architectures
architectures : string []
attention_ k_ eq_ v
attention_k_eq_v : boolean
block_ size
block_size : number
confidence_ head_ with_ markov
confidence_head_with_markov : boolean
confidence_ threshold
confidence_threshold : number
enable_ confidence_ head
enable_confidence_head : boolean
final_ logit_ softcapping
final_logit_softcapping : number | null
global_ head_ dim
global_head_dim : number
hidden_ activation
hidden_activation : string
hidden_ size
hidden_size : number
intermediate_size : number
markov_ head_ type
markov_head_type : string
markov_ rank
markov_rank : number
mask_ token_ id
mask_token_id : number
num_ attention_ heads
num_attention_heads : number
num_ global_ key_ value_ heads
num_global_key_value_heads : number
num_ hidden_ layers
num_hidden_layers : number
num_ target_ layers
num_target_layers : number
partial_ rotary_ factor
partial_rotary_factor : number
rms_ norm_ eps
rms_norm_eps : number
rope_ theta
rope_theta : number
target_ layer_ ids
target_layer_ids : number []
vocab_ size
vocab_size : number
Not in the checkpoint's config.json (draft_ops.py takes it as a call arg to build_dspark_proposal); default 0 = confidence head disabled functionally even when weights exist (threshold<=0 ⇒ full γ block).