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

↓ 2 callersFunctioncreate_marlin_buffers

Create properly-shaped Marlin buffers with random data.

tests/archive/test_marlin_gpu1.py:25

↓ 2 callersFunctioncreate_prefill_engine_for_server

( store: &mut GpuDecodeStore, max_context_tokens: usize, )

src/server.rs:237

↓ 2 callersFunctioncutlassGetStatusString

Convert cutlass status to status strings

src/cuda/flash_attn/cutlass/cutlass/cutlass.h:62

↓ 2 callersMethoddata

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:81

↓ 2 callersMethoddecode_step

Full decode step — runs entire layer loop in Rust. Replaces the Python step() method. One Python call per token. output_ptr: *mut f32 [vocab_size] —

src/decode.rs:3120

↓ 2 callersFunctiondelete_hqq_attention_pending_manifest

( model_path: str, cache_profile: Optional[str] = HQQ_CACHE_PROFILE_BASELINE, nbits: int = 4,

python/krasis/attention_backend.py:1032

↓ 2 callersFunctiondepth

src/cuda/flash_attn/cutlass/cute/layout.hpp:623

↓ 2 callersFunctiondequant_f16

(data: &[u8], n: usize)

src/gguf.rs:547

↓ 2 callersFunctiondequant_q4_0

Q4_0: 32 elements per block, 18 bytes = fp16 d + 16 bytes qs (nibbles). Each element = d * (nibble - 8).

src/gguf.rs:635

↓ 2 callersFunctiondequant_q4_k

Dequantize Q4_K blocks to FP32. Q4_K block (144 bytes per 256 elements): fp16 d (super-block scale), fp16 dmin (super-block min), 12 bytes packed 6-b

src/gguf.rs:681

↓ 2 callersFunctiondequant_q5_0

Q5_0: 32 elements per block, 22 bytes = fp16 d + 4 bytes qh + 16 bytes qs (nibbles). Each element = d * ((qs_nibble | (qh_bit << 4)) - 16).

src/gguf.rs:599

↓ 2 callersFunctiondequant_q5_k

Dequantize Q5_K blocks to FP32. Q5_K block (176 bytes per 256 elements): fp16 d, fp16 dmin, 12 bytes scales, 32 bytes qh (5th bits), 128 bytes qs (4-

src/gguf.rs:740

↓ 2 callersFunctiondequant_q6_k

Dequantize Q6_K blocks to FP32. Q6_K block (210 bytes per 256 elements): 128 bytes ql (lower 4 bits), 64 bytes qh (upper 2 bits), 16 bytes scales (in

src/gguf.rs:813

↓ 2 callersFunctiondequant_q8_0

Q8_0 block: fp16 scale + 32 x int8 weights. Block size = 32.

src/gguf.rs:574

↓ 2 callersFunctiondequantize_mxfp4_to_bf16

Dequantize MXFP4 blocks + E8M0 scales to BF16 for a single projection. # Arguments - `blocks`: contiguous [out_features, num_blocks, 16] u8 (each byt

src/weights/mod.rs:6225

↓ 2 callersFunctiondestination_for_repo

(models_dir: str, repo_id: str)

python/krasis/hf_downloader.py:444

↓ 2 callersFunctiondetect_linear_attention

Detect whether this model uses linear attention (recurrent state) layers. Models with linear attention (e.g. Gated DeltaNet in Qwen3-Coder-Next)

tests/reference_test.py:117

↓ 2 callersFunctiondetect_repetition

Detect if text contains repeating patterns (sign of degeneration). Returns True if a pattern of length >= min_pattern_len repeats >= min_repeats

tests/archive/test_v2lite_thorough.py:62

↓ 2 callersFunctiondiagnose_solve_scale_drift

(name: str, chunk: torch.Tensor, zero: torch.Tensor, scale_seed: torch.Tensor)

tests/test_hqq_rust_quantizer.py:1937

↓ 2 callersMethoddispatch_matmul_ext

Dispatch matmul to correct INT4/INT8 kernel (uses provided buffers).

src/decode.rs:1497

↓ 2 callersFunctiondomain_distribute

src/cuda/flash_attn/cutlass/cute/layout.hpp:1449

↓ 2 callersFunctiondrain_vram_pressure_for_state

( state: &mut ServerState, reason: &str, force_measure: bool, )

src/server.rs:131

↓ 2 callersFunctionemit_reference_generation_trace

( context: str, contract: Dict[str, Any], artifact_path: str, *, max_new_tokens: Optional[

tests/reference_contract.py:1279

↓ 2 callersMethodenable_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_triangular_matrix.h:758

↓ 2 callersMethodenable_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h:739

↓ 2 callersMethodend_epilogue

Called after all steps have been completed

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_2x.hpp:121

↓ 2 callersFunctionepilogue_predication

src/cuda/flash_attn/cutlass/cute/algorithm/cooperative_gemm.hpp:58

↓ 2 callersFunctionerror_metrics_from_diff

(source: torch.Tensor, diff: torch.Tensor)

tests/hqq_attention_diff.py:527

↓ 2 callersFunctionevaluate_ppl_with_skip

Run PPL evaluation, skipping attention for specified layers. When a layer is skipped, its attention output is set to zeros so the residual st

perplexity/ablate_attention.py:37

↓ 2 callersFunctionexchange_sort

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:586

↓ 2 callersMethodexpect_transaction

Performs an expected transaction bytes increment without doing an arrive operation

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:553

↓ 2 callersFunctionexpert_byte_sizes

(config: &ModelConfig, group_size: usize, num_bits: u8)

src/weights/mod.rs:1748

↓ 2 callersFunctionexpert_forward_bf16

BF16 reference expert forward: SiLU(gate(x)) * up(x) → down().

tests/test_rust_vs_python.py:77

↓ 2 callersFunctionextract_bundle

(path: Path, contracts: dict[str, dict[str, object]])

scripts/build_sidecars.py:442

↓ 2 callersFunctionextract_token

Extract a token string from tokenizer_config.json. Handles both `"bos_token": "<s>"` and `"bos_token": {"content": "<s>", ...}`.

src/chat_template.rs:296

↓ 2 callersFunctionfast_acos

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:823

↓ 2 callersFunctionfast_log

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:904

↓ 2 callersFunctionfast_silu_mul_avx2

(gate: &[f32], up: &[f32], output: &mut [f32], n: usize)

src/decode.rs:2053

↓ 2 callersFunctionfence_view_async_tmem_store

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:894

↓ 2 callersFunctionfield_stage

(layer: int, tensor: str)

tests/hqq_attention_diff.py:3507

↓ 2 callersFunctionfile_mtime

↓ 2 callersFunctionfind_log2

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:245

↓ 2 callersFunctionfind_reference_data

Find reference data JSON for a model.

tests/reference_test.py:155

↓ 2 callersMethodfind_shared_expert_tensors

Find shared expert tensor names for a given layer.

src/gguf.rs:517

↓ 2 callersFunctionflash_attn_cu_files

scripts/build_sidecars.py:150

↓ 2 callersFunctionfloat_exmy_base<T, Derived>

Floating point conversion

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:1023

↓ 2 callersFunctionformat_completion

Format a complete (non-streaming) chat completion response.

src/server.rs:648

↓ 2 callersFunctionformat_contract_report

(reference_contract: Dict[str, Any], runtime_contract: Dict[str, Any], validation: Dict[str, List[str]])

tests/reference_contract.py:1089

↓ 2 callersFunctionfpclassify

src/cuda/flash_attn/cutlass/cutlass/half.h:531

↓ 2 callersMethodfront

src/cuda/flash_attn/cutlass/cutlass/array.h:385

↓ 2 callersFunctiongather_ranges

(t: torch.Tensor, ranges: list[tuple[int, int]])

tests/hqq_attention_diff.py:2064

↓ 2 callersFunctiongcd_cxx11

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:194

↓ 2 callersFunctiongemm

src/cuda/flash_attn/cutlass/cute/algorithm/cooperative_gemm.hpp:519

↓ 2 callersFunctiongenerate

Generate with or without CUDA graph.

tests/test_la_graph.py:30

↓ 2 callersFunctiongenerate

Generate with original or inplace recurrent forward.

tests/archive/test_la_inplace.py:30

↓ 2 callersFunctiongenerate

Generate text from a prompt and return (text, stats_dict).

tests/archive/run_long_prompt_test.py:24

↓ 2 callersFunctiongenerate_hf_reference

Generate reference logits using HuggingFace transformers.

tests/archive/test_v2lite_thorough.py:176

↓ 2 callersFunctiongenerate_weight_perm_int4

tests/test_marlin_attn_shapes.rs:149

↓ 2 callersFunctiongenerate_weight_perm_int8

tests/test_marlin_attn_shapes.rs:176

↓ 2 callersMethodgetParams

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_tile_iterator.h:102

↓ 2 callersFunctiongetVersionBuild

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

↓ 2 callersMethodget_1d_coord

src/cuda/flash_attn/cutlass/cute/layout.hpp:272

↓ 2 callersMethodget_abs_max

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/epilogue_with_absmax.h:190

↓ 2 callersMethodget_callbacks

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_2x.hpp:134

↓ 2 callersMethodget_current_target

src/cuda/fla/compile_kernels.py:307

↓ 2 callersFunctionget_current_work_cta_m_n_in_cluster

Given raster order and current work tile linear index, reset cta m and n index in the cluster.

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:732

↓ 2 callersFunctionget_err_ratio

src/cuda/fla/utils.py:88

↓ 2 callersMethodget_grid_dims

Returns the grid extents in thread blocks to launch

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/params_universal_base.h:216

↓ 2 callersMethodget_hier_coord

src/cuda/flash_attn/cutlass/cute/layout.hpp:250

↓ 2 callersMethodget_layoutA_MK

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:430

↓ 2 callersMethodget_layoutB_NK

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:467

↓ 2 callersMethodget_layoutC_MN

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:397

↓ 2 callersMethodget_layoutC_TV

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:414

↓ 2 callersMethodget_mask

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_triangular_matrix.h:770

↓ 2 callersMethodget_mask

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h:751

↓ 2 callersFunctionget_model_details

(repo_or_url: str)

python/krasis/hf_downloader.py:437

↓ 2 callersMethodget_ptr_aux_output_abs_max

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_generic_with_scaling.h:305

↓ 2 callersMethodget_ptr_output_abs_max

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_generic_with_scaling.h:300

↓ 2 callersFunctionget_scale_min_k4

(j: usize, scales: &[u8])

src/gguf_kernels.rs:640

↓ 2 callersFunctionget_scale_min_k4_ptr

(j: usize, scales: *const u8)

src/gguf_kernels.rs:652

↓ 2 callersMethodget_shape_C

Get C extents. fprop: C extents array contains [N,Z,P,Q,K]. Turn that into ((Q,P,Z,N), (K)) dgrad: C extents array contains [N,D,H,W,C]. Turn that int

src/cuda/flash_attn/cutlass/cutlass/conv/convnd_problem_shape.hpp:457

↓ 2 callersMethodget_shared_expert_unified

Backward compat: returns CPU shared expert ref.

src/weights/mod.rs:4294

↓ 2 callersFunctionget_slice

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:370

↓ 2 callersFunctionget_stride

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:402

↓ 2 callersFunctionget_strided_dgrad_tile_m

////////////////////////////////////////////////////////////////////////////////////////////

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/threadblock_swizzle.h:53

↓ 2 callersFunctionget_tensor_decision

Get the quantization decision for a specific tensor. For v2 templates: input projections get "int4" (AWQ-scaled at load time), output project

python/krasis/awq_calibrate.py:968

↓ 2 callersFunctionget_tma_swizzle_base

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma_swizzle.hpp:110

↓ 2 callersFunctionget_tma_swizzle_bits

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma_swizzle.hpp:72

↓ 2 callersMethodget_work_k_tile_count

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm100_tile_scheduler.hpp:499

↓ 2 callersFunctiongithub_asset

(token: str | None, release: dict[str, object], name: str)

scripts/build_sidecars.py:590

↓ 2 callersFunctiongithub_repo

scripts/build_sidecars.py:505

↓ 2 callersFunctiongmem_wait

Wait until we have at least one completed global fetch stage

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_multistage.h:485

↓ 2 callersFunctiongmem_wait

Wait until we have at least one completed global fetch stage

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_pipelined.h:274

↓ 2 callersFunctiongpu_mem

tests/archive/decode_only_bench.py:43

↓ 2 callersFunctiongpu_mem

tests/archive/bench_qwen235b_pp2.py:47

↓ 2 callersFunctiongpu_mem

Return per-GPU allocated MB.

tests/archive/run_hcs_benchmark.py:52

↓ 2 callersFunctiongpu_mem

tests/archive/run_hcs_hybrid_bench.py:55

↓ 2 callersFunctiongpu_mem

tests/archive/bench_prefill_10k.py:44

↓ 2 callersFunctiongpu_mem

tests/archive/investigate_decode_regression.py:41

↓ 2 callersMethodgridDim

Commonly used utility functions

src/cuda/flash_attn/cutlass/cutlass/cluster_launch.hpp:95

← previous next →1,701–1,800 of 12,380, ranked by callers