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

↓ 37 callersFunctionload_with_pointer_offset

Loads a fragment from memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:201

↓ 37 callersMethodpartition_A

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:522

↓ 37 callersMethodto_nv_bfloat16

src/cuda/flash_attn/cutlass/cutlass/bfloat16.h:195

↓ 37 callersMethodvalue

src/cuda/flash_attn/cutlass/cute/numeric/arithmetic_tuple.hpp:242

↓ 36 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:959

↓ 36 callersMethodget_problem_shape

src/cuda/flash_attn/cutlass/cutlass/gemm/group_array_problem_shape.hpp:64

↓ 35 callersMethod_load_bf16

Read a weight tensor and place on GPU as BF16.

python/krasis/weight_loader.py:212

↓ 35 callersMethodproducer_try_acquire

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:1100

↓ 35 callersFunctionround_up

Returns the smallest value in the half-open range [a, a+b) that is a multiple of b

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:213

↓ 35 callersFunctionsample

Sample next token from logits. Args: logits: [batch_size, vocab_size] float32/bf16 temperature: Sampling temperature (0 = greedy)

python/krasis/sampler.py:8

↓ 35 callersFunctionstore_with_pointer_offset

Store a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h:237

↓ 34 callersMethodforward

Full forward pass with streaming attention architecture. ALL attention runs on GPU0. During decode (M=1): - Attention runs on GPU0 (w

python/krasis/model.py:5125

↓ 34 callersFunctiontma_partition

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:1352

↓ 33 callersMethodclose

Close all safetensors handles.

python/krasis/weight_loader.py:880

↓ 33 callersMethodraw_data

src/cuda/flash_attn/cutlass/cutlass/array.h:415

↓ 33 callersFunctionsynclog_alloc

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:65

↓ 33 callersFunctionsynclog_condition_emit

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:233

↓ 33 callersFunctionsynclog_emit_prefix

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:77

↓ 33 callersFunctionsynclog_print_prefix

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:253

↓ 32 callersFunction_html_escape

Escape HTML special characters.

tests/release_test.py:1455

↓ 32 callersFunction_linear

Dispatch to INT8 or BF16 linear based on weight type. INT8 weights are stored as (weight_int8, scale) tuples. BF16 weights are stored as plai

python/krasis/layer.py:59

↓ 32 callersFunction_stage_done

(case: str, variant: str, stage: str, start: float, **extra)

tests/test_hqq_rust_quantizer.py:1130

↓ 32 callersFunction_stage_start

(case: str, variant: str, stage: str)

tests/test_hqq_rust_quantizer.py:1125

↓ 32 callersFunctioncopy

src/cuda/flash_attn/fa2/utils.h:299

↓ 32 callersFunctionfp32_bytes

Convert tensor to raw FP32 bytes.

tests/generate_kernel_test_data.py:37

↓ 32 callersMethodpartition_B

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:533

↓ 31 callersMethodtensor_as_slice

Get tensor data as a typed slice (zero-copy). Panics if the tensor dtype doesn't match T's size.

src/weights/safetensors_io.rs:178

↓ 30 callersMethodadd

(self, name: str, elapsed: float)

tests/hqq_attention_diff.py:44

↓ 30 callersFunctioncomposition

src/cuda/flash_attn/cutlass/cute/layout_composed.hpp:361

↓ 30 callersMethoddivide

Alias for `div` to match the interface of FastDivmodU64

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:397

↓ 30 callersFunctionget

Unpacks an element from memory

src/cuda/flash_attn/cutlass/cutlass/subbyte_reference.h:180

↓ 30 callersMethodload_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.hpp:629

↓ 30 callersMethodphase

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:193

↓ 30 callersFunctiontma_descriptor_cp_fence_release

src/cuda/flash_attn/cutlass/cute/arch/copy_sm90_desc.hpp:423

↓ 30 callersFunctiontma_descriptor_fence_acquire

src/cuda/flash_attn/cutlass/cute/arch/copy_sm90_desc.hpp:457

↓ 30 callersFunctiontma_descriptor_replace_addr_in_shared_mem

Replace tensor pointer by bringing the tensormap from GMEM into the shared memory

src/cuda/flash_attn/cutlass/cute/arch/copy_sm90_desc.hpp:341

↓ 29 callersFunctionfill_tma_gmem_shape_stride

src/cuda/flash_attn/cutlass/cute/atom/copy_traits_sm90_tma.hpp:817

↓ 29 callersFunctionhqq_attention_manifest_path

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

python/krasis/attention_backend.py:430

↓ 29 callersMethodmoe_abs_layer

(&self, moe_idx: usize)

src/weights/mod.rs:88

↓ 29 callersFunctionmul

src/cuda/flash_attn/cutlass/cute/numeric/real.hpp:49

↓ 29 callersFunctionproduct_each

src/cuda/flash_attn/cutlass/cute/int_tuple.hpp:247

↓ 28 callersFunction_status

Print a highlighted status section header (also logged).

python/krasis/server.py:110

↓ 28 callersFunctionbasis_get

src/cuda/flash_attn/cutlass/cute/numeric/arithmetic_tuple.hpp:273

↓ 28 callersFunctionmake_tiled_copy

src/cuda/flash_attn/cutlass/cute/atom/copy_atom.hpp:523

↓ 28 callersFunctionreverse

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:1042

↓ 28 callersMethodstore

Stores a fragment to memory at the location pointed to by the iterator

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_iterator_wmma.h:777

↓ 28 callersFunctiontile_to_mma_shape

src/cuda/flash_attn/cutlass/cute/atom/mma_traits_sm100.hpp:114

↓ 28 callersFunctiontma_descriptor_replace_dims_strides_in_shared_mem

Replace tensor dims and strides for GEMMs by bringing the tensormap from GMEM into the shared memory

src/cuda/flash_attn/cutlass/cute/arch/copy_sm90_desc.hpp:358

↓ 27 callersMethodcount

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator_params.h:84

↓ 27 callersFunctionmake_fragment_like

src/cuda/flash_attn/cutlass/cute/layout.hpp:453

↓ 27 callersMethodpartition_C

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:511

↓ 27 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/conv/warp/mma_depthwise_simt_tile_iterator.h:244

↓ 26 callersFunctionTensorRef_aligned

src/cuda/flash_attn/cutlass/cutlass/tensor_ref.h:398

↓ 26 callersMethodfind

src/cuda/flash_attn/cutlass/cute/container/type_list.hpp:55

↓ 26 callersFunctionquantize_activation_int16_f32

INT16 per-group quantization from f32 activations directly (no BF16 intermediate). Same as `quantize_activation_int16` but takes f32 input, eliminati

src/kernel/avx2.rs:274

↓ 26 callersMethodset_iteration_index

Overrides the internal iteration index

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_access_iterator.h:238

↓ 26 callersFunctionsynclog_setup

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:297

↓ 26 callersFunctiontiled_product

src/cuda/flash_attn/cutlass/cute/layout.hpp:1712

↓ 26 callersFunctionwarn

tests/release_test.py:90

↓ 25 callersMethodadvance_to_next_work

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm100_tile_scheduler.hpp:439

↓ 25 callersFunctionapply

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:71

↓ 25 callersFunctioncp_async_wait<0>

src/cuda/flash_attn/cutlass/cutlass/arch/memory_sm80.h:460

↓ 25 callersFunctioninfo

tests/release_test.py:82

↓ 25 callersMethodinit

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:352

↓ 25 callersFunctionmake_coord_iterator

src/cuda/flash_attn/cutlass/cute/stride.hpp:556

↓ 25 callersFunctionsave_test

Save a test case as raw binary files + metadata JSON.

tests/generate_kernel_test_data.py:47

↓ 25 callersMethodset_k_partition

Functionally required for serial reduction in the epilogue

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/conversion_op.h:96

↓ 25 callersFunctiontransform_apply

src/cuda/flash_attn/cutlass/cute/algorithm/tuple_algorithms.hpp:125

↓ 25 callersFunctionwrite_json

(path: str, payload: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:378

↓ 24 callersFunction_dim

Print a dim info line (secondary details).

python/krasis/server.py:121

↓ 24 callersFunctionblock_rank_in_cluster

Get 1D ctaid in a cluster.

src/cuda/flash_attn/cutlass/cute/arch/cluster_sm90.hpp:154

↓ 24 callersFunctionfence_view_async_shared

Issue a shared memory fence for async operations

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:711

↓ 24 callersFunctioninfo

tests/generate_reference.py:109

↓ 24 callersMethoditers_per_tile

Number of iterations per output tile

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/threadblock_swizzle_streamk.h:180

↓ 24 callersMethodrelease

Releases ownership of the managed object, if any

src/cuda/flash_attn/cutlass/cutlass/platform/platform.h:749

↓ 24 callersFunctionrequire_complete_hqq_attention_manifest

Load a selected complete HQQ manifest, failing visibly without fallback.

python/krasis/attention_backend.py:1094

↓ 24 callersFunctionscale_if_needed

scale (multiply_add) the results from the MMA accumulators to main accumulator if needed.

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/fp8_accumulation.hpp:186

↓ 24 callersMethodsize

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:202

↓ 23 callersFunction_format_pairs

(pairs: Sequence[Tuple[str, Any]])

tests/trace_diff.py:701

↓ 23 callersMethod_move_hqq_tensor_to_device

( tensor: torch.Tensor, device: torch.device, keepalive: list )

python/krasis/model.py:2999

↓ 23 callersFunctionadd

src/cuda/flash_attn/cutlass/cute/numeric/real.hpp:40

↓ 23 callersMethodclear

src/cuda/flash_attn/cutlass/cute/container/array.hpp:179

↓ 23 callersMethodget_thr_layout_vmnk

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:233

↓ 23 callersMethodlaunch

This is the method we expect to use going forward

src/cuda/flash_attn/cutlass/cutlass/cluster_launch.hpp:215

↓ 23 callersFunctionlog_ram_ledger

Log opt-in process/system RAM checkpoints for startup residency diagnosis.

python/krasis/model.py:262

↓ 23 callersFunctionmake_coord_tensor

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:482

↓ 23 callersFunctionreduce

src/cuda/flash_attn/cutlass/cute/algorithm/tensor_reduce.hpp:45

↓ 23 callersMethodtransform

Transform the mma operands to the required types

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_simt.h:251

↓ 22 callersMethodadvance

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_tile_iterator.h:172

↓ 22 callersFunctioncompute_epilogue

Returns whether the block assigned this work should compute the epilogue for the corresponding output tile. For the case of stream-K, this should only

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:580

↓ 22 callersMethodis_available

src/cuda/fla/ops/backends/__init__.py:28

↓ 22 callersMethodis_producer_load_needed

src/cuda/flash_attn/cutlass/cutlass/epilogue/collective/detail.hpp:302

↓ 22 callersFunctionpipeline_check_is_consumer

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm90_pipeline.hpp:80

↓ 22 callersFunctionquantize_int8

Quantize a BF16 weight matrix to symmetric INT8 with per-group scales. Symmetric INT8: values in [-128, 127], scale chosen so that max(abs(group)) ma

src/weights/marlin.rs:73

↓ 21 callersFunction_file_sha256

python/krasis/hqq_self_calibrate.py:151

↓ 21 callersFunctionassert_raises_contains

(fn, text: str)

tests/test_hqq_self_calibrate.py:147

↓ 21 callersFunctionexp2

src/cuda/fla/ops/utils/op.py:15

↓ 21 callersMethodinitial_work_tile_info

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_group.hpp:579

↓ 21 callersMethodpartition_fragment_A

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:552

↓ 21 callersMethodreset_state

Reset recurrent and conv state (e.g. between sequences).

python/krasis/linear_attention.py:213

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