Functions — github.com/facebookresearch/xformers

CodeHub MCPcopy Create free account

hub / github.com/facebookresearch/xformers / functions

Functions1,445 in github.com/facebookresearch/xformers

⨍ Functions 1,445◇ Types & classes 346↳ Endpoints 17

Methodbuild_env

xformers/_cpp_lib.py:51

Methodbuild_extensions

xformers/benchmarks/benchmark_indexing.py:85

xformers/benchmarks/benchmark_indexing.py:134

xformers/benchmarks/benchmark_sp24.py:131

Functioncache_prefix

Take a prefix view of a larger cache. The original cache object remains of identical size and valid after the shrinked alias has been us

examples/llama_inference/model.py:371

Functioncausal_1d_pattern

(attn_size: int)

xformers/components/attention/attention_patterns.py:92

Methodcleanup_at_exit

xformers/fwbw_overlap.py:379

Functioncompute_once

xformers/__init__.py:32

Methodcopy_extensions_to_source

Used for `pip install -e .` Copies everything we built back into the source repo

MethodcreateTensors

xformers/csrc/sparse24/sparse24_metadata.h:244

MethodcreateTensors

xformers/csrc/sparse24/sparse24_metadata.h:269

Methodcreate_compressed_representation

xformers/csrc/sparse24/sparse24_metadata.h:65

Methodcreate_compressed_representation

xformers/csrc/sparse24/sparse24_metadata.h:157

Methodcuda_version

xformers/_cpp_lib.py:27

Methodcurrent_stream_wait

xformers/fwbw_overlap.py:35

Methodcurrent_stream_wait

xformers/fwbw_overlap.py:100

Functiondeprecated_function

xformers/_deprecation_warning.py:9

Functiondequant_func

xformers/ops/fmha/common.py:133

Functiondilated_2d_pattern

Returns a 2d pattern that samples 1 every k elements in the attention mask. Can be seen as a form of downsampling, where every pixel attends

xformers/components/attention/attention_patterns.py:155

Functionearly_config_prune

(configs, named_args, **kwargs)

xformers/ops/fmha/_triton/splitk_kernels.py:631

Functionefficient_attention_backward_ck

xformers/csrc/attention/hip_fmha/attention_backward_generic_ck_tiled.cpp:36

Functionefficient_attention_backward_ck_meta

xformers/csrc/attention/hip_fmha/attention_backward_generic_ck_tiled.cpp:548

Functionefficient_attention_forward_ck

There are 2 modes for using this function. (Mode BMHK) With all the heads having the same seqlen (Mode 1MHK) `batch=1` with all tokens across batc

xformers/csrc/attention/hip_fmha/attention_forward_generic_ck_tiled.cpp:50

Functionefficient_attention_forward_ck_meta

There are 2 modes for using this function. (Mode BMHK) With all the heads having the same seqlen (Mode 1MHK) `batch=1` with all tokens across batc

xformers/csrc/attention/hip_fmha/attention_forward_generic_ck_tiled.cpp:475

Functionefficient_attention_forward_decoder_splitk_ck

xformers/csrc/attention/hip_decoder/attention_forward_splitk.cpp:293

Functionefficient_attention_forward_decoder_splitk_ck_impl

xformers/csrc/attention/hip_decoder/attention_forward_splitk.cpp:251

Methodenable_if_t<Cond, Kargs> MakeKargs

xformers/csrc/attention/hip_fmha/ck_tiled_rand_uniform_kernel.h:88

Methodext_vector_t<float, vec_size> scalar_scale_acc

xformers/csrc/attention/hip_decoder/ck_tile_attention_forward_decoder_splitk.h:10

( q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, )

tests/test_fmha_merge_attentions.py:646

Functionfallback_dispatcher

(func, types, args, kwargs)

xformers/ops/sp24.py:236

Methodflash_version

xformers/_cpp_lib.py:43

Functionflush_pending_bw

Flush all BW chunks that we can compute recursively

xformers/fwbw_overlap.py:488

xformers/ops/fmha/cutlass_blackwell.py:435

(self, *args, **kwargs)

xformers/checkpoint.py:520

(ctx: torch.autograd.function.FunctionCtx, *tensors: torch.Tensor)

xformers/fwbw_overlap.py:147

( ctx: torch.autograd.function.FunctionCtx, holder: EventOverlapHolder, *tensors: torc

xformers/fwbw_overlap.py:176

( ctx: torch.autograd.function.FunctionCtx, holder: EventOverlapHolder, *tensors: torc

xformers/fwbw_overlap.py:206

( ctx: torch.autograd.function.FunctionCtx, boundary: PhaseBoundary, *x: torch.Tensor,

xformers/fwbw_overlap.py:399

(ctx, x: torch.Tensor, algo: str, gradient: str, backend: str)

xformers/ops/sp24.py:602

( ctx, x: torch.Tensor, algo: str, backend: str, bw_mul0: float,

xformers/ops/sp24.py:648

(ctx, x: torch.Tensor, pattern: Sparse24Tensor, gradient: str, backend: str)

xformers/ops/sp24.py:682

( # type: ignore[override] ctx, input_: torch.Tensor, process_group: torch.distributed.ProcessGroup

xformers/ops/differentiable_collectives.py:85

( # type: ignore[override] ctx, input_: torch.Tensor, process_group: torch.distributed.ProcessGroup

xformers/ops/differentiable_collectives.py:109

( # type: ignore[override] ctx, x: torch.Tensor, process_group: torch.distributed.ProcessGroup )

xformers/ops/differentiable_collectives.py:133

( # type: ignore[override] ctx, x: torch.Tensor, process_group: torch.distributed.ProcessGroup )

xformers/ops/differentiable_collectives.py:161

(self, input_: torch.Tensor)

xformers/ops/modpar_layers.py:94

(self, input_: torch.Tensor)

xformers/ops/modpar_layers.py:149

( ctx, x: torch.Tensor, index: torch.Tensor, source: torch.Tensor, sca

xformers/ops/indexing.py:46

( ctx, *args: torch.Tensor, )

xformers/ops/indexing.py:135

(self, x: torch.Tensor)

xformers/ops/rmsnorm.py:99

(cls, ctx, x, w1, b1, w2, b2, w3, b3)

xformers/ops/swiglu_op.py:39

Computes :attr:`swiglu` with the module's weights Args: x (torch.Tensor): A Tensor of shape ``[..., in_features]`` Retur

xformers/ops/swiglu_op.py:343

(ctx, x: torch.Tensor, dim: int)

xformers/ops/unbind.py:88

(ctx, dim: int, *tensors: torch.Tensor)

xformers/ops/unbind.py:105

( ctx: torch.autograd.function.FunctionCtx, query: torch.Tensor, key: torch.Tensor,

xformers/ops/fmha/merge_training.py:42

( ctx: torch.autograd.function.FunctionCtx, *inputs: torch.Tensor, )

xformers/ops/fmha/merge_training.py:89

(ctx, op_fw, op_bw, *args: Any)

xformers/ops/fmha/__init__.py:86

(self, input: torch.Tensor)

xformers/benchmarks/benchmark_sp24.py:102

tests/test_checkpoint.py:139

tests/test_checkpoint.py:257

tests/test_checkpoint.py:284

tests/test_rmsnorm.py:38

(self, input: torch.Tensor)

tests/test_sparsity24.py:772

(self, x: torch.Tensor)

tests/test_sparsity24.py:975

tests/test_sparsity24.py:1044

( ctx: Any, journal: list[str], name: str, tensor: torch.Tensor )

tests/test_fwbw_overlap.py:22

(ctx: Any, x: torch.Tensor)

tests/test_fwbw_overlap.py:136

(self, x: torch.Tensor)

examples/llama_inference/model.py:201

( self, x: torch.Tensor, cache: LayerCache, attn_bias: AttnBias, )

examples/llama_inference/model.py:243

( self, token_values: torch.Tensor, token_lengths: torch.Tensor, start_pos: to

examples/llama_inference/model.py:308

Methodfrom_seqlens

( cls, seqlens: Iterable[int], *, device: Optional[torch.device] = None )

xformers/ops/fmha/attn_bias.py:566

Methodfrom_seqlens

( cls, seqlens: Iterable[int], *, device: Optional[torch.device] = None )

xformers/ops/fmha/attn_bias.py:709

Methodfrom_seqlens

Creates a :attr:`BlockDiagonalMask` from a list of tensors lengths for query and key/value. Args: q_seqlen (Union[Sequence[int],

xformers/ops/fmha/attn_bias.py:849

Methodfrom_seqlens

Creates a :attr:`BlockDiagonalPaddedKeysMask` from a list of tensor lengths for query and key/value. Args: q_seqlen (Sequ

xformers/ops/fmha/attn_bias.py:1140

Methodfrom_seqlens

Creates a :attr:`BlockDiagonalCausalWithOffsetPaddedKeysMask` from a list of tensor lengths for query and key/value. Args:

xformers/ops/fmha/attn_bias.py:1246

Methodfrom_seqlens

Creates a :attr:`PagedBlockDiagonalPaddedKeysMask` from a list of tensor lengths for query and key/value. Args: q_seqlen

xformers/ops/fmha/attn_bias.py:1470

Methodfrom_seqlens

Creates a :attr:`BlockDiagonalGappyKeysMask` from a list of tensor lengths for query and key/value.

xformers/ops/fmha/attn_bias.py:1576

Methodfrom_seqlens

Creates a :attr:`PagedBlockDiagonalGappyKeysMask` from a list of tensor lengths for query and key/value. Note that unlike :attr:`Bloc

xformers/ops/fmha/attn_bias.py:1787

Methodfrom_seqlens_inplace

Perform in-place update. You can only update with the same shape. Can be useful with CUDA graphs.

xformers/ops/fmha/attn_bias.py:454

Methodfrom_seqlens_local

( cls, q_seqlen: Sequence[int], kv_padding: int, kv_seqlen: Sequence[int],

xformers/ops/fmha/attn_bias.py:1385

Methodfrom_seqlens_padded_inplace

Perform in-place update. You can only update with the same shape. Can be useful with CUDA graphs. Note: we don't update paddi

xformers/ops/fmha/attn_bias.py:602

Methodfrom_tree_choices_cached

( cls, tree_choices: Tuple[Tuple[int, ...]], dtype: Optional[torch.dtype] = None,

xformers/ops/tree_attention.py:133

Functionfunc_wrapper

xformers/__init__.py:35

Functionfused_group_contiguous_nan_clamp_copied_from_inductor

( _group_a_ptrs: "VAR_ARGS_ARRAY", # type: ignore # noqa: F821 XBLOCK: tl.constexpr, )

tests/test_triton_varargs.py:148

xformers/benchmarks/benchmark_attn_decoding.py:155

xformers/benchmarks/benchmark_attn_decoding.py:370

xformers/benchmarks/benchmark_indexing.py:76

xformers/benchmarks/benchmark_indexing.py:94

xformers/benchmarks/benchmark_indexing.py:131

xformers/benchmarks/benchmark_indexing.py:141

xformers/benchmarks/benchmark_sp24.py:77

xformers/benchmarks/benchmark_sp24.py:90

xformers/benchmarks/benchmark_sp24.py:136

xformers/benchmarks/benchmark_sp24.py:142

xformers/benchmarks/benchmark_sp24.py:148

xformers/benchmarks/benchmark_sp24.py:154

Functionfwd_splitkv_get_mtile_size

xformers/csrc/attention/hip_fmha/ck_tiled_fmha_fwd_splitkv_setting.h:153

Functionfwd_splitkv_smallq_get_mtile_size

xformers/csrc/attention/hip_fmha/ck_tiled_fmha_fwd_splitkv_smallq_setting.h:115

(q: torch.Tensor, k1: torch.Tensor, v1: torch.Tensor)

tests/test_fmha_merge_attentions.py:670

← previous next →801–900 of 1,445, ranked by callers