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 callersFunctionhqq_attention_cache_layer_bytes

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

python/krasis/attention_backend.py:2809

↓ 3 callersFunctionhqq_auto_candidate_from_records

(base_record: dict, promoted_record: dict)

python/krasis/attention_backend.py:621

↓ 3 callersFunctionhqq_num_groups

(cols: usize, group_size: usize)

↓ 3 callersFunctionhqq_padded_cols

(cols: usize, group_size: usize)

↓ 3 callersFunctionhsum_avx

src/gguf_kernels.rs:666

↓ 3 callersFunctionhsum_avx2

src/kernel/avx2.rs:168

↓ 3 callersMethodis_last_tile

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/static_tile_scheduler.hpp:196

↓ 3 callersFunctionis_mxfp4

Detect MXFP4 pre-quantized format (GPT OSS). These models store all experts in a single tensor per projection per layer, with _blocks/_scales suffixes

src/weights/mod.rs:6205

↓ 3 callersFunctionis_output_fresh

(inputs: &[&str], outputs: &[&str])

↓ 3 callersFunctionis_prequantized

Detect whether the model uses BF16 weights or pre-quantized compressed-tensors INT4. Returns true if pre-quantized (weight_packed tensors found).

src/weights/mod.rs:6139

↓ 3 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm90_epilogue_array_tma_warpspecialized.hpp:473

↓ 3 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm100_epilogue_tma_warpspecialized.hpp:447

↓ 3 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/sm100_epilogue_array_tma_warpspecialized.hpp:454

↓ 3 callersFunctionl2norm_bwd

( y: torch.Tensor, rstd: torch.Tensor, dy: torch.Tensor, eps: float = 1e-6, )

src/cuda/fla/modules/l2norm.py:184

↓ 3 callersFunctionl2norm_fwd

( x: torch.Tensor, eps: float = 1e-6, output_dtype: torch.dtype | None = None, )

src/cuda/fla/modules/l2norm.py:132

↓ 3 callersFunctionlaunch_kernel

( func: RawCuFunc, grid_x: u32, block_x: u32, params: &mut [*mut std::ffi::c_void], )

↓ 3 callersMethodlaunch_server

Write a temp config file and exec the Krasis server with --config.

python/krasis/launcher.py:2439

↓ 3 callersFunctionload_and_quantize_expert_weight_int4

( layer_idx: usize, expert_idx: usize, prefix: &str, proj_name: &str, weight_map: &HashMap

src/weights/mod.rs:6752

↓ 3 callersFunctionload_dataset_text

Load a dataset by name. Downloads and caches on first use. Args: dataset_name: Key from DATASETS registry (e.g. "wikitext-2") Return

perplexity/measure_ppl.py:79

↓ 3 callersFunctionload_expert_proj

(layer_idx, expert_idx, proj_name)

tests/test_rust_vs_python.py:51

↓ 3 callersFunctionload_expert_proj

Load one expert projection from compressed-tensors safetensors.

tests/archive/test_prequant_compare.py:96

↓ 3 callersFunctionload_hqq_search_func

(device: &Arc<CudaDevice>, name: &str)

↓ 3 callersFunctionload_input_rows_from_trace

( paths: list[Path], *, positions: set[int] | None, layer: int, tensor: str, )

tests/hqq_attention_diff.py:162

↓ 3 callersFunctionload_or_infer_reference_contract

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

tests/reference_contract.py:614

↓ 3 callersFunctionload_prompt_text_truncated

Load a Gutenberg prompt truncated to max_chars.

tests/prompt_utils.py:37

↓ 3 callersFunctionload_reference_model_and_tokenizer

Load the BF16 HF reference model through the shared capture path.

tests/generate_reference.py:1251

↓ 3 callersFunctionload_safetensors_tensor

tests/archive/test_prequant_compare.py:25

↓ 3 callersFunctionload_shared_expert_proj

Load shared expert projection as BF16 weight tensor.

tests/test_rust_vs_python.py:69

↓ 3 callersFunctionload_with_byte_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:310

↓ 3 callersMethodload_with_byte_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator.h:1465

↓ 3 callersFunctionload_with_pointer_offset

Loads a fragment from memory with additional logical offset

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_complex_tensor_op_tile_iterator_sm80.h:284

↓ 3 callersFunctionload_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_tensor_op.h:210

↓ 3 callersMethodload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:871

↓ 3 callersFunctionmake_prompt

benchmarks/bench_decode.py:53

↓ 3 callersFunctionmake_row

tests/hqq_attention_diff.py:670

↓ 3 callersFunctionmake_swizzle_ptr

src/cuda/flash_attn/cutlass/cute/pointer_swizzle.hpp:125

↓ 3 callersFunctionmake_tma_atom_im2col

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_im2col.hpp:563

↓ 3 callersMethodmantissa

Returns the mantissa

src/cuda/flash_attn/cutlass/cutlass/tfloat32.h:185

↓ 3 callersMethodmantissa

Returns the mantissa

src/cuda/flash_attn/cutlass/cutlass/half.h:487

↓ 3 callersMethodmantissa

Returns the mantissa

src/cuda/flash_attn/cutlass/cutlass/bfloat16.h:223

↓ 3 callersFunctionmarlin_cache_basename

(gpu_bits: int, group_size: Union[int, str], gpu_expert_int4_calib: str = "amax")

python/krasis/config.py:58

↓ 3 callersFunctionmarlin_cache_lock_is_live

(lock_path: &Path)

src/weights/mod.rs:1311

↓ 3 callersFunctionmarlin_make_workspace

Allocate Marlin GEMM workspace (SM count * max_blocks_per_sm ints).

python/krasis/marlin_utils.py:76

↓ 3 callersFunctionmatmul_int4_integer

(packed_int32, weight_scales, act_i16, a_scales, rows_, cols_, gs=32)

tests/archive/test_prequant_compare.py:235

↓ 3 callersFunctionmatmul_int4_marlin

Safe wrapper for Marlin-native INT4 matmul (sequential).

src/kernel/avx2.rs:2096

↓ 3 callersFunctionmatmul_int4_transposed_integer_parallel

Parallel AVX2 integer transposed INT4 matmul.

src/kernel/avx2.rs:1250

↓ 3 callersFunctionmatmul_int4_transposed_integer_tiled

Single-threaded tiled INT4 matmul. Data in [N/TILE, K/8, TILE] layout.

src/kernel/avx2.rs:1361

↓ 3 callersFunctionmatmul_int8_transposed_integer_parallel

Parallel AVX2 integer transposed INT8 matmul. Splits work by N columns across rayon threads (same strategy as INT4 transposed).

src/kernel/avx2.rs:1630

↓ 3 callersFunctionmatmul_int8_transposed_integer_tiled

Single-threaded tiled INT8 matmul. Data in [N/TILE, K, TILE] layout.

src/kernel/avx2.rs:1717

↓ 3 callersMethodmax_dim_index

Returns the index of the dimension with greatest value

src/cuda/flash_attn/cutlass/cutlass/coord.h:142

↓ 3 callersFunctionmetrics

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

tests/hqq_attention_diff.py:502

↓ 3 callersFunctionmetrics_from_quant

( source: torch.Tensor, quant: torch.Tensor, scales: torch.Tensor, zeros: torch.Tensor, gr

tests/hqq_gpu_search_probe.py:104

↓ 3 callersMethodnext_u64

src/decode.rs:4922

↓ 3 callersMethodnum_moe_layers

src/weights/mod.rs:94

↓ 3 callersFunctionnvcc_host_compiler_args

↓ 3 callersMethodoperator++

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:498

↓ 3 callersMethodoperator++

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op_sm80.h:588

↓ 3 callersFunctionparse_all_input_row_events

tests/hqq_attention_diff.py:228

↓ 3 callersFunctionparse_config

Parse a .conf file into a dict of key=value pairs.

tests/validate_model.py:162

↓ 3 callersFunctionparse_input_row_positions

tests/hqq_attention_diff.py:155

↓ 3 callersMethodpipeline

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler.hpp:62

↓ 3 callersFunctionpopcount

src/cuda/flash_attn/cutlass/cutlass/functional.h:67

↓ 3 callersFunctionprefetch_buf

(ptr: *const u8, bytes: usize)

src/moe.rs:1172

↓ 3 callersFunctionprefetch_weight

Prefetch a QuantWeight's data into L3 using NTA hints.

src/moe.rs:1319

↓ 3 callersMethodprefill_scratch_reservation_mb

(self, tokens: int)

tests/test_startup_calibration.py:100

↓ 3 callersFunctionpressure_slot

(device_id: i32)

src/vram_monitor.rs:38

↓ 3 callersFunctionproducer_commit

NOP for TMA based mainloop

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:560

↓ 3 callersMethodproducer_commit

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:1089

↓ 3 callersFunctionproducer_try_acquire

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:119

↓ 3 callersFunctionprofile_decode

Prefill then decode with per-token timing.

tests/test_gguf.py:20

↓ 3 callersFunctionprofile_decode

Prefill then decode with per-token timing.

tests/archive/test_qwen235b_bench.py:71

↓ 3 callersFunctionprofile_filename

(profile_id: str)

tests/reference_contract.py:447

↓ 3 callersFunctionquantize_hqq6_tensor

HQQ6 tensor quantizer with true packed 6-bit storage.

python/krasis/attention_backend.py:2294

↓ 3 callersFunctionquantize_with_affine

( source: torch.Tensor, scale: float, zero: float, )

tests/hqq_attention_diff.py:767

↓ 3 callersFunctionquery_free_bytes

Query free VRAM in bytes for a specific device.

src/vram_monitor.rs:143

↓ 3 callersMethodread_bytes

(&mut self, n: usize)

src/gguf.rs:274

↓ 3 callersFunctionread_quantized

( data: &[u8], offset: &mut usize, rows: usize, cols: usize, group_size: usize, num_bi

src/weights/mod.rs:5966

↓ 3 callersFunctionread_unified_expert_cpu_gated

( data: &[u8], offset: &mut usize, hidden_size: usize, intermediate_size: usize, group_siz

src/weights/mod.rs:5711

↓ 3 callersFunctionread_unified_expert_cpu_mixed_gated

( data: &[u8], offset: &mut usize, hidden_size: usize, intermediate_size: usize, group_siz

src/weights/mod.rs:5815

↓ 3 callersFunctionreduce_like_torch_f32

(values: impl Iterator<Item = f32>)

↓ 3 callersFunctionreference_candidate_filenames

(profile_id: Optional[str])

tests/reference_contract.py:523

↓ 3 callersMethodregister

Register a backend.

src/cuda/fla/ops/backends/__init__.py:71

↓ 3 callersFunctionrel

scripts/build_sidecars.py:87

↓ 3 callersFunctionreset_mempolicy

Reset process memory policy to MPOL_DEFAULT (local allocation).

src/numa.rs:341

↓ 3 callersFunctionresolve_model_name

(raw_model: str)

tests/generate_reference.py:205

↓ 3 callersFunctionresolve_model_name

Extract model name from config's model path.

tests/reference_test.py:110

↓ 3 callersFunctionrestore_store_after_rust_prefill

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

src/server.rs:260

↓ 3 callersMethodrun

Run the full benchmark suite. Returns results dict.

python/krasis/benchmark.py:792

↓ 3 callersMethodrun_all

Run all combos sequentially, skipping on failure.

python/krasis/suite.py:443

↓ 3 callersFunctionrun_config

Run V2-Lite with given config, return results dict.

tests/archive/test_quant_config.py:46

↓ 3 callersFunctionrun_mha_fwd

src/cuda/flash_attn/fa2/flash_api_reference.cpp:243

↓ 3 callersFunctionrun_status_timed

( mut cmd: std::process::Command, label: &str, )

↓ 3 callersFunctionrun_streaming_test

Run a streaming test and measure TTFT + ITL separately.

benchmarks/bench_decode.py:61

↓ 3 callersFunctionsafe_div

src/cuda/flash_attn/cutlass/cute/swizzle.hpp:365

↓ 3 callersFunctionsafe_div

src/cuda/flash_attn/cutlass/cute/numeric/math.hpp:307

↓ 3 callersFunctionsample_from_logits

Sample a token from logits using temperature, top-k, and top-p.

src/decode.rs:4279

↓ 3 callersFunctionsend_chat_request

Send a chat completion request and measure timing. Returns (prefill_tok_s, decode_tok_s, ttft, total_time, num_generated, text)

tests/archive/bench_network_vs_benchmark.py:35

↓ 3 callersFunctionset_block_rank

Set the destination block-ID in cluster for a given SMEM Address

src/cuda/flash_attn/cutlass/cute/arch/cluster_sm90.hpp:166

↓ 3 callersMethodset_iteration_index

Overrides the internal iteration index

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:866

↓ 3 callersMethodset_iteration_index

Overrides the internal iteration index

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op.h:289

← previous next →1,301–1,400 of 12,380, ranked by callers