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

↓ 2 callersFunction_validate_sidecar_correction_scale

python/krasis/hqq_self_calibrate.py:3198

↓ 2 callersFunction_variant_case_effects

(variant: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:4653

↓ 2 callersFunction_verify_hcs_vram_floor

Gate readiness on measured free VRAM, not just planned HCS budget.

python/krasis/server.py:2378

↓ 2 callersFunction_weight_bytes

(params: int, quantization: str)

python/krasis/vram_budget.py:119

↓ 2 callersFunction_weight_dtype_code

Map torch dtype to Rust GpuWeight dtype code: 0=BF16, 1=FP32, 2=FP16.

python/krasis/model.py:125

↓ 2 callersFunction_witness_summary_metrics

(summary_path: str)

python/krasis/hqq_self_calibrate.py:4437

↓ 2 callersFunction_write_hqq_attention_artifact_instrumented

( case: str, variant: str, *, model_path: str, layer_idx: int, layer_type: str, te

tests/test_hqq_rust_quantizer.py:489

↓ 2 callersFunction_write_real_artifact_layer_instrumented

( case: str, variant: str, model: KrasisModel, real_case: dict, *, use_rust_shadow: bo

tests/test_hqq_rust_quantizer.py:923

↓ 2 callersFunctionabs

src/cuda/flash_attn/cutlass/cutlass/quaternion.h:391

↓ 2 callersFunctionabsolute_value

src/cuda/flash_attn/cutlass/cutlass/fast_math.h:1073

↓ 2 callersFunctionadd_byte_offset_

Adds a pointer offset in units of element

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv3d_dgrad_output_gradient_tile_access_iterator_optimized.h:286

↓ 2 callersFunctionadd_byte_offset_

Adds a pointer offset in units of element

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv3d_fprop_activation_tile_access_iterator_optimized.h:281

↓ 2 callersFunctionadd_byte_offset_

Adds a pointer offset in units of element

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h:264

↓ 2 callersFunctionadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_access_iterator_pitch_linear_direct_conv.h:155

↓ 2 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.h:168

↓ 2 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_triangular_matrix.h:723

↓ 2 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h:704

↓ 2 callersMethodadd_pointer_offset

Adds a pointer offset in units of Element

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear_2dthreadtile.h:243

↓ 2 callersFunctionadd_ranked

(bucket: List[Dict[str, Any]], bucket_budget: int)

python/krasis/hqq_self_calibrate.py:2045

↓ 2 callersFunctionadd_samples

(item: dict[str, Any])

tests/hqq_attention_diff.py:746

↓ 2 callersMethodadd_tile_offset

Adds a tile offset

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear_2dthreadtile.h:249

↓ 2 callersFunctionadvance

Advances the iterator along the advance dimension

src/cuda/flash_attn/cutlass/cutlass/gemm/warp/mma_tensor_op_tile_access_iterator.h:242

↓ 2 callersFunctionadvance_smem_write_stage

Advance global memory read-iterators and shared memory write-iterators to the stage

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_multistage.h:263

↓ 2 callersFunctionadvance_to_next_work

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:323

↓ 2 callersFunctionaffine_candidate_score

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

tests/hqq_attention_diff.py:864

↓ 2 callersMethodapply

src/cuda/flash_attn/cutlass/cute/swizzle.hpp:74

↓ 2 callersFunctionapply_capture_template

(tokenizer: Any, messages: List[Dict[str, str]], capture_settings: Dict[str, Any])

tests/reference_contract.py:512

↓ 2 callersMethodapply_dropout

src/cuda/flash_attn/fa2/dropout.h:27

↓ 2 callersMethodapply_with_tools

Apply with optional tools array for accurate token estimation.

src/chat_template.rs:108

↓ 2 callersMethodapply_with_tools_inner

( &self, messages_json: &str, tools_json: &str, add_generation_prompt: bool,

src/chat_template.rs:141

↓ 2 callersFunctionarrive

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:218

↓ 2 callersFunctionas_position_independent_swizzle_layout

src/cuda/flash_attn/cutlass/cute/pointer_flagged.hpp:91

↓ 2 callersMethodas_slice

Get a typed slice view of the allocation. # Safety Caller must ensure the data has been properly initialized as type T.

↓ 2 callersMethodat

Returns a reference to the element at a given Coord

src/cuda/flash_attn/cutlass/cutlass/tensor_ref_planar_complex.h:286

↓ 2 callersFunctionawq_template_exists

Check if an AWQ calibration template exists for this model. Checks both the bundled templates directory and user cache.

tests/release_test.py:398

↓ 2 callersMethodbegin_epilogue

Called at the start of the epilogue just before iterating over accumulator slices

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/fusion/visitor_2x.hpp:63

↓ 2 callersFunctionbench_decode

Benchmark pure decode speed (timing OFF).

tests/archive/investigate_decode_regression.py:45

↓ 2 callersFunctionbenchmark_decode

Run NUM_RUNS decode benchmarks and return results.

tests/archive/test_qwen3_next_gpu_decode.py:44

↓ 2 callersFunctionbenchmark_decode

Benchmark decode: measure tok/s after prefill.

tests/archive/run_hcs_benchmark.py:63

↓ 2 callersFunctionbenchmark_decode

Benchmark: prefill + decode, measuring TTFT and decode speed separately.

tests/archive/run_hcs_hybrid_bench.py:65

↓ 2 callersFunctionbenchmark_short_decode

Benchmark decode with short prompt to isolate decode speed.

tests/archive/run_hcs_benchmark.py:100

↓ 2 callersFunctionbenchmark_short_decode

Short prompt to isolate decode speed.

tests/archive/run_hcs_hybrid_bench.py:102

↓ 2 callersFunctionbest_fit_affine_params

(source: torch.Tensor)

tests/hqq_attention_diff.py:831

↓ 2 callersFunctionbf16_to_f32

src/decode.rs:1989

↓ 2 callersFunctionbf16_word_to_f32

tests/hqq_attention_diff.py:151

↓ 2 callersFunctionbuild_base_store

( *, include_fused_config: bool, )

tests/test_hqq_fused_branch_runtime.py:176

↓ 2 callersFunctionbuild_invocation_metadata

(model_name: str, profile_id: str, max_new_tokens: int)

tests/generate_reference.py:184

↓ 2 callersFunctionbuild_prompt

Build a prompt of approximately target_tokens length using Gutenberg text.

tests/archive/run_hcs_benchmark.py:57

↓ 2 callersFunctionbuild_prompt

Build a prompt of approximately target_tokens length using Gutenberg text.

tests/archive/run_hcs_hybrid_bench.py:59

↓ 2 callersFunctionbuild_prompt

Build a prompt of approximately target_tokens length using Gutenberg text.

tests/archive/token_scaling_bench.py:29

↓ 2 callersFunctionbuild_qwen35_source

(model_dir: Path, layer: int, tensor_name: str)

tests/hqq_attention_diff.py:2203

↓ 2 callersFunctionbuild_reference_sanity_report

( reference: Dict[str, Any], *, expected_conversations: int, expected_turns: int, )

tests/reference_contract.py:683

↓ 2 callersFunctionbuild_runtime_contract

( conf_path: str, tokenizer: Any, *, max_new_tokens: int, effective_profile_id: Optional[s

tests/reference_contract.py:929

↓ 2 callersFunctionbuild_ssh_tunnel_command

Build the ssh command for a reverse tunnel without invoking a shell.

python/krasis/ssh_tunnel.py:49

↓ 2 callersFunctionbuild_token_audit_turns

(reference: Dict[str, Any])

tests/generate_reference.py:659

↓ 2 callersFunctionbundle_filename

(bundle_hash: str)

scripts/build_sidecars.py:433

↓ 2 callersFunctioncache_path_cpu

Cache file path for CPU-optimized transposed format (INT4 or INT8).

src/weights/mod.rs:1553

↓ 2 callersFunctioncache_path_gguf_avx2

Cache file path for GGUF-sourced AVX2 transposed CPU cache.

src/weights/mod.rs:1559

↓ 2 callersFunctioncalculate_umma_peer_mask

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:95

↓ 2 callersFunctioncall_prefill_logits

Call the /v1/internal/prefill_logits endpoint with raw token IDs.

tests/reference_test.py:442

↓ 2 callersFunctioncall_reference_test

Call the /v1/internal/reference_test endpoint.

tests/reference_test.py:360

↓ 2 callersMethodcan_use

src/cuda/fla/ops/backends/__init__.py:40

↓ 2 callersFunctioncandidate_regression_record

( *, scope: dict[str, Any], aggregate: dict[str, Any], candidate_name: str, )

tests/hqq_attention_diff.py:1811

↓ 2 callersFunctioncheck_health

Check if server is healthy.

tests/test_network.py:149

↓ 2 callersFunctionchunk_gated_delta_rule_fwd_h

( k: torch.Tensor, w: torch.Tensor, u: torch.Tensor, g: torch.Tensor | None = None, gk: to

src/cuda/fla/ops/common/chunk_delta_h.py:655

↓ 2 callersFunctionchunk_local_cumsum

( g: torch.Tensor, chunk_size: int, reverse: bool = False, scale: float = None, cu_seqlens

src/cuda/fla/ops/utils/cumsum.py:429

↓ 2 callersFunctioncleanup_marlin_cache_before_build

( model_dir: &Path, config: &ModelConfig, total_moe_layers: usize, config_hash: u64, gpu_b

src/weights/mod.rs:1355

↓ 2 callersMethodclear

< Efficiently disables all accesses guarded by mask

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator_affine.h:227

↓ 2 callersMethodclear

< Efficiently disables all accesses guarded by mask

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator_direct_conv.h:154

↓ 2 callersMethodclear

< Efficiently disables all accesses guarded by mask

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator_strided_dgrad.h:155

↓ 2 callersFunctionclear_mask

Clears the predicates

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h:243

↓ 2 callersFunctionclear_mask

Clears the predicates

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h:324

↓ 2 callersMethodclear_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_triangular_matrix.h:752

↓ 2 callersMethodclear_mask

Clears the predicate set efficiently

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h:733

↓ 2 callersFunctioncluster_arrive

src/cuda/flash_attn/cutlass/cute/arch/cluster_sm90.hpp:57

↓ 2 callersFunctioncoalesce_x

src/cuda/flash_attn/cutlass/cute/layout.hpp:824

↓ 2 callersFunctioncollect_failure_case

(*, path_name: str, use_artifacts: bool)

tests/test_hqq_fused_branch_runtime.py:777

↓ 2 callersFunctioncollect_real_model_success_case

( *, real_case: dict, path_name: str, include_fused_desc: bool, include_fused_config: bool

tests/test_hqq_fused_branch_runtime.py:870

↓ 2 callersFunctioncompact_order

src/cuda/flash_attn/cutlass/cute/stride.hpp:410

↓ 2 callersFunctioncompare_real_artifact_layer

(name: str, real_case: dict)

tests/test_hqq_rust_quantizer.py:1213

↓ 2 callersFunctioncompare_real_case_tensors

( name: str, real_case: dict, tensor_names: list[str], group_indices: list[int], )

tests/test_hqq_rust_quantizer.py:152

↓ 2 callersFunctioncompare_rmse

(name: str, chunk: torch.Tensor, q: torch.Tensor, scale: torch.Tensor, zero: torch.Tensor)

tests/test_hqq_rust_quantizer.py:77

↓ 2 callersFunctioncompare_solve

(name: str, chunk: torch.Tensor, zero: torch.Tensor, scale_seed: torch.Tensor)

tests/test_hqq_rust_quantizer.py:112

↓ 2 callersFunctioncomposition_impl

src/cuda/flash_attn/cutlass/cute/layout.hpp:1031

↓ 2 callersFunctioncompute_metrics

Compute all comparison metrics for one turn.

tests/reference_test.py:506

↓ 2 callersFunctioncompute_prefill_metrics

Compare our prefill logits against BF16 reference at sampled positions.

tests/reference_test.py:461

↓ 2 callersFunctioncompute_rmse

(chunk: &[f32], q: &[u8], scale: f32, zero: f32)

↓ 2 callersMethodconsumer_release

src/cuda/flash_attn/cutlass/cutlass/pipeline/sm100_pipeline.hpp:1097

↓ 2 callersMethodcontext_for

( &self, layer_idx: usize, expert_idx: usize, proj_name: &str, row_idx

src/weights/mod.rs:1870

↓ 2 callersFunctioncontinue_current_work_for_linear_idx

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp:305

↓ 2 callersFunctionconvert_from_float

src/cuda/flash_attn/cutlass/cutlass/exmy_base.h:1004

↓ 2 callersMethodconvert_to_float

src/cuda/flash_attn/cutlass/cutlass/float8.h:1083

↓ 2 callersFunctioncooperative_gemm_no_predication

src/cuda/flash_attn/cutlass/cute/algorithm/cooperative_gemm.hpp:279

↓ 2 callersFunctioncoprofile

src/cuda/flash_attn/cutlass/cute/layout.hpp:634

↓ 2 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_with_reduction_multistage.h:90

↓ 2 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_planar_complex_multistage.h:94

↓ 2 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_sparse_multistage.h:97

↓ 2 callersFunctioncopy_tiles_and_advance

src/cuda/flash_attn/cutlass/cutlass/gemm/threadblock/mma_blas3_multistage.h:94

↓ 2 callersFunctioncopy_to_package

scripts/build_sidecars.py:386

↓ 2 callersFunctioncpu_expert_byte_sizes

Compute per-expert byte sizes for CPU transposed format. INT4 has same sizes as Marlin (same u32 packing, different layout). INT8 has larger packed da

src/weights/mod.rs:1603

← previous next →1,601–1,700 of 12,380, ranked by callers