Datasets:

Akahsizrr
/

kernelstrain

Error code:   DatasetGenerationError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<M: int64, N: int64, H: int64, Hk: int64, D: int64>
to
{'M': Value('int64'), 'N': Value('int64')}
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<M: int64, N: int64, H: int64, Hk: int64, D: int64>
              to
              {'M': Value('int64'), 'N': Value('int64')}
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

episode_id string	op string	impl string	desc string	signature string	gpu string	arch string	dtype string	shape dict	objective string	baseline_ms float64	best dict	attempts dict	num_steps int64	steps list
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16406459	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.087301	{ "step": 13, "time_ms": 0.020898, "speedup": 4.177 }	{ "kept": 6, "reverted": 25, "errors": 0, "rebench": 0 }	31	[ { "tag": "v0_scalar", "action": "scalar two-pass, shared-mem atomic reduce", "kind": "baseline", "params": {}, "code": "\n// baseline: one block per row, scalar loads, two passes over global memory\n__global__ void rmsnorm_v0(const half* __restrict__ x,\n const half* __...
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16414378	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.087329	{ "step": 14, "time_ms": 0.02081, "speedup": 4.196 }	{ "kept": 6, "reverted": 26, "errors": 1, "rebench": 2 }	33	[ { "tag": "v0_scalar", "action": "scalar two-pass, shared-mem atomic reduce", "kind": "baseline", "params": {}, "code": "\n// baseline: one block per row, scalar loads, two passes over global memory\n__global__ void rmsnorm_v0(const half* __restrict__ x,\n const half* __...
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16422297	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.084062	{ "step": 14, "time_ms": 0.021175, "speedup": 3.97 }	{ "kept": 6, "reverted": 26, "errors": 1, "rebench": 2 }	33	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16430216	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.087286	{ "step": 8, "time_ms": 0.02157, "speedup": 4.047 }	{ "kept": 5, "reverted": 25, "errors": 1, "rebench": 0 }	31	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16438135	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.087822	{ "step": 17, "time_ms": 0.021341, "speedup": 4.115 }	{ "kept": 7, "reverted": 27, "errors": 0, "rebench": 3 }	34	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-fp16-s16446054	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	fp16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.08707	{ "step": 3, "time_ms": 0.022155, "speedup": 3.93 }	{ "kept": 4, "reverted": 20, "errors": 0, "rebench": 2 }	24	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-bf16-s5558987	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	bf16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.08381	{ "step": 14, "time_ms": 0.020939, "speedup": 4.003 }	{ "kept": 6, "reverted": 28, "errors": 0, "rebench": 3 }	34	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-bf16-s5566906	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	bf16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.088531	{ "step": 14, "time_ms": 0.020851, "speedup": 4.246 }	{ "kept": 6, "reverted": 29, "errors": 0, "rebench": 4 }	35	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-bf16-s5574825	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	bf16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.087611	{ "step": 14, "time_ms": 0.021114, "speedup": 4.149 }	{ "kept": 6, "reverted": 26, "errors": 0, "rebench": 1 }	32	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)
cuda_rmsnorm-a100-80g-M=2048,N=4096-bf16-s5582744	rmsnorm	cuda	RMSNorm over last dim, one row per sample	rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16	a100-80g	sm_80	bf16	{ "M": 2048, "N": 4096 }	minimize latency_ms	0.085783	{ "step": 17, "time_ms": 0.021389, "speedup": 4.011 }	{ "kept": 8, "reverted": 25, "errors": 0, "rebench": 0 }	33	[{"tag":"v0_scalar","action":"scalar two-pass, shared-mem atomic reduce","kind":"baseline","params":(...TRUNCATED)

End of preview.

KernelStrain

Long-horizon GPU kernel optimization trajectories for training small models to iterate on CUDA and Triton kernels.

KernelStrain is a large synthetic dataset of kernel-optimization episodes: given a kernel task (shapes, dtype, GPU target, baseline code + timing), a model proposes successive complete kernel candidates, observes simulated benchmark / correctness / compile feedback, and keeps improving over many steps — structural rewrites, parameter sweeps, joint configs, rebenchmarks, and endgame fine-tuning.

Built to teach a small model the loop of kernel work: propose → measure → regress → revert → recover → repeat, for tens of steps per task.

⚠️ All benchmark numbers are produced by an analytic roofline-style simulator — no kernel in this dataset was ever compiled or executed. Timings are plausible, not measured. Kernel source is realistic-looking but unverified; use it for training dynamics, not as production code.

At a glance


Episodes	11,514
SFT step rows	~282k
Multiturn conversations	11,514
Holdout tasks	185 (descriptions only, no solutions)
Op modules	57 (45 distinct ops × CUDA/Triton impls)
Avg steps / episode	~24.5
Step statuses	~94% ok, ~2.6% wrong-result, ~3.3% compile-error
GPU targets	A100 (sm_80), H100 (sm_90), RTX 4090, MI300X (gfx942)

Coverage

Kernels spanning AI inference and training:

Attention: flash prefill, paged decode, MLA decode (latent KV), sliding-window, cross-attention, tree attention (speculative decoding), flash attention backward
Norms: RMSNorm, LayerNorm, fused add+RMSNorm, LayerNorm backward
Activations/fusions: SwiGLU fwd/bwd, bias+GELU, dropout (Philox)
Matmul: fp16 GEMM/GEMV, int4/int8 GEMM, fp8 GEMM, blockwise-scaled fp8 GEMM (DeepGEMM-style), grouped GEMM
Training: AdamW, Adafactor, grad clipping, cross-entropy fwd/bwd, embedding fwd/bwd, reductions, prefix scan
Quant: fp8 tensor quant, int8 KV-cache quant
Sampling: top-k, nucleus (top-p) — sort, histogram-refine, and CDF-pick strategies
SSM/conv: Mamba selective scan (CUDA + Triton), causal Conv1D
MoE: scatter, grouped GEMM, combine
Misc: RoPE, transpose, KV-cache append, embedding gather

Optimization patterns exercised: vectorized loads (uint4/float4/half2), shared-memory staging, cp.async pipelines, TMA, warp shuffles, online softmax, split-K/split-KV with combine kernels, persistent kernels, tensor-core mma/wgmma, atomics, multi-kernel decompositions, occupancy/tile tuning — plus realistic failures: arch-gated compile errors (sm_90-only wgmma, no fp8 on A100, PTX on ROCm), shared-memory overflows, and wrong-result regressions that get reverted.

Files

file	format	contents
`episodes.jsonl`	jsonl	Raw episodes — every step with full source, params, status, timing, smem, keep/revert decisions
`sft_multiturn.jsonl`	chat	Multi-turn conversations; assistant turns contain complete kernel code only (no reasoning traces)
`sft_steps.jsonl`	chat	One row per transition: task + history + current-best → next candidate
`tasks_holdout.jsonl`	jsonl	Held-out task prompts without solutions (~8% of combos)
`stats.json`	json	Generation statistics

Feedback messages report best-so-far only — no future information leaks into context.

Usage

from datasets import load_dataset

# step-level SFT rows
ds = load_dataset("Akahsizrr/kernelstrain", "sft_steps", split="train")

# full multiturn episodes
mt = load_dataset("Akahsizrr/kernelstrain", "sft_multiturn", split="train")

Generation

Pure-stdlib simulator: time = max(memory_time, compute_time) + launch_overhead, with GPU-specific peaks, tunable efficiency scaling, arch/smem gates, and Gaussian rebench noise. Deterministic from seed. Source: the kernelstrain generator.

Caveats

Perf numbers are analytic-model-plausible, not measured — the simulator can occasionally rank variants differently than real hardware would.
Kernel code is unverified; some templates may contain bugs that a real compiler would catch.
Assistant outputs are intentionally code-only; there are no reasoning/CoT traces by design.

License

MIT

Downloads last month: 40