Functions — github.com/RightNow-AI/autokernel

CodeHub MCPcopy Create free account

hub / github.com/RightNow-AI/autokernel / functions

Functions293 in github.com/RightNow-AI/autokernel

⨍ Functions 293◇ Types & classes 28

(self, in_channels: int = 3, num_classes: int = 1000)

models/custom.py:19

(self, dim: int, n_heads: int, n_kv_heads: int)

models/llama_7b.py:52

(self, dim: int, hidden_dim: int)

models/llama_7b.py:91

(self, dim: int, n_heads: int, n_kv_heads: int, hidden_dim: int)

models/llama_7b.py:102

( self, vocab_size: int = 32000, dim: int = 768, n_layers: int = 12, n

models/llama_7b.py:122

( self, vocab_size: int = 32000, dim: int = 4096, n_layers: int = 32,

models/llama_7b.py:171

(self, hidden_size: int, num_heads: int, dropout: float = 0.0)

models/bert_base.py:17

(self, hidden_size: int, num_heads: int, intermediate_size: int, dropout: float = 0.0)

models/bert_base.py:56

( self, vocab_size: int = 30522, hidden_size: int = 768, num_layers: int = 12,

models/bert_base.py:74

Function_get_freqs

Get or compute cached cos/sin frequency tables.

kernels/cuda/rotary_embedding.py:157

(self, signum, frame)

(self, signum, frame)

kernelbench/bench_kb.py:92

Function_load_result_rows

Load all result rows for a kernel. Returns empty list if no file.

orchestrate.py:227

Function_output_shape_str

Get shape string from model output.

Function_ref_cross_entropy

Function_ref_flash_attention

Function_ref_fused_mlp

Function_ref_layernorm

Function_ref_matmul

Function_ref_reduce

Function_ref_rmsnorm

Function_ref_rotary_embedding

Function_ref_softmax

Method_thread_raise

kernelbench/bench_kb.py:95

Method_timeout_thread

Methodapplied_summary

Methodcache_path

kernelbench/bridge.py:61

Functioncompile_cuda_with_wrapper

Like compile_cuda, but auto-generates the C++ wrapper from arg_specs. Parameters ---------- cuda_src : str CUDA source that

kernels/cuda/_compile.py:258

Functioncross_entropy_kernel

Fused cross-entropy: log_softmax + nll_loss per row. One program per row (batch element).

kernels/cross_entropy.py:21

Functioncross_entropy_ref

Standard cross entropy loss.

reference.py:58

Functionflash_attention_kernel

Flash attention with online softmax. One program per (batch, head, query-block).

kernels/flash_attention.py:23

Functionflash_attention_ref

Standard scaled dot-product attention.

reference.py:32

(self, x: torch.Tensor)

(self, x: torch.Tensor)

(self, x: torch.Tensor)

(self, x: torch.Tensor)

models/gpt2.py:37

(self, x: torch.Tensor)

models/gpt2.py:65

(self, x: torch.Tensor)

models/gpt2.py:81

(self, input_ids: torch.Tensor)

models/gpt2.py:126

(self, x: torch.Tensor)

models/custom.py:39

(self, x: torch.Tensor)

models/llama_7b.py:28

(self, x: torch.Tensor, freqs_cis: torch.Tensor)

models/llama_7b.py:64

(self, x: torch.Tensor)

models/llama_7b.py:97

(self, x: torch.Tensor, freqs_cis: torch.Tensor)

models/llama_7b.py:109

(self, input_ids: torch.Tensor)

models/llama_7b.py:151

(self, input_ids: torch.Tensor)

models/llama_7b.py:199

(self, x: torch.Tensor, attention_mask: torch.Tensor = None)

models/bert_base.py:29

(self, x: torch.Tensor)

models/bert_base.py:51

(self, x: torch.Tensor)

models/bert_base.py:63

(self, input_ids: torch.Tensor)

models/bert_base.py:101

Functionfused_gate_up_kernel

Fused kernel: computes activation(X @ W_gate^T) * (X @ W_up^T). W_gate and W_up are [intermediate_size, hidden_size] (transposed access).

kernels/fused_mlp.py:26

Functionfused_mlp_ref

SwiGLU-style fused MLP: down(activation(gate(x)) * up(x)).

reference.py:45

Functiongen_cross_entropy_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_flash_attention_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_fused_mlp_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_layernorm_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_matmul_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_reduce_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_rmsnorm_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_rotary_embedding_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functiongen_softmax_inputs

(size: dict, dtype: torch.dtype, device: str, seed: int = 42)

Functionkernel_fn

Entry point called by bench.py. Must match reference.layernorm_ref signature.

kernels/layernorm.py:65

Functionkernel_fn

Entry point called by bench.py. Must match reference.fused_mlp_ref signature. SwiGLU MLP: hidden = activation(x @ w_gate.T) * (x @ w_u

kernels/fused_mlp.py:102

Functionkernel_fn

Entry point called by bench.py. Must match reference.rotary_embedding_ref signature. Args: x: [..., head_dim] tensor to apply rotary

kernels/rotary_embedding.py:75

Functionkernel_fn

Entry point called by bench.py. Must match reference.cross_entropy_ref signature. Args: logits: [batch_size, vocab_size] raw logits

kernels/cross_entropy.py:70

Functionkernel_fn

Entry point called by bench.py. Must match reference.flash_attention_ref signature. Args: Q: [batch, heads, seq_len, head_dim]

kernels/flash_attention.py:130

Functionkernel_fn

Entry point called by bench.py. Must match reference.softmax_ref signature.

kernels/softmax.py:57

Functionkernel_fn

Entry point called by bench.py. Must match reference.matmul_ref signature.

kernels/matmul.py:62

Functionkernel_fn

Entry point called by bench.py. Must match reference.reduce_sum_ref signature. Args: x: Input tensor of any shape dim: Dimen

kernels/reduce.py:51

Functionkernel_fn

Entry point called by bench.py. Must match reference.rmsnorm_ref signature.

kernels/rmsnorm.py:45

Functionkernel_fn

Entry point called by bench.py. Must match reference.layernorm_ref signature.

kernels/cuda/layernorm.py:320

Functionkernel_fn

Entry point called by bench.py. Must match reference.fused_mlp_ref signature.

kernels/cuda/fused_mlp.py:170

Functionkernel_fn

Entry point called by bench.py. Must match reference.rotary_embedding_ref signature.

kernels/cuda/rotary_embedding.py:173

Functionkernel_fn

Entry point called by bench.py. Must match reference.cross_entropy_ref signature.

kernels/cuda/cross_entropy.py:151

Functionkernel_fn

Entry point called by bench.py. Must match reference.flash_attention_ref signature.

kernels/cuda/flash_attention.py:248

Functionkernel_fn

Entry point called by bench.py. Must match reference.softmax_ref signature.

kernels/cuda/softmax.py:245

Functionkernel_fn

Entry point called by bench.py. Must match reference.matmul_ref signature.

kernels/cuda/matmul.py:246

Functionkernel_fn

Entry point called by bench.py. Must match reference.reduce_sum_ref signature.

kernels/cuda/reduce.py:111

Functionkernel_fn

Entry point called by bench.py. Must match reference.rmsnorm_ref signature.

kernels/cuda/rmsnorm.py:171

Functionlayernorm_kernel

Row-parallel layer normalization. One program per row.

kernels/layernorm.py:20

Functionlayernorm_ref

Layer normalization over last dimension.

reference.py:20

Functionmatmul_kernel

Basic tiled matmul. The agent improves this.

Functionmatmul_kernel

Basic tiled matmul. The agent improves this.

kernels/matmul.py:25

Functionreduce_max_ref

reference.py:76

Functionreduce_sum_kernel

Parallel sum reduction. One program per output element. Reduces over `reduce_size` elements with stride `stride_x_col`.

kernels/reduce.py:21

Functionreduce_sum_ref

reference.py:72

Functionrmsnorm_kernel

Row-parallel RMS normalization.

kernels/rmsnorm.py:14

Functionrmsnorm_ref

RMS normalization.

reference.py:26

Functionrotary_embedding_kernel

Apply rotary embeddings using interleaved (even/odd) decomposition. One program per row. Each row has head_dim elements. x1 = even-index

kernels/rotary_embedding.py:25

Functionrotary_embedding_ref

Apply rotary position embeddings.

reference.py:63

Functionsoftmax_kernel

Row-parallel online softmax. One program per row.

kernels/softmax.py:20

Functionsoftmax_ref

Standard softmax along dim.

reference.py:15

kernelbench/bridge.py:57

← previous201–293 of 293, ranked by callers