Datasets:
Upload folder using huggingface_hub
Browse files- .gitattributes +3 -0
- README.md +113 -0
- episodes.jsonl +3 -0
- sft_multiturn.jsonl +3 -0
- sft_steps.jsonl +3 -0
- stats.json +63 -0
- tasks_holdout.jsonl +185 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
episodes.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
sft_multiturn.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
sft_steps.jsonl filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pretty_name: KernelStrain
|
| 4 |
+
task_categories:
|
| 5 |
+
- text-generation
|
| 6 |
+
tags:
|
| 7 |
+
- code
|
| 8 |
+
- cuda
|
| 9 |
+
- triton
|
| 10 |
+
- gpu-kernels
|
| 11 |
+
- kernel-optimization
|
| 12 |
+
- synthetic-dataset
|
| 13 |
+
- sft
|
| 14 |
+
- pytorch
|
| 15 |
+
- llm-training
|
| 16 |
+
size_categories:
|
| 17 |
+
- 100K<n<1M
|
| 18 |
+
configs:
|
| 19 |
+
- config_name: episodes
|
| 20 |
+
data_files:
|
| 21 |
+
- split: train
|
| 22 |
+
path: episodes.jsonl
|
| 23 |
+
- config_name: sft_multiturn
|
| 24 |
+
data_files:
|
| 25 |
+
- split: train
|
| 26 |
+
path: sft_multiturn.jsonl
|
| 27 |
+
- config_name: sft_steps
|
| 28 |
+
data_files:
|
| 29 |
+
- split: train
|
| 30 |
+
path: sft_steps.jsonl
|
| 31 |
+
- config_name: holdout
|
| 32 |
+
data_files:
|
| 33 |
+
- split: test
|
| 34 |
+
path: tasks_holdout.jsonl
|
| 35 |
+
---
|
| 36 |
+
|
| 37 |
+
# KernelStrain
|
| 38 |
+
|
| 39 |
+
**Long-horizon GPU kernel optimization trajectories for training small models to iterate on CUDA and Triton kernels.**
|
| 40 |
+
|
| 41 |
+
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.
|
| 42 |
+
|
| 43 |
+
Built to teach a small model the *loop* of kernel work: propose → measure → regress → revert → recover → repeat, for tens of steps per task.
|
| 44 |
+
|
| 45 |
+
> ⚠️ **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.
|
| 46 |
+
|
| 47 |
+
## At a glance
|
| 48 |
+
|
| 49 |
+
| | |
|
| 50 |
+
|---|---|
|
| 51 |
+
| Episodes | 11,514 |
|
| 52 |
+
| SFT step rows | ~282k |
|
| 53 |
+
| Multiturn conversations | 11,514 |
|
| 54 |
+
| Holdout tasks | 185 (descriptions only, no solutions) |
|
| 55 |
+
| Op modules | 57 (45 distinct ops × CUDA/Triton impls) |
|
| 56 |
+
| Avg steps / episode | ~24.5 |
|
| 57 |
+
| Step statuses | ~94% ok, ~2.6% wrong-result, ~3.3% compile-error |
|
| 58 |
+
| GPU targets | A100 (sm_80), H100 (sm_90), RTX 4090, MI300X (gfx942) |
|
| 59 |
+
|
| 60 |
+
## Coverage
|
| 61 |
+
|
| 62 |
+
Kernels spanning AI **inference and training**:
|
| 63 |
+
|
| 64 |
+
- **Attention**: flash prefill, paged decode, MLA decode (latent KV), sliding-window, cross-attention, tree attention (speculative decoding), flash attention backward
|
| 65 |
+
- **Norms**: RMSNorm, LayerNorm, fused add+RMSNorm, LayerNorm backward
|
| 66 |
+
- **Activations/fusions**: SwiGLU fwd/bwd, bias+GELU, dropout (Philox)
|
| 67 |
+
- **Matmul**: fp16 GEMM/GEMV, int4/int8 GEMM, fp8 GEMM, blockwise-scaled fp8 GEMM (DeepGEMM-style), grouped GEMM
|
| 68 |
+
- **Training**: AdamW, Adafactor, grad clipping, cross-entropy fwd/bwd, embedding fwd/bwd, reductions, prefix scan
|
| 69 |
+
- **Quant**: fp8 tensor quant, int8 KV-cache quant
|
| 70 |
+
- **Sampling**: top-k, nucleus (top-p) — sort, histogram-refine, and CDF-pick strategies
|
| 71 |
+
- **SSM/conv**: Mamba selective scan (CUDA + Triton), causal Conv1D
|
| 72 |
+
- **MoE**: scatter, grouped GEMM, combine
|
| 73 |
+
- **Misc**: RoPE, transpose, KV-cache append, embedding gather
|
| 74 |
+
|
| 75 |
+
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.
|
| 76 |
+
|
| 77 |
+
## Files
|
| 78 |
+
|
| 79 |
+
| file | format | contents |
|
| 80 |
+
|---|---|---|
|
| 81 |
+
| `episodes.jsonl` | jsonl | Raw episodes — every step with full source, params, status, timing, smem, keep/revert decisions |
|
| 82 |
+
| `sft_multiturn.jsonl` | chat | Multi-turn conversations; assistant turns contain **complete kernel code only** (no reasoning traces) |
|
| 83 |
+
| `sft_steps.jsonl` | chat | One row per transition: task + history + current-best → next candidate |
|
| 84 |
+
| `tasks_holdout.jsonl` | jsonl | Held-out task prompts without solutions (~8% of combos) |
|
| 85 |
+
| `stats.json` | json | Generation statistics |
|
| 86 |
+
|
| 87 |
+
Feedback messages report **best-so-far** only — no future information leaks into context.
|
| 88 |
+
|
| 89 |
+
## Usage
|
| 90 |
+
|
| 91 |
+
```python
|
| 92 |
+
from datasets import load_dataset
|
| 93 |
+
|
| 94 |
+
# step-level SFT rows
|
| 95 |
+
ds = load_dataset("Akahsizrr/kernelstrain", "sft_steps", split="train")
|
| 96 |
+
|
| 97 |
+
# full multiturn episodes
|
| 98 |
+
mt = load_dataset("Akahsizrr/kernelstrain", "sft_multiturn", split="train")
|
| 99 |
+
```
|
| 100 |
+
|
| 101 |
+
## Generation
|
| 102 |
+
|
| 103 |
+
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.
|
| 104 |
+
|
| 105 |
+
## Caveats
|
| 106 |
+
|
| 107 |
+
- Perf numbers are analytic-model-plausible, **not measured** — the simulator can occasionally rank variants differently than real hardware would.
|
| 108 |
+
- Kernel code is unverified; some templates may contain bugs that a real compiler would catch.
|
| 109 |
+
- Assistant outputs are intentionally code-only; there are no reasoning/CoT traces by design.
|
| 110 |
+
|
| 111 |
+
## License
|
| 112 |
+
|
| 113 |
+
MIT
|
episodes.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6afba4039406e4b10ba723f0800e2f181ee4cdb2c1aa2d11bda70530177e6fda
|
| 3 |
+
size 796024084
|
sft_multiturn.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:69af3782ab774083bb8415f7063d9c28373266e06fb289eaf573db84cf0cdebd
|
| 3 |
+
size 1006019649
|
sft_steps.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7a1d1ff21838a2f949405104591d11139941dd51c5e0477b1a3c4c4a0dd2e531
|
| 3 |
+
size 1811158420
|
stats.json
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"episodes": 11514,
|
| 3 |
+
"sft_multiturn_rows": 11514,
|
| 4 |
+
"sft_step_rows": 282139,
|
| 5 |
+
"holdout_tasks": 185,
|
| 6 |
+
"ops": 57,
|
| 7 |
+
"step_status": {
|
| 8 |
+
"ok": 265440,
|
| 9 |
+
"wrong_result": 7363,
|
| 10 |
+
"compile_error": 9336
|
| 11 |
+
},
|
| 12 |
+
"avg_steps": 24.5,
|
| 13 |
+
"avg_speedup": 13.699,
|
| 14 |
+
"max_speedup": 320.458,
|
| 15 |
+
"per_op_episodes": {
|
| 16 |
+
"rmsnorm": 438,
|
| 17 |
+
"layernorm": 438,
|
| 18 |
+
"softmax": 444,
|
| 19 |
+
"swiglu": 444,
|
| 20 |
+
"rope": 450,
|
| 21 |
+
"gemv_fp16": 222,
|
| 22 |
+
"gemm_fp16": 204,
|
| 23 |
+
"attn_decode_paged": 210,
|
| 24 |
+
"int4_gemm": 216,
|
| 25 |
+
"topk": 204,
|
| 26 |
+
"transpose": 204,
|
| 27 |
+
"fused_add_rmsnorm": 228,
|
| 28 |
+
"moe_scatter": 222,
|
| 29 |
+
"matmul": 228,
|
| 30 |
+
"attn_decode": 174,
|
| 31 |
+
"adamw": 204,
|
| 32 |
+
"cross_entropy": 396,
|
| 33 |
+
"layernorm_backward": 432,
|
| 34 |
+
"reduce_sum": 204,
|
| 35 |
+
"scan_prefix": 120,
|
| 36 |
+
"flash_prefill": 420,
|
| 37 |
+
"fp8_quant": 438,
|
| 38 |
+
"gemm_fp8": 90,
|
| 39 |
+
"kv_cache_append": 210,
|
| 40 |
+
"embedding": 216,
|
| 41 |
+
"moe_grouped_gemm": 216,
|
| 42 |
+
"mla_decode": 216,
|
| 43 |
+
"swa_decode": 222,
|
| 44 |
+
"cross_attn": 222,
|
| 45 |
+
"flash_attn_bwd": 234,
|
| 46 |
+
"tree_attn": 234,
|
| 47 |
+
"swiglu_backward": 234,
|
| 48 |
+
"cross_entropy_backward": 228,
|
| 49 |
+
"embedding_backward": 228,
|
| 50 |
+
"dropout": 222,
|
| 51 |
+
"adafactor": 108,
|
| 52 |
+
"grad_clip": 234,
|
| 53 |
+
"int8_gemm": 96,
|
| 54 |
+
"fp8_blockwise_gemm": 90,
|
| 55 |
+
"nucleus_sample": 456,
|
| 56 |
+
"kv_int8_quant": 114,
|
| 57 |
+
"fused_bias_gelu": 222,
|
| 58 |
+
"mamba_scan": 432,
|
| 59 |
+
"conv1d_causal": 222,
|
| 60 |
+
"moe_combine": 228
|
| 61 |
+
},
|
| 62 |
+
"chars_total": 796001056
|
| 63 |
+
}
|
tasks_holdout.jsonl
ADDED
|
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"task_id": "cuda_rmsnorm-rtx4090-M=4096,N=8192-fp16-s15252588", "op": "rmsnorm", "impl": "cuda", "desc": "RMSNorm over last dim, one row per sample", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 2 |
+
{"task_id": "cuda_rmsnorm-mi300x-M=4096,N=8192-bf16-s13123706", "op": "rmsnorm", "impl": "cuda", "desc": "RMSNorm over last dim, one row per sample", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 3 |
+
{"task_id": "cuda_rmsnorm-mi300x-M=16384,N=5120-fp16-s3575975", "op": "rmsnorm", "impl": "cuda", "desc": "RMSNorm over last dim, one row per sample", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 16384, "N": 5120}, "objective": "minimize latency_ms"}
|
| 4 |
+
{"task_id": "cuda_rmsnorm-h100-sxm-M=1024,N=14336-fp16-s11332754", "op": "rmsnorm", "impl": "cuda", "desc": "RMSNorm over last dim, one row per sample", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|
| 5 |
+
{"task_id": "cuda_layernorm-mi300x-M=2048,N=4096-fp16-s7096340", "op": "layernorm", "impl": "cuda", "desc": "LayerNorm over last dim with affine gamma/beta, one row per sample", "signature": "layernorm(x[M,N]:fp16, gamma[N]:fp16, beta[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 6 |
+
{"task_id": "cuda_layernorm-a100-80g-M=4096,N=8192-fp16-s1527032", "op": "layernorm", "impl": "cuda", "desc": "LayerNorm over last dim with affine gamma/beta, one row per sample", "signature": "layernorm(x[M,N]:fp16, gamma[N]:fp16, beta[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 7 |
+
{"task_id": "cuda_layernorm-h100-sxm-M=4096,N=8192-fp16-s16576709", "op": "layernorm", "impl": "cuda", "desc": "LayerNorm over last dim with affine gamma/beta, one row per sample", "signature": "layernorm(x[M,N]:fp16, gamma[N]:fp16, beta[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 8 |
+
{"task_id": "cuda_layernorm-mi300x-M=4096,N=8192-fp16-s1861878", "op": "layernorm", "impl": "cuda", "desc": "LayerNorm over last dim with affine gamma/beta, one row per sample", "signature": "layernorm(x[M,N]:fp16, gamma[N]:fp16, beta[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 9 |
+
{"task_id": "cuda_layernorm-mi300x-M=1024,N=14336-fp16-s13965546", "op": "layernorm", "impl": "cuda", "desc": "LayerNorm over last dim with affine gamma/beta, one row per sample", "signature": "layernorm(x[M,N]:fp16, gamma[N]:fp16, beta[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|
| 10 |
+
{"task_id": "cuda_softmax-h100-sxm-M=4096,N=8192-fp16-s12030069", "op": "softmax", "impl": "cuda", "desc": "Numerically-stable softmax over last dim, one row per sample", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 11 |
+
{"task_id": "cuda_softmax-h100-sxm-M=16384,N=4096-bf16-s4982813", "op": "softmax", "impl": "cuda", "desc": "Numerically-stable softmax over last dim, one row per sample", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 16384, "N": 4096}, "objective": "minimize latency_ms"}
|
| 12 |
+
{"task_id": "cuda_softmax-rtx4090-M=16384,N=4096-bf16-s6641821", "op": "softmax", "impl": "cuda", "desc": "Numerically-stable softmax over last dim, one row per sample", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 16384, "N": 4096}, "objective": "minimize latency_ms"}
|
| 13 |
+
{"task_id": "cuda_softmax-mi300x-M=2048,N=14336-bf16-s2691185", "op": "softmax", "impl": "cuda", "desc": "Numerically-stable softmax over last dim, one row per sample", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 2048, "N": 14336}, "objective": "minimize latency_ms"}
|
| 14 |
+
{"task_id": "cuda_swiglu-mi300x-M=8192,N=8192-bf16-s9360460", "op": "swiglu", "impl": "cuda", "desc": "SwiGLU: silu(gate) * up, elementwise over flattened M*N", "signature": "swiglu(gate[M,N]:fp16, up[M,N]:fp16) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 8192, "N": 8192}, "objective": "minimize latency_ms"}
|
| 15 |
+
{"task_id": "cuda_swiglu-a100-80g-M=16384,N=4096-bf16-s1542285", "op": "swiglu", "impl": "cuda", "desc": "SwiGLU: silu(gate) * up, elementwise over flattened M*N", "signature": "swiglu(gate[M,N]:fp16, up[M,N]:fp16) -> y[M,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 16384, "N": 4096}, "objective": "minimize latency_ms"}
|
| 16 |
+
{"task_id": "cuda_swiglu-h100-sxm-M=16384,N=4096-bf16-s11916316", "op": "swiglu", "impl": "cuda", "desc": "SwiGLU: silu(gate) * up, elementwise over flattened M*N", "signature": "swiglu(gate[M,N]:fp16, up[M,N]:fp16) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 16384, "N": 4096}, "objective": "minimize latency_ms"}
|
| 17 |
+
{"task_id": "cuda_swiglu-h100-sxm-M=512,N=5632-bf16-s11987112", "op": "swiglu", "impl": "cuda", "desc": "SwiGLU: silu(gate) * up, elementwise over flattened M*N", "signature": "swiglu(gate[M,N]:fp16, up[M,N]:fp16) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 512, "N": 5632}, "objective": "minimize latency_ms"}
|
| 18 |
+
{"task_id": "cuda_rope-h100-sxm-M=2048,H=32,Hk=32,D=128-fp16-s7044144", "op": "rope", "impl": "cuda", "desc": "Rotary position embedding on consecutive pairs, in-place q and k", "signature": "rope(q[M,H,D]:fp16, k[M,Hk,D]:fp16, cos[D/2]:fp32, sin[D/2]:fp32) -> q[M,H,D]:fp16, k[M,Hk,D]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 2048, "H": 32, "Hk": 32, "D": 128}, "objective": "minimize latency_ms"}
|
| 19 |
+
{"task_id": "cuda_rope-a100-80g-M=1024,H=64,Hk=16,D=64-fp16-s11065459", "op": "rope", "impl": "cuda", "desc": "Rotary position embedding on consecutive pairs, in-place q and k", "signature": "rope(q[M,H,D]:fp16, k[M,Hk,D]:fp16, cos[D/2]:fp32, sin[D/2]:fp32) -> q[M,H,D]:fp16, k[M,Hk,D]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 1024, "H": 64, "Hk": 16, "D": 64}, "objective": "minimize latency_ms"}
|
| 20 |
+
{"task_id": "cuda_rope-mi300x-M=1024,H=64,Hk=16,D=64-fp16-s14576834", "op": "rope", "impl": "cuda", "desc": "Rotary position embedding on consecutive pairs, in-place q and k", "signature": "rope(q[M,H,D]:fp16, k[M,Hk,D]:fp16, cos[D/2]:fp32, sin[D/2]:fp32) -> q[M,H,D]:fp16, k[M,Hk,D]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 1024, "H": 64, "Hk": 16, "D": 64}, "objective": "minimize latency_ms"}
|
| 21 |
+
{"task_id": "cuda_gemv_fp16-a100-80g-N=8192,K=8192-bf16-s12281928", "op": "gemv_fp16", "impl": "cuda", "desc": "fp16 weight-stationary matvec for LLM decode", "signature": "gemv(y[N] = W[N,K] @ x[K])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 22 |
+
{"task_id": "cuda_gemv_fp16-rtx4090-N=10240,K=5120-bf16-s7680094", "op": "gemv_fp16", "impl": "cuda", "desc": "fp16 weight-stationary matvec for LLM decode", "signature": "gemv(y[N] = W[N,K] @ x[K])", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"N": 10240, "K": 5120}, "objective": "minimize latency_ms"}
|
| 23 |
+
{"task_id": "cuda_gemv_fp16-h100-sxm-N=57344,K=8192-bf16-s1121298", "op": "gemv_fp16", "impl": "cuda", "desc": "fp16 weight-stationary matvec for LLM decode", "signature": "gemv(y[N] = W[N,K] @ x[K])", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"N": 57344, "K": 8192}, "objective": "minimize latency_ms"}
|
| 24 |
+
{"task_id": "cuda_gemm_fp16-h100-sxm-M=8192,N=8192,K=8192-bf16-s8054962", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 8192, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 25 |
+
{"task_id": "cuda_gemm_fp16-mi300x-M=2048,N=14336,K=4096-fp16-s3849429", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 2048, "N": 14336, "K": 4096}, "objective": "minimize latency_ms"}
|
| 26 |
+
{"task_id": "cuda_gemm_fp16-rtx4090-M=128,N=8192,K=8192-bf16-s7216139", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 128, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 27 |
+
{"task_id": "cuda_gemm_fp16-a100-80g-M=16384,N=4096,K=4096-fp16-s3023072", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 16384, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 28 |
+
{"task_id": "cuda_gemm_fp16-a100-80g-M=16384,N=4096,K=4096-bf16-s15307849", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 16384, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 29 |
+
{"task_id": "cuda_gemm_fp16-mi300x-M=16384,N=4096,K=4096-fp16-s11262057", "op": "gemm_fp16", "impl": "cuda", "desc": "fp16 GEMM with fp32 accumulate, tensor-core trajectory", "signature": "gemm(C[M,N] = A[M,K] @ B[K,N])", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 16384, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 30 |
+
{"task_id": "cuda_attn_decode_paged-h100-sxm-B=1,H=32,Hkv=8,D=128,L=2048-bf16-s6663256", "op": "attn_decode_paged", "impl": "cuda", "desc": "GQA flash-decoding over paged KV cache", "signature": "attn_decode(q[B,H,D], kvcache pages, kv_len) -> o[B,H,D]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"B": 1, "H": 32, "Hkv": 8, "D": 128, "L": 2048}, "objective": "minimize latency_ms"}
|
| 31 |
+
{"task_id": "cuda_attn_decode_paged-mi300x-B=8,H=32,Hkv=8,D=128,L=8192-bf16-s16776373", "op": "attn_decode_paged", "impl": "cuda", "desc": "GQA flash-decoding over paged KV cache", "signature": "attn_decode(q[B,H,D], kvcache pages, kv_len) -> o[B,H,D]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 8, "H": 32, "Hkv": 8, "D": 128, "L": 8192}, "objective": "minimize latency_ms"}
|
| 32 |
+
{"task_id": "cuda_attn_decode_paged-rtx4090-B=64,H=40,Hkv=8,D=128,L=16384-bf16-s15031458", "op": "attn_decode_paged", "impl": "cuda", "desc": "GQA flash-decoding over paged KV cache", "signature": "attn_decode(q[B,H,D], kvcache pages, kv_len) -> o[B,H,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 64, "H": 40, "Hkv": 8, "D": 128, "L": 16384}, "objective": "minimize latency_ms"}
|
| 33 |
+
{"task_id": "cuda_attn_decode_paged-a100-80g-B=16,H=32,Hkv=4,D=128,L=32768-fp16-s1889514", "op": "attn_decode_paged", "impl": "cuda", "desc": "GQA flash-decoding over paged KV cache", "signature": "attn_decode(q[B,H,D], kvcache pages, kv_len) -> o[B,H,D]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 16, "H": 32, "Hkv": 4, "D": 128, "L": 32768}, "objective": "minimize latency_ms"}
|
| 34 |
+
{"task_id": "cuda_attn_decode_paged-mi300x-B=16,H=32,Hkv=4,D=128,L=32768-bf16-s7194732", "op": "attn_decode_paged", "impl": "cuda", "desc": "GQA flash-decoding over paged KV cache", "signature": "attn_decode(q[B,H,D], kvcache pages, kv_len) -> o[B,H,D]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 16, "H": 32, "Hkv": 4, "D": 128, "L": 32768}, "objective": "minimize latency_ms"}
|
| 35 |
+
{"task_id": "cuda_int4_gemm-rtx4090-M=64,N=14336,K=4096-bf16-s1575955", "op": "int4_gemm", "impl": "cuda", "desc": "Marlin-style int4 weight dequant + GEMM (gs=128)", "signature": "gemm(C[M,N] = A[M,K]fp16 @ dequant(W4[N,K], scales[N,K/gs]))", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 64, "N": 14336, "K": 4096}, "objective": "minimize latency_ms"}
|
| 36 |
+
{"task_id": "cuda_int4_gemm-mi300x-M=64,N=14336,K=4096-bf16-s11956440", "op": "int4_gemm", "impl": "cuda", "desc": "Marlin-style int4 weight dequant + GEMM (gs=128)", "signature": "gemm(C[M,N] = A[M,K]fp16 @ dequant(W4[N,K], scales[N,K/gs]))", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 64, "N": 14336, "K": 4096}, "objective": "minimize latency_ms"}
|
| 37 |
+
{"task_id": "cuda_int4_gemm-mi300x-M=1024,N=4096,K=14336-fp16-s1529906", "op": "int4_gemm", "impl": "cuda", "desc": "Marlin-style int4 weight dequant + GEMM (gs=128)", "signature": "gemm(C[M,N] = A[M,K]fp16 @ dequant(W4[N,K], scales[N,K/gs]))", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 1024, "N": 4096, "K": 14336}, "objective": "minimize latency_ms"}
|
| 38 |
+
{"task_id": "cuda_int4_gemm-rtx4090-M=4096,N=8192,K=8192-bf16-s3102820", "op": "int4_gemm", "impl": "cuda", "desc": "Marlin-style int4 weight dequant + GEMM (gs=128)", "signature": "gemm(C[M,N] = A[M,K]fp16 @ dequant(W4[N,K], scales[N,K/gs]))", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 4096, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 39 |
+
{"task_id": "cuda_topk-h100-sxm-M=8,V=128256,K=8-bf16-s6417607", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 8, "V": 128256, "K": 8}, "objective": "minimize latency_ms"}
|
| 40 |
+
{"task_id": "cuda_topk-rtx4090-M=8,V=128256,K=8-fp16-s9606264", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 8, "V": 128256, "K": 8}, "objective": "minimize latency_ms"}
|
| 41 |
+
{"task_id": "cuda_topk-mi300x-M=8,V=128256,K=8-bf16-s506881", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 8, "V": 128256, "K": 8}, "objective": "minimize latency_ms"}
|
| 42 |
+
{"task_id": "cuda_topk-a100-80g-M=64,V=128256,K=8-fp16-s11216100", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 64, "V": 128256, "K": 8}, "objective": "minimize latency_ms"}
|
| 43 |
+
{"task_id": "cuda_topk-mi300x-M=64,V=128256,K=8-bf16-s4055075", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 64, "V": 128256, "K": 8}, "objective": "minimize latency_ms"}
|
| 44 |
+
{"task_id": "cuda_topk-a100-80g-M=64,V=32000,K=8-fp16-s1096956", "op": "topk", "impl": "cuda", "desc": "Per-row top-K over vocab V, returns sorted vals and indices", "signature": "topk(x[M,V]:fp16) -> vals[M,K]:fp16, idx[M,K]:int32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 64, "V": 32000, "K": 8}, "objective": "minimize latency_ms"}
|
| 45 |
+
{"task_id": "cuda_transpose-a100-80g-M=8192,N=5120-fp16-s7099509", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 8192, "N": 5120}, "objective": "minimize latency_ms"}
|
| 46 |
+
{"task_id": "cuda_transpose-h100-sxm-M=8192,N=5120-fp16-s4365438", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 8192, "N": 5120}, "objective": "minimize latency_ms"}
|
| 47 |
+
{"task_id": "cuda_transpose-mi300x-M=2048,N=14336-fp16-s14014683", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 2048, "N": 14336}, "objective": "minimize latency_ms"}
|
| 48 |
+
{"task_id": "cuda_transpose-mi300x-M=2048,N=14336-bf16-s12635170", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 2048, "N": 14336}, "objective": "minimize latency_ms"}
|
| 49 |
+
{"task_id": "cuda_transpose-rtx4090-M=16384,N=4096-fp16-s16026673", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 16384, "N": 4096}, "objective": "minimize latency_ms"}
|
| 50 |
+
{"task_id": "cuda_transpose-h100-sxm-M=1024,N=8192-fp16-s13575347", "op": "transpose", "impl": "cuda", "desc": "2D matrix transpose y[n,m]=x[m,n], coalescing on both sides", "signature": "transpose(x[M,N]:fp16) -> y[N,M]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 1024, "N": 8192}, "objective": "minimize latency_ms"}
|
| 51 |
+
{"task_id": "cuda_fused_add_rmsnorm-rtx4090-M=2048,N=4096-fp16-s6488256", "op": "fused_add_rmsnorm", "impl": "cuda", "desc": "fused residual add + RMSNorm producing y and res_out", "signature": "fused_add_rmsnorm(x[M,N], res[M,N], w[N], eps) -> y[M,N], res_out[M,N]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 52 |
+
{"task_id": "cuda_fused_add_rmsnorm-mi300x-M=2048,N=4096-bf16-s10813464", "op": "fused_add_rmsnorm", "impl": "cuda", "desc": "fused residual add + RMSNorm producing y and res_out", "signature": "fused_add_rmsnorm(x[M,N], res[M,N], w[N], eps) -> y[M,N], res_out[M,N]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 53 |
+
{"task_id": "cuda_moe_scatter-a100-80g-T=8192,N=5120,K=2,E=16-fp16-s9139406", "op": "moe_scatter", "impl": "cuda", "desc": "counting-sort scatter of tokens to experts + row gather", "signature": "moe_scatter(x[T,N], expert_ids[T,K], E) -> perm[T*K], offsets[E+1], x_perm[T*K,N]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"T": 8192, "N": 5120, "K": 2, "E": 16}, "objective": "minimize latency_ms"}
|
| 54 |
+
{"task_id": "cuda_moe_scatter-rtx4090-T=8192,N=5120,K=2,E=16-bf16-s9468118", "op": "moe_scatter", "impl": "cuda", "desc": "counting-sort scatter of tokens to experts + row gather", "signature": "moe_scatter(x[T,N], expert_ids[T,K], E) -> perm[T*K], offsets[E+1], x_perm[T*K,N]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"T": 8192, "N": 5120, "K": 2, "E": 16}, "objective": "minimize latency_ms"}
|
| 55 |
+
{"task_id": "cuda_moe_scatter-rtx4090-T=2048,N=8192,K=8,E=128-bf16-s7007330", "op": "moe_scatter", "impl": "cuda", "desc": "counting-sort scatter of tokens to experts + row gather", "signature": "moe_scatter(x[T,N], expert_ids[T,K], E) -> perm[T*K], offsets[E+1], x_perm[T*K,N]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"T": 2048, "N": 8192, "K": 8, "E": 128}, "objective": "minimize latency_ms"}
|
| 56 |
+
{"task_id": "triton_rmsnorm-a100-80g-M=4096,N=4096-bf16-s3738789", "op": "rmsnorm", "impl": "triton", "desc": "RMSNorm over last dim, one program-tile per row (Triton)", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 4096, "N": 4096}, "objective": "minimize latency_ms"}
|
| 57 |
+
{"task_id": "triton_rmsnorm-mi300x-M=16384,N=5120-fp16-s3575975", "op": "rmsnorm", "impl": "triton", "desc": "RMSNorm over last dim, one program-tile per row (Triton)", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 16384, "N": 5120}, "objective": "minimize latency_ms"}
|
| 58 |
+
{"task_id": "triton_rmsnorm-h100-sxm-M=1024,N=14336-fp16-s11332754", "op": "rmsnorm", "impl": "triton", "desc": "RMSNorm over last dim, one program-tile per row (Triton)", "signature": "rmsnorm(x[M,N]:fp16, w[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|
| 59 |
+
{"task_id": "triton_layernorm-mi300x-M=2048,N=4096-fp16-s7096340", "op": "layernorm", "impl": "triton", "desc": "LayerNorm over last dim with affine weight/bias (Triton)", "signature": "layernorm(x[M,N]:fp16, w[N]:fp16, b[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 60 |
+
{"task_id": "triton_layernorm-mi300x-M=1024,N=14336-fp16-s13965546", "op": "layernorm", "impl": "triton", "desc": "LayerNorm over last dim with affine weight/bias (Triton)", "signature": "layernorm(x[M,N]:fp16, w[N]:fp16, b[N]:fp16, eps) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|
| 61 |
+
{"task_id": "triton_softmax-rtx4090-M=2048,N=8192-bf16-s12413972", "op": "softmax", "impl": "triton", "desc": "Softmax over last dim incl. vocab-sized rows (Triton)", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 2048, "N": 8192}, "objective": "minimize latency_ms"}
|
| 62 |
+
{"task_id": "triton_softmax-a100-80g-M=1,N=131072-fp16-s12045332", "op": "softmax", "impl": "triton", "desc": "Softmax over last dim incl. vocab-sized rows (Triton)", "signature": "softmax(x[M,N]:fp16) -> y[M,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 1, "N": 131072}, "objective": "minimize latency_ms"}
|
| 63 |
+
{"task_id": "triton_swiglu-mi300x-M=4096,N=2752-bf16-s8008743", "op": "swiglu", "impl": "triton", "desc": "SwiGLU on packed gate|up tensor x[M,2N] (Triton)", "signature": "swiglu(x[M,2N]:fp16) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 4096, "N": 2752}, "objective": "minimize latency_ms"}
|
| 64 |
+
{"task_id": "triton_swiglu-mi300x-M=8192,N=5504-bf16-s7105545", "op": "swiglu", "impl": "triton", "desc": "SwiGLU on packed gate|up tensor x[M,2N] (Triton)", "signature": "swiglu(x[M,2N]:fp16) -> y[M,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 8192, "N": 5504}, "objective": "minimize latency_ms"}
|
| 65 |
+
{"task_id": "triton_rope-mi300x-M=1,H=32,Hk=8,D=128-bf16-s7345295", "op": "rope", "impl": "triton", "desc": "In-place rotary embedding on q[M,H,D], k[M,Hk,D] with cos/sin[M,D/2] (Triton)", "signature": "rope(q[M,H,D]:fp16, k[M,Hk,D]:fp16, cos[M,D/2]:fp32, sin[M,D/2]:fp32) -> q,k in-place", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 1, "H": 32, "Hk": 8, "D": 128}, "objective": "minimize latency_ms"}
|
| 66 |
+
{"task_id": "triton_rope-a100-80g-M=64,H=32,Hk=8,D=128-bf16-s9538814", "op": "rope", "impl": "triton", "desc": "In-place rotary embedding on q[M,H,D], k[M,Hk,D] with cos/sin[M,D/2] (Triton)", "signature": "rope(q[M,H,D]:fp16, k[M,Hk,D]:fp16, cos[M,D/2]:fp32, sin[M,D/2]:fp32) -> q,k in-place", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 64, "H": 32, "Hk": 8, "D": 128}, "objective": "minimize latency_ms"}
|
| 67 |
+
{"task_id": "triton_matmul-h100-sxm-M=8192,N=8192,K=8192-fp16-s6039776", "op": "matmul", "impl": "triton", "desc": "C = A @ B with tl.dot tensor-core tiles (Triton)", "signature": "matmul(a[M,K]:fp16, b[K,N]:fp16) -> c[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 8192, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 68 |
+
{"task_id": "triton_matmul-h100-sxm-M=128,N=8192,K=8192-bf16-s5296188", "op": "matmul", "impl": "triton", "desc": "C = A @ B with tl.dot tensor-core tiles (Triton)", "signature": "matmul(a[M,K]:fp16, b[K,N]:fp16) -> c[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 128, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 69 |
+
{"task_id": "triton_attn_decode-h100-sxm-B=8,H=32,Hkv=8,D=128,L=4096-fp16-s7701575", "op": "attn_decode", "impl": "triton", "desc": "GQA flash-decode: single-token query vs paged KV (Triton)", "signature": "attn_decode(q[B,H,D]:fp16, k[B,Hkv,L,D]:fp16, v[B,Hkv,L,D]:fp16, kv_len) -> o[B,H,D]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"B": 8, "H": 32, "Hkv": 8, "D": 128, "L": 4096}, "objective": "minimize latency_ms"}
|
| 70 |
+
{"task_id": "triton_attn_decode-rtx4090-B=8,H=32,Hkv=8,D=128,L=4096-bf16-s1603789", "op": "attn_decode", "impl": "triton", "desc": "GQA flash-decode: single-token query vs paged KV (Triton)", "signature": "attn_decode(q[B,H,D]:fp16, k[B,Hkv,L,D]:fp16, v[B,Hkv,L,D]:fp16, kv_len) -> o[B,H,D]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 8, "H": 32, "Hkv": 8, "D": 128, "L": 4096}, "objective": "minimize latency_ms"}
|
| 71 |
+
{"task_id": "triton_attn_decode-a100-80g-B=32,H=64,Hkv=8,D=128,L=8192-fp16-s2989086", "op": "attn_decode", "impl": "triton", "desc": "GQA flash-decode: single-token query vs paged KV (Triton)", "signature": "attn_decode(q[B,H,D]:fp16, k[B,Hkv,L,D]:fp16, v[B,Hkv,L,D]:fp16, kv_len) -> o[B,H,D]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 32, "H": 64, "Hkv": 8, "D": 128, "L": 8192}, "objective": "minimize latency_ms"}
|
| 72 |
+
{"task_id": "cuda_adamw-h100-sxm-N=268435456-fp32-s4346943", "op": "adamw", "impl": "cuda", "desc": "Fused AdamW update on one flattened fp32 param tensor", "signature": "adamw(p[N]:fp32, g[N]:fp32, m[N]:fp32, v[N]:fp32, lr, b1, b2, eps, wd, t)", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp32", "shape": {"N": 268435456}, "objective": "minimize latency_ms"}
|
| 73 |
+
{"task_id": "cuda_cross_entropy-rtx4090-M=512,V=32000-bf16-s1914099", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 512, "V": 32000}, "objective": "minimize latency_ms"}
|
| 74 |
+
{"task_id": "cuda_cross_entropy-h100-sxm-M=8192,V=32000-bf16-s1440989", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 8192, "V": 32000}, "objective": "minimize latency_ms"}
|
| 75 |
+
{"task_id": "cuda_cross_entropy-h100-sxm-M=4096,V=262144-bf16-s2310311", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 76 |
+
{"task_id": "cuda_cross_entropy-rtx4090-M=4096,V=262144-fp16-s7514189", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 77 |
+
{"task_id": "cuda_cross_entropy-mi300x-M=4096,V=262144-fp16-s9919532", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 78 |
+
{"task_id": "cuda_cross_entropy-h100-sxm-M=32768,V=50257-fp16-s1683547", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 32768, "V": 50257}, "objective": "minimize latency_ms"}
|
| 79 |
+
{"task_id": "cuda_cross_entropy-mi300x-M=32768,V=50257-bf16-s10988698", "op": "cross_entropy", "impl": "cuda", "desc": "Row-wise cross-entropy loss = logsumexp - target logit", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int32) -> loss[M]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 32768, "V": 50257}, "objective": "minimize latency_ms"}
|
| 80 |
+
{"task_id": "cuda_layernorm_backward-rtx4090-M=2048,N=4096-fp16-s3471429", "op": "layernorm_backward", "impl": "cuda", "desc": "LayerNorm backward: row dx plus dgamma/dbeta column reductions", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, gamma[N]:fp16, mean[M]:fp16, rstd[M]:fp16) -> dx[M,N]:fp16, dgamma[N]:fp16, dbeta[N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 81 |
+
{"task_id": "cuda_layernorm_backward-rtx4090-M=4096,N=8192-fp16-s14871768", "op": "layernorm_backward", "impl": "cuda", "desc": "LayerNorm backward: row dx plus dgamma/dbeta column reductions", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, gamma[N]:fp16, mean[M]:fp16, rstd[M]:fp16) -> dx[M,N]:fp16, dgamma[N]:fp16, dbeta[N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 82 |
+
{"task_id": "cuda_layernorm_backward-rtx4090-M=4096,N=8192-bf16-s13910075", "op": "layernorm_backward", "impl": "cuda", "desc": "LayerNorm backward: row dx plus dgamma/dbeta column reductions", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, gamma[N]:fp16, mean[M]:fp16, rstd[M]:fp16) -> dx[M,N]:fp16, dgamma[N]:fp16, dbeta[N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 83 |
+
{"task_id": "cuda_layernorm_backward-a100-80g-M=1024,N=14336-bf16-s5253228", "op": "layernorm_backward", "impl": "cuda", "desc": "LayerNorm backward: row dx plus dgamma/dbeta column reductions", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, gamma[N]:fp16, mean[M]:fp16, rstd[M]:fp16) -> dx[M,N]:fp16, dgamma[N]:fp16, dbeta[N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|
| 84 |
+
{"task_id": "cuda_reduce_sum-a100-80g-N=134217728-fp16-s8165440", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"N": 134217728}, "objective": "minimize latency_ms"}
|
| 85 |
+
{"task_id": "cuda_reduce_sum-a100-80g-N=134217728-bf16-s8669431", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"N": 134217728}, "objective": "minimize latency_ms"}
|
| 86 |
+
{"task_id": "cuda_reduce_sum-a100-80g-N=268435456-bf16-s12198949", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"N": 268435456}, "objective": "minimize latency_ms"}
|
| 87 |
+
{"task_id": "cuda_reduce_sum-h100-sxm-N=268435456-fp16-s1536243", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"N": 268435456}, "objective": "minimize latency_ms"}
|
| 88 |
+
{"task_id": "cuda_reduce_sum-a100-80g-N=16777216-fp16-s12776652", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"N": 16777216}, "objective": "minimize latency_ms"}
|
| 89 |
+
{"task_id": "cuda_reduce_sum-rtx4090-N=16777216-bf16-s11047046", "op": "reduce_sum", "impl": "cuda", "desc": "Full-tensor sum to a single fp32 scalar, large N", "signature": "reduce_sum(x[N]:fp16) -> out[1]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"N": 16777216}, "objective": "minimize latency_ms"}
|
| 90 |
+
{"task_id": "cuda_flash_prefill-rtx4090-B=4,H=32,S=4096,D=64-fp16-s8515615", "op": "flash_prefill", "impl": "cuda", "desc": "causal self-attention forward, fp16, fp32 softmax state", "signature": "flash_prefill(q[B,H,S,D], k[B,H,S,D], v[B,H,S,D]) -> o[B,H,S,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 4, "H": 32, "S": 4096, "D": 64}, "objective": "minimize latency_ms"}
|
| 91 |
+
{"task_id": "cuda_flash_prefill-mi300x-B=2,H=32,S=8192,D=128-bf16-s12579973", "op": "flash_prefill", "impl": "cuda", "desc": "causal self-attention forward, fp16, fp32 softmax state", "signature": "flash_prefill(q[B,H,S,D], k[B,H,S,D], v[B,H,S,D]) -> o[B,H,S,D]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 2, "H": 32, "S": 8192, "D": 128}, "objective": "minimize latency_ms"}
|
| 92 |
+
{"task_id": "cuda_flash_prefill-a100-80g-B=8,H=16,S=4096,D=128-fp16-s635980", "op": "flash_prefill", "impl": "cuda", "desc": "causal self-attention forward, fp16, fp32 softmax state", "signature": "flash_prefill(q[B,H,S,D], k[B,H,S,D], v[B,H,S,D]) -> o[B,H,S,D]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 8, "H": 16, "S": 4096, "D": 128}, "objective": "minimize latency_ms"}
|
| 93 |
+
{"task_id": "cuda_flash_prefill-rtx4090-B=8,H=16,S=4096,D=128-bf16-s9283758", "op": "flash_prefill", "impl": "cuda", "desc": "causal self-attention forward, fp16, fp32 softmax state", "signature": "flash_prefill(q[B,H,S,D], k[B,H,S,D], v[B,H,S,D]) -> o[B,H,S,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 8, "H": 16, "S": 4096, "D": 128}, "objective": "minimize latency_ms"}
|
| 94 |
+
{"task_id": "cuda_flash_prefill-h100-sxm-B=1,H=8,S=16384,D=64-bf16-s15249661", "op": "flash_prefill", "impl": "cuda", "desc": "causal self-attention forward, fp16, fp32 softmax state", "signature": "flash_prefill(q[B,H,S,D], k[B,H,S,D], v[B,H,S,D]) -> o[B,H,S,D]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"B": 1, "H": 8, "S": 16384, "D": 64}, "objective": "minimize latency_ms"}
|
| 95 |
+
{"task_id": "cuda_fp8_quant-rtx4090-M=4096,N=4096-fp16-s5686453", "op": "fp8_quant", "impl": "cuda", "desc": "fp8 e4m3 quantization with per-row scale, fp16 in", "signature": "fp8_quant(x[M,N]fp16, scale[1 or M]) -> y[M,N]fp8, amax[1]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "N": 4096}, "objective": "minimize latency_ms"}
|
| 96 |
+
{"task_id": "cuda_fp8_quant-mi300x-M=4096,N=4096-fp16-s8786041", "op": "fp8_quant", "impl": "cuda", "desc": "fp8 e4m3 quantization with per-row scale, fp16 in", "signature": "fp8_quant(x[M,N]fp16, scale[1 or M]) -> y[M,N]fp8, amax[1]", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 4096, "N": 4096}, "objective": "minimize latency_ms"}
|
| 97 |
+
{"task_id": "cuda_gemm_fp8-a100-80g-M=2048,N=14336,K=4096-fp16-s12506212", "op": "gemm_fp8", "impl": "cuda", "desc": "fp8 e4m3 GEMM with fp16 accum + scalar descale", "signature": "gemm_fp8(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, descale)", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 2048, "N": 14336, "K": 4096}, "objective": "minimize latency_ms"}
|
| 98 |
+
{"task_id": "cuda_kv_cache_append-h100-sxm-T=64,Hkv=8,D=128,B=8,PP=1024-bf16-s14579799", "op": "kv_cache_append", "impl": "cuda", "desc": "scatter T new tokens' K,V into paged cache via slot_map", "signature": "kv_append(k_new[T,Hkv,D], v_new[T,Hkv,D], pages[B,PP], slot_map[T]) -> cache", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"T": 64, "Hkv": 8, "D": 128, "B": 8, "PP": 1024}, "objective": "minimize latency_ms"}
|
| 99 |
+
{"task_id": "cuda_kv_cache_append-a100-80g-T=512,Hkv=4,D=128,B=16,PP=4096-bf16-s13120742", "op": "kv_cache_append", "impl": "cuda", "desc": "scatter T new tokens' K,V into paged cache via slot_map", "signature": "kv_append(k_new[T,Hkv,D], v_new[T,Hkv,D], pages[B,PP], slot_map[T]) -> cache", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"T": 512, "Hkv": 4, "D": 128, "B": 16, "PP": 4096}, "objective": "minimize latency_ms"}
|
| 100 |
+
{"task_id": "cuda_kv_cache_append-h100-sxm-T=512,Hkv=4,D=128,B=16,PP=4096-fp16-s3265632", "op": "kv_cache_append", "impl": "cuda", "desc": "scatter T new tokens' K,V into paged cache via slot_map", "signature": "kv_append(k_new[T,Hkv,D], v_new[T,Hkv,D], pages[B,PP], slot_map[T]) -> cache", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"T": 512, "Hkv": 4, "D": 128, "B": 16, "PP": 4096}, "objective": "minimize latency_ms"}
|
| 101 |
+
{"task_id": "cuda_kv_cache_append-mi300x-T=2048,Hkv=8,D=64,B=32,PP=8192-fp16-s215217", "op": "kv_cache_append", "impl": "cuda", "desc": "scatter T new tokens' K,V into paged cache via slot_map", "signature": "kv_append(k_new[T,Hkv,D], v_new[T,Hkv,D], pages[B,PP], slot_map[T]) -> cache", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"T": 2048, "Hkv": 8, "D": 64, "B": 32, "PP": 8192}, "objective": "minimize latency_ms"}
|
| 102 |
+
{"task_id": "cuda_kv_cache_append-a100-80g-T=8192,Hkv=2,D=128,B=64,PP=4096-fp16-s15283443", "op": "kv_cache_append", "impl": "cuda", "desc": "scatter T new tokens' K,V into paged cache via slot_map", "signature": "kv_append(k_new[T,Hkv,D], v_new[T,Hkv,D], pages[B,PP], slot_map[T]) -> cache", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"T": 8192, "Hkv": 2, "D": 128, "B": 64, "PP": 4096}, "objective": "minimize latency_ms"}
|
| 103 |
+
{"task_id": "cuda_embedding-a100-80g-T=4096,N=4096,V=128256-bf16-s9443451", "op": "embedding", "impl": "cuda", "desc": "embedding row gather W[idx[t]] -> y[t], fused scale variant", "signature": "embedding(idx[T], W[V,N]) -> y[T,N]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"T": 4096, "N": 4096, "V": 128256}, "objective": "minimize latency_ms"}
|
| 104 |
+
{"task_id": "cuda_embedding-mi300x-T=4096,N=4096,V=128256-bf16-s15182017", "op": "embedding", "impl": "cuda", "desc": "embedding row gather W[idx[t]] -> y[t], fused scale variant", "signature": "embedding(idx[T], W[V,N]) -> y[T,N]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"T": 4096, "N": 4096, "V": 128256}, "objective": "minimize latency_ms"}
|
| 105 |
+
{"task_id": "cuda_embedding-a100-80g-T=8192,N=8192,V=50257-fp16-s15704166", "op": "embedding", "impl": "cuda", "desc": "embedding row gather W[idx[t]] -> y[t], fused scale variant", "signature": "embedding(idx[T], W[V,N]) -> y[T,N]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"T": 8192, "N": 8192, "V": 50257}, "objective": "minimize latency_ms"}
|
| 106 |
+
{"task_id": "cuda_embedding-h100-sxm-T=16384,N=5120,V=262144-bf16-s10122391", "op": "embedding", "impl": "cuda", "desc": "embedding row gather W[idx[t]] -> y[t], fused scale variant", "signature": "embedding(idx[T], W[V,N]) -> y[T,N]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"T": 16384, "N": 5120, "V": 262144}, "objective": "minimize latency_ms"}
|
| 107 |
+
{"task_id": "triton_flash_prefill-rtx4090-B=4,H=32,S=4096,D=64-fp16-s8515615", "op": "flash_prefill", "impl": "triton", "desc": "Causal flash-attention prefill fwd over [B,H,S,D] (Triton)", "signature": "flash_prefill(q[B,H,S,D]:fp16, k[B,H,S,D]:fp16, v[B,H,S,D]:fp16) -> o[B,H,S,D]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 4, "H": 32, "S": 4096, "D": 64}, "objective": "minimize latency_ms"}
|
| 108 |
+
{"task_id": "triton_flash_prefill-mi300x-B=2,H=32,S=8192,D=128-bf16-s12579973", "op": "flash_prefill", "impl": "triton", "desc": "Causal flash-attention prefill fwd over [B,H,S,D] (Triton)", "signature": "flash_prefill(q[B,H,S,D]:fp16, k[B,H,S,D]:fp16, v[B,H,S,D]:fp16) -> o[B,H,S,D]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 2, "H": 32, "S": 8192, "D": 128}, "objective": "minimize latency_ms"}
|
| 109 |
+
{"task_id": "triton_flash_prefill-a100-80g-B=8,H=16,S=4096,D=128-fp16-s635980", "op": "flash_prefill", "impl": "triton", "desc": "Causal flash-attention prefill fwd over [B,H,S,D] (Triton)", "signature": "flash_prefill(q[B,H,S,D]:fp16, k[B,H,S,D]:fp16, v[B,H,S,D]:fp16) -> o[B,H,S,D]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 8, "H": 16, "S": 4096, "D": 128}, "objective": "minimize latency_ms"}
|
| 110 |
+
{"task_id": "triton_flash_prefill-rtx4090-B=8,H=16,S=4096,D=128-bf16-s9283758", "op": "flash_prefill", "impl": "triton", "desc": "Causal flash-attention prefill fwd over [B,H,S,D] (Triton)", "signature": "flash_prefill(q[B,H,S,D]:fp16, k[B,H,S,D]:fp16, v[B,H,S,D]:fp16) -> o[B,H,S,D]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 8, "H": 16, "S": 4096, "D": 128}, "objective": "minimize latency_ms"}
|
| 111 |
+
{"task_id": "triton_flash_prefill-h100-sxm-B=1,H=8,S=16384,D=64-bf16-s15249661", "op": "flash_prefill", "impl": "triton", "desc": "Causal flash-attention prefill fwd over [B,H,S,D] (Triton)", "signature": "flash_prefill(q[B,H,S,D]:fp16, k[B,H,S,D]:fp16, v[B,H,S,D]:fp16) -> o[B,H,S,D]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"B": 1, "H": 8, "S": 16384, "D": 64}, "objective": "minimize latency_ms"}
|
| 112 |
+
{"task_id": "triton_moe_grouped_gemm-a100-80g-T=8192,N=4096,K=4096,E=8-fp16-s7406833", "op": "moe_grouped_gemm", "impl": "triton", "desc": "MoE grouped GEMM over expert-sorted tokens via expert_ptrs (Triton)", "signature": "moe_grouped_gemm(a[T,K]:fp16, w[E,K,N]:fp16, expert_ptrs[E+1]:int32) -> c[T,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"T": 8192, "N": 4096, "K": 4096, "E": 8}, "objective": "minimize latency_ms"}
|
| 113 |
+
{"task_id": "triton_moe_grouped_gemm-h100-sxm-T=16384,N=14336,K=4096,E=16-fp16-s2184313", "op": "moe_grouped_gemm", "impl": "triton", "desc": "MoE grouped GEMM over expert-sorted tokens via expert_ptrs (Triton)", "signature": "moe_grouped_gemm(a[T,K]:fp16, w[E,K,N]:fp16, expert_ptrs[E+1]:int32) -> c[T,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"T": 16384, "N": 14336, "K": 4096, "E": 16}, "objective": "minimize latency_ms"}
|
| 114 |
+
{"task_id": "triton_moe_grouped_gemm-mi300x-T=4096,N=8192,K=8192,E=8-bf16-s651491", "op": "moe_grouped_gemm", "impl": "triton", "desc": "MoE grouped GEMM over expert-sorted tokens via expert_ptrs (Triton)", "signature": "moe_grouped_gemm(a[T,K]:fp16, w[E,K,N]:fp16, expert_ptrs[E+1]:int32) -> c[T,N]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"T": 4096, "N": 8192, "K": 8192, "E": 8}, "objective": "minimize latency_ms"}
|
| 115 |
+
{"task_id": "triton_moe_grouped_gemm-rtx4090-T=65536,N=2048,K=5120,E=128-bf16-s3714133", "op": "moe_grouped_gemm", "impl": "triton", "desc": "MoE grouped GEMM over expert-sorted tokens via expert_ptrs (Triton)", "signature": "moe_grouped_gemm(a[T,K]:fp16, w[E,K,N]:fp16, expert_ptrs[E+1]:int32) -> c[T,N]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"T": 65536, "N": 2048, "K": 5120, "E": 128}, "objective": "minimize latency_ms"}
|
| 116 |
+
{"task_id": "triton_fp8_quant-rtx4090-M=4096,N=4096-fp16-s5686453", "op": "fp8_quant", "impl": "triton", "desc": "Per-row fp8 e4m3 quantization with fp32 row_scale (Triton)", "signature": "fp8_quant(x[M,N]:fp16) -> y[M,N]:fp8, row_scale[M]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "N": 4096}, "objective": "minimize latency_ms"}
|
| 117 |
+
{"task_id": "triton_fp8_quant-mi300x-M=4096,N=4096-fp16-s8786041", "op": "fp8_quant", "impl": "triton", "desc": "Per-row fp8 e4m3 quantization with fp32 row_scale (Triton)", "signature": "fp8_quant(x[M,N]:fp16) -> y[M,N]:fp8, row_scale[M]:fp32", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 4096, "N": 4096}, "objective": "minimize latency_ms"}
|
| 118 |
+
{"task_id": "triton_fp8_quant-a100-80g-M=16384,N=7168-fp16-s34023", "op": "fp8_quant", "impl": "triton", "desc": "Per-row fp8 e4m3 quantization with fp32 row_scale (Triton)", "signature": "fp8_quant(x[M,N]:fp16) -> y[M,N]:fp8, row_scale[M]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 16384, "N": 7168}, "objective": "minimize latency_ms"}
|
| 119 |
+
{"task_id": "triton_fp8_quant-mi300x-M=16384,N=7168-bf16-s14134392", "op": "fp8_quant", "impl": "triton", "desc": "Per-row fp8 e4m3 quantization with fp32 row_scale (Triton)", "signature": "fp8_quant(x[M,N]:fp16) -> y[M,N]:fp8, row_scale[M]:fp32", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 16384, "N": 7168}, "objective": "minimize latency_ms"}
|
| 120 |
+
{"task_id": "triton_fp8_quant-a100-80g-M=2048,N=5120-fp16-s11050237", "op": "fp8_quant", "impl": "triton", "desc": "Per-row fp8 e4m3 quantization with fp32 row_scale (Triton)", "signature": "fp8_quant(x[M,N]:fp16) -> y[M,N]:fp8, row_scale[M]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 2048, "N": 5120}, "objective": "minimize latency_ms"}
|
| 121 |
+
{"task_id": "triton_cross_entropy-rtx4090-M=512,V=32000-bf16-s1914099", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 512, "V": 32000}, "objective": "minimize latency_ms"}
|
| 122 |
+
{"task_id": "triton_cross_entropy-h100-sxm-M=8192,V=32000-bf16-s1440989", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 8192, "V": 32000}, "objective": "minimize latency_ms"}
|
| 123 |
+
{"task_id": "triton_cross_entropy-h100-sxm-M=4096,V=262144-bf16-s2310311", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 124 |
+
{"task_id": "triton_cross_entropy-rtx4090-M=4096,V=262144-fp16-s7514189", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 125 |
+
{"task_id": "triton_cross_entropy-mi300x-M=4096,V=262144-fp16-s9919532", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 4096, "V": 262144}, "objective": "minimize latency_ms"}
|
| 126 |
+
{"task_id": "triton_cross_entropy-h100-sxm-M=32768,V=50257-fp16-s1683547", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 32768, "V": 50257}, "objective": "minimize latency_ms"}
|
| 127 |
+
{"task_id": "triton_cross_entropy-mi300x-M=32768,V=50257-bf16-s10988698", "op": "cross_entropy", "impl": "triton", "desc": "Per-row cross-entropy over vocab-width logits (Triton)", "signature": "cross_entropy(logits[M,V]:fp16, labels[M]:int) -> loss[M]:fp32", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"M": 32768, "V": 50257}, "objective": "minimize latency_ms"}
|
| 128 |
+
{"task_id": "triton_adamw-h100-sxm-N=268435456-fp32-s4346943", "op": "adamw", "impl": "triton", "desc": "AdamW fused optimizer step on flat fp32 states (Triton)", "signature": "adamw(p[N]:fp32, g[N]:fp32, m[N]:fp32, v[N]:fp32, lr, b1, b2, eps, wd, t)", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp32", "shape": {"N": 268435456}, "objective": "minimize latency_ms"}
|
| 129 |
+
{"task_id": "cuda_mla_decode-a100-80g-B=8,H=64,L=4096,R=512,Dr=64-fp16-s27715", "op": "mla_decode", "impl": "cuda", "desc": "MLA decode over compressed latent KV cache (R=512 + Dr=64 rope)", "signature": "mla_decode(q_nope[B,H,R], q_rope[B,H,Dr], ckv[B,L,R+Dr]) -> o[B,H,R]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 8, "H": 64, "L": 4096, "R": 512, "Dr": 64}, "objective": "minimize latency_ms"}
|
| 130 |
+
{"task_id": "cuda_mla_decode-mi300x-B=8,H=64,L=4096,R=512,Dr=64-bf16-s2568420", "op": "mla_decode", "impl": "cuda", "desc": "MLA decode over compressed latent KV cache (R=512 + Dr=64 rope)", "signature": "mla_decode(q_nope[B,H,R], q_rope[B,H,Dr], ckv[B,L,R+Dr]) -> o[B,H,R]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 8, "H": 64, "L": 4096, "R": 512, "Dr": 64}, "objective": "minimize latency_ms"}
|
| 131 |
+
{"task_id": "cuda_mla_decode-mi300x-B=64,H=128,L=4096,R=512,Dr=64-fp16-s10699999", "op": "mla_decode", "impl": "cuda", "desc": "MLA decode over compressed latent KV cache (R=512 + Dr=64 rope)", "signature": "mla_decode(q_nope[B,H,R], q_rope[B,H,Dr], ckv[B,L,R+Dr]) -> o[B,H,R]", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"B": 64, "H": 128, "L": 4096, "R": 512, "Dr": 64}, "objective": "minimize latency_ms"}
|
| 132 |
+
{"task_id": "cuda_mla_decode-mi300x-B=64,H=128,L=4096,R=512,Dr=64-bf16-s399513", "op": "mla_decode", "impl": "cuda", "desc": "MLA decode over compressed latent KV cache (R=512 + Dr=64 rope)", "signature": "mla_decode(q_nope[B,H,R], q_rope[B,H,Dr], ckv[B,L,R+Dr]) -> o[B,H,R]", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 64, "H": 128, "L": 4096, "R": 512, "Dr": 64}, "objective": "minimize latency_ms"}
|
| 133 |
+
{"task_id": "cuda_swa_decode-rtx4090-B=32,H=64,Hkv=8,D=128,L=32768,W=8192-fp16-s703611", "op": "swa_decode", "impl": "cuda", "desc": "sliding-window decode attention over last W KV tokens, GQA", "signature": "swa_decode(q[B,H,D], kv[B,Hkv,L,D], W) -> o[B,H,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 32, "H": 64, "Hkv": 8, "D": 128, "L": 32768, "W": 8192}, "objective": "minimize latency_ms"}
|
| 134 |
+
{"task_id": "cuda_swa_decode-rtx4090-B=32,H=64,Hkv=8,D=128,L=32768,W=8192-bf16-s8663266", "op": "swa_decode", "impl": "cuda", "desc": "sliding-window decode attention over last W KV tokens, GQA", "signature": "swa_decode(q[B,H,D], kv[B,Hkv,L,D], W) -> o[B,H,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 32, "H": 64, "Hkv": 8, "D": 128, "L": 32768, "W": 8192}, "objective": "minimize latency_ms"}
|
| 135 |
+
{"task_id": "cuda_swa_decode-h100-sxm-B=64,H=32,Hkv=4,D=128,L=8192,W=2048-bf16-s15993046", "op": "swa_decode", "impl": "cuda", "desc": "sliding-window decode attention over last W KV tokens, GQA", "signature": "swa_decode(q[B,H,D], kv[B,Hkv,L,D], W) -> o[B,H,D]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"B": 64, "H": 32, "Hkv": 4, "D": 128, "L": 8192, "W": 2048}, "objective": "minimize latency_ms"}
|
| 136 |
+
{"task_id": "cuda_cross_attn-rtx4090-B=8,H=16,Sq=1,Se=8192,D=64,Hkv=16-bf16-s6641715", "op": "cross_attn", "impl": "cuda", "desc": "decoder cross-attention: small Sq against long encoder context", "signature": "cross_attn(q[B,H,Sq,D], kenc[B,Hkv,Se,D], venc) -> o[B,H,Sq,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 8, "H": 16, "Sq": 1, "Se": 8192, "D": 64, "Hkv": 16}, "objective": "minimize latency_ms"}
|
| 137 |
+
{"task_id": "cuda_cross_attn-h100-sxm-B=32,H=8,Sq=4,Se=16384,D=128,Hkv=8-bf16-s9621596", "op": "cross_attn", "impl": "cuda", "desc": "decoder cross-attention: small Sq against long encoder context", "signature": "cross_attn(q[B,H,Sq,D], kenc[B,Hkv,Se,D], venc) -> o[B,H,Sq,D]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"B": 32, "H": 8, "Sq": 4, "Se": 16384, "D": 128, "Hkv": 8}, "objective": "minimize latency_ms"}
|
| 138 |
+
{"task_id": "cuda_cross_attn-a100-80g-B=4,H=16,Sq=64,Se=32768,D=64,Hkv=16-fp16-s8743049", "op": "cross_attn", "impl": "cuda", "desc": "decoder cross-attention: small Sq against long encoder context", "signature": "cross_attn(q[B,H,Sq,D], kenc[B,Hkv,Se,D], venc) -> o[B,H,Sq,D]", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"B": 4, "H": 16, "Sq": 64, "Se": 32768, "D": 64, "Hkv": 16}, "objective": "minimize latency_ms"}
|
| 139 |
+
{"task_id": "cuda_flash_attn_bwd-rtx4090-B=8,H=8,S=4096,D=128-fp16-s1173672", "op": "flash_attn_bwd", "impl": "cuda", "desc": "causal flash-attention backward: recompute P, fused dq/dk/dv", "signature": "flash_attn_bwd(q,k,v,o,do[B,H,S,D]) -> dq,dk,dv", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 8, "H": 8, "S": 4096, "D": 128}, "objective": "minimize latency_ms"}
|
| 140 |
+
{"task_id": "cuda_tree_attn-rtx4090-B=2,H=32,T=48,Hkv=4,L=16384,D=128-fp16-s14386214", "op": "tree_attn", "impl": "cuda", "desc": "speculative-decoding tree attention: T draft tokens, bitmask ancestry", "signature": "tree_attn(q[B,H,T,D], kv[B,Hkv,L+T,D], tree_mask[T,L+T]) -> o[B,H,T,D]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 2, "H": 32, "T": 48, "Hkv": 4, "L": 16384, "D": 128}, "objective": "minimize latency_ms"}
|
| 141 |
+
{"task_id": "cuda_swiglu_backward-h100-sxm-M=2048,N=18944-fp16-s6905982", "op": "swiglu_backward", "impl": "cuda", "desc": "SwiGLU backward: dg = dy*u*silu'(g), du = dy*silu(g), elementwise", "signature": "swiglu_bwd(dy[M,N]:fp16, g[M,N]:fp16, u[M,N]:fp16) -> dg[M,N]:fp16, du[M,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 2048, "N": 18944}, "objective": "minimize latency_ms"}
|
| 142 |
+
{"task_id": "cuda_cross_entropy_backward-h100-sxm-M=512,V=32000-bf16-s6426696", "op": "cross_entropy_backward", "impl": "cuda", "desc": "CE backward: dlogits = softmax(logits) - onehot(labels), row-wise", "signature": "ce_bwd(logits[M,V]:fp16, labels[M]:int32) -> dlogits[M,V]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"M": 512, "V": 32000}, "objective": "minimize latency_ms"}
|
| 143 |
+
{"task_id": "cuda_cross_entropy_backward-rtx4090-M=2048,V=128256-bf16-s7105677", "op": "cross_entropy_backward", "impl": "cuda", "desc": "CE backward: dlogits = softmax(logits) - onehot(labels), row-wise", "signature": "ce_bwd(logits[M,V]:fp16, labels[M]:int32) -> dlogits[M,V]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 2048, "V": 128256}, "objective": "minimize latency_ms"}
|
| 144 |
+
{"task_id": "cuda_embedding_backward-h100-sxm-T=4096,N=8192,V=50257-bf16-s16757782", "op": "embedding_backward", "impl": "cuda", "desc": "embedding backward: scatter-add dy rows into fp32 dW by idx", "signature": "emb_bwd(dy[T,N]:fp16, idx[T]:int32) -> dW[V,N]:fp32", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"T": 4096, "N": 8192, "V": 50257}, "objective": "minimize latency_ms"}
|
| 145 |
+
{"task_id": "cuda_embedding_backward-a100-80g-T=65536,N=2048,V=128256-fp16-s10552441", "op": "embedding_backward", "impl": "cuda", "desc": "embedding backward: scatter-add dy rows into fp32 dW by idx", "signature": "emb_bwd(dy[T,N]:fp16, idx[T]:int32) -> dW[V,N]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"T": 65536, "N": 2048, "V": 128256}, "objective": "minimize latency_ms"}
|
| 146 |
+
{"task_id": "cuda_dropout-a100-80g-M=32768,N=8192-fp16-s12241962", "op": "dropout", "impl": "cuda", "desc": "dropout forward: philox RNG keep-mask + 1/(1-p) scale, bit mask out", "signature": "dropout(x[M,N]:fp16, p, seed) -> y[M,N]:fp16, mask[M,N]:u8", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 32768, "N": 8192}, "objective": "minimize latency_ms"}
|
| 147 |
+
{"task_id": "cuda_dropout-a100-80g-M=65536,N=2048-bf16-s6826149", "op": "dropout", "impl": "cuda", "desc": "dropout forward: philox RNG keep-mask + 1/(1-p) scale, bit mask out", "signature": "dropout(x[M,N]:fp16, p, seed) -> y[M,N]:fp16, mask[M,N]:u8", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 65536, "N": 2048}, "objective": "minimize latency_ms"}
|
| 148 |
+
{"task_id": "cuda_dropout-rtx4090-M=65536,N=2048-fp16-s16204844", "op": "dropout", "impl": "cuda", "desc": "dropout forward: philox RNG keep-mask + 1/(1-p) scale, bit mask out", "signature": "dropout(x[M,N]:fp16, p, seed) -> y[M,N]:fp16, mask[M,N]:u8", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 65536, "N": 2048}, "objective": "minimize latency_ms"}
|
| 149 |
+
{"task_id": "cuda_adafactor-a100-80g-R=2048,C=32768-fp32-s6171652", "op": "adafactor", "impl": "cuda", "desc": "Adafactor update: factored 2nd moment outer(rsum,csum)/rmean", "signature": "adafactor(p[R,C]:fp32, g[R,C]:fp32, rsum[R]:fp32, csum[C]:fp32, lr, b2, eps) -> p'[R,C]:fp32, rsum'[R]:fp32, csum'[C]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp32", "shape": {"R": 2048, "C": 32768}, "objective": "minimize latency_ms"}
|
| 150 |
+
{"task_id": "cuda_adafactor-rtx4090-R=65536,C=1024-fp32-s2138147", "op": "adafactor", "impl": "cuda", "desc": "Adafactor update: factored 2nd moment outer(rsum,csum)/rmean", "signature": "adafactor(p[R,C]:fp32, g[R,C]:fp32, rsum[R]:fp32, csum[C]:fp32, lr, b2, eps) -> p'[R,C]:fp32, rsum'[R]:fp32, csum'[C]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp32", "shape": {"R": 65536, "C": 1024}, "objective": "minimize latency_ms"}
|
| 151 |
+
{"task_id": "cuda_grad_clip-rtx4090-M=32768,N=14336-bf16-s11910327", "op": "grad_clip", "impl": "cuda", "desc": "global grad-norm reduce + clip scale on one flat fp16 grad tensor", "signature": "grad_clip(g[M,N]:fp16, max_norm) -> g'[M,N]:fp16, total_norm:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 32768, "N": 14336}, "objective": "minimize latency_ms"}
|
| 152 |
+
{"task_id": "cuda_int8_gemm-h100-sxm-M=16,N=8192,K=8192-fp16-s12666068", "op": "int8_gemm", "impl": "cuda", "desc": "W8A8 int8 mma GEMM, per-token x per-channel fp16 descale", "signature": "gemm_int8(C[M,N]fp16 = A[M,K]int8 @ B[N,K]int8, row_scale[M], col_scale[N])", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 16, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 153 |
+
{"task_id": "cuda_int8_gemm-h100-sxm-M=256,N=14336,K=4096-fp16-s589983", "op": "int8_gemm", "impl": "cuda", "desc": "W8A8 int8 mma GEMM, per-token x per-channel fp16 descale", "signature": "gemm_int8(C[M,N]fp16 = A[M,K]int8 @ B[N,K]int8, row_scale[M], col_scale[N])", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 256, "N": 14336, "K": 4096}, "objective": "minimize latency_ms"}
|
| 154 |
+
{"task_id": "cuda_int8_gemm-h100-sxm-M=1024,N=8192,K=8192-fp16-s11166932", "op": "int8_gemm", "impl": "cuda", "desc": "W8A8 int8 mma GEMM, per-token x per-channel fp16 descale", "signature": "gemm_int8(C[M,N]fp16 = A[M,K]int8 @ B[N,K]int8, row_scale[M], col_scale[N])", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 1024, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 155 |
+
{"task_id": "cuda_int8_gemm-a100-80g-M=4096,N=4096,K=4096-fp16-s16739557", "op": "int8_gemm", "impl": "cuda", "desc": "W8A8 int8 mma GEMM, per-token x per-channel fp16 descale", "signature": "gemm_int8(C[M,N]fp16 = A[M,K]int8 @ B[N,K]int8, row_scale[M], col_scale[N])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 4096, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 156 |
+
{"task_id": "cuda_fp8_blockwise_gemm-a100-80g-M=4096,N=4096,K=4096-fp16-s15326412", "op": "fp8_blockwise_gemm", "impl": "cuda", "desc": "DeepSeek-style block-scaled fp8 GEMM (1x128 x 128x128 scales)", "signature": "gemm_fp8b(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, sa[M,K/128], sb[K/128,N/128])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 4096, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 157 |
+
{"task_id": "cuda_fp8_blockwise_gemm-rtx4090-M=8192,N=8192,K=8192-fp16-s2304008", "op": "fp8_blockwise_gemm", "impl": "cuda", "desc": "DeepSeek-style block-scaled fp8 GEMM (1x128 x 128x128 scales)", "signature": "gemm_fp8b(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, sa[M,K/128], sb[K/128,N/128])", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 8192, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 158 |
+
{"task_id": "cuda_fp8_blockwise_gemm-mi300x-M=8192,N=8192,K=8192-fp16-s13099019", "op": "fp8_blockwise_gemm", "impl": "cuda", "desc": "DeepSeek-style block-scaled fp8 GEMM (1x128 x 128x128 scales)", "signature": "gemm_fp8b(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, sa[M,K/128], sb[K/128,N/128])", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"M": 8192, "N": 8192, "K": 8192}, "objective": "minimize latency_ms"}
|
| 159 |
+
{"task_id": "cuda_fp8_blockwise_gemm-a100-80g-M=16384,N=4096,K=4096-fp16-s11283645", "op": "fp8_blockwise_gemm", "impl": "cuda", "desc": "DeepSeek-style block-scaled fp8 GEMM (1x128 x 128x128 scales)", "signature": "gemm_fp8b(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, sa[M,K/128], sb[K/128,N/128])", "gpu": "a100-80g", "arch": "sm_80", "dtype": "fp16", "shape": {"M": 16384, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 160 |
+
{"task_id": "cuda_fp8_blockwise_gemm-rtx4090-M=16384,N=4096,K=4096-fp16-s6432777", "op": "fp8_blockwise_gemm", "impl": "cuda", "desc": "DeepSeek-style block-scaled fp8 GEMM (1x128 x 128x128 scales)", "signature": "gemm_fp8b(C[M,N]fp16 = A[M,K]fp8 @ B[K,N]fp8, sa[M,K/128], sb[K/128,N/128])", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 16384, "N": 4096, "K": 4096}, "objective": "minimize latency_ms"}
|
| 161 |
+
{"task_id": "cuda_nucleus_sample-rtx4090-M=64,V=262144-fp16-s2792639", "op": "nucleus_sample", "impl": "cuda", "desc": "top-p (nucleus) sampling over vocab logits, per row", "signature": "nucleus_sample(logits[M,V], p, temp) -> tok[M]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 64, "V": 262144}, "objective": "minimize latency_ms"}
|
| 162 |
+
{"task_id": "cuda_nucleus_sample-rtx4090-M=64,V=262144-bf16-s7268414", "op": "nucleus_sample", "impl": "cuda", "desc": "top-p (nucleus) sampling over vocab logits, per row", "signature": "nucleus_sample(logits[M,V], p, temp) -> tok[M]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 64, "V": 262144}, "objective": "minimize latency_ms"}
|
| 163 |
+
{"task_id": "cuda_kv_int8_quant-rtx4090-T=8192,Hkv=2,D=128,B=64,PP=4096-fp16-s8752143", "op": "kv_int8_quant", "impl": "cuda", "desc": "quantize fp16 K,V into paged int8 cache, per-(token,head) scale", "signature": "kvq(k[T,Hkv,D]fp16, v, pages, slot_map[T]) -> kq[T,Hkv,D]int8, kscale[T,Hkv]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"T": 8192, "Hkv": 2, "D": 128, "B": 64, "PP": 4096}, "objective": "minimize latency_ms"}
|
| 164 |
+
{"task_id": "cuda_fused_bias_gelu-rtx4090-M=16384,N=5120-bf16-s1145969", "op": "fused_bias_gelu", "impl": "cuda", "desc": "y = gelu(x + bias), tanh approx, GEMM epilogue elementwise", "signature": "fbg(x[M,N], bias[N]) -> y[M,N]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 16384, "N": 5120}, "objective": "minimize latency_ms"}
|
| 165 |
+
{"task_id": "cuda_fused_bias_gelu-h100-sxm-M=32768,N=3072-fp16-s9277615", "op": "fused_bias_gelu", "impl": "cuda", "desc": "y = gelu(x + bias), tanh approx, GEMM epilogue elementwise", "signature": "fbg(x[M,N], bias[N]) -> y[M,N]", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "fp16", "shape": {"M": 32768, "N": 3072}, "objective": "minimize latency_ms"}
|
| 166 |
+
{"task_id": "cuda_fused_bias_gelu-rtx4090-M=2048,N=14336-bf16-s12162075", "op": "fused_bias_gelu", "impl": "cuda", "desc": "y = gelu(x + bias), tanh approx, GEMM epilogue elementwise", "signature": "fbg(x[M,N], bias[N]) -> y[M,N]", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 2048, "N": 14336}, "objective": "minimize latency_ms"}
|
| 167 |
+
{"task_id": "cuda_mamba_scan-rtx4090-B=32,D=1536,L=2048,N=32-fp16-s2512921", "op": "mamba_scan", "impl": "cuda", "desc": "Mamba selective scan: per-channel exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 32, "D": 1536, "L": 2048, "N": 32}, "objective": "minimize latency_ms"}
|
| 168 |
+
{"task_id": "cuda_mamba_scan-rtx4090-B=32,D=1536,L=2048,N=32-bf16-s7412831", "op": "mamba_scan", "impl": "cuda", "desc": "Mamba selective scan: per-channel exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 32, "D": 1536, "L": 2048, "N": 32}, "objective": "minimize latency_ms"}
|
| 169 |
+
{"task_id": "cuda_mamba_scan-rtx4090-B=2,D=2048,L=16384,N=16-bf16-s3102731", "op": "mamba_scan", "impl": "cuda", "desc": "Mamba selective scan: per-channel exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 2, "D": 2048, "L": 16384, "N": 16}, "objective": "minimize latency_ms"}
|
| 170 |
+
{"task_id": "cuda_mamba_scan-mi300x-B=2,D=2048,L=16384,N=16-fp16-s3041292", "op": "mamba_scan", "impl": "cuda", "desc": "Mamba selective scan: per-channel exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"B": 2, "D": 2048, "L": 16384, "N": 16}, "objective": "minimize latency_ms"}
|
| 171 |
+
{"task_id": "cuda_conv1d_causal-a100-80g-B=8,D=2048,L=4096,Kw=4-bf16-s15415325", "op": "conv1d_causal", "impl": "cuda", "desc": "Causal depthwise conv1d, Kw taps, left-padded (Mamba mixer)", "signature": "conv1d(x[B,D,L]:fp16, w[D,Kw]:fp16, bias[D]:fp16) -> y[B,D,L]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"B": 8, "D": 2048, "L": 4096, "Kw": 4}, "objective": "minimize latency_ms"}
|
| 172 |
+
{"task_id": "cuda_conv1d_causal-mi300x-B=4,D=4096,L=2048,Kw=4-bf16-s3348676", "op": "conv1d_causal", "impl": "cuda", "desc": "Causal depthwise conv1d, Kw taps, left-padded (Mamba mixer)", "signature": "conv1d(x[B,D,L]:fp16, w[D,Kw]:fp16, bias[D]:fp16) -> y[B,D,L]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "bf16", "shape": {"B": 4, "D": 4096, "L": 2048, "Kw": 4}, "objective": "minimize latency_ms"}
|
| 173 |
+
{"task_id": "cuda_conv1d_causal-rtx4090-B=32,D=1536,L=4096,Kw=3-bf16-s7053370", "op": "conv1d_causal", "impl": "cuda", "desc": "Causal depthwise conv1d, Kw taps, left-padded (Mamba mixer)", "signature": "conv1d(x[B,D,L]:fp16, w[D,Kw]:fp16, bias[D]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 32, "D": 1536, "L": 4096, "Kw": 3}, "objective": "minimize latency_ms"}
|
| 174 |
+
{"task_id": "cuda_moe_combine-h100-sxm-T=32768,N=7168,K=2-bf16-s6088951", "op": "moe_combine", "impl": "cuda", "desc": "MoE weighted gather-add of K expert rows per token", "signature": "moe_combine(xperm[TK,N]:fp16, w[T,K]:fp32, perm[T,K]:int32) -> y[T,N]:fp16", "gpu": "h100-sxm", "arch": "sm_90a", "dtype": "bf16", "shape": {"T": 32768, "N": 7168, "K": 2}, "objective": "minimize latency_ms"}
|
| 175 |
+
{"task_id": "cuda_moe_combine-a100-80g-T=2048,N=8192,K=8-bf16-s10641628", "op": "moe_combine", "impl": "cuda", "desc": "MoE weighted gather-add of K expert rows per token", "signature": "moe_combine(xperm[TK,N]:fp16, w[T,K]:fp32, perm[T,K]:int32) -> y[T,N]:fp16", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"T": 2048, "N": 8192, "K": 8}, "objective": "minimize latency_ms"}
|
| 176 |
+
{"task_id": "triton_mamba_scan-rtx4090-B=32,D=1536,L=2048,N=32-fp16-s2512921", "op": "mamba_scan", "impl": "triton", "desc": "Mamba selective scan in Triton: exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"B": 32, "D": 1536, "L": 2048, "N": 32}, "objective": "minimize latency_ms"}
|
| 177 |
+
{"task_id": "triton_mamba_scan-rtx4090-B=32,D=1536,L=2048,N=32-bf16-s7412831", "op": "mamba_scan", "impl": "triton", "desc": "Mamba selective scan in Triton: exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 32, "D": 1536, "L": 2048, "N": 32}, "objective": "minimize latency_ms"}
|
| 178 |
+
{"task_id": "triton_mamba_scan-rtx4090-B=2,D=2048,L=16384,N=16-bf16-s3102731", "op": "mamba_scan", "impl": "triton", "desc": "Mamba selective scan in Triton: exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"B": 2, "D": 2048, "L": 16384, "N": 16}, "objective": "minimize latency_ms"}
|
| 179 |
+
{"task_id": "triton_mamba_scan-mi300x-B=2,D=2048,L=16384,N=16-fp16-s3041292", "op": "mamba_scan", "impl": "triton", "desc": "Mamba selective scan in Triton: exp recurrence, N-wide state", "signature": "mamba_scan(x[B,D,L]:fp16, dt[B,D,L]:fp32, A[D,N]:fp32, Bm[B,L,N]:fp16, Cm[B,L,N]:fp16) -> y[B,D,L]:fp16", "gpu": "mi300x", "arch": "gfx942", "dtype": "fp16", "shape": {"B": 2, "D": 2048, "L": 16384, "N": 16}, "objective": "minimize latency_ms"}
|
| 180 |
+
{"task_id": "triton_nucleus_sample-rtx4090-M=64,V=262144-fp16-s2792639", "op": "nucleus_sample", "impl": "triton", "desc": "Nucleus top-p sampling over vocab logits (Triton)", "signature": "nucleus_sample(logits[M,V]:fp16, p, temp) -> tok[M]:int32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 64, "V": 262144}, "objective": "minimize latency_ms"}
|
| 181 |
+
{"task_id": "triton_nucleus_sample-rtx4090-M=64,V=262144-bf16-s7268414", "op": "nucleus_sample", "impl": "triton", "desc": "Nucleus top-p sampling over vocab logits (Triton)", "signature": "nucleus_sample(logits[M,V]:fp16, p, temp) -> tok[M]:int32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 64, "V": 262144}, "objective": "minimize latency_ms"}
|
| 182 |
+
{"task_id": "triton_layernorm_backward-rtx4090-M=2048,N=4096-fp16-s3471429", "op": "layernorm_backward", "impl": "triton", "desc": "LayerNorm backward in Triton: row dx + column dgamma/dbeta", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, g[N]:fp16, mean[M]:fp32, rstd[M]:fp32) -> dx[M,N]:fp16, dg[N]:fp32, db[N]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 2048, "N": 4096}, "objective": "minimize latency_ms"}
|
| 183 |
+
{"task_id": "triton_layernorm_backward-rtx4090-M=4096,N=8192-fp16-s14871768", "op": "layernorm_backward", "impl": "triton", "desc": "LayerNorm backward in Triton: row dx + column dgamma/dbeta", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, g[N]:fp16, mean[M]:fp32, rstd[M]:fp32) -> dx[M,N]:fp16, dg[N]:fp32, db[N]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "fp16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 184 |
+
{"task_id": "triton_layernorm_backward-rtx4090-M=4096,N=8192-bf16-s13910075", "op": "layernorm_backward", "impl": "triton", "desc": "LayerNorm backward in Triton: row dx + column dgamma/dbeta", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, g[N]:fp16, mean[M]:fp32, rstd[M]:fp32) -> dx[M,N]:fp16, dg[N]:fp32, db[N]:fp32", "gpu": "rtx4090", "arch": "sm_89", "dtype": "bf16", "shape": {"M": 4096, "N": 8192}, "objective": "minimize latency_ms"}
|
| 185 |
+
{"task_id": "triton_layernorm_backward-a100-80g-M=1024,N=14336-bf16-s5253228", "op": "layernorm_backward", "impl": "triton", "desc": "LayerNorm backward in Triton: row dx + column dgamma/dbeta", "signature": "ln_bwd(dy[M,N]:fp16, x[M,N]:fp16, g[N]:fp16, mean[M]:fp32, rstd[M]:fp32) -> dx[M,N]:fp16, dg[N]:fp32, db[N]:fp32", "gpu": "a100-80g", "arch": "sm_80", "dtype": "bf16", "shape": {"M": 1024, "N": 14336}, "objective": "minimize latency_ms"}
|