Functions — github.com/brontoguana/krasis

CodeHub MCPcopy Create free account

hub / github.com/brontoguana/krasis / functions

Functions12,380 in github.com/brontoguana/krasis

⨍ Functions 12,380◇ Types & classes 7,782↳ Endpoints 10

↓ 83 callersMethodproducer_acquire

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1104

↓ 82 callersMethodpush

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/epilogue.h:452

↓ 81 callersFunctionbits_to_bytes

src/cuda/flash_attn/cutlass/cutlass/numeric_size.h:64

↓ 79 callersMethodlayout

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/output_iterator_parameter.h:74

↓ 79 callersFunctionmma

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized.hpp:677

↓ 77 callersMethodget_tile_offset

Obtains the threadblock offset (in units of threadblock-scoped tiles)

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle.h:131

↓ 77 callersFunctionzipped_divide

src/cuda/flash_attn/cutlass/cute/layout.hpp:1624

↓ 76 callersMethodoffset

Computes the offset of an index from the origin of the tensor

src/cuda/flash_attn/cutlass/cutlass/tensor_ref.h:315

↓ 75 callersMethodcompose

src/cuda/flash_attn/cutlass/cute/layout.hpp:190

↓ 75 callersFunctionelem_less

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:883

↓ 75 callersMethodlayout

Returns the layout object

src/cuda/flash_attn/cutlass/cutlass/thread/matrix.h:125

↓ 75 callersMethodload

(&mut self, model_dir: &str, group_size: Option<usize>, max_layers: Option<usize>, start_layer: Option<usize>,

src/moe.rs:1714

↓ 74 callersMethodproducer_tail

Prevents early exit of producer blocks in Cluster. This should be called once before kernel exits.

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1123

↓ 73 callersMethodelement_size

Bytes per element.

src/weights/safetensors_io.rs:58

↓ 73 callersMethodlayout_b

src/cuda/flash_attn/cutlass/cute/layout_composed.hpp:83

↓ 72 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_tile_iterator.h:114

↓ 71 callersMethodload

Load addend source fragment from global memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/epilogue.h:270

↓ 71 callersMethodwait

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:367

↓ 70 callersFunctionfor_each

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:172

↓ 70 callersMethodnorm

Returns the sum of squared elements

src/cuda/flash_attn/cutlass/cutlass/matrix.h:571

↓ 70 callersMethodvalues

src/cuda/flash_attn/cutlass/cute/container/tuple.hpp:486

↓ 69 callersMethodnon_const_ref

src/cuda/flash_attn/cutlass/cutlass/tensor_ref.h:244

↓ 69 callersFunctionprefetch_tma_descriptor

src/cuda/flash_attn/cutlass/cute/arch/copy_sm90_desc.hpp:302

↓ 69 callersFunctionswap

src/cuda/flash_attn/cutlass/cutlass/platform/platform.h:788

↓ 69 callersMethodvalid

Returns whether access is valid or not

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_access_iterator.h:295

↓ 68 callersMethodarrive

Barrier arrive on local smem

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:373

↓ 67 callersFunction_detail

Print a detail line under a status header (green, indented).

python/krasis/server.py:116

↓ 67 callersMethodempty

Create an empty (zero-size) QuantWeight for ungated experts. Used as a dummy gate_proj when the model has no gate projection.

src/weights/mod.rs:329

↓ 67 callersFunctionmake_identity_tensor

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:495

↓ 66 callersFunctionmake_tma_copy

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:1266

↓ 65 callersFunctionbf16_to_f32

src/weights/marlin.rs:19

↓ 65 callersFunctioncanonical_warp_idx_sync

Returns a warp-uniform value indicating the canonical warp index of the calling threads. Threads within the warp must be converged.

src/cuda/flash_attn/cutlass/cutlass/cutlass.h:125

↓ 64 callersMethodretile_D

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:425

↓ 62 callersMethodencode

(self, text: str, add_special_tokens: bool = True)

python/krasis/tokenizer.py:144

↓ 61 callersFunctionshape_div

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:384

↓ 60 callersFunctionbf16_bytes

Convert tensor to raw BF16 bytes (little-endian unsigned short).

tests/generate_kernel_test_data.py:32

↓ 59 callersMethodadvance

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/grouped_problem_visitor.h:149

↓ 59 callersMethodat

Returns a reference to the element at a given Coord

src/cuda/flash_attn/cutlass/cutlass/thread/matrix.h:164

↓ 59 callersFunctionfma

src/cuda/flash_attn/cutlass/cute/arch/simd_sm100.hpp:77

↓ 59 callersMethodmin

Min representable value for this scalar type. (accounting for bias if there is one)

src/cuda/marlin/scalar_type.hpp:256

↓ 59 callersFunctiontiled_divide

src/cuda/flash_attn/cutlass/cute/layout.hpp:1635

↓ 58 callersMethodtranspose

src/cuda/flash_attn/cutlass/cutlass/matrix.h:138

↓ 57 callersFunctionas_position_independent_swizzle_tensor

src/cuda/flash_attn/cutlass/cute/pointer_flagged.hpp:99

↓ 55 callersFunctionabs

src/cuda/flash_attn/cutlass/cutlass/complex.h:443

↓ 55 callersMethodset_kgroup_index

Notify the iterator which k-group it is currently pointing to. This does not advance the iterator. Rather, it overrides its internal tracking with co

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/tile_iterator_planar_complex.h:238

↓ 54 callersFunctioncopy_if

src/cuda/flash_attn/cutlass/cute/algorithm/copy.hpp:47

↓ 54 callersMethodlayout

src/cuda/flash_attn/cutlass/cute/layout.hpp:119

↓ 54 callersFunctionzip

src/cuda/flash_attn/cutlass/cute/layout.hpp:1531

↓ 53 callersFunctionfilter

src/cuda/flash_attn/cutlass/cute/layout.hpp:934

↓ 51 callersMethodforward

Run one token through the draft model. Updates KV cache. After this call, h_logits contains the output logits on host.

src/draft_model.rs:472

↓ 51 callersFunctionselect_cluster_shape

src/cuda/flash_attn/cutlass/cutlass/detail/cluster.hpp:49

↓ 50 callersFunctionconst_min

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:753

↓ 50 callersFunctiondie

tests/generate_reference.py:118

↓ 50 callersFunctionget_grid_shape

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:357

↓ 48 callersMethodget_host_problem_shape

src/cuda/flash_attn/cutlass/cutlass/gemm/group_array_problem_shape.hpp:70

↓ 48 callersFunctiontensor

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:508

↓ 47 callersFunctionflat_divide

src/cuda/flash_attn/cutlass/cute/layout.hpp:1649

↓ 47 callersMethodreset

Replaces the managed object, deleting the old object.

src/cuda/flash_attn/cutlass/cutlass/platform/platform.h:756

↓ 46 callersFunction_print

Print with flush=True by default so output appears immediately when piped.

python/krasis/suite.py:37

↓ 46 callersMethodfree

Release all pages back to the pool.

python/krasis/kv_cache.py:449

↓ 46 callersFunctioninitialize_workspace

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler.hpp:145

↓ 46 callersMethodlayout

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:192

↓ 46 callersMethodlayout_a

src/cuda/flash_attn/cutlass/cute/layout_composed.hpp:71

↓ 46 callersMethodstore

Stores a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:988

↓ 45 callersMethodat

Returns a reference to the element at a given Coord

src/cuda/flash_attn/cutlass/cutlass/tensor_ref.h:321

↓ 45 callersFunctionclear

src/cuda/flash_attn/cutlass/cute/container/array.hpp:352

↓ 45 callersFunctionexp

src/cuda/fla/ops/utils/op.py:13

↓ 44 callersMethodcontains

Determines whether a location is within a tensor

src/cuda/flash_attn/cutlass/cutlass/tensor_view.h:190

↓ 44 callersFunctionhqq_attention_cache_dir

Return the native HQQ attention artifact directory under the model cache.

python/krasis/attention_backend.py:411

↓ 44 callersMethodproducer_get_barrier

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1131

↓ 44 callersMethodsize

Static method that returns the packed logical size of any TensorExtent

src/cuda/flash_attn/cutlass/cutlass/conv/convnd_problem_shape.hpp:490

↓ 44 callersFunctionwarpgroup_arrive

src/cuda/flash_attn/cutlass/cute/arch/mma_sm90_gmma.hpp:47

↓ 44 callersFunctionwarpgroup_commit_batch

Marks the commit point for one or more sized batch of warpgroup MMAs.

src/cuda/flash_attn/cutlass/cute/arch/mma_sm90_gmma.hpp:74

↓ 43 callersFunctioncp_async_fence

Establishes an ordering w.r.t previously issued cp.async instructions. Does not block.

src/cuda/flash_attn/cutlass/cutlass/arch/memory_sm80.h:439

↓ 43 callersMethodproducer_commit

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1109

↓ 42 callersFunctionget_tiled_cta_shape_mnl

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:349

↓ 42 callersMethodopen

Open and parse a GGUF file.

src/gguf.rs:320

↓ 42 callersFunctionquantize_int4

Quantize a BF16 weight matrix to symmetric INT4 with per-group scales. Symmetric INT4: values in [-8, 7], scale chosen so that max(abs(group)) maps t

src/weights/marlin.rs:153

↓ 41 callersMethodadd_tile_offset

Advances an iterator along logical dimensions of matrix in units of whole threadblock tiles

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/predicated_scale_bias_vector_iterator.h:347

↓ 41 callersMethodid

unique id for this scalar type that can be computed at compile time for c++17 template specialization this is not needed once we migrate to c++20 and

src/cuda/marlin/scalar_type.hpp:129

↓ 41 callersFunctionlogical_divide

src/cuda/flash_attn/cutlass/cute/layout.hpp:1573

↓ 41 callersFunctionto_gemm_coord

src/cuda/flash_attn/cutlass/cutlass/gemm_coord.hpp:47

↓ 41 callersFunctionwait_on_dependent_grids

Issuing the griddepcontrol.wait instruction enforces no global memory access prior to this istruction. This ensures the correctness of global memory a

src/cuda/flash_attn/cutlass/cutlass/arch/grid_dependency_control.h:88

↓ 41 callersMethodx

Accesses the x element of the imaginary part of the quaternion

src/cuda/flash_attn/cutlass/cutlass/quaternion.h:152

↓ 41 callersMethody

Accesses the y element of the imaginary part of the quaternion

src/cuda/flash_attn/cutlass/cutlass/quaternion.h:164

↓ 41 callersMethodz

Accesses the z element of the imaginary part of the quaternion

src/cuda/flash_attn/cutlass/cutlass/quaternion.h:176

↓ 40 callersMethodget

Returns a pointer to the vector starting at the current coordinate

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv3d_fprop_filter_tile_access_iterator_optimized.h:233

↓ 40 callersMethodoffset

src/cuda/flash_attn/cutlass/cute/layout_composed.hpp:77

↓ 40 callersFunctionright_inverse

src/cuda/flash_attn/cutlass/cute/layout.hpp:1263

↓ 40 callersFunctionsend_json

Send a JSON response.

src/server.rs:311

↓ 40 callersMethodto_half

Bitcasts to CUDA's half type

src/cuda/flash_attn/cutlass/cutlass/half.h:445

↓ 40 callersMethodwork_tile_to_cta_coord

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm100_tile_scheduler.hpp:372

↓ 39 callersMethodload

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_tile_iterator.h:166

↓ 39 callersFunctionstride

Returns stride as MatrixCoord

src/cuda/flash_attn/cutlass/cutlass/conv/conv3d_problem_size.h:319

↓ 39 callersFunctionsync

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:230

↓ 39 callersMethodupdate

Lightweight update given a subset of arguments

src/cuda/flash_attn/cutlass/cutlass/gemm/device/trmm.h:459

↓ 38 callersMethodend

Exit of the subtile load loop.

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_tma_warpspecialized.hpp:154

↓ 38 callersMethodflush

Flush buffer at end of stream.

python/krasis/chat.py:233

↓ 38 callersMethodget_tiled_shape

Returns the GEMM volume in thread block tiles

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/params_universal_base.h:201

↓ 38 callersFunctionget_work_k_tile_count

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:689

← previous next →101–200 of 12,380, ranked by callers