Crate rlx_models_core 

Shared infrastructure for RLX model crates: HuggingFace config parsing, safetensors / GGUF weight loading, tier-1 compile profile helpers, and packed GGUF prefill guards (flow_bridge::packed_gguf_compile_guard, etc.).

Re-exports

pub use device_capabilities::STANDARD_DEVICE_NAMES;

pub use device_capabilities::STANDARD_DEVICES;

pub use device_capabilities::device_memory_for_moe_offload;

pub use device_capabilities::is_standard_device;

pub use device_capabilities::validate_sam_device;

pub use device_capabilities::validate_standard_device;

pub use gguf_config::DINOV2_GGUF_ARCHES;

pub use gguf_config::EMBED_GGUF_ARCHES;

pub use gguf_config::EmbedGgufKind;

pub use gguf_config::FLUX_GGUF_ARCHES;

pub use gguf_config::GgufMemoryFootprint;

pub use gguf_config::SAM_GGUF_ARCHES;

pub use gguf_config::SAM2_GGUF_ARCHES;

pub use gguf_config::SAM3_GGUF_ARCHES;

pub use gguf_config::VJEPA2_GGUF_ARCHES;

pub use gguf_config::W2V_BERT_GGUF_ARCHES;

pub use gguf_config::embed_gguf_kind;

pub use gguf_config::gguf_memory_footprint;

pub use gguf_config::gguf_meta_u32;

pub use gguf_config::gguf_runner_hint;

pub use gguf_config::is_dinov2_gguf_arch;

pub use gguf_config::is_embed_gguf_arch;

pub use gguf_config::is_flux_gguf_arch;

pub use gguf_config::is_sam_gguf_arch;

pub use gguf_config::is_sam2_gguf_arch;

pub use gguf_config::is_sam3_gguf_arch;

pub use gguf_config::is_vjepa2_gguf_arch;

pub use gguf_config::is_w2v_bert_gguf_arch;

pub use gguf_resolve::GgufTensorNameResolver;

pub use gguf_resolve::LlamaFamilyGgufResolver;

pub use gguf_resolve::PassThroughGgufResolver;

pub use gguf_resolve::PrefixStripGgufResolver;

pub use gguf_resolve::Qwen35NativeGgufResolver;

pub use gguf_resolve::register_gguf_tensor_resolver;

pub use gguf_resolve::resolve_gguf_tensor_name;

pub use gguf_support::GgufModelFamily;

pub use gguf_support::ResolveWeightsOptions;

pub use gguf_support::assert_gguf_family;

pub use gguf_support::gguf_architecture_from_path;

pub use gguf_support::gguf_architecture_str;

pub use gguf_support::gguf_f32_bytes_estimate;

pub use gguf_support::gguf_family_for_arch;

pub use gguf_support::gguf_safetensors_only_hint;

pub use gguf_support::gguf_split_hint;

pub use gguf_support::gguf_split_siblings;

pub use gguf_support::gguf_validate_arch;

pub use gguf_support::list_gguf_files_in_dir;

pub use gguf_support::load_gguf_file;

pub use gguf_support::resolve_weights_file;

pub use gguf_support::resolve_weights_file_with_options;

pub use autoregressive::compile_cache_ensure_graph;

pub use autoregressive::kv_from_prefill_outputs;

pub use autoregressive::past_kv_input_names;

pub use autoregressive::prefill_cache_key;

pub use autoregressive::run_bucketed_kv_decode;

pub use autoregressive::run_bucketed_kv_decode_hir;

pub use autoregressive::split_bucketed_decode_kv;

pub use autoregressive::split_decode_logits_kv;

pub use config::BertConfig;

pub use config::NomicBertConfig;

pub use config::NomicVisionConfig;

pub use flow_bridge::apply_compile_profile;

pub use flow_bridge::compile_graph_encoder;

pub use flow_bridge::compile_graph_gemma_decode;

pub use flow_bridge::compile_graph_gemma_prefill;

pub use flow_bridge::compile_graph_legacy;

pub use flow_bridge::compile_graph_llama32_decode;

pub use flow_bridge::compile_graph_llama32_prefill;

pub use flow_bridge::compile_graph_qwen3_decode;

pub use flow_bridge::compile_graph_qwen3_prefill;

pub use flow_bridge::compile_graph_qwen35_decode;

pub use flow_bridge::compile_graph_qwen35_prefill;

pub use flow_bridge::compile_graph_sam;

pub use flow_bridge::compile_graph_with_profile;

pub use flow_bridge::compile_options_for_packed_gguf_prefill;

pub use flow_bridge::compile_options_for_packed_gguf_prefill_with_profile;

pub use flow_bridge::compile_options_for_profile;

pub use flow_bridge::load_compile_profile;

pub use flow_bridge::packed_gguf_compile_guard;

pub use flow_bridge::packed_gguf_execution_device;

pub use flow_bridge::profile_near_weights;

pub use flow_util::WeightMapSource;

pub use flow_util::bucket_cache_ensure_built;

pub use flow_util::build_graph;

pub use flow_util::built_from_graph;

pub use flow_util::built_from_hir;

pub use flow_util::built_from_hir_with_profile;

pub use flow_util::compile_built;

pub use flow_util::compile_built_cpu;

pub use flow_util::compile_cache_ensure_built;

pub use flow_util::compile_graph_encoder_with_params;

pub use flow_util::compile_graph_gemma_decode_with_params;

pub use flow_util::compile_graph_gemma_prefill_with_params;

pub use flow_util::compile_graph_profile;

pub use flow_util::compile_graph_qwen3_prefill_with_params;

pub use flow_util::compile_graph_qwen35_decode_with_params;

pub use flow_util::compile_graph_qwen35_prefill_with_params;

pub use flow_util::compile_graph_sam_with_params;

pub use flow_util::compile_graph_with_kv_export_params;

pub use flow_util::graph_from_built;

pub use flow_util::graph_from_hir;

pub use gguf_resolve::ensure_builtin_resolvers;

pub use gguf_support::DEFAULT_GGUF_PREFER_SUBSTR;

pub use lm::FlowBuildExt;

pub use lm::into_compile_parts;

pub use weight_loader::GgufLoader;

pub use weight_loader::HfTranslatingLoader;

pub use weight_loader::WeightLoader;

pub use weight_loader::ggml_type_to_quant_scheme;

pub use weight_loader::gguf_to_hf_name;

pub use weight_loader::gguf_to_hf_name_for_arch;

pub use weight_loader::hf_to_gguf_name;

pub use weight_loader::is_mtp_weight;

pub use weight_loader::load_from_path;

pub use weight_map::WeightDrainPolicy;

pub use weight_map::WeightMap;

pub use weight_registry::LoadWeightsOptions;

pub use weight_registry::LoadedWeights;

pub use weight_registry::RegisteredFormat;

pub use weight_registry::WeightFormatRegistration;

pub use weight_registry::format_for_extension;

pub use weight_registry::list_registered_formats;

pub use weight_registry::load_weight_map_resolved;

pub use weight_registry::load_weights_resolved;

pub use weight_registry::open_weight_loader;

pub use weight_registry::register_weight_format;

pub use weight_registry::registered_extensions_hint;

pub use weights::GgufDirGuide;

pub use weights::LoadOpts;

pub use weights::ResolveOpts;

pub use weights::default_resolve_opts;

pub use weights::gguf_dir_guide;

pub use weights::init;

pub use weights::load_weight_map;

pub use weights::open;

pub use weights::open_map;

pub use weights::open_map_with;

pub use weights::open_with;

pub use weights::pick;

pub use weights::pick_default;

Modules

arch_registry

Architecture registry (plan #82).

autoregressive

Shared helpers for autoregressive decode loops (KV cache + bucketed compile cache).

config

Model configuration structs — parsed from HuggingFace config.json.

dataprocessing

Reusable batch-prep utilities (plan #83).

device_capabilities

Shared backend policy for RLX model crates.

flow_bridge

Bridge between rlx-models loaders/runtime and rlx-flow.

flow_util

Shared helpers for tier-0 model flow migration.

gguf_config

Read HuggingFace-shaped config fields from GGUF metadata ({arch}.* keys).

gguf_resolve

Pluggable GGUF tensor-name resolution per general.architecture.

gguf_support

Shared GGUF helpers for LM runners (architecture checks, path resolution).

lm

Shared causal-LM flow helpers — re-export tier-0 surface for model authors.

moe_weights

GGUF MoE expert-stack loader.

vision_ops_ir

Shared HIR builders for NCHW vision ops (Conv, ConvTranspose2d, LayerNorm2d, bias broadcast). Used by SAM / SAM2 / SAM3.

weight_loader

Pluggable weight loader trait (plan #56).

weight_map

Safetensors weight loading — standalone, no framework dependency.

weight_registry

Extensible weight-format registry — register custom loaders for new extensions.

weights

Model-agnostic weight I/O — paths, formats, drain policy only.

Structs

KvCacheState

Layer-wise past K/V tensors in row-major [past_len * kv_dim] layout per layer.