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

↓ 3 callersFunction_unsupported_torch_devices

Return visible devices whose SM arch is missing from installed torch.

python/krasis/setup.py:366

↓ 3 callersFunction_variant_metric

(variant: Dict[str, Any], *keys: str, default: float = 0.0)

python/krasis/hqq_self_calibrate.py:4670

↓ 3 callersFunction_vector_delta_summary

( expected: Optional[Sequence[float]], actual: Optional[Sequence[float]], )

tests/trace_diff.py:307

↓ 3 callersMethod_visible_config_options

Return config options visible in the current TUI mode.

python/krasis/launcher.py:1495

↓ 3 callersFunction_visible_len

Length of string with ANSI escape codes stripped.

python/krasis/launcher.py:71

↓ 3 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:870

↓ 3 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op.h:293

↓ 3 callersMethodadd_tile_offset

Advances an iterator along logical dimensions of matrix in units of whole tiles

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:877

↓ 3 callersMethodadd_tile_offset

Adds a tile offset

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op.h:305

↓ 3 callersFunctionaggregate_candidate_groups

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

tests/hqq_attention_diff.py:1787

↓ 3 callersFunctionany_of

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:304

↓ 3 callersFunctionappend_safety_limit_dump

( kind: &str, device_id: i32, free_mb: u64, safety_margin_mb: u64, deficit_mb: u64, )

src/vram_monitor.rs:242

↓ 3 callersMethodapply_saved

Apply loaded config values.

python/krasis/launcher.py:607

↓ 3 callersMethodarrive_and_expect_tx

Performs an arrive operation + expected transaction bytes increment

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:541

↓ 3 callersFunctionarrive_internal

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:301

↓ 3 callersMethodas_bf16

Return as BF16 ref (panics if not BF16).

src/weights/mod.rs:311

↓ 3 callersFunctionattention_quant_label

(attention_quant: str)

python/krasis/attention_backend.py:255

↓ 3 callersFunctionbasis_value

src/cuda/flash_attn/cutlass/cute/numeric/arithmetic_tuple.hpp:260

↓ 3 callersMethodbegin_row

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_store.hpp:315

↓ 3 callersFunctionbenchmark_decode_detailed

Detailed decode benchmark with per-token timings.

tests/archive/test_v2lite_timing.py:77

↓ 3 callersFunctionblock

src/cuda/flash_attn/cutlass/cute/util/debug.hpp:121

↓ 3 callersFunctionbuild_contract

( model_name: str, model_path: str, tokenizer: Any, *, max_new_tokens: int, add_genera

tests/reference_contract.py:533

↓ 3 callersFunctionbuild_prefill_diagnostic_from_logits

( tokenizer: Any, prompt_input_ids: Any, generated_token_ids: List[int], step_logits: Any,

tests/generate_reference.py:1017

↓ 3 callersFunctionbuild_prompt

Build a prompt of approximately target_tokens length using Gutenberg text.

tests/archive/test_qwen235b_bench.py:46

↓ 3 callersFunctionbuild_reference_artifact_metadata

( reference: Dict[str, Any], reference_contract: Dict[str, Any], reference_path: Optional[str], )

tests/reference_contract.py:667

↓ 3 callersFunctionbuild_teacher_forced_per_token_data

( model: Any, tokenizer: Any, prompt_input_ids: Any, generated_token_ids: List[int], *,

tests/generate_reference.py:959

↓ 3 callersFunctionbuild_token_diagnostic_entry

( tokenizer: Any, step_logits: Any, *, expected_token_id: int, logit_pos: int, prev_to

tests/generate_reference.py:913

↓ 3 callersFunctioncapture_prefill_logits

(model: Any, prompt_input_ids: Any, *, use_cache: Optional[bool] = None)

tests/generate_reference.py:1007

↓ 3 callersFunctioncapture_settings_for_profile

(profile_id: str)

tests/reference_contract.py:462

↓ 3 callersMethodclear_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:924

↓ 3 callersFunctioncommand_output

(args: list[str])

scripts/build_sidecars.py:111

↓ 3 callersFunctioncompare

(name, cpu_t, gpu_t, n=8)

tests/test_gemv_verify.py:34

↓ 3 callersFunctioncompare

(name, ref_t, test_t, n=8)

tests/archive/test_moe_accum_verify.py:31

↓ 3 callersFunctioncompare_init

(name: str, chunk: torch.Tensor)

tests/test_hqq_rust_quantizer.py:94

↓ 3 callersFunctioncompare_real_artifact_registration_layer

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

tests/test_hqq_rust_quantizer.py:1558

↓ 3 callersFunctioncomplement

src/cuda/flash_attn/cutlass/cute/layout.hpp:1178

↓ 3 callersFunctioncompute_hqq4_rmse_rust

Rust shadow RMSE path for one HQQ4 group chunk.

python/krasis/attention_backend.py:1421

↓ 3 callersMethodconvert

src/cuda/flash_attn/cutlass/cutlass/half.h:195

↓ 3 callersMethodconvert_from_float

src/cuda/flash_attn/cutlass/cutlass/float8.h:1071

↓ 3 callersFunctioncooperative_gemm_predication

src/cuda/flash_attn/cutlass/cute/algorithm/cooperative_gemm.hpp:166

↓ 3 callersFunctioncopy_unpack

src/cuda/flash_attn/cutlass/cute/atom/copy_traits.hpp:111

↓ 3 callersFunctioncrd2crd

src/cuda/flash_attn/cutlass/cute/stride.hpp:254

↓ 3 callersFunctioncreate_bundle

scripts/build_sidecars.py:477

↓ 3 callersFunctioncreate_run_dir

(run_type: str, parent: Optional[os.PathLike] = None)

python/krasis/run_paths.py:35

↓ 3 callersFunctioncuda_alloc_f32

↓ 3 callersFunctioncuda_free

(ptr: cuda_sys::CUdeviceptr)

↓ 3 callersFunctiondequant_bf16

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

src/gguf.rs:559

↓ 3 callersFunctiondequant_ct

Vectorized dequantize compressed-tensors INT4 to FP32.

tests/test_rust_vs_python.py:40

↓ 3 callersFunctiondequant_ct_vectorized

Vectorized dequantization of compressed-tensors INT4. Each int32 word contains 8 nibbles: nibble_j at bits [4j, 4j+3]. Dequantized value = (n

tests/archive/test_prequant_compare.py:44

↓ 3 callersFunctiondequant_f32

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

src/gguf.rs:536

↓ 3 callersFunctiondequantize_raw_data

Dequantize raw GGUF-format bytes to FP32. Accepts raw byte data + GGML type + element count. Used by the GGUF→AVX2 cache builder to dequantize indivi

src/gguf.rs:872

↓ 3 callersMethoddispatch_matmul

Dispatch matmul to correct INT4/INT8 kernel.

src/decode.rs:1502

↓ 3 callersMethoddiv

Computes integer division using precomputed values. This is computationally inexpensive.

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:389

↓ 3 callersFunctiondtype_name

(tensor: torch.Tensor)

tests/test_hqq_fused_branch_runtime.py:33

↓ 3 callersMethodell_add_mask

add mask for small tiles in ELL

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:940

↓ 3 callersFunctionemit_contract_trace

( context: str, reference_contract: Dict[str, Any], runtime_contract: Dict[str, Any], validati

tests/reference_contract.py:1142

↓ 3 callersMethodenable

< CUTLASS_HOST_DEVICE enables all accesses guarded by mask

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:179

↓ 3 callersMethodenable_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:928

↓ 3 callersMethodend_row

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_store.hpp:341

↓ 3 callersMethodend_step

Called after all accumulator elements have been visited

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_2x.hpp:111

↓ 3 callersMethodevict_page_cache

Evict all pages from page cache. Call after all data has been copied out.

src/weights/safetensors_io.rs:210

↓ 3 callersFunctionevidence_trace_log_path

(output_path: str)

python/krasis/hqq_self_calibrate.py:387

↓ 3 callersFunctionexpected_marlin_cache_size

Expected total Marlin cache file size.

src/weights/mod.rs:1719

↓ 3 callersFunctionexpert_matmul_int4

( packed: *const u32, // [N, K/8] packed INT4 scales: *const u16, // [N, K/group_size] BF16 sc

src/kernel/avx2.rs:105

↓ 3 callersFunctionexpert_matmul_int4_integer

( packed: *const u32, // [N, K/8] packed INT4 weight_scales: *const u16, // [N, K/group_size] B

src/kernel/avx2.rs:366

↓ 3 callersFunctionexpert_matmul_int4_marlin

AVX2 Marlin-native INT4 matmul — production kernel. Processes Marlin tiles: for each (K-group × N-chunk-of-64), unpacks 1024 INT4 values from 128 u32

src/kernel/avx2.rs:1965

↓ 3 callersFunctionexpert_matmul_int4_transposed

( packed: *const u32, // [K/8, n_stride] scales: *const u16, // [K/group_size, n_stride] a

src/kernel/avx2.rs:868

↓ 3 callersFunctionexpert_matmul_int8_integer

( data: *const i8, // [N, K] raw INT8 weights weight_scales: *const u16, // [N, K/group_size

src/kernel/avx2.rs:656

↓ 3 callersFunctionextract_server_error

Extract the tail of the server log for error reporting.

tests/release_test.py:938

↓ 3 callersFunctionf32_to_bf16

tests/test_marlin_attn_shapes.rs:22

↓ 3 callersFunctionfast_exp_avx2_inline

src/decode.rs:2004

↓ 3 callersFunctionfill_workspace

src/cuda/flash_attn/cutlass/cutlass/workspace.h:97

↓ 3 callersFunctionfind

src/cuda/flash_attn/cutlass/cute/container/tuple.hpp:254

↓ 3 callersFunctionfind_krasis_command

Find the installed krasis command. Fails if not found. Prefers the conda env krasis because that's the environment with GPU dependencies (PyT

tests/release_test.py:541

↓ 3 callersFunctionfind_nvcc

↓ 3 callersMethodfinish

↓ 3 callersFunctionflatten

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:582

↓ 3 callersFunctionfnv1a

FNV-1a hash for cache invalidation.

src/weights/mod.rs:1215

↓ 3 callersMethodforward

Forward pass for one layer. Args: hidden: [M, hidden_size] BF16 residual: [M, hidden_size] BF16 or None (first layer)

python/krasis/layer.py:313

↓ 3 callersFunctiongemm

src/cuda/flash_attn/fa2/utils.h:139

↓ 3 callersFunctiongen

(prompt, num_tokens=10)

tests/test_gguf_native.py:33

↓ 3 callersFunctiongenerate

(model, prompt, max_new_tokens=64, temperature=0.6, top_k=50, top_p=0.95)

tests/archive/run_length_sweep.py:20

↓ 3 callersFunctionget

src/cuda/flash_attn/cutlass/cute/layout.hpp:497

↓ 3 callersMethodget

Returns a pointer

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:883

↓ 3 callersMethodget

Returns a pointer

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op.h:299

↓ 3 callersFunctiongetParams

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h:177

↓ 3 callersMethodget_batch_idx

Obtains the batch index

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:653

↓ 3 callersFunctionget_grid_shape

src/cuda/flash_attn/cutlass/cutlass/transform/kernel/sm90_sparse_gemm_compressor.hpp:183

↓ 3 callersMethodget_grid_shape

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm100_tile_scheduler.hpp:204

↓ 3 callersMethodget_mask

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:936

↓ 3 callersMethodget_mask

< Sets the mask

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:1060

↓ 3 callersFunctionget_max_cta_occupancy

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/tile_scheduler_params.h:54

↓ 3 callersFunctionget_scalar_type

Get ScalarType for Marlin quantization config.

python/krasis/marlin_utils.py:66

↓ 3 callersFunctionget_shard_for_tensor

(model_path, tensor_name)

tests/archive/test_prequant_compare.py:36

↓ 3 callersMethodget_stride

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

↓ 3 callersFunctiongguf_matvec_int

INT16-activation matvec dispatch (uses AVX2 for Q4_K, Q8_0, Q4_0).

src/gguf_kernels.rs:177

↓ 3 callersFunctiongptq_marlin_gemm

Call vendored Marlin GEMM kernel via ctypes. Matches sgl_kernel.gptq_marlin_gemm API for drop-in replacement.

python/krasis/marlin_utils.py:137

↓ 3 callersFunctiongpu_mem_mb

Current GPU memory usage per device in MB.

tests/archive/meta_optimiser.py:76

↓ 3 callersMethodhas_cpu_weights

Whether CPU decode weights (transposed format) have been populated.

src/weights/mod.rs:4250

↓ 3 callersFunctionhf_login

python/krasis/hf_downloader.py:271

← previous next →1,201–1,300 of 12,380, ranked by callers