data_source large_stringclasses 1
value | prompt listlengths 1 1 | ability large_stringclasses 1
value | reward_model dict | extra_info dict |
|---|---|---|---|---|
cuda_llm | [
{
"content": "You write custom Triton-Ascend kernels to replace the pytorch operators in the given architecture to get speedups.\n\nYou have complete freedom to choose which operators to replace. You may replace multiple operators, fuse operators (combine several into one kernel, e.g. matmul+relu), or change th... | kernel_optimization | {
"ground_truth": "import torch\nimport torch_npu\nimport torch.nn as nn\n\n\nclass Model(nn.Module):\n def __init__(self, in_channels, out_channels, kernel_size, stride, padding, bn_num_features):\n super().__init__()\n self.identity = nn.Identity()\n self.conv3d = nn.Conv3d(in_channels, out_... | {
"difficulty_level": "L3",
"difficulty_score": 9.85,
"entry_point": "Model",
"has_3d": true,
"heavy_ops": 2,
"level": "0",
"module_name": "Model",
"num_ops": 3,
"ops": "[\"nn.Identity\", \"nn.Conv3d\", \"nn.BatchNorm3d\"]",
"original_prompt": [
{
"content": "You are an expert in PyTorch a... |
cuda_llm | [
{
"content": "You write custom Triton-Ascend kernels to replace the pytorch operators in the given architecture to get speedups.\n\nYou have complete freedom to choose which operators to replace. You may replace multiple operators, fuse operators (combine several into one kernel, e.g. matmul+relu), or change th... | kernel_optimization | {
"ground_truth": "import torch\nimport torch_npu\n\n\nclass Model(torch.nn.Module):\n def __init__(self, max_value):\n super(Model, self).__init__()\n self.max_value = max_value\n\n def forward(self, x, y):\n return torch.fmin(x, torch.tensor(self.max_value))\n\n\ndef get_inputs():\n re... | {
"difficulty_level": "L1",
"difficulty_score": 1.9,
"entry_point": "Model",
"has_3d": false,
"heavy_ops": 0,
"level": "0",
"module_name": "Model",
"num_ops": 1,
"ops": "[\"torch.fmin\"]",
"original_prompt": [
{
"content": "You are an expert in PyTorch and CUDA programming. You will be giv... |
cuda_llm | [
{
"content": "You write custom Triton-Ascend kernels to replace the pytorch operators in the given architecture to get speedups.\n\nYou have complete freedom to choose which operators to replace. You may replace multiple operators, fuse operators (combine several into one kernel, e.g. matmul+relu), or change th... | kernel_optimization | {
"ground_truth": "import torch\nimport torch_npu\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\nclass Model(nn.Module):\n def __init__(self, unfold_kernel_size, unfold_dilation, unfold_padding, unfold_stride, num_classes, embedding_dim):\n super().__init__()\n self.unfold = nn.Unfold(\... | {
"difficulty_level": "L2",
"difficulty_score": 6.1,
"entry_point": "Model",
"has_3d": false,
"heavy_ops": 0,
"level": "0",
"module_name": "Model",
"num_ops": 4,
"ops": "[\"nn.Unfold\", \"torch.std\", \"F.one_hot\", \"nn.ParameterDict\"]",
"original_prompt": [
{
"content": "You are an expe... |
cuda_llm | [
{
"content": "You write custom Triton-Ascend kernels to replace the pytorch operators in the given architecture to get speedups.\n\nYou have complete freedom to choose which operators to replace. You may replace multiple operators, fuse operators (combine several into one kernel, e.g. matmul+relu), or change th... | kernel_optimization | {
"ground_truth": "import torch\nimport torch_npu\n\n\nclass Model(torch.nn.Module):\n def __init__(self, k):\n super(Model, self).__init__()\n self.k = k\n\n def forward(self, x):\n x, indices = torch.topk(x, self.k)\n x = torch.softmax(x, dim=0)\n x = torch.gather(x, dim=0, ... | {
"difficulty_level": "L2",
"difficulty_score": 4.1,
"entry_point": "Model",
"has_3d": false,
"heavy_ops": 0,
"level": "0",
"module_name": "Model",
"num_ops": 3,
"ops": "[\"torch.topk\", \"torch.softmax\", \"torch.gather\"]",
"original_prompt": [
{
"content": "You are an expert in PyTorch ... |
cuda_llm | [
{
"content": "You write custom Triton-Ascend kernels to replace the pytorch operators in the given architecture to get speedups.\n\nYou have complete freedom to choose which operators to replace. You may replace multiple operators, fuse operators (combine several into one kernel, e.g. matmul+relu), or change th... | kernel_optimization | {
"ground_truth": "import torch\nimport torch_npu\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\nclass Model(nn.Module):\n def __init__(self, num_embeddings, embedding_dim, padding_idx=None):\n super().__init__()\n self.embedding = nn.Embedding(num_embeddings, embedding_dim, padding_idx... | {
"difficulty_level": "L2",
"difficulty_score": 6.5,
"entry_point": "Model",
"has_3d": false,
"heavy_ops": 0,
"level": "0",
"module_name": "Model",
"num_ops": 5,
"ops": "[\"nn.Embedding\", \"F.softplus\", \"torch.addcmul\", \"torch.log_softmax\", \"torch.tril\"]",
"original_prompt": [
{
"c... |
cuda_llm | [{"content":"You write custom Triton-Ascend kernels to replace the pytorch operators in the given ar(...TRUNCATED) | kernel_optimization | {"ground_truth":"import torch\nimport torch_npu\nimport torch.nn as nn\n\n\nclass Model(nn.Module):\(...TRUNCATED) | {"difficulty_level":"L2","difficulty_score":5.55,"entry_point":"Model","has_3d":false,"heavy_ops":1,(...TRUNCATED) |
cuda_llm | [{"content":"You write custom Triton-Ascend kernels to replace the pytorch operators in the given ar(...TRUNCATED) | kernel_optimization | {"ground_truth":"import torch\nimport torch_npu\n\n\nclass Model(torch.nn.Module):\n def __init__(...TRUNCATED) | {"difficulty_level":"L2","difficulty_score":4.1,"entry_point":"Model","has_3d":false,"heavy_ops":0,"(...TRUNCATED) |
cuda_llm | [{"content":"You write custom Triton-Ascend kernels to replace the pytorch operators in the given ar(...TRUNCATED) | kernel_optimization | {"ground_truth":"import torch\nimport torch_npu\nimport torch.nn as nn\nimport torch.nn.functional a(...TRUNCATED) | {"difficulty_level":"L3","difficulty_score":7.8,"entry_point":"Model","has_3d":false,"heavy_ops":0,"(...TRUNCATED) |
cuda_llm | [{"content":"You write custom Triton-Ascend kernels to replace the pytorch operators in the given ar(...TRUNCATED) | kernel_optimization | {"ground_truth":"import torch\nimport torch_npu\nfrom torch import nn\n\n\nclass Model(nn.Module):\n(...TRUNCATED) | {"difficulty_level":"L1","difficulty_score":2.3,"entry_point":"Model","has_3d":false,"heavy_ops":0,"(...TRUNCATED) |
cuda_llm | [{"content":"You write custom Triton-Ascend kernels to replace the pytorch operators in the given ar(...TRUNCATED) | kernel_optimization | {"ground_truth":"import torch\nimport torch_npu\nimport torch.nn as nn\nimport torch.nn.functional a(...TRUNCATED) | {"difficulty_level":"L2","difficulty_score":6.7,"entry_point":"Model","has_3d":false,"heavy_ops":1,"(...TRUNCATED) |
End of preview. Expand in Data Studio
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
dr-kernel-RL-v4 (Ascend-prompt edition)
Derived from AhNr/dr-kernel-RL.
Rows, labels, reward_model, and extra_info are identical; the only change is the
shared prompt preamble, edited to teach the Ascend NPU execution model so the policy can
act on msprof feedback.
What changed (uniform across all rows):
- Fixed the few-shot example to the NPU idiom:
grid = (min(triton.cdiv(N, BLOCK_SIZE), 48),), an in-core grid-stride loop,BLOCK_SIZE=8192, andmultibuffer=True(was CUDA-stylegrid = cdiv(N, BLOCK),BLOCK_SIZE=128, one-program-per-tile). - Added an "Ascend NPU execution model" rules block: grid = physical cores (48 vector / 24 cube),
192 KB UB tiling constraint,
multibuffer=Truefor memory-bound kernels, vectorize +tl.constexpr.
Net prompt cost: ~+138 tokens/row.
- Downloads last month
- 16