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

↓ 15 callersMethodtensor_data

Get a raw byte slice for a named tensor (zero-copy from mmap).

src/weights/safetensors_io.rs:160

↓ 14 callersFunction_clear_screen

python/krasis/launcher.py:130

↓ 14 callersFunction_linear

Dispatch to Marlin GEMM, INT8, or BF16 linear based on weight type.

python/krasis/linear_attention.py:111

↓ 14 callersFunction_load_index_tensor

(model_dir: Path, name: str)

python/krasis/hqq_self_calibrate.py:704

↓ 14 callersFunction_require_torch

python/krasis/attention_backend.py:40

↓ 14 callersMethoddivmod

Computes the quotient and remainder in a single method.

src/cuda/flash_attn/cutlass/cutlass/uint128.h:211

↓ 14 callersFunctionfast_cos

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:787

↓ 14 callersFunctionfill

src/cuda/flash_attn/cutlass/cute/algorithm/fill.hpp:45

↓ 14 callersFunctionfind_divisor

* Find divisor, using find_log2 */

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:257

↓ 14 callersFunctionfind_tmem_tensor_col_offset

src/cuda/flash_attn/cutlass/cutlass/detail/sm100_tmem_helper.hpp:48

↓ 14 callersFunctionget_current_device

src/cuda/flash_attn/fa2/hardware_info.h:26

↓ 14 callersMethodgroups

src/cuda/flash_attn/cutlass/cutlass/gemm/group_array_problem_shape.hpp:60

↓ 14 callersMethodinitialize

Initializes members. This variant of the method should only be used when problem_shape and tile_shape contain modes of only rank 1.

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/tile_scheduler_params.h:108

↓ 14 callersFunctionis_hqq_auto_attention

(attention_quant: str)

python/krasis/attention_backend.py:311

↓ 14 callersFunctionmake_tiled_copy_D

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:613

↓ 14 callersFunctionmax_common_layout

src/cuda/flash_attn/cutlass/cute/layout.hpp:1382

↓ 14 callersFunctionok

tests/release_test.py:86

↓ 14 callersMethodpack

(member, bit_width)

python/krasis/marlin_utils.py:44

↓ 14 callersMethodsubmit_forward

( &self, _py: Python<'_>, moe_layer_idx: usize, activation_bf16: &[u8],

src/moe.rs:2980

↓ 14 callersMethodsync_forward

Wait for async MoE forward to complete and return BF16 output. Must be called after `submit_forward()`. Blocks until the worker finishes. Returns: B

src/moe.rs:3066

↓ 14 callersFunctiontapply

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:96

↓ 14 callersFunctiontransform_leaf

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:250

↓ 14 callersMethodvisit

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_store.hpp:321

↓ 14 callersFunctionwrite_vec_u16

Write a Vec<u16> as raw bytes to a writer.

src/weights/mod.rs:5438

↓ 14 callersFunctionwrite_vec_u32

Write a Vec<u32> as raw bytes to a writer.

src/weights/mod.rs:5430

↓ 13 callersFunctionRematerializeBlockIdxY

Helper to rematerialize block Idx. Reduces register liveness.

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/index_remat.h:72

↓ 13 callersFunction_read_key

Read a single keypress in raw mode. Returns key constant or char.

python/krasis/launcher.py:183

↓ 13 callersFunctionaccum

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized_blockwise_scaling.hpp:1052

↓ 13 callersFunctioncanonical_warp_group_idx

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

src/cuda/flash_attn/cutlass/cutlass/cutlass.h:147

↓ 13 callersFunctionconstexpr

Input A: float_e4m3_t ; Input B: float_e5m2_t

src/cuda/flash_attn/cutlass/cute/arch/mma_sm90.hpp:685

↓ 13 callersFunctionfloat_exmy_base

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:966

↓ 13 callersMethodis_host_problem_shape_available

src/cuda/flash_attn/cutlass/cutlass/gemm/group_array_problem_shape.hpp:75

↓ 13 callersFunctionisnan

src/cuda/flash_attn/cutlass/cutlass/half.h:505

↓ 13 callersFunctionload_with_pointer_offset

Loads a fragment from memory with additional logical offset

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator_sm70.h:329

↓ 13 callersFunctionmake_tiled_copy_S

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:602

↓ 13 callersFunctionmake_tmem_copy

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm100.hpp:256

↓ 13 callersFunctionmoe_forward_unified

Full MoE forward using unified transposed weights. Same logic as `moe_forward` but uses `expert_forward_unified` with the combined w13 (gate+up) tran

↓ 13 callersMethodones

Constructs a matrix from a uniform element 1

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

↓ 13 callersFunctionpartition_shape_B

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:647

↓ 13 callersFunctionpartition_shape_C

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:607

↓ 13 callersFunctionprepare_chunk_indices

( cu_seqlens: torch.LongTensor, chunk_size: int, cu_seqlens_cpu: torch.LongTensor | None = None, )

src/cuda/fla/ops/utils/index.py:112

↓ 13 callersMethodreport_event

Log a named event with current VRAM snapshot. No-op if report not enabled.

src/vram_monitor.rs:784

↓ 13 callersFunctionstore_with_byte_offset

Stores a fragment to memory with additional pointer offset

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator.h:4767

↓ 13 callersFunctionsynclog_emit_tma_store

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:837

↓ 13 callersFunctionunwrap

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:514

↓ 12 callersFunctionRematerializeBlockIdxX

Helper to rematerialize block Idx. Reduces register liveness.

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/index_remat.h:66

↓ 12 callersFunction_event_values

( event: Dict[str, Any], ignore_fields: set[str], )

tests/trace_diff.py:685

↓ 12 callersFunction_run

Run a command, printing it first.

python/krasis/setup.py:32

↓ 12 callersFunction_tensor_hash

(tensor: torch.Tensor)

python/krasis/hqq_self_calibrate.py:140

↓ 12 callersFunction_timed_bf16_load

( loader: "WeightLoader", *, layer_idx: int, tensor_name: str, device: torch.device, s

python/krasis/weight_loader.py:30

↓ 12 callersFunctionbf16_to_f32

tests/test_marlin_attn_shapes.rs:17

↓ 12 callersMethodcurrent_free_mb

Get current free VRAM on a device (in MB) — live reading, not tracked min.

src/vram_monitor.rs:740

↓ 12 callersFunctionfast_sin

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:805

↓ 12 callersFunctionfence_view_async_tmem_load

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:882

↓ 12 callersMethodget_expert_unified

Backward compat: returns CPU expert ref (used by moe_forward_unified).

src/weights/mod.rs:4285

↓ 12 callersFunctionget_work_k_tile_start

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:695

↓ 12 callersMethodhas_unified

Whether unified weights are loaded.

src/moe.rs:2122

↓ 12 callersMethodisinf

Is infinite implementation

src/cuda/flash_attn/cutlass/cutlass/float8.h:204

↓ 12 callersFunctionload_index_tensor

(model_dir: Path, name: str)

tests/hqq_attention_diff.py:56

↓ 12 callersFunctionload_with_byte_offset

Loads a fragment from memory with additional logical offset

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator_sm70.h:301

↓ 12 callersFunctionload_with_pointer_offset

Loads a fragment from memory at the location pointed to by the iterator. (scalar loads)

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_simt_tile_iterator.h:472

↓ 12 callersMethodname

Human-readable name.

↓ 12 callersFunctionsimt_transpose_padding

Computes padding in shared memory to perform efficient transpose without bank conflicts.

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/default_mma_core_simt.h:73

↓ 12 callersFunctionsize

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:636

↓ 12 callersMethodsk_blocks_per_region

Number of SK blocks per region (splitting factor)

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:201

↓ 12 callersMethodsk_regions

Number of SK regions

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:194

↓ 12 callersMethodstride

Returns the stride of the layout

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm70.h:176

↓ 12 callersMethodstride

Returns the stride of the layout

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm75.h:207

↓ 12 callersMethodtiled_shape

Returns the GEMM volume in thread block tiles

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:170

↓ 12 callersFunctionto_string

src/cuda/flash_attn/cutlass/cute/arch/mma_sm90_desc.hpp:55

↓ 12 callersMethodtransform

Transform the mma operands to the required types

src/cuda/flash_attn/cutlass/cutlass/conv/warp/mma_depthwise_simt.h:368

↓ 12 callersFunctionvalid_warpgroup_in_work_tile

When the work tile is not special for reduction, it's valid. Otherwise need to skip global loading that producer warpgroup do, also math computation t

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:395

↓ 12 callersMethodwait

Wait on a stage to be unlocked

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1325

↓ 11 callersFunction_accumulate_hqq4_core_timing

(timing_stats: Optional[dict], key: str, elapsed_s: float)

python/krasis/attention_backend.py:1526

↓ 11 callersFunction_hide_cursor

python/krasis/launcher.py:135

↓ 11 callersFunction_hqq_num_groups

(cols: int, group_size: int)

python/krasis/attention_backend.py:1233

↓ 11 callersFunction_hqq_padded_cols

(cols: int, group_size: int)

python/krasis/attention_backend.py:1237

↓ 11 callersMethod_rand_bf16

(self, shape: Tuple[int, ...], *, device: torch.device)

tests/decode_harness.py:133

↓ 11 callersMethod_shared_expert_forward

Shared expert MLP on GPU, with optional sigmoid gate (Qwen3-Next).

python/krasis/layer.py:602

↓ 11 callersFunction_show_cursor

python/krasis/launcher.py:140

↓ 11 callersMethodadd_tile_offset

Adds a tile offset in the unit of tile.

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_scale_bias_vector_access_iterator.h:152

↓ 11 callersFunctioncache_dir_for_model

Return the cache directory for a model: ~/.krasis/cache/<model_folder_name>/. On first call, migrates any existing .krasis_cache/ from the model

python/krasis/config.py:29

↓ 11 callersFunctioncopy_aligned

src/cuda/flash_attn/cutlass/cute/algorithm/copy.hpp:334

↓ 11 callersFunctiondie

tests/release_test.py:94

↓ 11 callersFunctionflatten

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:564

↓ 11 callersFunctiongcd

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:174

↓ 11 callersFunctionget_transformed_problem_shape_MNKL

src/cuda/flash_attn/cutlass/cutlass/conv/detail.hpp:62

↓ 11 callersMethodhas_gpu_weights

Whether GPU prefill weights (Marlin format) have been populated.

src/weights/mod.rs:4255

↓ 11 callersFunctionhqq_layout_for_nbits

python/krasis/attention_backend.py:347

↓ 11 callersMethodis_C_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp:308

↓ 11 callersFunctionisolated_home

tests/test_hqq_self_calibrate.py:32

↓ 11 callersFunctionload_hqq_attention_artifact

( model_path: str, entry: dict, expected_nbits: int, device: str = "cpu", cache_profile: O

python/krasis/attention_backend.py:2749

↓ 11 callersMethodlog_vram_ledger_residency

Log opt-in categorized CUDA tensor residency for startup VRAM diagnosis.

python/krasis/model.py:3720

↓ 11 callersFunctionmake_mixed_bits

src/cuda/flash_attn/cutlass/cute/swizzle.hpp:165

↓ 11 callersFunctionmake_tma_copy_B_sm90

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:1495

↓ 11 callersFunctionmarlin_repack

Repack our QuantizedInt4 into Marlin GPU format. Follows vLLM's Python reference: unpack → transpose → tile permute → repack. Our format: packed `[N,

src/weights/marlin.rs:430

↓ 11 callersFunctionok

tests/generate_reference.py:112

↓ 11 callersFunctionparse_meminfo_value

Parse a value from /proc/meminfo (format: "FieldName: 12345 kB").

src/syscheck.rs:167

↓ 11 callersMethodpartition

src/cuda/flash_attn/cutlass/cute/atom/partitioner.hpp:82

↓ 11 callersFunctionprefetch_tma_descriptors

Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance

src/cuda/flash_attn/cutlass/cutlass/conv/collective/sm100_implicit_gemm_umma_warpspecialized.hpp:662

← previous next →401–500 of 12,380, ranked by callers