Expand description
inference category (reorg).
Re-exports§
pub use inference_agent as llm_agent;pub use inference_awq as llm_awq;pub use inference_bench as llm_bench;pub use inference_eval as llm_eval;pub use inference_gpu_profiler as llm_gpu_profiler;pub use inference_kernel_parity as llm_kernel_parity;pub use inference_modes::active_inference_mode;pub use inference_modes::apply_mode_toggles;pub use inference_modes::bootstrap_inference_mode;pub use inference_modes::fast_verify_html_default;pub use inference_modes::post_turn_verify_enabled;pub use inference_modes::prefer_tensor_core_gemm;pub use inference_modes::quant_graph_grounding_enabled;pub use inference_modes::rights_mode_enabled;pub use inference_modes::sentinel_mid_decode_enabled;pub use inference_modes::set_inference_mode;pub use inference_modes::InferenceMode;pub use application_profile::active_application_profile;pub use application_profile::apply_application_profile;pub use application_profile::bootstrap_application_profile;pub use application_profile::set_application_profile;pub use application_profile::ApplicationProfile;pub use post_turn_verify::maybe_verify_turn;pub use post_turn_verify::return_html_as_text;pub use post_turn_verify::verify_and_heal_turn;pub use post_turn_verify::VerifiedTurn;pub use post_turn_verify::VerifyCheck;pub use inference_path_selector::apply_inference_path_plan;pub use inference_path_selector::bootstrap_optimal_inference_path;pub use inference_path_selector::format_path_plan;pub use inference_path_selector::last_inference_path_plan;pub use inference_path_selector::path_auto_enabled;pub use inference_path_selector::resolve_inference_path_plan;pub use inference_path_selector::run_path_select_cli;pub use inference_path_selector::ComputeLane;pub use inference_path_selector::InferencePathPlan;pub use inference_path_selector::QuantProfile;pub use quant_graph_grounding::export_fact_quins;pub use quant_graph_grounding::fact_count;pub use quant_graph_grounding::ground_generation;pub use quant_graph_grounding::load_facts_from_tsv;pub use quant_graph_grounding::lookup_capital_object;pub use quant_graph_grounding::maybe_ground_generation;pub use quant_graph_grounding::register_capital_fact;pub use quant_graph_grounding::register_fact;pub use quant_graph_grounding::reset_fact_store_to_defaults;pub use quant_graph_grounding::seed_facts_from_bundled;pub use quant_graph_grounding::GroundingFact;pub use quant_graph_grounding::GroundingResult;pub use quant_graph_grounding::CTX_GROUNDING;pub use quant_graph_grounding::P_CAPITAL_OF;pub use qualia_hybrid::apply_graph_logit_bias;pub use qualia_hybrid::force_fact_tokens;pub use qualia_hybrid::graph_force_enabled;pub use qualia_hybrid::prepare_hybrid_decode;pub use qualia_hybrid::propose_best_draft;pub use qualia_hybrid::propose_fact_draft;pub use qualia_hybrid::publish_graph_route_from_prompt;pub use qualia_hybrid::publish_grounding_obligation;pub use qualia_hybrid::publish_prompt_query_tensor;pub use qualia_hybrid::GRAPH_LOGIT_BIAS;pub use cuda_lane_stub as cuda_lane;pub use speculative_decode::DominoMasker;pub use speculative_decode::GrammarState;pub use speculative_decode::GrammarStateMachine;pub use speculative_decode::TokenTrie;pub use speculative_decode::MAX_TRIE_DEPTH;pub use semantic_skills::Scratchpad;pub use semantic_skills::ScratchpadEntry;pub use semantic_skills::ScratchpadError;pub use semantic_skills::SearchResult;pub use semantic_skills::StoredVector;pub use semantic_skills::TextEmbedder;pub use semantic_skills::Vector;pub use semantic_skills::VectorStore;pub use semantic_skills::DEFAULT_SCRATCHPAD_TTL;pub use semantic_skills::EMBED_DIM;pub use semantic_skills::MAX_SCRATCHPAD_ENTRIES;pub use semantic_skills::MAX_VECTORS;pub use etalon_tensor_bridge::build_forward_arena_on;pub use etalon_tensor_bridge::build_fused_attention_on;pub use etalon_tensor_bridge::build_topk_pipeline_on;
Modules§
- agent
- Agent-identifier model (#16): resolve an identifier’s AGENT TYPE and check grounding.
- ambient_
orchestration - Ambient Sub-Threshold Orchestration Implementation
- application_
profile - Application profiles — how inference is used, not only which GPU path.
- compute_
universe - Qualia-native compute universe fabric (Track B2).
- conditioning
- Portable conditioning compiler (Prompt Precision P1C).
- conditioning_
eval - Cold, bounded evaluation API for prompt precision (Prompt Precision P5).
- conditioning_
opt - Model mapping and precision optimization pipeline.
- cuda_
lane_ stub - Stub when CUDA feature is off — CUDA lane APIs return false / no-op.
- domino_
gbnf - DOMINO GBNF Constrained LLM Decoding for VibeScript.
- etalon_
tensor_ bridge - Bridge between
etalon-tensoraccelerated operators and QualiaDB inference. - ggml_
quants - Canonical numerical encodings and descriptor helpers, with legacy Qualia paths.
- gguf_
sharder - Q-GGUF Hybrid Packaging
Parses monolithic
.gguffiles: vocabulary (KV section) and tensor names/offsets (tensor-info section) are extracted into native Rust types; multi-gigabyte tensor payloads are left on disk for direct VRAM mapping viagguf_bridge.rs. - inference_
agent - inference_
awq - W1/AWQ — activation-statistics capture for Activation-aware Weight Quantization (no external libs).
- inference_
bench - A0 — native LLM benchmark harness (STELLAR §A; decisions D17 + D22).
- inference_
eval - W1 — in-project quality oracle (no external libs): perplexity, KL-divergence, and a coherence ratio, measured by teacher-forcing an eval corpus through the engine and comparing a candidate model’s output distribution against a higher-fidelity reference’s.
- inference_
gpu_ profiler - Per-kernel GPU timing for the native LLM forward/decode path (W2 / D17).
- inference_
kernel_ parity - W3 — in-project GPU↔CPU kernel-parity oracle (no external LLM libs).
- inference_
modes - Multi-mode inference — keep parallel approaches, do not dump one pipeline.
- inference_
path_ selector - Device-optimal inference path selection.
- kv_
capture - W5b — KV-vector capture hook for sparse-dictionary calibration.
- kv_dict
- W5b Phase 4b — the KV-dictionary type and its sparse codec, in CORE.
- kv_
dict_ runtime - W5b Phase 4b — runtime KV-dictionary install + reconstruction, in CORE (engine-side, no forge dep).
- lab
- Inference superiority lab — instruments for evidence-based methodology.
- metal_
lane - Shared-device and host-policy adapter for the canonical Metal lane.
- moe
- Mixture-of-Experts (MoE) execution and offload engine (Work Packages F9, F10).
- neuro_
symbolic_ sieve - Grammar-constrained FSM sieve for neuro-symbolic LLM output (zero-heap hot path).
- operator_
converter - Compatibility forwarding to the canonical Etalon owner.
- operator_
package - Compatibility forwarding to the canonical Etalon owner.
- operator_
runtime - Compatibility forwarding to the canonical Etalon owner.
- orchestrator
- The Orchestration Sieve — LLM Sub-Agent Dispatch Layer
- paged_
kv - Compatibility exports for the production paged-KV runtime.
- post_
turn_ verify - Post-turn verification — “generate first, heal second”.
- prompt_
lookup - W6a — prompt-lookup (n-gram) speculative decoding: the proposer.
- qualia_
hybrid - Qualia-unique hybrid inference: graph + manifold + modes + deontic.
- quant_
graph_ grounding - QuantGraph mode — selective grounding / repair against a NQuin fact graph.
- qwen
- Qwen hybrid architecture execution and state management (Work Package F11).
- reranker
- Cross-encoder reranker for candidate re-scoring.
- residency_
planner - STELLAR §A AH-track H2 — residency + device-priority planner (decisions D24/D25/D30/D31).
- resident_
model - Process-wide resident GGUF mmap — released explicitly on model eviction.
- runtime
- Prepared native inference runtime.
- safetensor
- Phase 6 / task #12 — safetensor (+ MLX) source parsing + dtype gate for the streaming
transcoder (
p64_weight::transcode_safetensor_to_p64). - safetensor_
loader - Safetensor model loader — native zero-copy adoption of Safetensors models (including BF16).
- sampler
- W2 — exact CPU sampling chain for decode.
- semantic_
culler - Semantic Culler - Pre-GPU pipeline for agency-driven data filtering
- semantic_
skills - A9 — Semantic Skills: Vectors, Embeddings & Scratchpads.
- sparse_
cache - spatial_
sieve - speculative_
decode - A2 — Speculative Constrained Decoding (DOMINO).
- tensor_
roles - Canonical tensor identities and the preserved FFN precision selection policy.
- ternary
- BitNet b1.58 ternary quantization codec — STELLAR §A compression (task #12).
- ternary_
gpu - Task #12 / STELLAR §A — native GPU dispatch of the ternary GEMM + on-device parity.
- thermal_
telemetry - W7 — real GPU thermal/power telemetry + a detect-and-recommend thermal governor.
- thermal_
wal - topk
- topk_
gpu - STELLAR §A A1a — native GPU dispatch of the top-K reduction + on-device parity.
Structs§
- Attention
Params - Legacy attention configuration. Use
FusedAttentionParamsfor GPU uniform bindings. - Cuda
Mega Pass Executor - Scalar transformer reference executor with a resident KV cache.
- Direct
MlCompute Lane - DirectML / DX12 compute executor lane supporting both tiled host execution and Etalon GPU compute pipeline dispatch.
- DmlGemm
Op - DirectML 2D matrix multiplication operator ($C = A_{m \times k} \times B_{k \times n}$).
- DmlGemm
Params - Forward
Arena - Unified forward execution arena on an Etalon device.
- Fused
Attention Pipeline - GPU compute pipeline for fused FlashAttention dispatch.
- Logit
Mask Header - 80-byte std140-aligned logit mask header containing the 512-bit sieve bitmask and autoregressive sampling hyper-parameters.
- Paged
KvBlock Table - Tensor
View - Topk
Params - Uniform parameters matching the WGSL
Paramslayout intopk_reduction.wgsl. - Topk
Reduction Pipeline - Complete GPU compute pipeline for block top-k / argmax reduction.
Enums§
- Forward
Mode - Operational mode of the forward execution arena.
- KvCache
Layout - Quant
Scheme
Constants§
- FUSED_
ATTENTION_ WGSL - Raw WGSL shader source for the fused FlashAttention kernel with guarded all-masked row handling.
- TOPK_
REDUCTION_ WGSL - Raw WGSL shader source for the GPU block top-k reduction kernel.
Functions§
- online_
softmax_ rescale - Numerical online-softmax rescale calculation with guarded all-masked row handling.