qualia_core_db/gguf_bridge/qwen35/ftw/
mod.rs1pub mod admission;
4pub mod attention;
5mod bridge;
6pub mod dense_host;
7pub(crate) mod nvfp4_rows;
8pub mod session;
9pub(crate) mod session_ffn;
10pub mod weights;
11
12use crate::inference::moe::ftw_loader::FtwModelPackage;
13use etalon_model::inference::{
14 qwen_attention::{QkNorm, Qwen35AttentionGeometry},
15 qwen_roles::{Qwen35Geometry, Qwen35Packing},
16 TensorInfo,
17};
18use std::sync::Arc;
19use weights::Nvfp4Projection;
20
21#[derive(Clone, Debug)]
22pub struct Profile {
23 pub(crate) hidden: usize,
24 pub(crate) layers: usize,
25 pub(crate) vocab: usize,
26 pub(crate) intermediate: usize,
27 pub(crate) experts: usize,
28 pub(crate) topk: usize,
29 pub(crate) max_context: usize,
30 pub(crate) epsilon: f32,
31 pub(crate) linear_geometry: Qwen35Geometry,
32 pub(crate) attention_geometry: Qwen35AttentionGeometry,
33 pub(crate) rotary_inv_freq: Vec<f32>,
34 pub(crate) packing: Qwen35Packing,
35 pub(crate) ordinary_norm: QkNorm,
36}
37
38pub struct LayerPlan {
39 pub(crate) attention: attention::AttentionPlan,
40 pub(crate) input_norm: TensorInfo,
41 pub(crate) post_norm: TensorInfo,
42 pub(crate) router: TensorInfo,
43 pub(crate) shared_gate: TensorInfo,
44 pub(crate) shared_up: Nvfp4Projection,
45 pub(crate) shared_down: Nvfp4Projection,
46}
47
48pub struct Plan {
49 pub(crate) pkg: Arc<FtwModelPackage>,
50 pub(crate) profile: Profile,
51 pub(crate) layers: Vec<LayerPlan>,
52 pub(crate) output: Nvfp4Projection,
53 pub(crate) final_norm: TensorInfo,
54 pub(crate) embedding: TensorInfo,
55 pub(crate) reserved_bytes: usize,
56}
57impl Plan {
58 pub(crate) fn reserved_bytes(&self) -> usize {
59 self.reserved_bytes
60 }
61}