Struct LayerTensors
pub struct LayerTensors {Show 59 fields
pub layer_idx: u32,
pub attn_norm: Option<TensorInfo>,
pub attn_q: Option<TensorInfo>,
pub attn_k: Option<TensorInfo>,
pub attn_v: Option<TensorInfo>,
pub attn_q_norm: Option<TensorInfo>,
pub attn_k_norm: Option<TensorInfo>,
pub attn_qkv: Option<TensorInfo>,
pub attn_gate: Option<TensorInfo>,
pub attn_output: Option<TensorInfo>,
pub ffn_norm: Option<TensorInfo>,
pub ffn_gate: Option<TensorInfo>,
pub ffn_up: Option<TensorInfo>,
pub ffn_down: Option<TensorInfo>,
pub moe_router: Option<TensorInfo>,
pub moe_gate_exps: Option<TensorInfo>,
pub moe_up_exps: Option<TensorInfo>,
pub moe_gate_up_exps: Option<TensorInfo>,
pub moe_down_exps: Option<TensorInfo>,
pub moe_shared_gate: Option<TensorInfo>,
pub moe_shared_up: Option<TensorInfo>,
pub moe_shared_down: Option<TensorInfo>,
pub moe_shared_gate_input: Option<TensorInfo>,
pub ssm_alpha: Option<TensorInfo>,
pub ssm_beta: Option<TensorInfo>,
pub ssm_in: Option<TensorInfo>,
pub ssm_conv1d: Option<TensorInfo>,
pub ssm_conv1d_bias: Option<TensorInfo>,
pub ssm_dt_bias: Option<TensorInfo>,
pub ssm_a: Option<TensorInfo>,
pub ssm_norm: Option<TensorInfo>,
pub ssm_out: Option<TensorInfo>,
pub hc_attn_norm: Option<TensorInfo>,
pub hc_attn_down: Option<TensorInfo>,
pub hc_attn_up: Option<TensorInfo>,
pub hc_attn_inject: Option<TensorInfo>,
pub hc_ffn_norm: Option<TensorInfo>,
pub hc_ffn_down: Option<TensorInfo>,
pub hc_ffn_up: Option<TensorInfo>,
pub hc_ffn_inject: Option<TensorInfo>,
pub ple_key: Option<TensorInfo>,
pub ple_value: Option<TensorInfo>,
pub ple_norm_key: Option<TensorInfo>,
pub ple_norm_query: Option<TensorInfo>,
pub ple_norm_conv: Option<TensorInfo>,
pub ple_conv1d: Option<TensorInfo>,
pub indexer_q: Option<TensorInfo>,
pub indexer_k: Option<TensorInfo>,
pub indexer_q_norm: Option<TensorInfo>,
pub indexer_k_norm: Option<TensorInfo>,
pub post_attention_norm: Option<TensorInfo>,
pub post_ffw_norm: Option<TensorInfo>,
pub post_ffw_norm_1: Option<TensorInfo>,
pub post_ffw_norm_2: Option<TensorInfo>,
pub pre_ffw_norm_2: Option<TensorInfo>,
pub ple_inp_gate: Option<TensorInfo>,
pub ple_proj: Option<TensorInfo>,
pub ple_post_norm: Option<TensorInfo>,
pub layer_output_scale: Option<TensorInfo>,
}Expand description
Per-layer transformer weight metadata (all Option — absent tensors skipped).
Fields§
§layer_idx: u32§attn_norm: Option<TensorInfo>§attn_q: Option<TensorInfo>§attn_k: Option<TensorInfo>§attn_v: Option<TensorInfo>§attn_q_norm: Option<TensorInfo>Per-head Qwen4Exp QSA RMSNorm weights (256 values = 2×128 heads).
attn_k_norm: Option<TensorInfo>§attn_qkv: Option<TensorInfo>Qwen hybrid models store Q/K/V as one row-concatenated matrix.
attn_gate: Option<TensorInfo>Qwen gated-attention scalar/vector gate.
attn_output: Option<TensorInfo>§ffn_norm: Option<TensorInfo>§ffn_gate: Option<TensorInfo>§ffn_up: Option<TensorInfo>§ffn_down: Option<TensorInfo>§moe_router: Option<TensorInfo>§moe_gate_exps: Option<TensorInfo>§moe_up_exps: Option<TensorInfo>§moe_gate_up_exps: Option<TensorInfo>Gemma MoE packs gate and up rows consecutively in one expert tensor.
moe_down_exps: Option<TensorInfo>Scalar gate applied to the shared-expert output in Qwen-style MoE.
ssm_alpha: Option<TensorInfo>§ssm_beta: Option<TensorInfo>§ssm_in: Option<TensorInfo>Granite hybrid input projection; Qwen uses attn_qkv instead.
ssm_conv1d: Option<TensorInfo>§ssm_conv1d_bias: Option<TensorInfo>§ssm_dt_bias: Option<TensorInfo>§ssm_a: Option<TensorInfo>Per-head learned negative decay used by Qwen4Exp GatedDeltaNet.
Unlike the other SSM tensors this is a vector named ssm_a.
ssm_norm: Option<TensorInfo>§ssm_out: Option<TensorInfo>§hc_attn_norm: Option<TensorInfo>§hc_attn_down: Option<TensorInfo>§hc_attn_up: Option<TensorInfo>§hc_attn_inject: Option<TensorInfo>§hc_ffn_norm: Option<TensorInfo>§hc_ffn_down: Option<TensorInfo>§hc_ffn_up: Option<TensorInfo>§hc_ffn_inject: Option<TensorInfo>§ple_key: Option<TensorInfo>§ple_value: Option<TensorInfo>§ple_norm_key: Option<TensorInfo>§ple_norm_query: Option<TensorInfo>§ple_norm_conv: Option<TensorInfo>§ple_conv1d: Option<TensorInfo>§indexer_q: Option<TensorInfo>§indexer_k: Option<TensorInfo>§indexer_q_norm: Option<TensorInfo>§indexer_k_norm: Option<TensorInfo>§post_attention_norm: Option<TensorInfo>§post_ffw_norm: Option<TensorInfo>§post_ffw_norm_1: Option<TensorInfo>§post_ffw_norm_2: Option<TensorInfo>§pre_ffw_norm_2: Option<TensorInfo>§ple_inp_gate: Option<TensorInfo>§ple_proj: Option<TensorInfo>§ple_post_norm: Option<TensorInfo>§layer_output_scale: Option<TensorInfo>Implementations§
§impl LayerTensors
impl LayerTensors
pub fn is_hybrid_ssm_layer(&self) -> bool
pub fn is_hybrid_ssm_layer(&self) -> bool
A Qwen hybrid SSM layer is identified by its required recurrent tensors, not merely by a
fused attn_qkv name. The latter must not be routed through scaled-dot-product attention.
pub fn has_full_attention_qkv(&self) -> bool
pub fn has_full_attention_qkv(&self) -> bool
Full attention requires independently projected Q/K/V matrices. Qwen’s SSM qkv matrix intentionally does not satisfy this predicate.
pub fn has_qwen_sparse_attention(&self) -> bool
pub fn has_qwen_sparse_attention(&self) -> bool
QSA is executable only when all indexer projections/norms accompany a full attention layer. Do not silently replace it with dense attention.
pub fn has_qwen_hyper_connection(&self, ffn: bool) -> bool
pub fn has_qwen_hyper_connection(&self, ffn: bool) -> bool
The Gated Residual block needs all four learned paths for the specified sub-block. Absence is an unsupported graph, not a harmless residual.
pub fn has_qwen_ple(&self) -> bool
pub fn has_qwen_ple(&self) -> bool
A layer that receives PLE must supply every projection/state tensor.
pub fn qkv_views(
&self,
q_rows: usize,
kv_rows: usize,
) -> Option<(TensorInfo, TensorInfo, TensorInfo)>
pub fn qkv_views( &self, q_rows: usize, kv_rows: usize, ) -> Option<(TensorInfo, TensorInfo, TensorInfo)>
Resolve separate Q/K/V projections or split the row-concatenated Qwen fused-QKV tensor.
The row layout is [Q; K; V], with q_rows = n_head * head_dim and each KV projection
occupying kv_rows = n_kv_head * head_dim. Returning copies of metadata keeps the
execution ABI zero-copy: the weight bytes stay in the mapped GGUF file.
Trait Implementations§
§impl Clone for LayerTensors
impl Clone for LayerTensors
§fn clone(&self) -> LayerTensors
fn clone(&self) -> LayerTensors
1.0.0 (const: unstable) · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read moreimpl Copy for LayerTensors
§impl Debug for LayerTensors
impl Debug for LayerTensors
§impl Default for LayerTensors
impl Default for LayerTensors
§fn default() -> LayerTensors
fn default() -> LayerTensors
impl Eq for LayerTensors
§impl PartialEq for LayerTensors
impl PartialEq for LayerTensors
impl StructuralPartialEq for LayerTensors
Auto Trait Implementations§
impl Freeze for LayerTensors
impl RefUnwindSafe for LayerTensors
impl Send for LayerTensors
impl Sync for LayerTensors
impl Unpin for LayerTensors
impl UnsafeUnpin for LayerTensors
impl UnwindSafe for LayerTensors
Blanket Implementations§
§impl<S, A> Aggregate<Result<S, Error>> for Awhere
A: Aggregate<S>,
impl<S, A> Aggregate<Result<S, Error>> for Awhere
A: Aggregate<S>,
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
impl<ST, DT> CastableFrom<ST, Initialized, Initialized> for DT
impl<ST, DT> CastableFrom<ST, Uninit, Uninit> for DT
Source§impl<T> CloneToUninit for Twhere
T: Clone,
impl<T> CloneToUninit for Twhere
T: Clone,
§impl<Q, K> Equivalent<K> for Q
impl<Q, K> Equivalent<K> for Q
§fn equivalent(&self, key: &K) -> bool
fn equivalent(&self, key: &K) -> bool
§impl<Q, K> Equivalent<K> for Q
impl<Q, K> Equivalent<K> for Q
§fn equivalent(&self, key: &K) -> bool
fn equivalent(&self, key: &K) -> bool
key and return true if they are equal.§impl<Q, K> Equivalent<K> for Q
impl<Q, K> Equivalent<K> for Q
§fn equivalent(&self, key: &K) -> bool
fn equivalent(&self, key: &K) -> bool
§impl<T> Instrument for T
impl<T> Instrument for T
§fn instrument(self, span: Span) -> Instrumented<Self> ⓘ
fn instrument(self, span: Span) -> Instrumented<Self> ⓘ
Source§impl<T> IntoEither for T
impl<T> IntoEither for T
Source§fn into_either(self, into_left: bool) -> Either<Self, Self> ⓘ
fn into_either(self, into_left: bool) -> Either<Self, Self> ⓘ
self into a Left variant of Either<Self, Self>
if into_left is true.
Converts self into a Right variant of Either<Self, Self>
otherwise. Read moreSource§fn into_either_with<F>(self, into_left: F) -> Either<Self, Self> ⓘ
fn into_either_with<F>(self, into_left: F) -> Either<Self, Self> ⓘ
self into a Left variant of Either<Self, Self>
if into_left(&self) returns true.
Converts self into a Right variant of Either<Self, Self>
otherwise. Read more