Datasets:

Akahsizrr
/

kernelstrain2

session_id string	op string	impl string	desc string	signature string	gpu string	arch string	dtype string	shape unknown	objective string	baseline_ms float64	best dict	attempts dict	num_steps int64	elapsed_s int64	elapsed_hms string	trace list	messages list
cuda_swiglu_act-a10-M=256,N=5632-bf16-s14559839	swiglu_act	cuda	fused SwiGLU on packed gate_up[M,2N]: y = silu(gate) * up	swiglu_act(gu[M,2N] fp16) -> y[M,N] fp16	a10	sm_86	bf16	{ "M": 256, "N": 5632 }	minimize latency_ms	0.069604	{ "step": 20, "time_ms": 0.020091, "speedup": 3.464 }	{ "kept": 7, "reverted": 71, "errors": 1, "rebench": 3 }	79	6,989	1:56:29	[{"step":0,"tag":"v0_scalar","kind":"baseline","params":{},"status":"ok","time_ms":0.069604,"kept":t(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_minp_sample-rtx_5090-B=128,V=50257-fp16-s1448045	minp_sample	cuda	min-p sampling: threshold = p*pmax, filtered cdf	minp(logits[B,V] fp16, p[B]) -> tok[B]	rtx_5090	sm_120	fp16	{ "B": 128, "V": 50257 }	minimize latency_ms	0.149129	{ "step": 21, "time_ms": 0.021544, "speedup": 6.922 }	{ "kept": 10, "reverted": 22, "errors": 1, "rebench": 2 }	33	4,625	1:17:05	[{"step":0,"tag":"v0_serial","kind":"baseline","params":{},"status":"ok","time_ms":0.149129,"kept":t(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_topk_sample-a10-B=1,V=128256-fp32-s1740829	topk_sample	cuda	top-k sampling over logits[B,V] with per-row k	topk(logits[B,V] fp16, k[B]) -> tok[B]	a10	sm_86	fp32	{ "B": 1, "V": 128256 }	minimize latency_ms	0.022085	{ "step": 17, "time_ms": 0.007021, "speedup": 3.146 }	{ "kept": 8, "reverted": 31, "errors": 0, "rebench": 2 }	39	3,723	1:02:03	[{"step":0,"tag":"v0_insertion","kind":"baseline","params":{},"status":"ok","time_ms":0.022085,"kept(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
triton_swa_decode-rtx_a6000-B=32,H=32,Hkv=8,D=128,L=4096,W=512,SNK=4-bf16-s8466411	swa_decode	triton	sliding-window paged decode attention with sink tokens (Triton)	"swa_decode(q[B,H,D], kc[NP,Hkv,PB,D], vc[NP,Hkv,PB,D], bt[B,LP] int32, lens[B] int32, W, SNK) -> o[(...TRUNCATED)	rtx_a6000	sm_86	bf16	{ "B": 32, "H": 32, "Hkv": 8, "D": 128, "L": 4096, "W": 512, "SNK": 4 }	minimize latency_ms	2.806251	{ "step": 17, "time_ms": 0.102731, "speedup": 27.316 }	{ "kept": 5, "reverted": 126, "errors": 69, "rebench": 14 }	200	13,542	3:45:42	[{"step":0,"tag":"v0_maskall","kind":"baseline","params":{},"status":"ok","time_ms":2.806251,"kept":(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_nucleus_sample-h100-B=1,V=128256-fp32-s16996168	nucleus_sample	cuda	top-p sampling over logits[B,V]	nucleus(logits[B,V] fp16, p[B]) -> tok[B]	h100	sm_90a	fp32	{ "B": 1, "V": 128256 }	minimize latency_ms	0.015444	{ "step": 26, "time_ms": 0.005351, "speedup": 2.886 }	{ "kept": 6, "reverted": 47, "errors": 2, "rebench": 3 }	55	4,893	1:21:33	[{"step":0,"tag":"v0_greedy_sort","kind":"baseline","params":{},"status":"ok","time_ms":0.015444,"ke(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_rmsnorm-rtx_a6000-M=1,N=4096-fp16-s13170310	rmsnorm	cuda	RMSNorm rows for decode/prefill: y = x/rms(x) * w	rmsnorm(x[M,N] fp16, w[N] fp16, eps f32) -> y[M,N] fp16	rtx_a6000	sm_86	fp16	{ "M": 1, "N": 4096 }	minimize latency_ms	0.005243	{ "step": 22, "time_ms": 0.003389, "speedup": 1.547 }	{ "kept": 5, "reverted": 76, "errors": 3, "rebench": 10 }	84	7,032	1:57:12	[{"step":0,"tag":"v0_scalar","kind":"baseline","params":{},"status":"ok","time_ms":0.005243,"kept":t(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
triton_nucleus_sample-rtx_a6000-B=1,V=32000-fp32-s1268690	nucleus_sample	triton	nucleus top-p sampling: histogram-refined mass threshold + draw	nucleus_sample(logits[B,V]:fp16, p, u[B]:f32, temp) -> tok[B]:int32	rtx_a6000	sm_86	fp32	{ "B": 1, "V": 32000 }	minimize latency_ms	0.065615	{ "step": 142, "time_ms": 0.006416, "speedup": 10.227 }	{ "kept": 9, "reverted": 169, "errors": 22, "rebench": 7 }	200	16,750	4:39:10	[{"step":0,"tag":"v0_bitonic","kind":"baseline","params":{},"status":"ok","time_ms":0.065615,"kept":(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_rmsnorm-rtx_5090-M=1,N=4096-fp16-s6664408	rmsnorm	cuda	RMSNorm rows for decode/prefill: y = x/rms(x) * w	rmsnorm(x[M,N] fp16, w[N] fp16, eps f32) -> y[M,N] fp16	rtx_5090	sm_120	fp16	{ "M": 1, "N": 4096 }	minimize latency_ms	0.005168	{ "step": 30, "time_ms": 0.003321, "speedup": 1.556 }	{ "kept": 10, "reverted": 76, "errors": 2, "rebench": 6 }	88	7,779	2:09:39	[{"step":0,"tag":"v0_scalar","kind":"baseline","params":{},"status":"ok","time_ms":0.005168,"kept":t(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
triton_gemm_fp16-rtx_a6000-M=1024,N=32000,K=4096-fp16-s6987342	gemm_fp16	triton	fp16 GEMM for prefill: C = A @ B with tl.dot tiles (Triton)	gemm(a[M,K] fp16, b[K,N] fp16) -> c[M,N] fp16	rtx_a6000	sm_86	fp16	{ "M": 1024, "N": 32000, "K": 4096 }	minimize latency_ms	11.38006	{ "step": 23, "time_ms": 2.109104, "speedup": 5.396 }	{ "kept": 7, "reverted": 126, "errors": 67, "rebench": 13 }	200	13,877	3:51:17	[{"step":0,"tag":"v0_naive","kind":"baseline","params":{},"status":"ok","time_ms":11.38006,"kept":tr(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)
cuda_embedding_gather-h200-T=512,D=8192,V=256000-bf16-s10973639	embedding_gather	cuda	token embedding gather emb[idx[t]] -> y[t,:]	emb_gather(emb[V,D] fp16, idx[T] int32) -> y[T,D] fp16	h200	sm_90a	bf16	{ "T": 512, "D": 8192, "V": 256000 }	minimize latency_ms	0.018761	{ "step": 11, "time_ms": 0.007302, "speedup": 2.569 }	{ "kept": 7, "reverted": 32, "errors": 1, "rebench": 3 }	40	3,731	1:02:11	[{"step":0,"tag":"v0_scalar","kind":"baseline","params":{},"status":"ok","time_ms":0.018761,"kept":t(...TRUNCATED)	[{"role":"system","content":"You are a kernel optimization agent. Given an optimization task and the(...TRUNCATED)

End of preview.

kernelstrain2 — kernelinfer

Synthetic long-horizon GPU kernel-optimization trajectories, inference-only.

Each row is a complete optimization session (~30–200 steps, simulating 1–5 hours of work): the model receives an inference-kernel task, emits a full CUDA or Triton candidate, and gets simulated compile / correctness / benchmark feedback — iterating, self-correcting, and tuning launch parameters until convergence. No reasoning traces; assistant turns are kernel code only.

Scope

Inference workloads only — no training kernels. Operations cover:

decode/prefill attention: flash_decode, flash_prefill, swa_decode, mla_decode, tree_attn, chunked_prefill, cross_attn, fused_qkv_rope
quantized GEMM/GEMV: fp16, int8, int4, fp4, fp8 (incl. blockwise DeepGEMM style), split-K, wgmma/TMA persistent variants
KV-cache ops: paged_kv_copy, kv_append, kv_int8_quant
samplers: topk, nucleus (top-p), min-p, logits_softmax
MoE: scatter, combine, grouped_gemm
SSM: mamba_scan, mamba_step
misc: rmsnorm, rope, swiglu_act, embedding_gather, conv1d_causal, conv2d_nhwc, gemv_fp16

41 ops, CUDA + Triton implementations.

Hardware targets

H100 (sm_90a), H200 (sm_90a), A100 (sm_80), A10 (sm_86), RTX A6000 (sm_86), RTX 5090 (sm_120). Arch-gating is modeled correctly (wgmma/TMA only on sm_90a; fp8 on sm_89+/sm_90/sm_120; fp4 on sm_120; per-block smem caps 99/164/227KB).

Format

JSONL, one session per line:

messages — training data: system + user (task, then [t+H:MM:SS] RESULT feedback turns) + assistant (full kernel source, no prose, no CoT)
trace — compact per-step metadata (tag/kind/params/status/time/kept)
top-level: op, impl (cuda|triton), gpu, arch, dtype, shape, signature, desc, baseline_ms, best{step,time_ms,speedup}, attempts, elapsed_s, elapsed_hms, num_steps

batches/ holds 100-session shards; sessions.jsonl is the merged view; tasks_holdout.jsonl = 195 held-out task specs (no solutions).

Stats

1,899 sessions, all unique (op, gpu, shape, dtype) combos
~180K total steps; avg 94.6 steps/session
avg 2.18h simulated (min 1.03h, max 4.93h); avg 13.4x speedup
step statuses: 151.8K ok / 22.4K compile_error / 5.3K wrong_result
~1.5GB

Provenance

Fully synthetic: benchmark/correctness numbers come from an analytic roofline-style model — no kernels were ever compiled or executed. Intended for training small models on long-horizon iterative optimization behavior. Per-batch QA reports in the generating repo (kernelinfer/qa_reports/); all 19 batches scored 10/10 on mechanical checks and ≥8.5/10 on qualitative review after fixes.

Downloads last month: 27

Total file size:

1.53 GB