Skip to main content

qualia_core_db/inference/qwen/
streamed_moe.rs

1//! Host tile adapter for the canonical Etalon streamed MoE engine.
2use crate::gguf_sharder::LayerTensors;
3use super::{TrunkNvmeReader, QwenExpertTileReader};
4pub use etalon_inference::qwen::streamed_moe::{execute_streamed_moe, select_top_experts, normalize_top_experts, StreamedMoeError, QWEN4EXP_EXPERT_COUNT, QWEN4EXP_TOP_EXPERTS};
5#[allow(clippy::too_many_arguments)]
6pub fn execute_streamed_moe_with_tiles(
7    reader: &mut TrunkNvmeReader,
8    tensors: &LayerTensors,
9    input: &[f32],
10    out: &mut [f32],
11    router_logits: &mut [f32],
12    top_indices: &mut [usize; QWEN4EXP_TOP_EXPERTS],
13    top_logits: &mut [f32; QWEN4EXP_TOP_EXPERTS],
14    top_weights: &mut [f32; QWEN4EXP_TOP_EXPERTS],
15    cached_tiles: &[Option<&QwenExpertTileReader>; QWEN4EXP_TOP_EXPERTS],
16    gate: &mut [f32],
17    up: &mut [f32],
18    activation: &mut [f32],
19    expert_out: &mut [f32],
20    raw_row: &mut [u8],
21    dequantized_row: &mut [f32],
22) -> Result<(), StreamedMoeError> {
23    let cached = cached_tiles.map(|tile| tile.map(|reader| &**reader));
24    etalon_inference::qwen::execute_streamed_moe_with_tiles(reader, tensors, input, out, router_logits, top_indices, top_logits, top_weights, &cached, gate, up, activation, expert_out, raw_row, dequantized_row)
25}