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 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/gemm/warp/mma_complex_tensor_op_tile_iterator_sm80.h:568

↓ 3 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/gemm/warp/mma_tensor_op_tile_iterator_sm70.h:1115

↓ 3 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/gemm/warp/mma_tensor_op_tile_iterator.h:2118

↓ 3 callersMethodset_mask

Sets the predicate mask, overriding value stored in predicate iterator

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:932

↓ 3 callersFunctionset_params_splitkv

src/cuda/flash_attn/fa2/flash_api_reference.cpp:299

↓ 3 callersFunctionshiftr

src/cuda/flash_attn/cutlass/cute/swizzle.hpp:349

↓ 3 callersFunctionsidecar_variant_dir

(model_path: str, target_profile: str, variant_name: str, mode: str)

python/krasis/hqq_self_calibrate.py:372

↓ 3 callersFunctionsilu_quantize_int16_avx2

( gate_inout: *mut f32, up: *const f32, hidden_int16: *mut i16, hidden_scales: *mut f32, n

src/kernel/avx2.rs:2310

↓ 3 callersFunctionsize

src/cuda/flash_attn/cutlass/cute/tensor_impl.hpp:550

↓ 3 callersFunctionslice_accumulator

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized_blockwise_scaling.hpp:557

↓ 3 callersFunctionslice_accumulator

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm100_mma_array_warpspecialized.hpp:465

↓ 3 callersFunctionslice_and_offset

src/cuda/flash_attn/cutlass/cute/layout.hpp:698

↓ 3 callersFunctionsoftplus

python/krasis/mamba2_ops/selective_state_update.py:21

↓ 3 callersFunctionsolve_hqq4_fixed_zero_rust

Rust shadow fixed-zero solve path for one HQQ4 group chunk.

python/krasis/attention_backend.py:1393

↓ 3 callersFunctionsource_slice_names

(source_meta: dict[str, Any])

tests/hqq_attention_diff.py:604

↓ 3 callersFunctionstore_with_byte_offset

Stores a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator.h:385

↓ 3 callersMethodstore_with_byte_offset

Store a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/predicated_tile_iterator.h:1483

↓ 3 callersFunctionstore_with_pointer_offset

src/cuda/flash_attn/cutlass/cutlass/epilogue/warp/tile_iterator_tensor_op.h:192

↓ 3 callersMethodstore_with_pointer_offset

Store a fragment to memory

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_iterator.h:899

↓ 3 callersMethodstride

Returns the layout object's stride vector

src/cuda/flash_attn/cutlass/cutlass/tensor_ref.h:291

↓ 3 callersFunctionsummarize_float_values

(values: list[float])

tests/hqq_attention_diff.py:3276

↓ 3 callersFunctionsymmetric_affine_params

(source: torch.Tensor)

tests/hqq_attention_diff.py:787

↓ 3 callersFunctionsynclog_print

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:996

↓ 3 callersFunctionsynclog_print_wgmma_desc

src/cuda/flash_attn/cutlass/cutlass/arch/synclog.hpp:286

↓ 3 callersMethodtensormaps_fence_acquire

src/cuda/flash_attn/cutlass/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized.hpp:751

↓ 3 callersFunctiontest_coherent_response

Test that a prompt produces a coherent (non-garbage) response.

tests/test_network.py:198

↓ 3 callersFunctiontest_multi_chunk

Test fused_marlin_moe with chunked expert processing (mimics GPU1 crash scenario).

tests/archive/test_marlin_gpu1.py:94

↓ 3 callersMethodthrfrg_C

src/cuda/flash_attn/cutlass/cute/atom/mma_atom.hpp:250

↓ 3 callersMethodthrottle_pipeline

src/cuda/flash_attn/cutlass/cutlass/gemm/kernel/sm90_tile_scheduler.hpp:63

↓ 3 callersFunctiontimed_phase

(label: &'static str, f: F)

↓ 3 callersMethodto_save_dict

Convert to dict for saving or launch config serialization.

python/krasis/launcher.py:827

↓ 3 callersFunctionty

tests/release_test.py:1544

↓ 3 callersFunctionumma_arrive

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:762

↓ 3 callersFunctionumma_arrive_multicast_2x1SM

UMMA arrive for MMA_2x1SM + TMA_LOAD_MULTICAST combination

src/cuda/flash_attn/cutlass/cutlass/arch/barrier.h:810

↓ 3 callersFunctionupdate_best_cuda

( chunk: torch.Tensor, zero_grid: torch.Tensor, scale_seed: torch.Tensor, qmax: float, ite

tests/hqq_gpu_search_probe.py:186

↓ 3 callersMethodvalid

Returns whether access is valid or not

src/cuda/flash_attn/cutlass/cutlass/transform/threadblock/ell_predicated_tile_access_iterator.h:946

↓ 3 callersFunctionvalidate_contracts

(reference_contract: Dict[str, Any], runtime_contract: Dict[str, Any])

tests/reference_contract.py:973

↓ 3 callersFunctionvalidate_image_only_messages

(messages: &serde_json::Value)

src/text_only_messages.rs:71

↓ 3 callersFunctionvariant

(name, row, effects)

tests/test_hqq_self_calibrate.py:1529

↓ 3 callersMethodwarmup_cuda_runtime

Trigger ALL lazy CUDA runtime allocations on ALL devices before HCS expert loading.

python/krasis/model.py:1388

↓ 3 callersFunctionwrite_quantized

(w: &mut W, q: &QuantWeight)

src/weights/mod.rs:5910

↓ 3 callersFunctionwrite_reference_case_boundary

Write non-production case span markers into the preserved server log.

tests/reference_test.py:286

↓ 3 callersMethodwrite_summary

Write a markdown summary table. Returns path to summary file.

python/krasis/suite.py:499

↓ 3 callersFunctionwrite_tokenizer_config

(template: &str)

src/chat_template.rs:417

↓ 3 callersFunctionwrite_tokenizer_config

(template: &str, name: &str)

tests/test_chat_template.rs:7

↓ 3 callersFunctionwrite_witness_summary

(path: Path, deltas, *, first_match=True, top_overlap=9)

tests/test_hqq_self_calibrate.py:406

↓ 2 callersFunctionLaneId

Computes laneId within a warp

src/cuda/flash_attn/cutlass/cutlass/arch/arch.h:50

↓ 2 callersFunctionLinearCombinationClampIsHeavy

Single source of truth for whether to unroll for `LinearCombinationClamp()`

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_clamp.h:56

↓ 2 callersFunctionLinearCombinationRelu0IsHeavy

Single source of truth for whether to unroll for `LinearCombinationClamp()`

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_relu0.h:60

↓ 2 callersFunctionLinearCombinationReluIsHeavy

Single source of truth for whether to unroll for `LinearCombinationClamp()`

src/cuda/flash_attn/cutlass/cutlass/epilogue/thread/linear_combination_relu.h:57

↓ 2 callersFunctionOutputTileThreadMapDesc

src/cuda/flash_attn/cutlass/cutlass/epilogue/threadblock/predicated_tile_iterator_params.h:126

↓ 2 callersFunctionPredicatedScaleBiasVectorAccessIterator

Construct a PredicatedTileAccessIterator with zero threadblock offset

src/cuda/flash_attn/cutlass/cutlass/conv/threadblock/predicated_scale_bias_vector_access_iterator.h:194

↓ 2 callersFunction_activation_evidence_policy

(layer: int, activation_tensor: str)

python/krasis/hqq_self_calibrate.py:1040

↓ 2 callersFunction_activation_fields_by_key

(evidence: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:1306

↓ 2 callersMethod_apply_rope

(x, positions, rotary_dim, rope_theta)

python/krasis/awq_calibrate.py:295

↓ 2 callersFunction_argv_has_option

(argv: list[str], option: str)

python/krasis/server.py:54

↓ 2 callersFunction_artifact_tensor_bytes

python/krasis/hqq_self_calibrate.py:159

↓ 2 callersFunction_artifact_tensor_bytes

python/krasis/attention_backend.py:495

↓ 2 callersFunction_artifact_weights_from_real_case

(real_case: dict)

tests/test_hqq_rust_quantizer.py:198

↓ 2 callersFunction_assert_heatmap_prompts_are_held_out

(prompts: list[str])

python/krasis/server.py:250

↓ 2 callersFunction_bf16_hex_to_floats

(hex_text: str, width: int)

python/krasis/hqq_self_calibrate.py:129

↓ 2 callersFunction_cache_mode_to_value

(cache_mode: str)

tests/generate_reference.py:441

↓ 2 callersFunction_check_p2p

Test if direct GPU-to-GPU transfer works.

python/krasis/model.py:138

↓ 2 callersMethod_chunked_inner

Nilpotent correction + recurrent chunk loop (shared by all chunked paths). Replaces ~600 kernel launches (63-iter nilpotent loop + 16-chunk r

python/krasis/linear_attention.py:590

↓ 2 callersFunction_cleanup_cuda

Release all CUDA contexts to prevent zombie GPU memory.

python/krasis/server.py:810

↓ 2 callersFunction_clear_screen

python/krasis/chat.py:161

↓ 2 callersFunction_collect_eos_ids

Collect all unique EOS token IDs from config.json and generation_config.json. generation_config.json is authoritative for stop tokens (often has

python/krasis/config.py:65

↓ 2 callersFunction_collect_eos_ids

(config_json: Dict[str, Any], generation_json: Dict[str, Any])

tests/reference_contract.py:236

↓ 2 callersFunction_count_total_prompts

Count total number of prompts across all conversations.

python/krasis/chat.py:899

↓ 2 callersFunction_cpu_device_warning

src/cuda/fla/utils.py:420

↓ 2 callersMethod_cycle_value

Cycle a config value left/right.

python/krasis/launcher.py:2089

↓ 2 callersFunction_decode_piece

(tokenizer: Any, token_id: int)

tests/reference_backfill_topk.py:314

↓ 2 callersFunction_detect_distro

Detect Linux distro for package manager selection. Returns (family, version_id) e.g. ("debian", "24.04") or ("rhel", "39").

python/krasis/setup.py:237

↓ 2 callersFunction_detect_gpu_vram_bytes

Auto-detect per-GPU VRAM in bytes via nvidia-smi.

python/krasis/vram_budget.py:86

↓ 2 callersMethod_discover_prefill_files

Discover available prefill_prompt_N files.

python/krasis/benchmark.py:309

↓ 2 callersFunction_divergence_window

( events: Sequence[Dict[str, Any]], index: int, *, radius: int, )

tests/trace_diff.py:1361

↓ 2 callersFunction_do_self_update

python/krasis/launcher.py:2915

↓ 2 callersFunction_entry_tuple_from_dict

(entry: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:4657

↓ 2 callersFunction_estimate_expert_ram_gb

Estimate system RAM needed for CPU expert weights in GB. Calculates per-expert byte sizes for the unified weight format: w13 (gate+up conca

python/krasis/model.py:314

↓ 2 callersFunction_estimate_message_tokens

Rough token estimate for a full message list.

python/krasis/chat.py:612

↓ 2 callersFunction_evidence_trace_path

(evidence_path: str, evidence: Dict[str, Any])

python/krasis/hqq_self_calibrate.py:1278

↓ 2 callersFunction_expected_heatmap_metadata

(model: KrasisModel, args, prompts: list[str])

python/krasis/server.py:278

↓ 2 callersFunction_expected_token_stats

(logits: Any, expected_token_id: int)

tests/reference_backfill_topk.py:412

↓ 2 callersFunction_expert_bytes_per_expert

Expert buffer size for Marlin INT4/INT8 on GPU.

python/krasis/vram_budget.py:277

↓ 2 callersFunction_family_packets

( events: Sequence[Dict[str, Any]], )

tests/trace_diff.py:1011

↓ 2 callersFunction_file_sha256

python/krasis/attention_backend.py:781

↓ 2 callersFunction_file_size

python/krasis/hf_downloader.py:302

↓ 2 callersFunction_find_nvidia_smi

Find nvidia-smi, including the WSL2 driver mount.

python/krasis/launcher.py:375

↓ 2 callersFunction_first_runtime_mismatch

(py_runtime: dict, rust_runtime: dict)

tests/test_hqq_rust_quantizer.py:1168

↓ 2 callersFunction_flush_range

(end_group_idx: int)

tests/test_hqq_rust_quantizer.py:858

↓ 2 callersMethod_format_hf_candidate

(self, candidate: Any)

python/krasis/launcher.py:1830

↓ 2 callersFunction_format_time

Format milliseconds as ms or s depending on magnitude.

python/krasis/chat.py:621

↓ 2 callersFunction_format_value

Format a config value for display.

python/krasis/launcher.py:961

↓ 2 callersMethod_forward_chunked

Chunked prefill: matches HF torch_chunk_gated_delta_rule. Uses the parallel-within-chunk, recurrent-across-chunks formulation.

python/krasis/linear_attention.py:692

↓ 2 callersFunction_generate_first_token_logits

( model: Any, tokenizer: Any, input_ids: Any, )

tests/reference_backfill_topk.py:437

↓ 2 callersFunction_get_chunk_step

Get the chunk step function (compiled if enabled, eager otherwise).

python/krasis/linear_attention.py:63

↓ 2 callersMethod_get_sorted_backends

Get backends sorted by priority (lower number = higher priority). Backends with the same priority are sorted by registration order.

src/cuda/fla/ops/backends/__init__.py:78

↓ 2 callersFunction_get_weight_tensor

Get the original BF16 weight tensor for an attention projection.

python/krasis/awq_calibrate.py:549

↓ 2 callersFunction_headline

Format a compact headline for server-prefixed benchmark output.

python/krasis/benchmark.py:42

↓ 2 callersMethod_hf_detail_screen

(self, candidate: Any)

python/krasis/launcher.py:1897

← previous next →1,401–1,500 of 12,380, ranked by callers