The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
median_us: double
cov: double
samples: list<item: double>
child 0, item: double
shapes: list<item: int64>
child 0, item: int64
iters: int64
seeds: list<item: int64>
child 0, item: int64
kernel: string
cos: double
arch: string
pack: int64
to
{'kernel': Value('string'), 'pack': Value('int64'), 'arch': Value('string'), 'shapes': List(Value('int64')), 'seeds': List(Value('int64')), 'iters': Value('int64'), 'cos': Value('float64')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
median_us: double
cov: double
samples: list<item: double>
child 0, item: double
shapes: list<item: int64>
child 0, item: int64
iters: int64
seeds: list<item: int64>
child 0, item: int64
kernel: string
cos: double
arch: string
pack: int64
to
{'kernel': Value('string'), 'pack': Value('int64'), 'arch': Value('string'), 'shapes': List(Value('int64')), 'seeds': List(Value('int64')), 'iters': Value('int64'), 'cos': Value('float64')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
AVO Release 30 — 可复测的 GPU 内核优化产物
本数据集收录 AVO(assembly-level kernel optimization agent)在 AMD MI308X(gfx942)上完成的 26 个算子的优化产物。Level 1 与 Level 2 均有覆盖。
与一般的优化实验记录不同,这里的每个算子都自带一套 pair harness:把当初真实发生的那次 GPU dispatch 完整抓了下来(内核二进制、kernarg、全部内存区域),任何人拿到之后可以在自己的 gfx942 卡上原样回放,独立复核正确性与加速比,不需要原始框架、模型或数据。
快速开始
cd level1/<算子名>/harness
make -j8
python run_harness.py
输出 result.json,包含正确性判定与加速比。
每个算子的目录结构
level{1,2}/<算子名>/
├── harness/ # 可独立复测的 pair harness
│ ├── capture/ # 真实 dispatch 的完整快照
│ ├── ref.co # 基准内核(= capture/kernel.hsaco,字节相同)
│ ├── opt.co # 优化后内核
│ ├── harness_manifest.json # 所有输入的 sha256
│ ├── result.json # 我们这边跑出的结果
│ ├── run_harness.py / Makefile / CMakeLists.txt
│ └── vendor/ # 编译所需第三方头文件
└── avo_kernel_<ts>/ # AVO 原始优化过程记录(溯源用)
└── kernel/ 或 baseline_co/
├── director_validation.json
├── avo_state/attempts.jsonl
└── ...
index.json 是全量索引,每个算子一条,含 harness 与 Director 两套结果。
两个口径,以 harness 为准
数据集里同时存在两个加速比,它们不可直接比较:
| Harness | Director | |
|---|---|---|
| 基准二进制 | 真实被 dispatch 的内核(capture/kernel.hsaco) |
AVO 自己重编的基准 |
| 负载 | 真实抓到的那次 dispatch | AVO 侧构造的形状 |
| 方法 | 同 capture 内 ABBA 交错,speedup = sqrt(fwd/rev) |
AVO 自有成对采样 |
| 26 个算子几何平均 | 1.3637x | 1.5101x |
复测请以 harness 为准。 它是这个数据集里唯一可被第三方独立复现的口径;Director 的数字保留在
director_validation.json 里仅作溯源参考。
harness 每次测量同时给出 slot_bias = sqrt(fwd·rev),用于检测两个测量槽位的系统性偏差。
本批 26 个算子的 slot_bias 落在 0.9914 ~ 1.0089,说明槽位对称、比值可信。
只报比值,不报绝对时间。 同卡邻居的功耗争用与 capture 虚拟地址的分配差异会让绝对微秒数 出现数十个百分点的漂移,只有同一次 ABBA 内部的比值是稳定的。
result.json里的*_us字段仅供调试,不应跨机器比较。
正确性判据
| 判据 | 算子数 | 含义 |
|---|---|---|
| 逐字节 | 21 | 优化后内核写出的每一个输出字节与基准完全相同 |
| 容差 | 5 | 仅对声明的字节范围放宽,范围之外仍要求逐字节相同 |
采用容差的算子,其 harness/capture/validation_policy.json 明确写出了放宽的范围、判据与阈值。
其中 4 个用余弦相似度 ≥ 0.9999;37_FrobeniusNorm_ 的输出是单个 float32 标量,余弦对一元向量
恒为 1、没有意义,因此改用相对误差上界 rtol=1e-6(约 8 个 float32 eps,实测 4.33e-07)。
容差判据的已知局限: 余弦相似度对"换算法"是不敏感的。例如把一个超越函数替换成多项式 近似,可能在保持 cos ≥ 0.9999 的同时引入单侧系统性误差。若你的场景对数值精度敏感, 请自行检查
opt.co的反汇编,不要只依赖本数据集的通过标记。
关于被排除的算子
原始实验集合共 30 个算子,本数据集发布其中 26 个。另外 4 个未通过 harness 正确性判据,未达发布标准。
这些算子的实验记录仍保留在上游仓库 puyuan-yang/asm_co_bench_ext 与 jinglwan/avo_exp 中。
本数据集为增量发布,后续通过复核的算子会按相同目录约定追加。
环境
- GPU:AMD MI308X,gfx942
- ROCm / HIP 运行时,编译需要
hipcc与cmake - harness 不依赖 PyTorch、Triton 或任何上游框架
- Downloads last month
- 13