aprender-serve 0.32.0

//! Check final hidden state

use realizar::gguf::{MappedGGUFModel, OwnedQKVWeights, OwnedQuantizedModel};
use realizar::quantize::{fused_q4k_parallel_matvec, fused_q6k_parallel_matvec};
use realizar::rms_norm;

const GGUF_TYPE_Q4_K: u32 = 12;
const GGUF_TYPE_Q6_K: u32 = 14;

fn l2_norm(v: &[f32]) -> f32 {
    (v.iter().map(|x| x * x).sum::<f32>()).sqrt()
}

fn silu(x: &mut [f32]) {
    for v in x.iter_mut() {
        *v = *v / (1.0 + (-*v).exp());
    }
}

fn fused_matmul(input: &[f32], data: &[u8], qtype: u32, in_dim: usize, out_dim: usize) -> Vec<f32> {
    match qtype {
        GGUF_TYPE_Q4_K => fused_q4k_parallel_matvec(data, input, in_dim, out_dim).expect("test"),
        GGUF_TYPE_Q6_K => fused_q6k_parallel_matvec(data, input, in_dim, out_dim).expect("test"),
        _ => panic!("Unsupported qtype"),
    }
}

fn main() {
    let path = "/tmp/parity-bench/tinyllama-1.1b-q4_k_m.gguf";
    let mapped = MappedGGUFModel::from_path(path).expect("Failed");
    let model = OwnedQuantizedModel::from_mapped(&mapped).expect("test");

    let hidden_dim = model.config().hidden_dim;
    let intermediate_dim = model.config().intermediate_dim;
    let eps = model.config().eps;

    // Token 450
    let start = 450 * hidden_dim;
    let mut hidden: Vec<f32> = model.token_embedding()[start..start + hidden_dim].to_vec();

    // Process all layers
    for layer_idx in 0..model.config().num_layers {
        let layer = &model.layers()[layer_idx];

        let normed = rms_norm(&hidden, &layer.attn_norm_weight, eps);

        let OwnedQKVWeights::Separate {
            q: q_weight,
            k: k_weight,
            v: v_weight,
        } = &layer.qkv_weight
        else {
            panic!("Expected separate")
        };

        let _q = fused_matmul(
            &normed,
            &q_weight.data,
            q_weight.qtype,
            q_weight.in_dim,
            q_weight.out_dim,
        );
        let _k = fused_matmul(
            &normed,
            &k_weight.data,
            k_weight.qtype,
            k_weight.in_dim,
            k_weight.out_dim,
        );
        let v = fused_matmul(
            &normed,
            &v_weight.data,
            v_weight.qtype,
            v_weight.in_dim,
            v_weight.out_dim,
        );

        let head_dim = hidden_dim / model.config().num_heads;
        let group_size = model.config().num_heads / model.config().num_kv_heads;
        let mut attn_out = Vec::with_capacity(hidden_dim);
        for h in 0..model.config().num_heads {
            let kv_head = h / group_size;
            let start = kv_head * head_dim;
            attn_out.extend_from_slice(&v[start..start + head_dim]);
        }

        let attn_proj = fused_matmul(
            &attn_out,
            &layer.attn_output_weight.data,
            layer.attn_output_weight.qtype,
            layer.attn_output_weight.in_dim,
            layer.attn_output_weight.out_dim,
        );

        for i in 0..hidden_dim {
            hidden[i] += attn_proj[i];
        }

        let ffn_input = if let Some(ref norm) = layer.ffn_norm_weight {
            rms_norm(&hidden, norm, eps)
        } else {
            hidden.clone()
        };

        if let Some(ref gate_weight) = layer.ffn_gate_weight {
            let ffn_up = fused_matmul(
                &ffn_input,
                &layer.ffn_up_weight.data,
                layer.ffn_up_weight.qtype,
                layer.ffn_up_weight.in_dim,
                layer.ffn_up_weight.out_dim,
            );

            let mut ffn_gate = fused_matmul(
                &ffn_input,
                &gate_weight.data,
                gate_weight.qtype,
                gate_weight.in_dim,
                gate_weight.out_dim,
            );
            silu(&mut ffn_gate);

            let mut ffn_hidden = vec![0.0f32; intermediate_dim];
            for i in 0..intermediate_dim {
                ffn_hidden[i] = ffn_gate[i] * ffn_up[i];
            }

            let ffn_out = fused_matmul(
                &ffn_hidden,
                &layer.ffn_down_weight.data,
                layer.ffn_down_weight.qtype,
                layer.ffn_down_weight.in_dim,
                layer.ffn_down_weight.out_dim,
            );

            for i in 0..hidden_dim {
                hidden[i] += ffn_out[i];
            }
        }
    }

    println!("Final hidden state (before norm):");
    println!("  L2: {:.4}", l2_norm(&hidden));
    println!("  First 5: {:?}", &hidden[..5]);

    println!("\nHuggingFace reference:");
    println!("  L2: 72.4048");
    println!("  First 5: [-0.488, -0.185, 1.411, -0.863, -0.380]");

    let final_hidden = rms_norm(&hidden, model.output_norm_weight(), eps);
    println!("\nAfter final norm:");
    println!("  L2: {:.4}", l2_norm(&final_hidden));

    println!("\nHuggingFace after norm:");
    println!("  L2: 90.5235");
}