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 callersFunctiongroup_diagnostics

(row_idx: int, group_idx: int)

tests/hqq_attention_diff.py:1385

↓ 2 callersFunctiongroup_events

(events: list[dict[str, Any]])

tests/hqq_attention_diff.py:3541

↓ 2 callersFunctionhandle_chat_completion

Handle /v1/chat/completions request.

src/server.rs:868

↓ 2 callersFunctionhandle_prefill_logits

Handle /v1/internal/prefill_logits endpoint. Runs a full prefill pass and extracts top-k logprobs at sampled positions.

src/server.rs:1571

↓ 2 callersFunctionhandle_reference_test

Handle /v1/internal/reference_test endpoint. Accepts raw input_token_ids, runs greedy prefill + decode, returns output tokens with logprobs. Used for

src/server.rs:1724

↓ 2 callersMethodhas_merged_experts

Check if this GGUF uses merged expert tensors (ffn_gate_exps).

src/gguf.rs:512

↓ 2 callersFunctionhf_auth_status

python/krasis/hf_downloader.py:259

↓ 2 callersFunctionhqq46_auto_budget_bytes_from_mib

(value: Optional[int])

python/krasis/attention_backend.py:572

↓ 2 callersFunctionhqq46_tensor_nbits

(tensor_name: str)

python/krasis/attention_backend.py:343

↓ 2 callersFunctionhqq_search_cuda_tensor_ptr

( input_ptr: usize, rows: usize, cols: usize, group_size: usize, nbits: usize, global_

↓ 2 callersFunctionimplicit_gemm_k_iterations_per_channel

src/cuda/flash_attn/cutlass/cutlass/conv/conv2d_problem_size.h:493

↓ 2 callersMethodinf_with_sign

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:603

↓ 2 callersFunctioninfer_capture_profile

(reference: Dict[str, Any], tokenizer: Any)

tests/reference_contract.py:361

↓ 2 callersFunctioninfo

tests/reference_inventory.py:53

↓ 2 callersMethodinit_params

Initialize params member

src/cuda/flash_attn/cutlass/cutlass/gemm/device/gemm_universal_base.h:207

↓ 2 callersMethodinverse

Computes the inverse of a 2-by-2 matrix given the matrix's determinant

src/cuda/flash_attn/cutlass/cutlass/matrix.h:3240

↓ 2 callersFunctioninverse_marlin_repack

Convert Marlin-packed INT4 weights to standard GPTQ packed format. Args: w_marlin: [..., K//16, 2*N] int32 (Marlin format, any leading ba

python/krasis/triton_moe.py:73

↓ 2 callersFunctioninverse_scale_permute

Un-permute Marlin scales to standard GPTQ format. Args: s_marlin: [..., K//gs, N] bf16 (Marlin permuted + transposed scales) Returns

python/krasis/triton_moe.py:136

↓ 2 callersMethodis_C_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_compute_tma_warpspecialized.hpp:143

↓ 2 callersFunctionis_auxiliary_prefix

python/krasis/config.py:224

↓ 2 callersMethodis_f16c_supported

src/cuda/flash_attn/cutlass/cutlass/half.h:149

↓ 2 callersMethodis_full_attention_layer

True if this layer uses standard full attention (GQA/MLA).

python/krasis/config.py:779

↓ 2 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_compute_tma_warpspecialized.hpp:138

↓ 2 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm90_epilogue_tma_warpspecialized.hpp:413

↓ 2 callersMethodis_reduction_unit

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:115

↓ 2 callersFunctionis_same_row_or_col

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:382

↓ 2 callersMethodis_source1_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_tensor_broadcast.hpp:200

↓ 2 callersMethodis_source_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm70_epilogue_vectorized.hpp:262

↓ 2 callersFunctionis_stacked_experts

Detect stacked expert format (Qwen3.5, Mistral 4). These models store all experts in stacked 3D tensors per layer: experts.gate_up_proj [E, 2*inter, h

src/weights/mod.rs:6198

↓ 2 callersFunctionis_valid

src/cuda/flash_attn/cutlass/cute/util/type_traits.hpp:262

↓ 2 callersMethodis_zero

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp:313

↓ 2 callersFunctionisfinite

src/cuda/flash_attn/cutlass/cutlass/tfloat32.h:208

↓ 2 callersFunctionisfinite

src/cuda/flash_attn/cutlass/cutlass/half.h:510

↓ 2 callersFunctionisinf

src/cuda/flash_attn/cutlass/cutlass/half.h:521

↓ 2 callersFunctionjudge_overall

Return overall PASS, WARN, or FAIL.

tests/reference_test.py:767

↓ 2 callersFunctionjudge_prompt

Return PASS, WARN, or FAIL for a single prompt. Primary metric: prefill top-10 containment (same input, only quant differs). Secondary: first

tests/reference_test.py:727

↓ 2 callersFunctionkill_server

Kill server process group.

tests/validate_model.py:317

↓ 2 callersFunctionkill_stale_krasis_processes

Kill any lingering krasis Python processes (excluding ourselves).

tests/release_test.py:641

↓ 2 callersMethodkn

Obtains a Coord<2> from GemmCoord

src/cuda/flash_attn/cutlass/cutlass/gemm_coord.h:186

↓ 2 callersFunctionl2_normalize_expand_avx2

( src: &[f32], dst: &mut [f32], src_offset: usize, dk: usize, nv: usize, hr: usize, scale: f32,

src/decode.rs:4470

↓ 2 callersFunctionl2norm

(x, dim=-1, eps=1e-6)

tests/hf_layer0_compare.py:22

↓ 2 callersFunctionlcm

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:183

↓ 2 callersFunctionlcm_cxx11

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

↓ 2 callersFunctionlinear_attention_recurrent_avx2

( state: &mut [f32], q: &[f32], k: &[f32], v: &[f32], g: &[f32], beta: &[f32],

src/decode.rs:1653

↓ 2 callersFunctionlist_available_references

List all available reference data directories.

tests/reference_test.py:172

↓ 2 callersFunctionload

Loads a fragment

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/shared_load_iterator.h:210

↓ 2 callersMethodload_ab_init

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:638

↓ 2 callersMethodload_ab_tail

Perform a Producer Epilogue to prevent early exit of ctas in a Cluster

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:896

↓ 2 callersFunctionload_and_quantize_weight

Load a BF16 or FP8 weight tensor and quantize it to INT4. For FP8 tensors, loads the per-tensor `weight_scale_inv` and dequantizes to BF16 first.

src/weights/mod.rs:6547

↓ 2 callersFunctionload_and_time

(label, threshold, expert_divisor=-1)

tests/archive/test_qwen3_next_decode_timing.py:15

↓ 2 callersFunctionload_bf16_tensor_as_f32

Load BF16 bias tensor from safetensors as f32 slice.

src/weights/mod.rs:6287

↓ 2 callersFunctionload_case_map

tests/hqq_attention_diff.py:346

↓ 2 callersFunctionload_case_summary_doc

tests/hqq_attention_diff.py:327

↓ 2 callersMethodload_config

Parse TOML config and return all model × config combinations.

python/krasis/suite.py:148

↓ 2 callersMethodload_final_norm

Load final RMSNorm weight (BF16, tiny).

python/krasis/weight_loader.py:228

↓ 2 callersFunctionload_hqq_attention_pending_manifest

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

python/krasis/attention_backend.py:1006

↓ 2 callersFunctionload_init

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized.hpp:501

↓ 2 callersMethodload_lm_head

Load LM head weight. Returns (weight_int8, scale) if INT8, or plain BF16 tensor if BF16.

python/krasis/weight_loader.py:282

↓ 2 callersFunctionload_model

Load V2-Lite with specified threshold.

tests/test_gpu_decode.py:25

↓ 2 callersFunctionload_model

Load V2-Lite with dual-format cache.

tests/archive/test_v2lite_dual_format.py:19

↓ 2 callersFunctionload_model

Load Qwen3-Coder-Next PP=2 with active_only + static_pin.

tests/archive/test_qwen3_next_gpu_decode.py:22

↓ 2 callersFunctionload_mxfp4_layer_experts

Load all experts for a single layer from MXFP4 format, dequantize to BF16, then quantize to INT4/INT8. MXFP4 stores all experts in a single tensor pe

src/weights/mod.rs:6366

↓ 2 callersFunctionload_prompt_text

Load a full Gutenberg prompt by index (0-5). Cached after first load.

tests/prompt_utils.py:25

↓ 2 callersMethodload_sf

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:916

↓ 2 callersMethodload_sf_init

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:704

↓ 2 callersMethodload_sf_tail

Perform a Producer Epilogue to prevent early exit of ctas in a Cluster

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:977

↓ 2 callersFunctionload_stacked_layer_experts

Load all experts from stacked 3D tensors (Qwen3.5/Mistral 4 format) and quantize. Stacked format: experts.gate_up_proj [E, 2*inter, hidden], experts.

src/weights/mod.rs:6997

↓ 2 callersMethodload_tail

Perform a Producer Epilogue to prevent early exit of blocks in a Cluster

src/cuda/flash_attn/cutlass/cutlass/conv/collective/sm90_implicit_gemm_gmma_ss_warpspecialized.hpp:634

↓ 2 callersFunctionload_template

Load a template for a model, if one exists. Searches: 0. Explicit override from KRASIS_AWQ_TEMPLATE_PATH 1. Bundled templates in template

python/krasis/awq_calibrate.py:899

↓ 2 callersFunctionload_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_volta_tensor_op.h:216

↓ 2 callersFunctionlog_build_timing

(label: &str, elapsed: std::time::Duration)

↓ 2 callersMethodmake_fragment_C

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:130

↓ 2 callersFunctionmake_layout_like

src/cuda/flash_attn/cutlass/cute/layout.hpp:439

↓ 2 callersFunctionmake_linspace

(start: f32, end: f32, steps: usize)

↓ 2 callersFunctionmake_ordered_layout

src/cuda/flash_attn/cutlass/cute/layout.hpp:423

↓ 2 callersFunctionmake_swizzle_strides

src/cuda/flash_attn/cutlass/cute/swizzle_layout.hpp:175

↓ 2 callersFunctionmake_tiled_copy_C_atom

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:481

↓ 2 callersFunctionmake_tiled_mma

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:575

↓ 2 callersFunctionmake_tma_copy_C_sm90

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:1531

↓ 2 callersFunctionmake_tmem_warp_partitioner

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm100.hpp:296

↓ 2 callersFunctionmanual_mla_attention

Manual MLA attention using torch.matmul (no FlashInfer). Args: q_nope_absorbed: [M, H, D_ckv] query with w_kc absorbed q_pe: [M,

tests/archive/test_attn_verify.py:21

↓ 2 callersFunctionmarlin_cache_config_hash

( config_str: &str, gpu_bits: u8, expert_int4_calib_mode: ExpertInt4CalibMode, expert_int4_cal

src/weights/mod.rs:1516

↓ 2 callersFunctionmarlin_expert_byte_sizes

Compute per-expert byte sizes for Marlin GPU format. For INT4: pack_factor=8, packed shape [K/16, 2*N] per tile. For INT8: pack_factor=4, packed shape

src/weights/mod.rs:1584

↓ 2 callersFunctionmatmul_bf16_scalar

Scalar BF16 matrix-vector multiply (reference). Computes output[n] = sum_k(weight[n, k] * activation[k]).

src/kernel/avx2.rs:27

↓ 2 callersFunctionmatmul_int4_marlin_scalar

Scalar Marlin-native INT4 matmul for correctness verification. Reads Marlin-packed `[K/16, 2*N]` weights and Marlin-permuted `[K/gs, N]` scales. Comp

src/kernel/avx2.rs:1884

↓ 2 callersFunctionmatmul_int4_transposed_integer_parallel_tiled

Parallel tiled INT4 matmul. Data in [N/TILE, K/8, TILE] layout.

src/kernel/avx2.rs:1406

↓ 2 callersFunctionmatmul_int4_transposed_scalar

Scalar INT4 transposed matmul (correctness reference). Weight layout: packed[K/8, N], scales[K/group_size, N]. Computes output[n] = sum_k(activation[

src/kernel/avx2.rs:818

↓ 2 callersFunctionmatmul_int8_integer

Safe wrapper for the AVX2 integer INT8 matmul kernel.

src/kernel/avx2.rs:708

↓ 2 callersFunctionmatmul_int8_transposed_integer_parallel_tiled

Parallel tiled INT8 matmul. Data in [N/TILE, K, TILE] layout.

src/kernel/avx2.rs:1762

↓ 2 callersFunctionmatvec_q4_0_scalar

(weights: &[u8], input: &[f32], n: usize, k: usize, output: &mut [f32])

src/gguf_kernels.rs:552

↓ 2 callersFunctionmatvec_q8_0_scalar

(weights: &[u8], input: &[f32], n: usize, k: usize, output: &mut [f32])

src/gguf_kernels.rs:534

↓ 2 callersFunctionmeasure_gpu_memory

Return (allocated_mb, reserved_mb).

tests/archive/test_quant_config.py:38

↓ 2 callersMethodmk

Obtains a Coord<2> from GemmCoord

src/cuda/flash_attn/cutlass/cutlass/gemm_coord.h:168

↓ 2 callersFunctionmla_attn_dot_fp16_avx2

( q: &[f32], cache: &[u16], dim: usize, )

src/decode.rs:4847

↓ 2 callersFunctionmoe_worker

Background worker for async MoE computation. Owns its own scratch buffers. Processes batches of tokens sequentially, with expert-level parallelism wi

src/moe.rs:1394

↓ 2 callersMethodnext_u32

src/decode.rs:4929

↓ 2 callersFunctionnorm

src/cuda/flash_attn/cutlass/cutlass/quaternion.h:410

↓ 2 callersFunctionnormalize_hqq_auto_budget_pct

(value: Optional[float], attention_quant: str = "hqq_auto")

python/krasis/attention_backend.py:583

↓ 2 callersFunctionnormalize_selected_gpus

(raw: Optional[str])

tests/reference_contract.py:903

↓ 2 callersFunctionnullspace

src/cuda/flash_attn/cutlass/cute/layout.hpp:1506

← previous next →1,801–1,900 of 12,380, ranked by callers