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

↓ 5 callersFunctionhash_bytes

tools/expert_dedup_analysis.py:77

↓ 5 callersFunctionhqq_cache_algorithm_for_nbits

python/krasis/attention_backend.py:356

↓ 5 callersFunctioni32_bytes

Convert tensor to raw INT32 bytes.

tests/generate_kernel_test_data.py:42

↓ 5 callersFunctionimplicit_gemm_k_iterations

Determine the number of gemm_k iterations for conv2d problem using implicit gemm algorithm

src/cuda/flash_attn/cutlass/cutlass/conv/conv3d_problem_size.h:372

↓ 5 callersMethodinverse

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm70.h:376

↓ 5 callersMethodinverse

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm75.h:303

↓ 5 callersMethodis_moe_layer

(self, layer_idx: int)

python/krasis/config.py:811

↓ 5 callersMethodis_numa

Whether NUMA-aware placement is meaningful (>1 node).

src/numa.rs:197

↓ 5 callersFunctionleft_inverse

src/cuda/flash_attn/cutlass/cute/layout.hpp:1322

↓ 5 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_sm80.h:2359

↓ 5 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_sm80.h:301

↓ 5 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_sm80.h:2366

↓ 5 callersMethodload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:507

↓ 5 callersMethodload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator.h:1459

↓ 5 callersMethodload_with_pointer_offset

Loads a fragment

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear.h:339

↓ 5 callersFunctionmake_tensor

(rows: int, cols: int, start: int)

tests/test_hqq_attention_carry.py:15

↓ 5 callersFunctionmake_tensor_like

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:422

↓ 5 callersFunctionmake_tma_atom

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:1330

↓ 5 callersFunctionmatmul_int4_parallel

Parallel AVX2 INT4 matmul — splits output rows across rayon threads. Each thread processes a chunk of rows independently. The activation vector is sh

src/kernel/avx2.rs:537

↓ 5 callersFunctionmax

src/cuda/flash_attn/cutlass/cute/numeric/math.hpp:48

↓ 5 callersMethodmin_free_mb

Get the minimum free VRAM observed on a device (in MB).

src/vram_monitor.rs:684

↓ 5 callersFunctionminimal_manifest

(complete=True, profile=HQQ_CACHE_PROFILE_BASELINE)

tests/test_hqq_cache_profile.py:57

↓ 5 callersFunctionminmax_affine_params

(source: torch.Tensor)

tests/hqq_attention_diff.py:779

↓ 5 callersFunctionmoe_forward_gguf

Scratch buffers for GGUF expert computation (reused across calls). Full MoE forward using native GGUF weights (no dequant→requant). Uses raw GGUF blo

src/moe.rs:1045

↓ 5 callersFunctionprefetch_expert_configurable

(expert: &UnifiedExpertWeights, stride: usize, hint: u8)

src/moe.rs:1211

↓ 5 callersMethodproduct

Returns the product of all elements

src/cuda/flash_attn/cutlass/cutlass/coord.h:320

↓ 5 callersFunctionprofile_decode

Profile decode with per-token timing after prefill.

tests/archive/test_v2lite_profile.py:98

↓ 5 callersFunctionquantize_hqq4_tensor_rust

Experimental Rust shadow implementation for HQQ4 tensor quantization.

python/krasis/attention_backend.py:1684

↓ 5 callersFunctionquantizer_candidate_metrics

( name: str, source: torch.Tensor, input_group: torch.Tensor, *, scale: float, zero: f

tests/hqq_attention_diff.py:1001

↓ 5 callersMethodraw

Obtains raw bits

src/cuda/flash_attn/cutlass/cutlass/tfloat32.h:161

↓ 5 callersFunctionrelatively_equal_float

src/cuda/flash_attn/cutlass/cutlass/relatively_equal.h:57

↓ 5 callersMethodrows

src/weights/mod.rs:269

↓ 5 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/mma_tensor_op_tile_iterator_sm80.h:2427

↓ 5 callersFunctionset_params_alibi

src/cuda/flash_attn/fa2/flash_api_reference.cpp:331

↓ 5 callersMethodset_sign_bit

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:523

↓ 5 callersMethodsetup_gpu_decode_store

Create and configure a GpuDecodeStore for Rust-native GPU decode. Registers all attention weights, MoE layers, RoPE tables, and KV cache.

python/krasis/model.py:6607

↓ 5 callersFunctionshared_load

src/cuda/flash_attn/cutlass/cutlass/arch/memory_sm75.h:231

↓ 5 callersMethodsignificand_hidden_bits

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:629

↓ 5 callersFunctionsize

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:265

↓ 5 callersFunctionslice

src/cuda/flash_attn/cutlass/cute/layout.hpp:689

↓ 5 callersFunctionsource_prefix_candidates

(layer: int, family: str)

tests/hqq_attention_diff.py:2261

↓ 5 callersFunctionstage_hqq

(store: GpuDecodeStore, layer_idx: int, tensor_name: str, weight: torch.Tensor)

tests/test_hqq_fused_branch_runtime.py:147

↓ 5 callersMethodstart

(self, startup_timeout: float = 12.0)

python/krasis/ssh_tunnel.py:117

↓ 5 callersMethodstart

(self, name: str)

tests/hqq_attention_diff.py:35

↓ 5 callersMethodstore_init

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/detail.hpp:370

↓ 5 callersFunctionstore_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_simt.h:195

↓ 5 callersFunctionstore_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_tensor_op_mixed.h:203

↓ 5 callersMethodstore_with_pointer_offset

Store a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:519

↓ 5 callersMethodstore_with_pointer_offset

Store a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator.h:1477

↓ 5 callersMethodstore_with_pointer_offset

Stores a fragment

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear.h:357

↓ 5 callersFunctionstrided_dgrad_tile_m_per_filter

/////////////////////////////////////////////////////////////////////////////////////////////// Strided dgrad helper functions

src/cuda/flash_attn/cutlass/cutlass/conv/conv2d_problem_size.h:617

↓ 5 callersFunctiontensor_bytes

(shapes: list[tuple[str, int, int]], nbits_for_name)

python/krasis/vram_budget.py:482

↓ 5 callersMethodtensormaps_cp_fence_release

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized.hpp:735

↓ 5 callersMethodtest_wait

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:357

↓ 5 callersFunctiontimed_run

(args: list[str], label: str)

scripts/build_sidecars.py:123

↓ 5 callersMethodtop_k

Number of experts selected per token (top-k).

src/moe.rs:2102

↓ 5 callersFunctiontrait_ratio

src/cuda/flash_attn/cutlass/cute/numeric/integral_ratio.hpp:271

↓ 5 callersFunctiontx

tests/release_test.py:1541

↓ 5 callersFunctionvalidate_hqq_attention_tensors

( tensors: Dict[str, torch.Tensor], *, expected_nbits: Optional[int] = None, expected_axis: in

python/krasis/attention_backend.py:1259

↓ 5 callersMethodvisit

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_tma_warpspecialized.hpp:597

↓ 5 callersFunctionwrite_manifest

(model_path: Path, profile: str, manifest: dict, nbits: int = 4)

tests/test_hqq_cache_profile.py:82

↓ 4 callersFunctionOffsetBytes

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:712

↓ 4 callersFunction_candidate_metrics

( name: str, source: torch.Tensor, input_rows: torch.Tensor, *, scale: float, zero: fl

python/krasis/hqq_self_calibrate.py:2685

↓ 4 callersFunction_csv_filter

(value: Optional[str])

python/krasis/hqq_self_calibrate.py:4024

↓ 4 callersFunction_dense_artifact_tensor_specs_from_real_case

( real_case: dict, selected_tensor_names: list[str] | None = None, )

tests/test_hqq_rust_quantizer.py:214

↓ 4 callersMethod_estimate_attention_vram

Estimate total GPU VRAM used by attention weights across all layers. Returns bytes consumed by all attention-related parameters: norm

python/krasis/model.py:3661

↓ 4 callersMethod_export_kv_to_rust

Set KV cache position for Rust decode after prefill. The KV cache is shared — Rust prefill already wrote FP8 data into the same GPU b

python/krasis/model.py:8591

↓ 4 callersFunction_family_identity

(event: Optional[Dict[str, Any]])

tests/trace_diff.py:744

↓ 4 callersFunction_family_member_key

(event: Dict[str, Any])

tests/trace_diff.py:755

↓ 4 callersFunction_find_nvidia_smi

Find nvidia-smi, including the WSL2 driver mount.

python/krasis/setup.py:59

↓ 4 callersFunction_format_timing

Format server timing into compact single-line format. Example: (pp: 21t 682ms 31t/s - tg 980t>12t 12s 89.2t/s)

python/krasis/chat.py:628

↓ 4 callersFunction_format_tokens

Format token count as e.g. '72K' or '1.2M'.

python/krasis/launcher.py:1081

↓ 4 callersFunction_get_nvcc_version

Get nvcc major.minor version, or None if not found.

python/krasis/setup.py:200

↓ 4 callersFunction_gpu_cleanup

Force GPU cleanup between tests to avoid stale CUDA state.

tests/archive/test_gpu_prefill.py:261

↓ 4 callersFunction_headline

Print a compact headline that stays readable when stdout is log-prefixed.

python/krasis/server.py:131

↓ 4 callersMethod_init_state

Initialize conv and recurrent states to zero.

python/krasis/linear_attention.py:222

↓ 4 callersFunction_load_activation_rows

(trace_path: str, *, layer: int, tensor: str)

python/krasis/hqq_self_calibrate.py:2480

↓ 4 callersFunction_load_trace_events

( path: Path, *, include_events: Optional[set[str]], include_components: Optional[set[str]],

tests/trace_diff.py:644

↓ 4 callersFunction_minmax_params

(source: torch.Tensor)

python/krasis/hqq_self_calibrate.py:2512

↓ 4 callersFunction_normalized_manifest_entry

tests/test_hqq_rust_quantizer.py:1060

↓ 4 callersFunction_normalized_runtime_entry

tests/test_hqq_rust_quantizer.py:1066

↓ 4 callersMethod_pack_expert_matrix

(self, weight_cpu: torch.Tensor)

tests/decode_harness.py:201

↓ 4 callersFunction_pack_hqq_quant

(quant: torch.Tensor, nbits: int)

python/krasis/attention_backend.py:1211

↓ 4 callersFunction_package_version

(package_name: str)

tests/generate_reference.py:123

↓ 4 callersFunction_parse_ranked_pairs

tests/trace_diff.py:269

↓ 4 callersFunction_real_model_quant_cfg

tests/test_hqq_fused_branch_runtime.py:395

↓ 4 callersFunction_require_hqq4_rust_symbols

python/krasis/attention_backend.py:67

↓ 4 callersFunction_select_int8_exception_candidates

( candidate_report: Dict[str, Any], *, requested_groups: Optional[set[Tuple[int, str, int]]],

python/krasis/hqq_self_calibrate.py:1979

↓ 4 callersMethod_set_interactive_attention_quant

Apply an interactive attention preset. The TUI exposes four simple attention presets: HQQ4, HQQ4+10%, HQQ6, and HQQ6+10%.

python/krasis/launcher.py:1399

↓ 4 callersMethod_stream_attn_prefetch

Start async DMA for a layer's attention weights on the DMA stream. Call this while CPU is busy with MoE experts. Before using the pre

python/krasis/model.py:4320

↓ 4 callersFunction_trace_enabled

(component: str)

tests/reference_contract.py:1128

↓ 4 callersFunction_truncate_ansi

Truncate text to a visible width while preserving ANSI color codes.

python/krasis/launcher.py:87

↓ 4 callersFunction_validate_int8_exception_byte_cap

(value: Optional[int])

python/krasis/hqq_self_calibrate.py:1925

↓ 4 callersFunction_vram_ledger_enabled

python/krasis/model.py:208

↓ 4 callersFunction_write_real_artifact_layer_with_writer

( model: KrasisModel, real_case: dict, *, writer, selected_tensor_names: list[str] | None

tests/test_hqq_rust_quantizer.py:418

↓ 4 callersMethodaccum

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:1079

↓ 4 callersMethodaccum_init

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized_blockwise_scaling.hpp:823

↓ 4 callersFunctionadd_byte_offset_

Adds a pointer offset in units of element

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_dgrad_output_gradient_tile_access_iterator_optimized.h:673

↓ 4 callersFunctionadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op_sm80.h:163

↓ 4 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:808

← previous next →901–1,000 of 12,380, ranked by callers