Module config 

Structs

GgufConfigInputs

Inputs to config_from_gguf: the metadata accessor plus two facts the builder needs from the tensor section (not the metadata KV block).

ModelConfig

Model configuration parsed from HuggingFace config.json.

PeftAdapterConfig

Parsed subset of a PEFT adapter_config.json that Atlas consumes.

QuantizationConfig

Advertised weight-quantization layout, as declared in the HF config.json’s quantization_config block (or a sibling hf_quant_config.json). This is the authoritative signal for format dispatch — the QuantFormat trait prefers this over tensor-name sniffing, matching the dispatch model used by vLLM / TensorRT-LLM / SGLang.

VisionConfig

Vision encoder configuration for Qwen3-VL models.

Enums

Glm5NextRouterMode

Which dtype ladder GLM-5.3’s MoE router runs in.

LayerType

Layer type in a hybrid transformer model.

Constants

PEFT_SUPPORTED_TARGET_MODULES

v0 target-module allow-list. Deltas apply on full-attention layers (holo-3.1-0.8b: layer indices 3,7,11,15,19,23) plus the dense SwiGLU FFN. q_proj IS supported: on attn_output_gate=true models the raw projection emits the interleaved [Q|gate] at width 2·q_heads·head_dim — the FULL width the PEFT lora_B was trained against (verified [8192,16] on holo-3.1-35b), so the delta folds onto the raw interleaved basis exactly like k/v/o (the deinterleave is deferred past the fold). GDN/linear-attention modules stay rejected (no exact-replay parity harness for the recurrence yet).

Traits

GgufMeta

Typed read access to GGUF metadata. Implemented by the spark-runtime GGUF parser over its parsed key/value table. All getters return None when the key is absent or holds a different value type — the builder decides whether absence is fatal or has a derivation rule.

Functions

allow_partial_targets

Per-module-name allow-list gate. PEFT entries may be bare names ("k_proj") or full paths ("model.layers.3.self_attn.k_proj"); both validate on the final .-segment. Per-LayerType enforcement (deltas land on full-attention layers only) is the weight loader’s job — this is the name-level gate. ATLAS_LORA_ALLOW_PARTIAL=1 — load an adapter naming target modules Atlas cannot apply, skipping those and applying the rest.

config_from_gguf

Build a validated ModelConfig from GGUF metadata.

glm5_next_mtp_layer_index

num_hidden_layers counts text layers only; the MTP layer sits at index num_hidden_layers (45) and is NOT included in that count.

parse_config

Parse a checkpoint config.json into a ModelConfig.

parse_mistral_params

parse_peft_adapter_config

Parse a PEFT adapter_config.json payload.

parse_quantization_config