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

↓ 7 callersMethodrun

Execute the full two-phase optimisation.

tests/archive/meta_optimiser.py:288

↓ 7 callersMethodset_mask

Sets the predicate mask, overriding value stored in predicate iterator

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator.h:1447

↓ 7 callersMethodslice_accumulator

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_warpspecialized.hpp:459

↓ 7 callersFunctionsource_row_info

(row: int, source_meta: dict[str, Any])

tests/hqq_attention_diff.py:608

↓ 7 callersMethodswap

src/cuda/flash_attn/cutlass/cute/container/array.hpp:185

↓ 7 callersMethodtensormaps_perform_update

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized.hpp:714

↓ 7 callersFunctiontensormaps_replace_global_tensor_properties

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized.hpp:784

↓ 7 callersFunctiontest_known_answer

Test with a prompt that has a known expected substring in the answer.

tests/test_network.py:159

↓ 7 callersFunctiontranspose_int4

Transpose a QuantizedInt4 from [N, K/8] layout to [K/8, N] layout. Used to convert from the original row-major format to the transposed format needed

src/kernel/avx2.rs:2195

↓ 7 callersMethodtry_wait

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:362

↓ 7 callersFunctiontuple_cat

src/cuda/flash_attn/cutlass/cute/container/tuple.hpp:291

↓ 7 callersFunctionvalid

Returns whether access is valid or not

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/predicated_scale_bias_vector_access_iterator.h:274

↓ 7 callersFunctionwarn

tests/generate_reference.py:115

↓ 6 callersFunction_attach_source_contract

( entry: Dict[str, Any], source_contract_lookup: Dict[Tuple[int, str], Dict[str, Any]], )

python/krasis/hqq_self_calibrate.py:931

↓ 6 callersFunction_build_shadow_real_artifact_model

(real_case: dict, model_path: str)

tests/test_hqq_rust_quantizer.py:345

↓ 6 callersMethod_copy_weights_dict

Deep-copy a layer weights dict, moving all tensors to device.

python/krasis/model.py:1653

↓ 6 callersFunction_diag_vector

(diag: dict, key: str)

tests/test_hqq_fused_branch_runtime.py:66

↓ 6 callersFunction_edit_value

Inline edit: shows current value, user types new one.

python/krasis/launcher.py:1268

↓ 6 callersMethod_engine_request

Single request through the Rust engine (no HTTP/SSE). Calls RustServer.benchmark_request() which runs the full evict -> prefill -> re

python/krasis/benchmark.py:396

↓ 6 callersFunction_event_brief

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

tests/trace_diff.py:705

↓ 6 callersFunction_event_key

(event: Dict[str, Any])

tests/trace_diff.py:676

↓ 6 callersMethod_fix_query_key_value_ordering

Un-interleave QKVZ and BA projections. The weight matrices are structured so that the output is interleaved per key-head group. This

python/krasis/linear_attention.py:362

↓ 6 callersMethod_hqq_attention_tensor_map

(self, layer_type: str, weights: dict)

python/krasis/model.py:1706

↓ 6 callersFunction_load_hqq_tensor

(cache_dir: str, entry: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:847

↓ 6 callersFunction_load_json

tests/reference_contract.py:60

↓ 6 callersMethod_maybe_write_hqq_attention_artifacts

(self, layer_idx: int, layer_type: str, weights: dict)

python/krasis/model.py:1940

↓ 6 callersFunction_normalized_staging_json

(store: GpuDecodeStore)

tests/test_hqq_rust_quantizer.py:1105

↓ 6 callersFunction_parse_float_list

tests/trace_diff.py:233

↓ 6 callersMethod_register_hqq_attention_layers_on_store

( self, store, target_device: torch.device, keepalive: list, )

python/krasis/model.py:3224

↓ 6 callersFunction_require_hf

python/krasis/hf_downloader.py:169

↓ 6 callersFunction_rmsnorm

RMSNorm using torch.

python/krasis/layer.py:32

↓ 6 callersFunction_safe_ratio

(numer: float, denom: float)

python/krasis/hqq_self_calibrate.py:1375

↓ 6 callersFunction_sidecar_entry_key_to_str

(entry: Tuple[int, str, int, int])

python/krasis/hqq_self_calibrate.py:3193

↓ 6 callersFunction_similarity_stats

(reference: torch.Tensor, candidate: torch.Tensor)

tests/test_hqq_fused_branch_runtime.py:41

↓ 6 callersFunction_source_contract_lookup_from_audit

(source_contract_audit_path: str)

python/krasis/hqq_self_calibrate.py:906

↓ 6 callersFunction_tensor_stats

tests/reference_backfill_topk.py:270

↓ 6 callersFunction_vram_ledger_enabled

python/krasis/server.py:164

↓ 6 callersFunction_wsl_cuda_dir

python/krasis/launcher.py:358

↓ 6 callersFunctionadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:166

↓ 6 callersMethodapply

Apply the chat template to a list of messages. `messages_json` is a JSON array of {role, content} objects. `tools_json` is an optional JSON array of

src/chat_template.rs:98

↓ 6 callersMethodblock_bytes

Bytes per block.

↓ 6 callersMethodblock_size

Block size (number of elements per quantization block).

↓ 6 callersFunctionbundle_path

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

scripts/build_sidecars.py:437

↓ 6 callersMethodcall

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:92

↓ 6 callersFunctioncan_implement

Determines whether the Implicit GEMM can execute the given problem.

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv3d_fprop_filter_tile_access_iterator_analytic.h:239

↓ 6 callersMethodcapacity

Compute the number of contiguous elements needed to store a tensor with the given size

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm70.h:188

↓ 6 callersMethodcapacity

Compute the number of contiguous elements needed to store a tensor with the given size

src/cuda/flash_attn/cutlass/cutlass/layout/tensor_op_multiplicand_sm75.h:216

↓ 6 callersFunctionclear_gpu

tests/archive/run_hcs_benchmark.py:45

↓ 6 callersFunctioncompare

(name, hf_t, our_t)

tests/test_linear_attn_compare.py:87

↓ 6 callersFunctioncompare

Compare two tensors and report.

tests/archive/test_layer0_detail.py:24

↓ 6 callersFunctioncompatible

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:486

↓ 6 callersFunctioncompute_lower_corner_whd

src/cuda/flash_attn/cutlass/cutlass/conv/collective/detail.hpp:188

↓ 6 callersFunctioncompute_source_not_needed_

Streams the result to global memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/epilogue_direct_store.h:275

↓ 6 callersFunctioncompute_upper_corner_whd

src/cuda/flash_attn/cutlass/cutlass/conv/collective/detail.hpp:212

↓ 6 callersFunctionconsider

(zero_values: torch.Tensor, scale_seed: float)

python/krasis/hqq_self_calibrate.py:2561

↓ 6 callersMethodconsumer_wait

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:1093

↓ 6 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_multistage.h:287

↓ 6 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/implicit_gemm_multistage.h:84

↓ 6 callersFunctioncp_async_fence

src/cuda/marlin/marlin_vendor_common.h:111

↓ 6 callersMethoddata_bytes

Total bytes of weight data (packed + scales).

src/weights/mod.rs:294

↓ 6 callersMethoddequantize_tensor

Dequantize a tensor to FP32.

src/gguf.rs:466

↓ 6 callersFunctiondice

src/cuda/flash_attn/cutlass/cute/layout.hpp:706

↓ 6 callersFunctionexpert_forward_integer

Compute a single expert's output using the integer kernel (_mm256_madd_epi16). Uses pre-quantized INT16 activations for gate/up projections (shared a

↓ 6 callersMethodexponent_biased

Returns the biased exponent

src/cuda/flash_attn/cutlass/cutlass/tfloat32.h:173

↓ 6 callersMethodexponent_biased

Returns the biased exponent

src/cuda/flash_attn/cutlass/cutlass/half.h:475

↓ 6 callersMethodexponent_biased

Returns the biased exponent

src/cuda/flash_attn/cutlass/cutlass/bfloat16.h:211

↓ 6 callersMethodfileno

python/krasis/server.py:1245

↓ 6 callersFunctionfind_nvcc

scripts/build_sidecars.py:91

↓ 6 callersFunctionfma

src/cuda/flash_attn/cutlass/cute/numeric/real.hpp:58

↓ 6 callersFunctionfp16x8_to_f32x8

(src: *const u16)

src/decode.rs:5004

↓ 6 callersMethodget_current_work

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/static_tile_scheduler.hpp:163

↓ 6 callersMethodget_grid_shape

Computes the grid shape

src/cuda/flash_attn/cutlass/cutlass/conv/device/conv_universal_adapter.hpp:174

↓ 6 callersMethodget_stride

Returns a stride

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:837

↓ 6 callersFunctiongguf_type_to_cpu_bits

Map GGUF quantization type to target CPU bit width for AVX2 transposed format. Returns (target_bits, is_exact) where is_exact indicates whether the c

src/weights/mod.rs:29

↓ 6 callersMethodhas_gguf

Whether native GGUF weights are loaded (for CPU decode).

src/moe.rs:2129

↓ 6 callersFunctionhqq_attention_pending_manifest_path

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

python/krasis/attention_backend.py:439

↓ 6 callersFunctionhqq_auto_budget_bytes_from_pct

(value: float, promotion_span_bytes: int, attention_quant: str = "hqq_auto")

python/krasis/attention_backend.py:598

↓ 6 callersFunctionhqq_backend_name

python/krasis/attention_backend.py:271

↓ 6 callersFunctionimplicit_gemm_tensor_c_size

Returns ImplicitGemm tensor C size in number of elements

src/cuda/flash_attn/cutlass/cutlass/conv/conv3d_problem_size.h:502

↓ 6 callersFunctioninit_masks

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:347

↓ 6 callersFunctionload_fp8_scale

Load a per-tensor FP8 scale_inv value. Handles both BF16 scalar and FP32 scalar formats.

src/weights/mod.rs:6967

↓ 6 callersMethodload_with_ell_index

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:888

↓ 6 callersMethodload_with_ell_index_fast

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:893

↓ 6 callersFunctionload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/predicated_scale_bias_vector_iterator.h:185

↓ 6 callersMethodload_with_pointer_offset

Loads a fragment

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear_2dthreadtile.h:148

↓ 6 callersFunctionlog

src/cuda/fla/ops/utils/op.py:17

↓ 6 callersFunctionlogical_product

src/cuda/flash_attn/cutlass/cute/layout.hpp:1667

↓ 6 callersFunctionmake_arithmetic_tuple

src/cuda/flash_attn/cutlass/cute/numeric/arithmetic_tuple.hpp:70

↓ 6 callersFunctionmake_hook

tests/archive/test_layer0_detail.py:64

↓ 6 callersFunctionmake_im2col_tma_copy

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_im2col.hpp:811

↓ 6 callersFunctionmake_layout

src/cuda/flash_attn/cutlass/cute/swizzle_layout.hpp:68

↓ 6 callersFunctionmake_tensor_descriptor

( base, shape, strides, block_shape, _builder=None, )

src/cuda/fla/ops/utils/op.py:57

↓ 6 callersFunctionmake_tiled_copy_C

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:470

↓ 6 callersFunctionmarlin_w2_padded_n

Compute padded output dimension for Marlin w2 (down_proj). The `moe_wna16_marlin_gemm` kernel's thread config lookup fails when both K and N dimensio

src/weights/mod.rs:1571

↓ 6 callersMethodmatmul

Matrix-vector multiply: output[N] = W[N,K] @ input[K] Input is f32, internally quantized to INT16. Output is f32.

src/decode.rs:688

↓ 6 callersFunctionmatmul_int4_integer

Safe wrapper for the AVX2 integer INT4 matmul kernel.

src/kernel/avx2.rs:432

↓ 6 callersFunctionmatmul_int4_transposed_integer

Safe wrapper for AVX2 integer transposed INT4 matmul. For large N, tiles the N dimension to keep scratch buffer in L1 cache.

src/kernel/avx2.rs:1211

↓ 6 callersFunctionmerge_desc_sorted_arrays

src/cuda/flash_attn/cutlass/cutlass/epilogue/fusion/sm90_visitor_topk_softmax.hpp:228

↓ 6 callersFunctionnormalize_hqq_attention_cache_profile

(cache_profile: Optional[str])

python/krasis/attention_backend.py:375

↓ 6 callersMethodnum_experts

Total number of routed experts per layer.

src/moe.rs:2095

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