MCPcopy Create free account

hub / github.com/HazyResearch/HipKittens / functions

Functions1,340 in github.com/HazyResearch/HipKittens

⨍ Functions 1,340◇ Types & classes 190↳ Endpoints 1
↓ 123 callersMethodwrite
(self, sizes, doc_idx)
training/llama/train/datamodules/datasets/indexed_dataset.py:380
↓ 113 callersMethodget
Retrieves a single item from the dataset with the option to only return a portion of the item. get(idx) is the same as [idx] but get(
training/llama/train/datamodules/datasets/indexed_dataset.py:531
↓ 63 callersFunctionprefill_swizzled_offsets
kernels/gemm/fp8fp32/FP8_4wave/utils.cpp:107
↓ 62 callersFunctionexp2
kernels/attn/gqa_backwards/attn_fwd_non_causal.cpp:55
↓ 51 callersFunctionceil_div
kernels/gemm/bf16fp32/micros/hint_based/kernel.cpp:26
↓ 51 callersFunctionload_lds_reg
kernels/gemm/bf16fp32/archive/utils.cpp:376
↓ 47 callersMethodempty
Create empty tensor (returns IrisTensor)
distributed-kernels/iris_py.cpp:72
↓ 38 callersFunctionexp2
analysis/attn/bkwd/benchmark/attn_fwd_causal.cpp:56
↓ 35 callersMethodexists
(path)
training/llama/train/datamodules/datasets/indexed_dataset.py:217
↓ 22 callersMethodnumel
distributed-kernels/iris_py.cpp:32
↓ 21 callersMethodsize
(self, index)
training/llama/train/datamodules/datasets/indexed_dataset.py:213
↓ 20 callersFunctionefficiency
Calculate efficiency in TFLOPS given time in ms and flop count in FLOPS.
analysis/attn/bkwd/plot.py:11
↓ 20 callersFunctionflops
(batch, seqlen, nheads, headdim, causal, mode="bwd")
analysis/attn/bkwd/plot.py:6
↓ 19 callersFunctionsafe_ratio
(n, d)
analysis/bf16_gemm/mi350x/analyze_prof.py:76
↓ 19 callersFunctionsafe_ratio
(n, d)
docs/profiling/analyze_pmc_counter_output.py:76
↓ 18 callersFunctionexp2
analysis/attn/bkwd/benchmark/attn_fwd_non_causal.cpp:55
↓ 18 callersFunctionexp2
kernels/attn/gqa_causal_backwards/attn_fwd_causal.cpp:56
↓ 18 callersFunctionexp2
training/llama/csrc/attn_fwd_causal.cpp:56
↓ 17 callersFunctionexp2
analysis/attn/fwd/benchmark/attn_fwd_causal.cpp:59
↓ 17 callersFunctionexp2
analysis/attn/fwd/benchmark/attn_fwd_causal_d64.cpp:59
↓ 17 callersFunctionexp2
analysis/attn/fwd/benchmark/attn_fwd_non_causal.cpp:57
↓ 17 callersFunctionexp2
analysis/attn/fwd/benchmark/attn_fwd_non_causal_d64.cpp:58
↓ 17 callersFunctionexp2
kernels/attn/gqa/kernel.cpp:58
↓ 17 callersFunctionexp2
kernels/attn/gqa/kernel_d64.cpp:58
↓ 17 callersFunctionexp2
kernels/attn/gqa_causal/kernel.cpp:59
↓ 17 callersFunctionexp2
kernels/attn/gqa_causal/kernel_d64.cpp:59
↓ 10 callersFunctionrun_bash_command
(commandstring, capture=True)
analysis/baselines/gemm/utils/utils.py:54
↓ 9 callersMethod__init__
TD [2021-10-27] act_fn takes precedence over act_layer if set. This is to support Pytorch 1.10 Transformer interface that construct the activa
training/llama/llama/models/seq_common.py:210
↓ 8 callersMethodbackward
(ctx, do)
analysis/baselines/attn/triton_baseline_v01.py:480
↓ 8 callersMethodbackward
(ctx, do)
training/llama/llama/models/rotary.py:73
↓ 7 callersMethodK
kernels/gemm/bf16fp32/gfx1250/common.h:42
↓ 7 callersMethodM
kernels/gemm/bf16fp32/gfx1250/common.h:40
↓ 7 callersMethodN
kernels/gemm/bf16fp32/gfx1250/common.h:41
↓ 7 callersMethodblock
kernels/gemm/bf16fp32/gfx1250/common.h:44
↓ 7 callersFunctionceil_div
analysis/paper_experiments/grid_micro/kernel_8192_w0.cpp:24
↓ 7 callersFunctiondata_file_path
(prefix_path)
training/llama/train/datamodules/datasets/indexed_dataset.py:124
↓ 7 callersFunctionget_wandb_logger
Safely get Weights&Biases logger from Trainer.
training/llama/train/callbacks/wandb_callbacks.py:16
↓ 7 callersMethodgrid
kernels/gemm/bf16fp32/gfx1250/common.h:43
↓ 7 callersFunctionprint_rank_zero
(*args, **kwargs)
training/llama/train/utils/utils.py:131
↓ 6 callersFunctionindex_file_path
(prefix_path)
training/llama/train/datamodules/datasets/indexed_dataset.py:120
↓ 6 callersFunctioninit_randn
(shape, dtype, device, scale=1)
kernels/gemm/bf16fp32/utils.py:6
↓ 5 callersMethodgrid
analysis/paper_experiments/phases/ds_read_b64/kernel.cpp:13
↓ 5 callersFunctionmask_kv_tile
analysis/attn/bkwd/benchmark/attn_fwd_causal.cpp:95
↓ 5 callersFunctionmask_kv_tile
analysis/attn/fwd/benchmark/attn_fwd_causal.cpp:116
↓ 5 callersFunctionmask_kv_tile
analysis/attn/fwd/benchmark/attn_fwd_causal_d64.cpp:116
↓ 5 callersFunctionmask_kv_tile
kernels/attn/gqa_causal_backwards/attn_fwd_causal.cpp:95
↓ 5 callersFunctionmask_kv_tile
kernels/attn/gqa_causal/kernel.cpp:116
↓ 5 callersFunctionmask_kv_tile
kernels/attn/gqa_causal/kernel_d64.cpp:116
↓ 5 callersFunctionmask_kv_tile
training/llama/csrc/attn_fwd_causal.cpp:95
↓ 5 callersMethodneed_causal
(self)
analysis/baselines/attn/triton_baseline_v02.py:92
↓ 5 callersFunctionprocess_data
Separate numeric values and OOM indices
analysis/bf16_gemm/plot.py:33
↓ 4 callersFunctionapply_rotary
Arguments: x: (batch, seqlen, nheads, headdim) if cu_seqlens is None else (total_seqlen, nheads, headdim). cos: (seql
training/llama/llama/ops/triton/rotary.py:144
↓ 4 callersMethodbackward
(ctx, *gradients)
analysis/baselines/attn/triton_baseline_v02.py:1180
↓ 4 callersMethoddtype
(self)
training/llama/train/datamodules/datasets/indexed_dataset.py:454
↓ 4 callersFunctionefficiency
Calculate efficiency in TFLOPS.
analysis/rotary/mi350x/test_python.py:31
↓ 4 callersFunctionefficiency
Calculate efficiency in TFLOPS.
kernels/rotary/test_python.py:29
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
analysis/baselines/attn/attn_bwd_baselines.py:144
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
analysis/attn/bkwd/benchmark/test_python.py:148
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
kernels/attn/gqa_backwards/test_python.py:147
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
kernels/attn/gqa_backwards/archive/test_python.py:147
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
kernels/attn/gqa_causal_backwards/test_python.py:147
↓ 4 callersFunctiongenerate_tensor
(shape, mean, std, dtype, device)
training/llama/csrc/test.py:147
↓ 4 callersFunctioninput_helper
(Z, HQ, HK, N_CTX_Q, N_CTX_K, D_HEAD, dtype, layout, requires_grad=True)
analysis/baselines/attn/triton_baseline_v02.py:1317
↓ 4 callersFunctionis_cuda
()
analysis/baselines/gemm/triton_gemm_v01.py:158
↓ 4 callersFunctionkittens_store
distributed-kernels/bf16_gemm/kernel.cpp:25
↓ 4 callersFunctionload_fn
(ptrs, offset_first, offset_second, boundary_first, boundary_second)
analysis/baselines/attn/triton_baseline_v02.py:173
↓ 4 callersFunctionpooling
(x, pooling_mode='CLS', key_padding_mask=None, batch_first=True)
training/llama/llama/models/seq_common.py:15
↓ 4 callersFunctionread_longs
(f, n)
training/llama/train/datamodules/datasets/indexed_dataset.py:91
↓ 4 callersFunctionrobustness_check
(ref, pred)
analysis/attn/bkwd/benchmark/test_python.py:57
↓ 4 callersFunctionrobustness_check
(ref, pred)
kernels/attn/gqa_backwards/test_python.py:56
↓ 4 callersFunctionrobustness_check
(ref, pred)
kernels/attn/gqa_backwards/archive/test_python.py:56
↓ 4 callersFunctionrobustness_check
(ref, pred)
kernels/attn/gqa_causal_backwards/test_python.py:56
↓ 4 callersFunctionrobustness_check
(ref, pred)
training/llama/csrc/test.py:56
↓ 4 callersFunctionwrite_longs
(f, a)
training/llama/train/datamodules/datasets/indexed_dataset.py:97
↓ 3 callersMethod__init__
(self, config: GPT2Config, device=None, dtype=None)
training/llama/llama/models/gpt.py:175
↓ 3 callersMethod_collate
(cls, batch, *args, **kwargs)
training/llama/train/datamodules/base.py:50
↓ 3 callersMethod_collate_callback
Modify the behavior of the default _collate method.
training/llama/train/datamodules/base.py:30
↓ 3 callersMethod_data_loader
(self, dataset: Dataset, batch_size: int, shuffle: bool = False, sampler=None)
training/llama/train/datamodules/language_modeling_hf.py:280
↓ 3 callersMethod_eval_dataloader
(self, dataset, **kwargs)
training/llama/train/datamodules/base.py:229
↓ 3 callersMethodbarrier
distributed-kernels/iris_py.cpp:99
↓ 3 callersFunctionbench_gemm
(gemm_params, gemm_func, transpose_B=False, num_warmup=500, num_iter=500)
kernels/gemm/bf16fp32/utils.py:20
↓ 3 callersFunctioncompute_alibi_block
(alibi_slope, seqlen_q, seqlen_k, offs_m, offs_n, transpose=False)
analysis/baselines/attn/triton_baseline_v02.py:199
↓ 3 callersFunctioncompute_alibi_tensor
(alibi_slopes, seqlen_q, seqlen_k)
analysis/baselines/attn/triton_baseline_v02.py:231
↓ 3 callersFunctioncreate_data_loader
(df: pd.DataFrame, tokenizer, max_len: int, batch_size: int)
training/bert/tasks.py:112
↓ 3 callersMethoddynamic_shared_memory
kernels/gemm/bf16fp32/gfx1250/common.h:50
↓ 3 callersFunctionefficiency
Calculate efficiency in TFLOPS.
analysis/layernorm/mi350x/test_python.py:31
↓ 3 callersFunctionefficiency
Calculate efficiency in TFLOPS.
kernels/layernorm/test_python.py:31
↓ 3 callersFunctioneval_model
( model: nn.Module, data_loader: DataLoader, loss_fn, device, n_examples: int, limit:
training/bert/tasks.py:268
↓ 3 callersMethodforward
(self, *args, **kwargs)
training/llama/train/tasks/seq.py:98
↓ 3 callersFunctionget_logger
Initializes multi-GPU-friendly python logger.
training/llama/train/utils/utils.py:37
↓ 3 callersFunctionget_shape_from_layout
(q, k, metadata)
analysis/baselines/attn/triton_baseline_v02.py:1046
↓ 3 callersFunctioninit_empty
(shape, dtype, device)
kernels/gemm/bf16fp32/utils.py:9
↓ 3 callersMethodload_state_dict
(self, checkpoint)
training/llama/train/datamodules/language_modeling_hf.py:293
↓ 3 callersFunctionmake_iris_tensor
Create iris tensor and wrap as PyTorch tensor (HACK using __cuda_array_interface__)
distributed-kernels/bf16_gemm/example.py:9
↓ 3 callersFunctionmatmul
(a, b, activation="")
analysis/baselines/gemm/triton_gemm_v01.py:330
↓ 3 callersMethodneed_alibi
(self, alibi_slopes, batch, nheads)
analysis/baselines/attn/triton_baseline_v02.py:85
↓ 3 callersFunctionprint_title
(title, len=30)
kernels/gemm/bf16fp32/utils.py:15
↓ 3 callersFunctionprocess_data
Separate numeric values and OOM indices
analysis/attn/bkwd/plot.py:103
↓ 3 callersFunctionprocess_data
Separate numeric values and OOM indices
analysis/attn/fwd/plot_d64.py:108
↓ 3 callersFunctionprocess_data
Separate numeric values and OOM indices
analysis/attn/fwd/plot.py:113
next →1–100 of 1,340, ranked by callers