Skip to main content

qualia_core_db/gguf_bridge/
projection_adapter.rs

1use super::*;
2use etalon_inference::projections::ProjectionProvider;
3impl ProjectionProvider for QTensorEngine {
4    type Index = crate::gguf_sharder::GgufTensorIndex;
5    type Mask = crate::neuro_symbolic_sieve::SieveStateMask;
6    fn logits(index: &Self::Index) -> Option<GgufTensorInfo> {
7        index.logits_projection_info().copied()
8    }
9    fn output_norm(index: &Self::Index) -> Option<GgufTensorInfo> {
10        index.output_norm_info().copied()
11    }
12    fn data_start(index: &Self::Index) -> u64 {
13        index.tensor_data_start
14    }
15    fn embedding_dim(index: &Self::Index) -> usize {
16        index.hyperparams.n_embd as usize
17    }
18    fn model_bytes(&self) -> Option<&[u8]> {
19        QTensorEngine::model_bytes(self)
20    }
21    fn nvfp4_output(&self) -> Option<(&[u8], &[u8], &[u8])> {
22        #[cfg(not(target_arch = "wasm32"))]
23        if let Some(pkg) = self.ftw_package.as_ref() {
24            return Some((
25                pkg.fetch_tensor_bytes("lm_head.weight")?,
26                pkg.fetch_tensor_bytes("lm_head.weight_scale")?,
27                pkg.fetch_tensor_bytes("lm_head.weight_global")?,
28            ));
29        }
30        None
31    }
32    fn fetch_tensor_raw_bytes(&self, info: &GgufTensorInfo) -> Option<&[u8]> {
33        QTensorEngine::fetch_tensor_raw_bytes(self, info)
34    }
35    fn gemm(
36        &self,
37        info: &GgufTensorInfo,
38        raw: &[u8],
39        input: &[f32],
40        out: &mut [f32],
41        n_in: usize,
42        n_out: usize,
43    ) -> bool {
44        self.dispatch_gemm_raw_into(info, raw, input, out, n_in, n_out)
45    }
46    fn allows(mask: &Self::Mask, token: u32) -> bool {
47        mask.allows(token)
48    }
49    fn scrub(&self, values: &mut [f32], n: usize) {
50        scrub_f32_volatile(values, n)
51    }
52}