "description": "Total number of query tokens."}, "num_qo_heads": {"type": "const", "value": 32}, "num_kv_heads": {"type": "const", "value": 8}, "head_dim": {"type": "const", "value": 128}, "num_pages": {"type": "var"}, "page_size": {"type": "const", "value": 1}, "len_indptr": {"type": "va...
{"q": {"shape": ["batch_size", "num_qo_heads", "head_dim"], "dtype": "bfloat16"}, "k_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16"}, "v_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16"}, "kv_indptr": {"shape": ["len_indptr"], "...
{"output": {"shape": ["batch_size", "num_qo_heads", "head_dim"], "dtype": "bfloat16"}, "lse": {"shape": ["batch_size", "num_qo_heads"], "dtype": "float32", "description": "The 2-based log-sum-exp of attention logits."}}
import torch import math @torch.no_grad() def run(q, k_cache, v_cache, kv_indptr, kv_indices, sm_scale): batch_size, num_qo_heads, head_dim = q.shape _, page_size, num_kv_heads, _ = k_cache.shape len_indptr = kv_indptr.shape[0] num_kv_indices = kv_indices.shape[0] # Check constants assert num...
null
[{"uuid": "e2142798-a0ce-49dd-9eaa-72727ad69f0f", "axes": {"batch_size": 1, "num_pages": 9316, "len_indptr": 2, "num_kv_indices": 73}, "inputs": {"q": {"type": "random"}, "k_cache": {"type": "random"}, "v_cache": {"type": "random"}, "kv_indptr": {"type": "safetensors", "path": "data/flashinfer-trace/blob/workloads/gqa_...
014_gqa_paged_prefill_causal_h32_kv4_d128_ps1
Batched Grouped Query Attention prefill with a paged KV cache. Causal mask is applied. Captured from Qwen3-30B-A3B during incremental prefill.
{"num_qo_heads": {"type": "const", "value": 32}, "num_kv_heads": {"type": "const", "value": 4}, "head_dim": {"type": "const", "value": 128}, "page_size": {"type": "const", "value": 1}, "len_indptr": {"type": "var", "description": "Length of indptr arrays. Should be the same for qo_indptr and kv_indptr (batch_size + 1)....
{"q": {"shape": ["total_q", "num_qo_heads", "head_dim"], "dtype": "bfloat16"}, "k_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16"}, "v_cache": {"shape": ["num_pages", "page_size", "num_kv_heads", "head_dim"], "dtype": "bfloat16"}, "qo_indptr": {"shape": ["len_indptr"], "dty...
{"output": {"shape": ["total_q", "num_qo_heads", "head_dim"], "dtype": "bfloat16"}, "lse": {"shape": ["total_q", "num_qo_heads"], "dtype": "float32", "description": "The 2-based log-sum-exp of attention logits."}}
import torch import math @torch.no_grad() def run(q, k_cache, v_cache, qo_indptr, kv_indptr, kv_indices, sm_scale): total_q, num_qo_heads, head_dim = q.shape num_pages, page_size, num_kv_heads, _ = k_cache.shape len_indptr = qo_indptr.shape[0] num_kv_indices = kv_indices.shape[0] # Check constant...
null
[{"uuid": "c3c5535c-1829-4618-b629-129c0190dfc4", "axes": {"len_indptr": 2, "total_q": 1, "num_kv_indices": 33, "num_pages": 50}, "inputs": {"q": {"type": "random"}, "k_cache": {"type": "random"}, "v_cache": {"type": "random"}, "qo_indptr": {"type": "safetensors", "path": "data/flashinfer-trace/blob/workloads/gqa_paged...
End of preview. Expand in Data Studio

nvidia/SOL-ExecBench(複製)

データの内容

オープンソースの HuggingFace モデルから導いた、CUDA カーネルの計算仕様と参照実装のベンチマークのデータ。1 行が 1 つのカーネル問題で、計算の説明 description、軸と入出力の仕様 axes・inputs・outputs、PyTorch の参照実装 reference、具体的な入力条件 workloads を持つ。HuggingFace の Transformers と Diffusers からモデル構造を集め、LLM エージェントで計算のサブグラフを抜き出して JSON の定義に直した。教師ラベルや人手のアノテーションは含まない。

データの概要

項目 値
行数 235(4 構成・4 分割の合計)
列数 9
容量 888.59 KB(ファイル 10 本)
ファイルの形式 parquet 4、py 2
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
FlashInfer-Bench train 26 9
L1 train 94 9
L2 train 82 9
Quant train 33 9

列(構成 FlashInfer-Bench・分割 train。構成ごとに列が違う)

列 型 先頭の行の値
name string 001_fused_add_rmsnorm_h2048
description string Fused Add + RMSNorm with hidden_size=2048 for Qw…
hf_id string
axes string {"batch_size": {"type": "var"}, "hidden_size": {…
inputs string {"hidden_states": {"shape": ["batch_size", "hidd…
outputs string {"output": {"shape": ["batch_size", "hidden_size…
reference string import torch @torch.no_grad() def run(hidden_st…
custom_inputs_entrypoint null null
workloads string [{"uuid": "906efdc6-aa3b-4b4f-a291-3a2be509d354"…

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

nvidia/SOL-ExecBench をリビジョン 63699402f003496acc3af4eb534a5304a8ac1ea9 の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

アノテーションの方法

データセットの作成日

2026 年 1 月

微調整に使うときの注意

  • 調査メモの台帳での区分: D(記載なし・other・unknown・資料間の食い違い)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
nvidia_nemotron.md 付録. 取得した nvidia 組織のデータセット全件の簡易一覧 (L381) other 確認済 確認印なし(各ファイルの凡例により確認済扱い) / 原文注記: × 数学・コード・科学・計測

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


データセットの説明

SOL(Speed Of Light)ExecBench は、オープンソースの HuggingFace のモデルアーキテクチャから導出された、235 件のカーネル単位の計算ワークロード仕様から成る、実世界の CUDA カーネルのベンチマーク用データセットである。問題は、テキスト、画像、音声モデルの順伝播と逆伝播の計算を含む、AI モデルのワークロードの幅広い範囲にわたる。また、行列積、畳み込み、attention の各種、Mixture-of-Experts、正規化(norm)といった中核的なアルゴリズムを、FP32、BF16、FP16、FP8、NVFP4 のデータ型にわたって含む。実行時間は参照実装の速度と比較されるほか、ハードウェアの理論上の「光速」(ハードウェアのピーク演算性能とメモリ帯域の上限が与える、カーネルが実行できる最速の速度)とも比較される。

本データセットは、HuggingFace の Transformers と Diffusers のリポジトリからモデルアーキテクチャを収集し、LLM エージェントを使ってカーネルに関係する計算のサブグラフを抽出して構築された。抽出されたサブグラフは、構造化された JSON の定義に変換されている。

各エントリは、次を含む計算カーネルのワークロードを定義する:

  • 記号的な軸の定義(定数、変数、式による定義)
  • テンソルの入力と出力の仕様(形状、dtype、制約)
  • 必須の run() エントリポイントを持つ PyTorch の参照実装
  • 具体的な軸の値と入力の生成方法を指定するワークロードのインスタンス化

評価は、専用の評価器(GitHub)と 公開リーダーボード で行える。光速スコアは SOLAR で解析的に計算され、公開リーダーボードで閲覧できる。

詳細は 技術報告 を参照のこと。

本データセットは商用利用可能である。

データセットの所有者

NVIDIA Corporation

データセットの作成日

2026 年 1 月

ライセンス/利用規約

NVIDIA Evaluation Dataset License Agreement

想定される用途

SOL ExecBench は、AI ベースのカーネル生成システム、コンパイラエージェント、自動化された性能最適化ツールを開発する研究者とエンジニアを対象とする。

主な用途は次のとおりである:

  • AI ベースのカーネル生成モデルのベンチマーク
  • エージェントベースのコード合成システムの評価
  • コンパイラの最適化戦略の検証
  • 生成された CUDA カーネルの正しさと性能の測定
  • 自動化された性能エンジニアリングシステムの研究開発

本データセットは主として、計算カーネルを生成または最適化する AI システムを評価するための研究用ベンチマークとして想定されている。

データセットの特徴づけ

データ収集の方法

  • ハイブリッド: 自動(HuggingFace のリポジトリからの収集)、合成(フロンティア LLM エージェントによって生成されたサブグラフの抽出)

教師ラベルや人手によるアノテーションは含まれない。本データセットは、構造化されたカーネルの定義と、実行可能な参照実装から成る。

本データセットには、個人データ、利用者データ、機密情報は含まれない。元となる資料はすべて、公開されているオープンソースの HuggingFace のモデルリポジトリに由来する。

データセットの形式

本データセットは 2 つの形式で提供される:

Parquet(HuggingFace Dataset Viewer 用)

data/ にある事前に作られた Parquet ファイルは、カーネルごとに 1 行を持ち、次の列を含む:

列 型 説明
name string カーネルの識別子
description string 演算の自然言語による説明
axes JSON string 記号的な軸の定義(const / var / expr)
inputs JSON string 入力テンソルの仕様(形状、dtype)
outputs JSON string 出力テンソルの仕様(形状、dtype)
reference string run() エントリポイントを持つ Python の参照実装
custom_inputs_entrypoint string (nullable) 存在する場合は、カスタム入力の生成関数の名前
workloads JSON string 具体的なワークロードのインスタンス化のリスト(軸の値、入力の仕様、許容誤差)

これらのファイルは 4 つのサブセット(config)に分けられている: FlashInfer-Bench、L1、L2、Quant。

モダリティ: コード/構造化テキスト

データセットの規模

レコード数: 235 件のカーネル問題

  • FlashInfer-Bench 1.0 由来 26 件
  • Level 1(L1)94 件
  • Level 2(L2)82 件
  • 量子化されたグラフ 33 件

特徴の数(問題ごと):

  • カーネル名
  • 演算子の種類
  • 軸の仕様(const / var / expr)
  • 入力テンソルの仕様
  • 出力テンソルの仕様
  • 参照実装のコード
  • ワークロードのインスタンス化(軸の値と入力の仕様)

データの総容量: 小規模なデータセット(1000 ファイル未満、JSON ベース、1 GB 未満)

参考資料

引用

@misc{lin2026solexecbench,
      title={SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits}, 
      author={Edward Lin, Sahil Modi, Siva Kumar Sastry Hari, Qijing Huang, Zhifan Ye, Nestor Qin, Fengzhe Zhou, Yuan Zhang, Jingquan Wang, Sana Damani, Dheeraj Peri, Ouye Xie, Aditya Kane, Moshe Maor, Michael Behar, Triston Cao, Rishabh Mehta, Vartika Singh, Vikram Sharma Mailthody, Terry Chen, Zihao Ye, Hanfeng Chen, Tianqi Chen, Vinod Grover, Wei Chen, Wei Liu, Eric Chung, Luis Ceze, Roger Bringmann, Cyril Zeller, Michael Lightstone, Christos Kozyrakis, Humphrey Shi},
      year={2026},
      eprint={2603.19173},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2603.19173}, 
}

倫理的配慮

NVIDIA は、信頼できる AI は共同の責任であると考えており、幅広い AI アプリケーションの開発を可能にするための方針と慣行を定めている。利用規約に従ってダウンロードまたは利用する場合、開発者は社内の開発チームと協力して、本データセットが該当する業界と用途の要件を満たし、予期しない製品の誤用に対処するようにすべきである。

品質、リスク、セキュリティ上の脆弱性、NVIDIA AI に関する懸念は こちら から報告すること。

原文(英語)

Dataset Description

SOL (Speed Of Light) ExecBench is a real-world CUDA kernel benchmarking dataset of 235 kernel-level computational workload specifications derived from open-source HuggingFace model architectures. The problems span a wide range of AI model workloads — covering text, vision, and speech models' forward and backward passes — and include core algorithms such as matrix multiplications, convolutions, attention variants, mixture-of-experts, and norms across FP32, BF16, FP16, FP8, and NVFP4 data types. Runtimes will be compared against reference implementation speed but also theoretical hardware "speed of light": the fastest that kernels can run given the hardware's peak compute and memory bandwidth limits.

The dataset was constructed by scraping model architectures from the HuggingFace Transformers and Diffusers repositories and using an LLM agent to extract kernel-relevant computational subgraphs, which are then transformed into structured JSON definitions.

Each entry defines a computational kernel workload including:

  • Symbolic axis definitions (constant, variable, expression-based)
  • Tensor input/output specifications (shape, dtype, constraints)
  • A PyTorch reference implementation with a required run() entrypoint
  • Workload instantiations specifying concrete axis values and input generation strategies

Evaluation is supported via a custom evaluator (GitHub) and a public leaderboard. Speed-of-light scores are analytically computed with SOLAR and are available on the public leaderboard.

More information can be found in our technical report.

This dataset is ready for commercial use.

Dataset Owner

NVIDIA Corporation

Dataset Creation Date

January 2026

License/Terms of Use

NVIDIA Evaluation Dataset License Agreement

Intended Usage

SOL ExecBench is intended for researchers and engineers developing AI-based kernel generation systems, compiler agents, and automated performance optimization tools.

Primary use cases include:

  • Benchmarking AI-based kernel generation models
  • Evaluating agent-based code synthesis systems
  • Testing compiler optimization strategies
  • Measuring correctness and performance of generated CUDA kernels
  • Research and development of automated performance engineering systems

The dataset is intended primarily as a research benchmark for evaluation of AI systems that generate or optimize computational kernels.

Dataset Characterization

Data Collection Method

  • Hybrid: Automated (HuggingFace repository scraping), Synthetic (Frontier-LLM- agent-generated subgraph extraction)

No supervised labels or human annotations are included. The dataset consists of structured kernel definitions and executable reference implementations.

The dataset does not contain personal data, user data, or confidential information. All source material is derived from publicly available open-source HuggingFace model repositories.

Dataset Format

The dataset is provided in two formats:

Parquet (for HuggingFace Dataset Viewer)

Pre-built Parquet files in data/ provide one row per kernel with the following columns:

Column Type Description
name string Kernel identifier
description string Natural-language description of the operation
axes JSON string Symbolic axis definitions (const / var / expr)
inputs JSON string Input tensor specifications (shape, dtype)
outputs JSON string Output tensor specifications (shape, dtype)
reference string Python reference implementation with a run() entrypoint
custom_inputs_entrypoint string (nullable) Name of custom input generation function, if any
workloads JSON string List of concrete workload instantiations (axes values, input specs, tolerances)

These files are split into four subsets (configs): FlashInfer-Bench, L1, L2, and Quant.

Modality: Code / Structured Text

Dataset Quantification

Record Count: 235 kernel problems

  • 26 from FlashInfer-Bench 1.0
  • 94 Level 1 (L1)
  • 82 Level 2 (L2)
  • 33 Quantized graphs

Feature Count (per problem):

  • Kernel name
  • Operator type
  • Axis specifications (const / var / expr)
  • Input tensor specifications
  • Output tensor specifications
  • Reference implementation code
  • Workload instantiations (axes values + input specifications)

Total Data Storage: Small-scale dataset (< 1000 files; JSON-based; < 1 GB)

References

Citation

@misc{lin2026solexecbench,
      title={SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits}, 
      author={Edward Lin, Sahil Modi, Siva Kumar Sastry Hari, Qijing Huang, Zhifan Ye, Nestor Qin, Fengzhe Zhou, Yuan Zhang, Jingquan Wang, Sana Damani, Dheeraj Peri, Ouye Xie, Aditya Kane, Moshe Maor, Michael Behar, Triston Cao, Rishabh Mehta, Vartika Singh, Vikram Sharma Mailthody, Terry Chen, Zihao Ye, Hanfeng Chen, Tianqi Chen, Vinod Grover, Wei Chen, Wei Liu, Eric Chung, Luis Ceze, Roger Bringmann, Cyril Zeller, Michael Lightstone, Christos Kozyrakis, Humphrey Shi},
      year={2026},
      eprint={2603.19173},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2603.19173}, 
}

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Downloads last month
70

Paper for HYUKI5/nvidia__SOL-ExecBench