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

↓ 4 callersFunctionapply_softcap

src/cuda/flash_attn/fa2/utils.h:396

↓ 4 callersMethodat

Accesses a bit within the predicate vector.

src/cuda/flash_attn/cutlass/cutlass/predicate_vector.h:477

↓ 4 callersMethodbegin

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:82

↓ 4 callersMethodbegin_step

Called at the start of one step before starting accumulator exchange

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_2x.hpp:73

↓ 4 callersFunctionbf16_to_f32

src/gguf_kernels.rs:759

↓ 4 callersFunctionbuild_marlin_scale_map

Build the inverse scale permutation for grouped quantization.

src/kernel/avx2.rs:1859

↓ 4 callersFunctionbuild_marlin_tile_map

Build the Marlin tile map from the weight permutation table. The weight permutation (from `generate_weight_perm_int4()`) maps packed positions to til

src/kernel/avx2.rs:1840

↓ 4 callersFunctionbuild_prompt

Build a prompt of approximately target_tokens length using Gutenberg text.

tests/archive/test_v2lite_timing.py:36

↓ 4 callersFunctionbuild_real_model_store

( real_case: dict, *, include_fused_config: bool, )

tests/test_hqq_fused_branch_runtime.py:478

↓ 4 callersFunctionclamp

(v: f32, lo: f32, hi: f32)

↓ 4 callersFunctionclear_gpu

tests/archive/run_hcs_hybrid_bench.py:48

↓ 4 callersFunctionclear_gpu

tests/archive/run_235b_benchmark.py:39

↓ 4 callersFunctionclear_gpu

tests/archive/investigate_decode_regression.py:34

↓ 4 callersFunctionclear_mask

Clears the predicates

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_dgrad_output_gradient_tile_access_iterator_optimized.h:386

↓ 4 callersMethodclear_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:841

↓ 4 callersFunctioncolex_less

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:846

↓ 4 callersFunctioncollect_success_case

( *, path_name: str, include_fused_desc: bool, include_fused_config: bool, expected_mode:

tests/test_hqq_fused_branch_runtime.py:728

↓ 4 callersFunctioncompare_quantized

(name: str, weight: torch.Tensor)

tests/test_hqq_rust_quantizer.py:130

↓ 4 callersMethodcomplete_transaction

Performs an expected transaction bytes decrement without doing an arrive operation

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:559

↓ 4 callersFunctioncompute_lower_srt

src/cuda/flash_attn/cutlass/cutlass/conv/collective/detail.hpp:242

↓ 4 callersFunctioncompute_model_hash

Compute a fast fingerprint of a model for template matching. Hashes config.json, attention tensor names/shapes/dtypes, and first/last 64 byte

python/krasis/awq_calibrate.py:54

↓ 4 callersFunctioncompute_pp_partition

Compute PP partition — always PP=1 (all layers on primary GPU). Multi-GPU uses Expert Parallelism (EP) instead of Pipeline Parallelism. PP>1

python/krasis/config.py:820

↓ 4 callersMethodcompute_predicates_

Computes predicates based on internally tracked per-thread offset.

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_access_iterator.h:133

↓ 4 callersMethodcompute_routing

Compute MoE routing without expert dispatch. Returns: (topk_ids [M, top_k] int32, topk_weights [M, top_k] float32)

python/krasis/layer.py:626

↓ 4 callersFunctionconvert_layout_acc_rowcol

src/cuda/flash_attn/fa2/utils.h:188

↓ 4 callersFunctioncountr_zero

src/cuda/flash_attn/cutlass/cute/numeric/math.hpp:248

↓ 4 callersFunctiondequantize_int8

Dequantize INT8 weights back to f32 for verification.

src/weights/marlin.rs:125

↓ 4 callersFunctiondequantize_marlin

Dequantize Marlin-repacked weights back to f32 for verification. Reverses the permutation and packing to recover the original [N, K] f32 values.

src/weights/marlin.rs:596

↓ 4 callersMethoddivmod

Computes integer division and modulus using precomputed values. This is computationally inexpensive. Simply returns the quotient

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:420

↓ 4 callersMethodempty

src/cuda/flash_attn/cutlass/cute/container/array.hpp:153

↓ 4 callersMethodenable_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:847

↓ 4 callersFunctionevenly_divides

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:518

↓ 4 callersFunctionexpert_forward_gguf

Full expert forward with native GGUF weights using optimal precision. Pipeline: quantize input → gate/up matvec → SiLU → quantize hidden → down matve

src/gguf_kernels.rs:690

↓ 4 callersFunctionf16_to_f32_ptr

(ptr: *const u8)

src/gguf_kernels.rs:679

↓ 4 callersFunctionf32_slice_to_fp16

(src: &[f32], dst: &mut [u16])

src/decode.rs:5025

↓ 4 callersFunctionfinalize_slot

(slot: dict[str, Any])

tests/hqq_attention_diff.py:3582

↓ 4 callersFunctionfind_if

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:282

↓ 4 callersMethodforward_moe_direct

Synchronous MoE forward — runs compute directly on the calling thread. Unlike submit_forward + sync_forward, this: - Makes ONE PyO3 boundary crossing

src/moe.rs:3100

↓ 4 callersFunctiongemm_rs

src/cuda/flash_attn/fa2/utils.h:166

↓ 4 callersFunctiongenerate

(model, prompt, max_new_tokens=64)

tests/archive/run_qcn_long_isolated.py:20

↓ 4 callersFunctiongenerate_scale_perms

tests/test_marlin_attn_shapes.rs:203

↓ 4 callersMethodget

src/cuda/flash_attn/cutlass/cutlass/subbyte_reference.h:1361

↓ 4 callersFunctionget_attention_backend_spec

(attention_quant: str)

python/krasis/attention_backend.py:176

↓ 4 callersFunctionget_available_device

src/cuda/fla/utils.py:437

↓ 4 callersMethodget_block_idx

Obtains calling threadblock's linear threadblock index

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:703

↓ 4 callersMethodget_cluster_size

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/tile_scheduler_params.h:512

↓ 4 callersFunctionget_expert_components

Return (w13_packed, w13_scales, w2_packed, w2_scales) as bytes views.

tools/expert_dedup_analysis.py:66

↓ 4 callersMethodget_expert_gguf

Get native GGUF expert weights for a given MoE layer and expert index.

src/weights/mod.rs:4308

↓ 4 callersMethodget_first_block_idx

Obtains calling linear threadblock index of the first block to work on the given tile

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:782

↓ 4 callersMethodget_iter_extents

Obtains iteration extends for the given SK block index

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:755

↓ 4 callersFunctionget_layer_scales

Get per-channel AWQ scales for a layer from a v2 template. Returns: Tensor[hidden_size] of per-channel scales, or None if layer not in te

python/krasis/awq_calibrate.py:946

↓ 4 callersFunctionget_linearized_problem_shape_MNKL

src/cuda/flash_attn/cutlass/cutlass/conv/detail.hpp:118

↓ 4 callersMethodget_mask

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:859

↓ 4 callersMethodget_producer_load_callbacks

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp:352

↓ 4 callersFunctionget_shard_for_tensor

(model_path, tensor_name)

tests/test_rust_vs_python.py:32

↓ 4 callersMethodget_sk_tile_idx

Obtains tile index for the given sk iteration

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:645

↓ 4 callersFunctiongguf_matvec_f32

f32-activation matvec dispatch (scalar fallback for all formats).

src/gguf_kernels.rs:239

↓ 4 callersMethodgroup_by_node

Group expert indices by NUMA node. Returns a map from node_id → Vec<(original_position, expert_idx)>.

src/numa.rs:278

↓ 4 callersFunctionhas_gate_proj_experts

Check if experts have gate_proj (standard gated MoE) or just up_proj (Nemotron).

src/weights/mod.rs:6101

↓ 4 callersFunctionhas_shared_gate_proj

Check whether shared experts include a gate projection. Stacked routed experts use `experts.gate_up_proj`, so routed gate detection cannot be reused

src/weights/mod.rs:6115

↓ 4 callersFunctionhqq_attention_tensor_path

( model_path: str, layer_idx: int, tensor_name: str, nbits: int = 4, cache_profile: Option

python/krasis/attention_backend.py:456

↓ 4 callersFunctioninput_hash

(contract: dict[str, object])

scripts/build_sidecars.py:229

↓ 4 callersFunctionint8_linear

W8A8 INT8 matmul with on-the-fly activation quantization. Both weights and activations are INT8, matmul accumulates in INT32 via torch._int_m

python/krasis/weight_loader.py:74

↓ 4 callersMethodinverse

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm80.h:142

↓ 4 callersFunctionis_C_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp:1057

↓ 4 callersFunctionis_hqq_attention

(attention_quant: str)

python/krasis/attention_backend.py:267

↓ 4 callersMethodis_nan

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:487

↓ 4 callersMethodis_source_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm70_epilogue_vectorized_array.hpp:167

↓ 4 callersMethodis_source_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/default_epilogue_array.hpp:150

↓ 4 callersMethodis_valid

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:107

↓ 4 callersFunctionisolated_home

tests/test_hqq_cache_profile.py:44

↓ 4 callersFunctionjson_escape

Escape a string for embedding inside a JSON string value.

src/server.rs:765

↓ 4 callersFunctionkill_server

Kill the server process group cleanly.

tests/release_test.py:613

↓ 4 callersFunctionld_without_cache

src/cuda/flash_attn/cutlass/cutlass/experimental/distributed/kernel/detail.hpp:48

↓ 4 callersFunctionlex_less

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:809

↓ 4 callersFunctionload

Loads a fragment from memory at the location pointed to by the iterator.

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator.h:378

↓ 4 callersMethodload

Loads a fragment from memory at the location pointed to by the iterator.

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_complex_tensor_op_tile_iterator_sm80.h:965

↓ 4 callersMethodload

Loads a fragment from memory at the location pointed to by the iterator.

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator_sm70.h:1306

↓ 4 callersMethodload

Loads a fragment from memory at the location pointed to by the iterator.

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator.h:3348

↓ 4 callersMethodload_ab

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:851

↓ 4 callersMethodload_auxiliary

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp:578

↓ 4 callersFunctionload_cuda_fns

Load cudaSetDevice + cudaMemGetInfo from the already-loaded libcudart. Returns None if the library isn't loaded or symbols aren't found.

src/vram_monitor.rs:109

↓ 4 callersMethodload_embedding

Load embedding table (BF16, ~2.2 GB for Kimi K2.5).

python/krasis/weight_loader.py:219

↓ 4 callersFunctionload_model

(expert_divisor=1)

tests/archive/test_v2lite_profile.py:21

↓ 4 callersFunctionload_prompt

(idx, max_chars=None)

tests/archive/run_long_prompt_test.py:126

↓ 4 callersFunctionload_prompt_file

Load a prompt file from the krasis prompts directory.

tests/archive/bench_network_vs_benchmark.py:24

↓ 4 callersFunctionload_safetensors_tensor

tests/test_rust_vs_python.py:23

↓ 4 callersFunctionload_tokenizer_with_compat

(model_path: str)

tests/reference_contract.py:105

↓ 4 callersFunctionload_u8_tensor

Load raw bytes for a tensor from mmapped safetensors shards.

src/weights/mod.rs:6262

↓ 4 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_sparse.h:283

↓ 4 callersFunctionload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/shared_load_iterator.h:163

↓ 4 callersFunctionload_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_simt.h:227

↓ 4 callersMethodload_with_pointer_offset

Loads a fragment from memory with additional logical offset

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_complex_tensor_op_tile_iterator_sm80.h:971

↓ 4 callersMethodload_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:1312

↓ 4 callersMethodload_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.h:3354

↓ 4 callersMethodload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op.h:315

↓ 4 callersFunctionmake_cotiled_copy

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:556

↓ 4 callersFunctionmake_tma_copy_im2col

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_im2col.hpp:709

↓ 4 callersFunctionmake_transpose_operand_b

src/cuda/flash_attn/cutlass/cutlass/transform/collective/sm90_wgmma_transpose.hpp:713

↓ 4 callersFunctionmake_weight

(rows: int, cols: int, *, scale: float, offset: float = 0.0, flip_cols: bool = False)

tests/test_hqq_fused_branch_runtime.py:140

← previous next →1,001–1,100 of 12,380, ranked by callers