rlx-sam3 0.2.5

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! IR-lowered detector decoder.
//!
//! The per-layer compute (self-attn, text cross-attn, image cross-attn
//! with explicit boxRPB add, FFN) is expressed as native `HirModule` +
//! `HirMut`, compiled once per layer on the requested device. Box-refinement /
//! sineembed / presence-token concatenation stay in Rust because they're
//! iterative and small.

use super::detector_decoder::{
    Mlp2, Mlp3, Sam3DecoderLayerWeights, Sam3DecoderOutput, Sam3DecoderWeights, mlp2_forward,
    mlp2_forward_into, mlp3_forward, mlp3_forward_into,
};
use super::packed_gguf::packed_linear;
use anyhow::{Result, ensure};
use rlx_flow::CompileProfile;
use rlx_flow::{GgufPackedLinear, GgufPackedParams};
use rlx_ir::hir::{HirGraphExt, HirModule, HirMut, HirNodeId};
use rlx_ir::op::{Activation, MaskKind, Op};
use rlx_ir::shape;
use rlx_ir::{DType, Shape};
use rlx_runtime::{CompiledGraph, Device};
use std::collections::HashMap;

const D_MODEL: usize = 256;
const DIM_FF: usize = 2048;
const N_HEADS: usize = 8;
const HEAD_DIM: usize = D_MODEL / N_HEADS;
const NUM_QUERIES: usize = 200;
const N_LAYERS: usize = 6;

type LayerHirParts = (
    HirModule,
    HashMap<String, Vec<f32>>,
    Vec<(String, Vec<u8>, DType)>,
);
type LayerRunOut = (Vec<f32>, Vec<f32>, Vec<f32>, Vec<f32>);

fn dec_layer_key(base: &str, li: usize, suffix: &str) -> String {
    format!("{base}.layers.{li}.{suffix}")
}

fn gguf_weight_param(
    g: &mut HirMut<'_>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    ir_name: &str,
    p: &GgufPackedLinear,
) -> HirNodeId {
    if let Some(&id) = cache.get(ir_name) {
        return id;
    }
    let id = g.param(ir_name, Shape::new(&[p.w_q.len()], DType::U8));
    typed.push((ir_name.to_string(), p.w_q.clone(), DType::U8));
    cache.insert(ir_name.to_string(), id);
    id
}

fn linear_gguf_matmul(
    g: &mut HirMut<'_>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    ir_stem: &str,
    p: &GgufPackedLinear,
    input: HirNodeId,
    in_dim: usize,
    out_dim: usize,
) -> Result<HirNodeId> {
    ensure!(
        p.in_dim == in_dim && p.out_dim == out_dim,
        "packed linear {ir_stem}: shape {}x{} vs {in_dim}x{out_dim}",
        p.in_dim,
        p.out_dim
    );
    let w_name = format!("{ir_stem}.w");
    let w_id = gguf_weight_param(g, typed, cache, &w_name, p);
    let cur = g.shape(input);
    let mut dims: Vec<usize> = cur.dims().iter().map(|d| d.unwrap_static()).collect();
    *dims.last_mut().unwrap() = out_dim;
    let out_shape = Shape::new(&dims, DType::F32);
    Ok(g.add_node(
        Op::DequantMatMul { scheme: p.scheme },
        vec![input, w_id],
        out_shape,
    ))
}

fn add_f32_bias(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    name: &str,
    input: HirNodeId,
    bias: &[f32],
) -> HirNodeId {
    if bias.iter().all(|&v| v == 0.0) {
        return input;
    }
    let out_dim = bias.len();
    let b_id = add_param(
        g,
        params,
        name,
        bias.to_vec(),
        Shape::new(&[out_dim], DType::F32),
    );
    g.add(input, b_id)
}

fn linear_gguf_bias(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    ir_stem: &str,
    p: &GgufPackedLinear,
    input: HirNodeId,
    bias: &[f32],
    in_dim: usize,
    out_dim: usize,
) -> Result<HirNodeId> {
    let y = linear_gguf_matmul(g, typed, cache, ir_stem, p, input, in_dim, out_dim)?;
    Ok(add_f32_bias(g, params, &format!("{ir_stem}.b"), y, bias))
}

fn in_proj_qkv(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    gguf_packed: Option<&GgufPackedParams>,
    gguf_key: &str,
    ir_stem: &str,
    layer_w_t: &[f32],
    layer_b: &[f32],
    input_q: HirNodeId,
    input_k: HirNodeId,
    input_v: HirNodeId,
    d: usize,
) -> Result<(HirNodeId, HirNodeId, HirNodeId)> {
    if let Some(p) = gguf_packed.and_then(|m| packed_linear(m, gguf_key)) {
        let qkv_q = linear_gguf_bias(
            g,
            params,
            typed,
            cache,
            ir_stem,
            p,
            input_q,
            layer_b,
            d,
            3 * d,
        )?;
        let qkv_k = linear_gguf_bias(
            g,
            params,
            typed,
            cache,
            ir_stem,
            p,
            input_k,
            layer_b,
            d,
            3 * d,
        )?;
        let qkv_v = linear_gguf_bias(
            g,
            params,
            typed,
            cache,
            ir_stem,
            p,
            input_v,
            layer_b,
            d,
            3 * d,
        )?;
        let axis = g.shape(qkv_q).rank().saturating_sub(1);
        let q = g.narrow_(qkv_q, axis, 0, d);
        let k = g.narrow_(qkv_k, axis, d, d);
        let v = g.narrow_(qkv_v, axis, 2 * d, d);
        return Ok((q, k, v));
    }
    let (wq, wk, wv) = split_qkv(layer_w_t, d);
    let bq = layer_b[0..d].to_vec();
    let bk = layer_b[d..2 * d].to_vec();
    let bv = layer_b[2 * d..3 * d].to_vec();
    let batch_q = g.shape(input_q).dims()[0].unwrap_static();
    let seq_q = g.shape(input_q).dims()[1].unwrap_static();
    let batch_k = g.shape(input_k).dims()[0].unwrap_static();
    let seq_k = g.shape(input_k).dims()[1].unwrap_static();
    let batch_v = g.shape(input_v).dims()[0].unwrap_static();
    let seq_v = g.shape(input_v).dims()[1].unwrap_static();
    let q = linear_bias_shaped(
        g,
        params,
        &format!("{ir_stem}.q"),
        input_q,
        wq,
        bq,
        d,
        d,
        Some(batch_q),
        Some(seq_q),
    );
    let k = linear_bias_shaped(
        g,
        params,
        &format!("{ir_stem}.k"),
        input_k,
        wk,
        bk,
        d,
        d,
        Some(batch_k),
        Some(seq_k),
    );
    let v = linear_bias_shaped(
        g,
        params,
        &format!("{ir_stem}.v"),
        input_v,
        wv,
        bv,
        d,
        d,
        Some(batch_v),
        Some(seq_v),
    );
    Ok((q, k, v))
}

fn linear_fused_or_gguf(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    gguf_packed: Option<&GgufPackedParams>,
    gguf_key: &str,
    ir_stem: &str,
    input: HirNodeId,
    w_t: Vec<f32>,
    bias: Vec<f32>,
    in_dim: usize,
    out_dim: usize,
) -> Result<HirNodeId> {
    if let Some(p) = gguf_packed.and_then(|m| packed_linear(m, gguf_key)) {
        return linear_gguf_bias(
            g, params, typed, cache, ir_stem, p, input, &bias, in_dim, out_dim,
        );
    }
    Ok(linear_bias(
        g, params, ir_stem, input, w_t, bias, in_dim, out_dim,
    ))
}

fn split_qkv(w_t: &[f32], e: usize) -> (Vec<f32>, Vec<f32>, Vec<f32>) {
    let mut wq = vec![0f32; e * e];
    let mut wk = vec![0f32; e * e];
    let mut wv = vec![0f32; e * e];
    for i in 0..e {
        for j in 0..e {
            wq[i * e + j] = w_t[i * 3 * e + j];
            wk[i * e + j] = w_t[i * 3 * e + e + j];
            wv[i * e + j] = w_t[i * 3 * e + 2 * e + j];
        }
    }
    (wq, wk, wv)
}

fn add_param(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    name: &str,
    data: Vec<f32>,
    shape: Shape,
) -> HirNodeId {
    let id = g.param(name, shape);
    params.insert(name.to_string(), data);
    id
}

fn linear_bias(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    name: &str,
    input: HirNodeId,
    w: Vec<f32>,
    b: Vec<f32>,
    in_dim: usize,
    out_dim: usize,
) -> HirNodeId {
    linear_bias_shaped(g, params, name, input, w, b, in_dim, out_dim, None, None)
}

fn linear_bias_shaped(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    name: &str,
    input: HirNodeId,
    w: Vec<f32>,
    b: Vec<f32>,
    in_dim: usize,
    out_dim: usize,
    batch: Option<usize>,
    seq: Option<usize>,
) -> HirNodeId {
    let f = DType::F32;
    let w_id = add_param(
        g,
        params,
        &format!("{name}.w"),
        w,
        Shape::new(&[in_dim, out_dim], f),
    );
    let b_id = add_param(
        g,
        params,
        &format!("{name}.b"),
        b,
        Shape::new(&[out_dim], f),
    );
    let out_shape = if let (Some(batch), Some(seq)) = (batch, seq) {
        Shape::new(&[batch, seq, out_dim], f)
    } else {
        let cur = g.shape(input);
        let mut out_dims: Vec<usize> = cur.dims().iter().map(|d| d.unwrap_static()).collect();
        *out_dims.last_mut().unwrap() = out_dim;
        Shape::new(&out_dims, f)
    };
    g.add_node(
        Op::FusedMatMulBiasAct { activation: None },
        vec![input, w_id, b_id],
        out_shape,
    )
}

fn fused_matmul_bias_act(
    g: &mut HirMut<'_>,
    input: HirNodeId,
    w: HirNodeId,
    b: HirNodeId,
    activation: Option<Activation>,
    out_shape: Shape,
) -> HirNodeId {
    g.add_node(
        Op::FusedMatMulBiasAct { activation },
        vec![input, w, b],
        out_shape,
    )
}

fn attention_bias(
    g: &mut HirMut<'_>,
    q: HirNodeId,
    k: HirNodeId,
    v: HirNodeId,
    bias: HirNodeId,
    num_heads: usize,
    head_dim: usize,
) -> HirNodeId {
    let attn_shape = shape::attention_shape(g.shape(q));
    g.add_node(
        Op::Attention {
            num_heads,
            head_dim,
            mask_kind: MaskKind::Bias,
            score_scale: None,
            attn_logit_softcap: None,
        },
        vec![q, k, v, bias],
        attn_shape,
    )
}

/// Build the boxRPB MLP + outer-add subgraph. Takes log-normed deltas
/// (cheap geometry computed on host) and emits the `[B, H, nq+1, hw]`
/// additive log-bias used by the image cross-attention. Replaces the
/// host-side `boxrpb_log_full_into` for the per-layer hot path.
///
/// `deltas_x: [B, nq, w, 2]`, `deltas_y: [B, nq, h, 2]` → bias.
#[allow(clippy::too_many_arguments)]
fn mlp2_relu_pair_gguf(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    cache: &mut HashMap<String, HirNodeId>,
    gguf_packed: Option<&GgufPackedParams>,
    mlp: &Mlp2,
    stem: &str,
    input: HirNodeId,
    rows: usize,
    hidden_dim: usize,
    out_dim: usize,
) -> Result<HirNodeId> {
    let h = if let Some(p) = mlp
        .w0_gguf_key
        .as_deref()
        .and_then(|key| gguf_packed.and_then(|m| super::packed_gguf::packed_linear(m, key)))
    {
        let y = linear_gguf_bias(
            g,
            params,
            typed,
            cache,
            &format!("{stem}.fc0"),
            p,
            input,
            &mlp.b0,
            mlp.in_dim,
            hidden_dim,
        )?;
        g.relu(y)
    } else {
        let w_id = add_param(
            g,
            params,
            &format!("{stem}.w0"),
            mlp.w0_t.clone(),
            Shape::new(&[mlp.in_dim, hidden_dim], DType::F32),
        );
        let b_id = add_param(
            g,
            params,
            &format!("{stem}.b0"),
            mlp.b0.clone(),
            Shape::new(&[hidden_dim], DType::F32),
        );
        fused_matmul_bias_act(
            g,
            input,
            w_id,
            b_id,
            Some(Activation::Relu),
            Shape::new(&[rows, hidden_dim], DType::F32),
        )
    };
    if let Some(p) = mlp
        .w1_gguf_key
        .as_deref()
        .and_then(|key| gguf_packed.and_then(|m| super::packed_gguf::packed_linear(m, key)))
    {
        return linear_gguf_bias(
            g,
            params,
            typed,
            cache,
            &format!("{stem}.fc1"),
            p,
            h,
            &mlp.b1,
            hidden_dim,
            out_dim,
        );
    }
    let w_id = add_param(
        g,
        params,
        &format!("{stem}.w1"),
        mlp.w1_t.clone(),
        Shape::new(&[hidden_dim, out_dim], DType::F32),
    );
    let b_id = add_param(
        g,
        params,
        &format!("{stem}.b1"),
        mlp.b1.clone(),
        Shape::new(&[out_dim], DType::F32),
    );
    Ok(fused_matmul_bias_act(
        g,
        h,
        w_id,
        b_id,
        None,
        Shape::new(&[rows, out_dim], DType::F32),
    ))
}

fn build_boxrpb_subgraph(
    g: &mut HirMut<'_>,
    params: &mut HashMap<String, Vec<f32>>,
    typed: &mut Vec<(String, Vec<u8>, DType)>,
    gguf_cache: &mut HashMap<String, HirNodeId>,
    gguf_packed: Option<&GgufPackedParams>,
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    deltas_x: HirNodeId,
    deltas_y: HirNodeId,
    batch: usize,
    nq: usize,
    nh: usize,
    h: usize,
    w: usize,
) -> Result<HirNodeId> {
    let f = DType::F32;
    let hidden_x = boxrpb_x.hidden;
    let hidden_y = boxrpb_y.hidden;
    assert_eq!(boxrpb_x.in_dim, 2);
    assert_eq!(boxrpb_y.in_dim, 2);
    assert_eq!(boxrpb_x.out_dim, nh);
    assert_eq!(boxrpb_y.out_dim, nh);

    let dx_flat = g.reshape_(deltas_x, vec![(batch * nq * w) as i64, 2]);
    let dx_o = mlp2_relu_pair_gguf(
        g,
        params,
        typed,
        gguf_cache,
        gguf_packed,
        boxrpb_x,
        "boxrpb_x",
        dx_flat,
        batch * nq * w,
        hidden_x,
        nh,
    )?;
    let dx_4d = g.reshape_(dx_o, vec![batch as i64, nq as i64, w as i64, nh as i64]);
    let dx_perm = g.transpose_(dx_4d, vec![0, 3, 1, 2]);
    let dx_bc = g.reshape_(
        dx_perm,
        vec![batch as i64, nh as i64, nq as i64, 1, w as i64],
    );

    let dy_flat = g.reshape_(deltas_y, vec![(batch * nq * h) as i64, 2]);
    let dy_o = mlp2_relu_pair_gguf(
        g,
        params,
        typed,
        gguf_cache,
        gguf_packed,
        boxrpb_y,
        "boxrpb_y",
        dy_flat,
        batch * nq * h,
        hidden_y,
        nh,
    )?;
    let dy_4d = g.reshape_(dy_o, vec![batch as i64, nq as i64, h as i64, nh as i64]);
    let dy_perm = g.transpose_(dy_4d, vec![0, 3, 1, 2]);
    let dy_bc = g.reshape_(
        dy_perm,
        vec![batch as i64, nh as i64, nq as i64, h as i64, 1],
    );

    let rpb_q = g.add(dx_bc, dy_bc);
    let rpb_q_flat = g.reshape_(
        rpb_q,
        vec![batch as i64, nh as i64, nq as i64, (h * w) as i64],
    );

    let hw = h * w;
    let _lq = nq + 1;
    let zero_pres = add_param(
        g,
        params,
        "rpb_zero_presence",
        vec![0f32; batch * nh * hw],
        Shape::new(&[batch, nh, 1, hw], f),
    );
    Ok(g.concat_(vec![zero_pres, rpb_q_flat], 2))
}

struct DecoderLayerHirParts {
    params: HashMap<String, Vec<f32>>,
    typed_params: Vec<(String, Vec<u8>, DType)>,
}

/// Build the per-layer decoder HIR body.
#[allow(clippy::too_many_arguments)]
fn build_layer_body(
    hir: &mut HirModule,
    layer: &Sam3DecoderLayerWeights,
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    norm_w: &[f32],
    norm_b: &[f32],
    dec_base: &str,
    li: usize,
    batch: usize,
    h: usize,
    w: usize,
    seq: usize,
    use_bias_attn: bool,
    boxrpb_in_ir: bool,
    gguf_packed: Option<&GgufPackedParams>,
) -> Result<DecoderLayerHirParts> {
    let hw = h * w;
    let mut g = HirMut::new(hir);
    let mut params: HashMap<String, Vec<f32>> = HashMap::new();
    let mut typed_params = Vec::new();
    let mut gguf_w_cache: HashMap<String, HirNodeId> = HashMap::new();
    let f = DType::F32;
    let d = D_MODEL;
    let nh = N_HEADS;
    let dh = HEAD_DIM;
    let nq = NUM_QUERIES;
    let lq = nq + 1;

    let tgt = g.input("tgt", Shape::new(&[batch, nq, d], f));
    let query_pos = g.input("query_pos", Shape::new(&[batch, nq, d], f));
    let presence = g.input("presence", Shape::new(&[batch, 1, d], f));
    let memory = g.input("memory", Shape::new(&[batch, hw, d], f));
    let memory_pos = g.input("memory_pos", Shape::new(&[batch, hw, d], f));
    let text = g.input("text", Shape::new(&[batch, seq, d], f));
    let text_kpm_inv = g.input("text_kpm_inv", Shape::new(&[batch, seq], f));
    let rpb_bias = if boxrpb_in_ir {
        let dx = g.input("deltas_x", Shape::new(&[batch, nq, w, 2], f));
        let dy = g.input("deltas_y", Shape::new(&[batch, nq, h, 2], f));
        build_boxrpb_subgraph(
            &mut g,
            &mut params,
            &mut typed_params,
            &mut gguf_w_cache,
            gguf_packed,
            boxrpb_x,
            boxrpb_y,
            dx,
            dy,
            batch,
            nq,
            nh,
            h,
            w,
        )?
    } else {
        g.input("rpb_bias", Shape::new(&[batch, nh, lq, hw], f))
    };

    let sa_x = g.concat_(vec![presence, tgt], 1);
    let zero_pos = add_param(
        &mut g,
        &mut params,
        "zero_presence_pos",
        vec![0f32; batch * d],
        Shape::new(&[batch, 1, d], f),
    );
    let sa_pos = g.concat_(vec![zero_pos, query_pos], 1);
    let sa_qk = g.add(sa_x, sa_pos);

    let (q_sa, k_sa, v_sa) = in_proj_qkv(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "self_attn.in_proj_weight"),
        "sa.in_proj",
        &layer.self_attn_in_w_t,
        &layer.self_attn_in_b,
        sa_qk,
        sa_qk,
        sa_x,
        d,
    )?;
    let sa_attn = g.attention_kind(
        q_sa,
        k_sa,
        v_sa,
        nh,
        dh,
        MaskKind::None,
        shape::attention_shape(g.shape(q_sa)),
    );
    let sa_proj = linear_fused_or_gguf(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "self_attn.out_proj.weight"),
        "sa.out",
        sa_attn,
        layer.self_attn_out_w_t.clone(),
        layer.self_attn_out_b.clone(),
        d,
        d,
    )?;
    let sa_res = g.add(sa_x, sa_proj);
    let n2_w = add_param(
        &mut g,
        &mut params,
        "norm2.w",
        layer.norm2_w.clone(),
        Shape::new(&[d], f),
    );
    let n2_b = add_param(
        &mut g,
        &mut params,
        "norm2.b",
        layer.norm2_b.clone(),
        Shape::new(&[d], f),
    );
    let sa_normed = g.ln(sa_res, n2_w, n2_b, 1e-5);
    let presence_after_sa = g.narrow_(sa_normed, 1, 0, 1);
    let queries_after_sa = g.narrow_(sa_normed, 1, 1, nq);

    let q_text_in = g.add(queries_after_sa, query_pos);
    let (q_text, k_text, v_text) = in_proj_qkv(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "ca_text.in_proj_weight"),
        "ca_text.in_proj",
        &layer.ca_text_in_w_t,
        &layer.ca_text_in_b,
        q_text_in,
        text,
        text,
        d,
    )?;
    let ca_text_attn = g.attention(
        q_text,
        k_text,
        v_text,
        text_kpm_inv,
        nh,
        dh,
        shape::attention_shape(g.shape(q_text)),
    );
    let ca_text_proj = linear_fused_or_gguf(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "ca_text.out_proj.weight"),
        "ca_text.out",
        ca_text_attn,
        layer.ca_text_out_w_t.clone(),
        layer.ca_text_out_b.clone(),
        d,
        d,
    )?;
    let after_ca_text_res = g.add(queries_after_sa, ca_text_proj);
    let cat_w = add_param(
        &mut g,
        &mut params,
        "catext_norm.w",
        layer.catext_norm_w.clone(),
        Shape::new(&[d], f),
    );
    let cat_b = add_param(
        &mut g,
        &mut params,
        "catext_norm.b",
        layer.catext_norm_b.clone(),
        Shape::new(&[d], f),
    );
    let after_ca_text = g.ln(after_ca_text_res, cat_w, cat_b, 1e-5);

    let ca_in = g.concat_(vec![presence_after_sa, after_ca_text], 1);
    let ca_q_in = g.add(ca_in, sa_pos);
    let k_mem_in = g.add(memory, memory_pos);

    let (q_img, k_img, v_img) = in_proj_qkv(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "cross_attn.in_proj_weight"),
        "ca_img.in_proj",
        &layer.cross_attn_in_w_t,
        &layer.cross_attn_in_b,
        ca_q_in,
        k_mem_in,
        memory,
        d,
    )?;

    let attn_flat = if use_bias_attn {
        attention_bias(&mut g, q_img, k_img, v_img, rpb_bias, nh, dh)
    } else {
        let q_4d = g.reshape_(q_img, vec![batch as i64, lq as i64, nh as i64, dh as i64]);
        let q_perm = g.transpose_(q_4d, vec![0, 2, 1, 3]);
        let k_4d = g.reshape_(k_img, vec![batch as i64, hw as i64, nh as i64, dh as i64]);
        let k_perm = g.transpose_(k_4d, vec![0, 2, 1, 3]);
        let v_4d = g.reshape_(v_img, vec![batch as i64, hw as i64, nh as i64, dh as i64]);
        let v_perm = g.transpose_(v_4d, vec![0, 2, 1, 3]);
        let k_t = g.transpose_(k_perm, vec![0, 1, 3, 2]);
        let scores = g.mm(q_perm, k_t);
        let scale_val = 1.0f32 / (HEAD_DIM as f32).sqrt();
        let scale_node = add_param(
            &mut g,
            &mut params,
            "img.scale",
            vec![scale_val],
            Shape::new(&[1], f),
        );
        let scores_scaled = g.mul(scores, scale_node);
        let scores_biased = g.add(scores_scaled, rpb_bias);
        let probs = g.sm(scores_biased, -1);
        let attn_out = g.mm(probs, v_perm);
        let attn_perm = g.transpose_(attn_out, vec![0, 2, 1, 3]);
        g.reshape_(attn_perm, vec![batch as i64, lq as i64, d as i64])
    };
    let ca_img_proj = linear_fused_or_gguf(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "cross_attn.out_proj.weight"),
        "ca_img.out",
        attn_flat,
        layer.cross_attn_out_w_t.clone(),
        layer.cross_attn_out_b.clone(),
        d,
        d,
    )?;
    let ca_img_res = g.add(ca_in, ca_img_proj);
    let n1_w = add_param(
        &mut g,
        &mut params,
        "norm1.w",
        layer.norm1_w.clone(),
        Shape::new(&[d], f),
    );
    let n1_b = add_param(
        &mut g,
        &mut params,
        "norm1.b",
        layer.norm1_b.clone(),
        Shape::new(&[d], f),
    );
    let after_ca_img = g.ln(ca_img_res, n1_w, n1_b, 1e-5);

    let ff1 = linear_fused_or_gguf(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "linear1.weight"),
        "ffn.fc1",
        after_ca_img,
        layer.linear1_w_t.clone(),
        layer.linear1_b.clone(),
        d,
        DIM_FF,
    )?;
    let relud = g.relu(ff1);
    let ff2 = linear_fused_or_gguf(
        &mut g,
        &mut params,
        &mut typed_params,
        &mut gguf_w_cache,
        gguf_packed,
        &dec_layer_key(dec_base, li, "linear2.weight"),
        "ffn.fc2",
        relud,
        layer.linear2_w_t.clone(),
        layer.linear2_b.clone(),
        DIM_FF,
        d,
    )?;
    let ffn_res = g.add(after_ca_img, ff2);
    let n3_w = add_param(
        &mut g,
        &mut params,
        "norm3.w",
        layer.norm3_w.clone(),
        Shape::new(&[d], f),
    );
    let n3_b = add_param(
        &mut g,
        &mut params,
        "norm3.b",
        layer.norm3_b.clone(),
        Shape::new(&[d], f),
    );
    let after_ffn = g.ln(ffn_res, n3_w, n3_b, 1e-5);

    let new_presence = g.narrow_(after_ffn, 1, 0, 1);
    let new_tgt = g.narrow_(after_ffn, 1, 1, nq);

    let dec_norm_w = add_param(
        &mut g,
        &mut params,
        "dec.norm.w",
        norm_w.to_vec(),
        Shape::new(&[d], f),
    );
    let dec_norm_b = add_param(
        &mut g,
        &mut params,
        "dec.norm.b",
        norm_b.to_vec(),
        Shape::new(&[d], f),
    );
    let out_norm = g.ln(new_tgt, dec_norm_w, dec_norm_b, 1e-5);

    g.set_outputs(vec![new_tgt, new_presence, out_norm]);
    let _ = (q_img, k_img, v_img, ca_img_proj);
    Ok(DecoderLayerHirParts {
        params,
        typed_params,
    })
}

/// Build a per-layer native HIR module.
fn build_layer_hir(
    layer: &Sam3DecoderLayerWeights,
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    norm_w: &[f32],
    norm_b: &[f32],
    dec_base: &str,
    li: usize,
    batch: usize,
    h: usize,
    w: usize,
    seq: usize,
    use_bias_attn: bool,
    boxrpb_in_ir: bool,
    gguf_packed: Option<&GgufPackedParams>,
) -> Result<LayerHirParts> {
    let mut hir = HirModule::new("sam3_dec_layer");
    let parts = build_layer_body(
        &mut hir,
        layer,
        boxrpb_x,
        boxrpb_y,
        norm_w,
        norm_b,
        dec_base,
        li,
        batch,
        h,
        w,
        seq,
        use_bias_attn,
        boxrpb_in_ir,
        gguf_packed,
    )?;
    Ok((hir, parts.params, parts.typed_params))
}

/// Compile-once-per-layer decoder, runnable across many frames.
pub struct Sam3CompiledDecoder {
    layers: Vec<CompiledGraph>,
    bbox_embed: Mlp3,
    ref_point_head: Mlp2,
    boxrpb_x: Mlp2,
    boxrpb_y: Mlp2,
    initial_query_embed: Vec<f32>,
    initial_reference_points: Vec<f32>,
    cached_layer0_query_pos: Vec<f32>,
    /// Layer-0 cached inputs for the boxRPB IR subgraph path (GPU
    /// backends). Constant geometry, ~115KB each.
    cached_layer0_deltas_x: Option<Vec<f32>>,
    cached_layer0_deltas_y: Option<Vec<f32>>,
    /// Layer-0 cached `rpb_bias [B, H, lq, hw]` for the host-MLP path
    /// (CPU backend). Constant boxRPB tensor, ~66MB.
    cached_layer0_rpb: Option<Vec<f32>>,
    #[allow(dead_code)]
    cached_initial_ref_boxes: Vec<f32>,
    boxrpb_in_ir: bool,
    presence_token: Vec<f32>,
    presence_head: Mlp3,
    presence_norm_w: Vec<f32>,
    presence_norm_b: Vec<f32>,
    /// Per-call delta scratch (geometry only — the MLP forward and
    /// outer-add run inside the IR graph for GPU backends).
    scratch_deltas_x: Vec<f32>,
    scratch_deltas_y: Vec<f32>,
    /// Host-MLP path scratch buffers (CPU backend). Allocated only
    /// when boxrpb_in_ir is false.
    scratch_rpb: Option<Vec<f32>>,
    scratch_dx_thq: Option<Vec<f32>>,
    scratch_dy_thq: Option<Vec<f32>>,
    scratch_boxrpb_x_hidden: Option<Vec<f32>>,
    scratch_boxrpb_y_hidden: Option<Vec<f32>>,
    scratch_boxrpb_x_feats: Option<Vec<f32>>,
    scratch_boxrpb_y_feats: Option<Vec<f32>>,
    /// Scratch for ref_point_head sineembed + MLP intermediates.
    scratch_sine: Vec<f32>,
    scratch_rph_hidden: Vec<f32>,
    /// Output of ref_point_head MLP for layers 1..N (layer 0 uses cache).
    scratch_query_pos: Vec<f32>,
    /// Scratch for box-refinement `mlp3_forward(bbox_embed)`.
    scratch_bbox_h0: Vec<f32>,
    scratch_bbox_h1: Vec<f32>,
    scratch_bbox_out: Vec<f32>,
    pub batch: usize,
    pub hw: usize,
    pub seq: usize,
    gguf_packed: Option<GgufPackedParams>,
}

impl Sam3CompiledDecoder {
    pub fn new(
        weights: &Sam3DecoderWeights,
        batch: usize,
        hw: usize,
        seq: usize,
        device: Device,
    ) -> Result<Self> {
        Self::new_with_profile(weights, batch, hw, seq, device, &CompileProfile::sam3())
    }

    pub fn new_with_profile(
        weights: &Sam3DecoderWeights,
        batch: usize,
        hw: usize,
        seq: usize,
        device: Device,
        profile: &CompileProfile,
    ) -> Result<Self> {
        Self::new_with_profile_and_gguf(weights, batch, hw, seq, device, profile, None)
    }

    pub fn new_with_profile_and_gguf(
        weights: &Sam3DecoderWeights,
        batch: usize,
        hw: usize,
        seq: usize,
        device: Device,
        profile: &CompileProfile,
        gguf_packed: Option<&GgufPackedParams>,
    ) -> Result<Self> {
        ensure!(weights.loaded, "decoder weights not loaded");
        let nq = NUM_QUERIES;
        let d = D_MODEL;
        let h_w = (hw as f64).sqrt().round() as usize;
        ensure!(
            h_w * h_w == hw,
            "boxRPB cache requires square spatial grid; got hw={hw}"
        );
        let mut layers = Vec::with_capacity(N_LAYERS);
        // Metal: opt-in to bias-mask SDPA via env var. The default
        // routes Metal through the MPSGraph manual-decomp because the
        // bias-aware SDPA kernels (sdpa_long, sdpa_fa_f32) currently
        // produce incorrect output — needs debugging.
        let use_bias_attn = if matches!(device, Device::Metal) {
            rlx_ir::env::flag("RLX_SAM3_METAL_BIAS_SDPA")
        } else {
            true
        };
        // MLX: boxRPB in-graph (saves per-call rpb_bias upload). CPU: opt-in
        // via `RLX_SAM3_BOXRPB_IR` (host BLAS is faster for tiny K=2 GEMMs).
        // Metal: 5D broadcast outer-add in boxRPB subgraph is wrong on Metal IR
        // — keep host boxRPB until fixed (`RLX_SAM3_BOXRPB_IR` ignored on Metal).
        let boxrpb_in_ir = matches!(device, Device::Mlx)
            || (matches!(device, Device::Cpu) && rlx_ir::env::flag("RLX_SAM3_BOXRPB_IR"));
        let dec_base = &weights.prefix;
        for (li, layer) in weights.layers.iter().enumerate() {
            let (hir, params, typed) = build_layer_hir(
                layer,
                &weights.boxrpb_x,
                &weights.boxrpb_y,
                &weights.norm_w,
                &weights.norm_b,
                dec_base,
                li,
                batch,
                h_w,
                h_w,
                seq,
                use_bias_attn,
                boxrpb_in_ir,
                gguf_packed,
            )?;
            let mut compiled =
                rlx_core::flow_bridge::compile_hir_with_profile(device, hir, profile)?;
            rlx_core::flow_util::attach_built_params(&mut compiled, params, &typed);
            layers.push(compiled);
        }
        // Precompute layer-0 inputs that depend only on constant model
        // weights (initial reference_points → cached query_pos and
        // boxRPB deltas). The boxRPB MLP+outer-add now runs inside the
        // graph, so we only cache the cheap geometry deltas.
        let mut cached_initial_ref_boxes = vec![0f32; batch * nq * 4];
        for b in 0..batch {
            for q in 0..nq {
                for k in 0..4 {
                    let v = weights.reference_points[q * 4 + k];
                    cached_initial_ref_boxes[(b * nq + q) * 4 + k] = sigmoid(v);
                }
            }
        }
        let sine = sineembed_4d(&cached_initial_ref_boxes, batch, nq, d);
        let cached_layer0_query_pos =
            mlp2_forward(&weights.ref_point_head, &sine, batch * nq, gguf_packed)?;
        let lq = nq + 1;
        let nh = N_HEADS;
        let (cached_layer0_deltas_x, cached_layer0_deltas_y, cached_layer0_rpb) = if boxrpb_in_ir {
            let mut dx = vec![0f32; batch * nq * h_w * 2];
            let mut dy = vec![0f32; batch * nq * h_w * 2];
            compute_deltas_into(
                &cached_initial_ref_boxes,
                batch,
                nq,
                h_w,
                h_w,
                &mut dx,
                &mut dy,
            );
            (Some(dx), Some(dy), None)
        } else {
            let rpb = boxrpb_log_full(
                &weights.boxrpb_x,
                &weights.boxrpb_y,
                &cached_initial_ref_boxes,
                batch,
                nq,
                h_w,
                h_w,
                gguf_packed,
            )?;
            (None, None, Some(rpb))
        };
        Ok(Self {
            layers,
            bbox_embed: weights.bbox_embed.clone(),
            ref_point_head: weights.ref_point_head.clone(),
            boxrpb_x: weights.boxrpb_x.clone(),
            boxrpb_y: weights.boxrpb_y.clone(),
            initial_query_embed: weights.query_embed.clone(),
            initial_reference_points: weights.reference_points.clone(),
            cached_layer0_query_pos,
            cached_layer0_deltas_x,
            cached_layer0_deltas_y,
            cached_layer0_rpb,
            cached_initial_ref_boxes,
            boxrpb_in_ir,
            presence_token: weights.presence_token.clone(),
            presence_head: weights.presence_token_head.clone(),
            presence_norm_w: weights.presence_token_out_norm_w.clone(),
            presence_norm_b: weights.presence_token_out_norm_b.clone(),
            scratch_deltas_x: if boxrpb_in_ir {
                vec![0f32; batch * nq * h_w * 2]
            } else {
                Vec::new()
            },
            scratch_deltas_y: if boxrpb_in_ir {
                vec![0f32; batch * nq * h_w * 2]
            } else {
                Vec::new()
            },
            scratch_rpb: (!boxrpb_in_ir).then(|| vec![0f32; batch * nh * lq * hw]),
            scratch_dx_thq: (!boxrpb_in_ir).then(|| vec![0f32; nh * nq * h_w]),
            scratch_dy_thq: (!boxrpb_in_ir).then(|| vec![0f32; nh * nq * h_w]),
            scratch_boxrpb_x_hidden: (!boxrpb_in_ir)
                .then(|| vec![0f32; nq * h_w * weights.boxrpb_x.hidden]),
            scratch_boxrpb_y_hidden: (!boxrpb_in_ir)
                .then(|| vec![0f32; nq * h_w * weights.boxrpb_y.hidden]),
            scratch_boxrpb_x_feats: (!boxrpb_in_ir)
                .then(|| vec![0f32; nq * h_w * weights.boxrpb_x.out_dim]),
            scratch_boxrpb_y_feats: (!boxrpb_in_ir)
                .then(|| vec![0f32; nq * h_w * weights.boxrpb_y.out_dim]),
            scratch_sine: vec![0f32; batch * nq * 2 * d],
            scratch_rph_hidden: vec![0f32; batch * nq * weights.ref_point_head.hidden],
            scratch_query_pos: vec![0f32; batch * nq * weights.ref_point_head.out_dim],
            scratch_bbox_h0: vec![0f32; batch * nq * weights.bbox_embed.hidden],
            scratch_bbox_h1: vec![0f32; batch * nq * weights.bbox_embed.hidden],
            scratch_bbox_out: vec![0f32; batch * nq * weights.bbox_embed.out_dim],
            batch,
            hw,
            seq,
            gguf_packed: gguf_packed.cloned(),
        })
    }

    /// Run the decoder. Inputs are batch-first: `memory [B, hw, D]`,
    /// `memory_pos [B, hw, D]`, `text [B, seq, D]` (note: text is
    /// batch-first here, not seq-first), `text_kpm` (1 = PAD).
    pub fn run(
        &mut self,
        memory: &[f32],
        memory_pos: &[f32],
        text_seq_first: &[f32],
        text_kpm: &[u8],
        h: usize,
        w: usize,
    ) -> Result<LayerRunOut> {
        let hw = h * w;
        ensure!(hw == self.hw);
        let batch = self.batch;
        let nq = NUM_QUERIES;
        let d = D_MODEL;
        let nh = N_HEADS;
        let lq = nq + 1;
        let seq = self.seq;

        // Initial tgt = query_embed expanded to batch.
        let mut tgt = vec![0f32; batch * nq * d];
        for b in 0..batch {
            tgt[b * nq * d..(b + 1) * nq * d].copy_from_slice(&self.initial_query_embed);
        }
        // Initial ref_boxes = sigmoid(reference_points).
        let mut ref_boxes = vec![0f32; batch * nq * 4];
        for b in 0..batch {
            for q in 0..nq {
                for k in 0..4 {
                    let v = self.initial_reference_points[q * 4 + k];
                    ref_boxes[(b * nq + q) * 4 + k] = sigmoid(v);
                }
            }
        }
        let mut presence = vec![0f32; batch * d];
        for b in 0..batch {
            presence[b * d..(b + 1) * d].copy_from_slice(&self.presence_token);
        }

        // Text → batch-first.
        let mut text_bf = vec![0f32; batch * seq * d];
        for b in 0..batch {
            for l in 0..seq {
                let s = (l * batch + b) * d;
                let dst = (b * seq + l) * d;
                text_bf[dst..dst + d].copy_from_slice(&text_seq_first[s..s + d]);
            }
        }
        let text_kpm_inv: Vec<f32> = text_kpm
            .iter()
            .map(|&v| if v == 0 { 1.0 } else { 0.0 })
            .collect();

        let mut intermediate = Vec::with_capacity(N_LAYERS);
        let mut intermediate_ref_boxes = Vec::with_capacity(N_LAYERS);
        intermediate_ref_boxes.push(ref_boxes.clone());
        let mut presence_logits = Vec::with_capacity(N_LAYERS);

        let profile = rlx_ir::env::flag("RLX_SAM3_PROFILE");
        let mut t_qpos = 0u128;
        let mut t_rpb = 0u128;
        let mut t_graph = 0u128;
        let mut t_box = 0u128;
        let mut t_other = 0u128;
        for li in 0..N_LAYERS {
            let tq = std::time::Instant::now();
            // Compute query_pos = ref_point_head(sineembed(ref_boxes)).
            // Layer 0's ref_boxes is the constant `sigmoid(reference_points)`,
            // so its query_pos and boxRPB are precomputed once at
            // construction and reused per call.
            // For layer 0, use the cached query_pos slice directly and
            // the cached rpb buffer. For other layers, recompute into
            // pre-allocated scratch buffers so we don't malloc 33MB/layer.
            let query_pos_slice: &[f32];
            let rpb_slice: &[f32];
            let deltas_x_slice: &[f32];
            let deltas_y_slice: &[f32];
            if li == 0 {
                query_pos_slice = &self.cached_layer0_query_pos;
                if self.boxrpb_in_ir {
                    deltas_x_slice = self.cached_layer0_deltas_x.as_ref().unwrap();
                    deltas_y_slice = self.cached_layer0_deltas_y.as_ref().unwrap();
                    rpb_slice = &[];
                } else {
                    rpb_slice = self.cached_layer0_rpb.as_ref().unwrap();
                    deltas_x_slice = &[];
                    deltas_y_slice = &[];
                }
            } else {
                sineembed_4d_into(&ref_boxes, batch, nq, d, &mut self.scratch_sine);
                mlp2_forward_into(
                    &self.ref_point_head,
                    &self.scratch_sine,
                    batch * nq,
                    &mut self.scratch_rph_hidden,
                    &mut self.scratch_query_pos,
                    self.gguf_packed.as_ref(),
                )?;
                query_pos_slice = &self.scratch_query_pos;
                if self.boxrpb_in_ir {
                    compute_deltas_into(
                        &ref_boxes,
                        batch,
                        nq,
                        h,
                        w,
                        &mut self.scratch_deltas_x,
                        &mut self.scratch_deltas_y,
                    );
                    deltas_x_slice = &self.scratch_deltas_x;
                    deltas_y_slice = &self.scratch_deltas_y;
                    rpb_slice = &[];
                } else {
                    let mut host_deltas_x = vec![0f32; nq * w * 2];
                    let mut host_deltas_y = vec![0f32; nq * h * 2];
                    boxrpb_log_full_into(
                        &self.boxrpb_x,
                        &self.boxrpb_y,
                        &ref_boxes,
                        batch,
                        nq,
                        h,
                        w,
                        self.scratch_rpb.as_mut().unwrap(),
                        self.scratch_dx_thq.as_mut().unwrap(),
                        self.scratch_dy_thq.as_mut().unwrap(),
                        &mut host_deltas_x,
                        &mut host_deltas_y,
                        self.scratch_boxrpb_x_hidden.as_mut().unwrap(),
                        self.scratch_boxrpb_y_hidden.as_mut().unwrap(),
                        self.scratch_boxrpb_x_feats.as_mut().unwrap(),
                        self.scratch_boxrpb_y_feats.as_mut().unwrap(),
                        self.gguf_packed.as_ref(),
                    )?;
                    rpb_slice = self.scratch_rpb.as_ref().unwrap();
                    deltas_x_slice = &[];
                    deltas_y_slice = &[];
                }
            }
            if profile {
                t_qpos += tq.elapsed().as_micros();
            }

            let tr = std::time::Instant::now();
            if profile {
                t_rpb += tr.elapsed().as_micros();
            }

            // Run graph.
            let tg = std::time::Instant::now();
            let outputs = if self.boxrpb_in_ir {
                self.layers[li].run(&[
                    ("tgt", tgt.as_slice()),
                    ("query_pos", query_pos_slice),
                    ("presence", presence.as_slice()),
                    ("memory", memory),
                    ("memory_pos", memory_pos),
                    ("text", text_bf.as_slice()),
                    ("text_kpm_inv", text_kpm_inv.as_slice()),
                    ("deltas_x", deltas_x_slice),
                    ("deltas_y", deltas_y_slice),
                ])
            } else {
                self.layers[li].run(&[
                    ("tgt", tgt.as_slice()),
                    ("query_pos", query_pos_slice),
                    ("presence", presence.as_slice()),
                    ("memory", memory),
                    ("memory_pos", memory_pos),
                    ("text", text_bf.as_slice()),
                    ("text_kpm_inv", text_kpm_inv.as_slice()),
                    ("rpb_bias", rpb_slice),
                ])
            };
            if profile {
                t_graph += tg.elapsed().as_micros();
            }
            ensure!(outputs.len() == 3, "decoder layer expected 3 outputs");
            tgt = outputs[0].clone();
            presence = outputs[1].clone();
            let out_norm = outputs[2].clone();

            let tb = std::time::Instant::now();
            // Box refinement: delta = bbox_embed(out_norm); ref = sigmoid(inv_sig(ref) + delta).
            mlp3_forward_into(
                &self.bbox_embed,
                &out_norm,
                batch * nq,
                &mut self.scratch_bbox_h0,
                &mut self.scratch_bbox_h1,
                &mut self.scratch_bbox_out,
                self.gguf_packed.as_ref(),
            )?;
            let delta: &[f32] = &self.scratch_bbox_out;
            if profile {
                t_box += tb.elapsed().as_micros();
            }
            let to = std::time::Instant::now();
            let _ = to;
            let _ = &mut t_other;
            let mut new_ref = vec![0f32; batch * nq * 4];
            for q in 0..nq {
                for b in 0..batch {
                    let cur = &ref_boxes[(b * nq + q) * 4..(b * nq + q + 1) * 4];
                    let dl = &delta[(b * nq + q) * 4..(b * nq + q + 1) * 4];
                    for k in 0..4 {
                        new_ref[(b * nq + q) * 4 + k] = sigmoid(inv_sigmoid(cur[k]) + dl[k]);
                    }
                }
            }
            ref_boxes = new_ref;
            if li != N_LAYERS - 1 {
                intermediate_ref_boxes.push(ref_boxes.clone());
            }

            // Intermediate output in seq-first convention.
            let mut out_seq_first = vec![0f32; nq * batch * d];
            for q in 0..nq {
                for b in 0..batch {
                    let src = (b * nq + q) * d;
                    let dst = (q * batch + b) * d;
                    out_seq_first[dst..dst + d].copy_from_slice(&out_norm[src..src + d]);
                }
            }
            intermediate.push(out_seq_first);

            // Presence logits.
            let p_norm =
                layer_norm_host(&presence, &self.presence_norm_w, &self.presence_norm_b, d);
            let p_logit = mlp3_forward(
                &self.presence_head,
                &p_norm,
                batch,
                self.gguf_packed.as_ref(),
            )?;
            presence_logits.push(p_logit);
        }
        if profile {
            let to_ms = |us: u128| us as f32 / 1000.0;
            eprintln!(
                "  decoder per-stage (6 layers total): qpos={:.1}ms  rpb={:.1}ms  graph={:.1}ms  box={:.1}ms",
                to_ms(t_qpos),
                to_ms(t_rpb),
                to_ms(t_graph),
                to_ms(t_box)
            );
        }

        // Stack.
        let mut int_stack = vec![0f32; N_LAYERS * nq * batch * d];
        for (li, l) in intermediate.iter().enumerate() {
            int_stack[li * nq * batch * d..(li + 1) * nq * batch * d].copy_from_slice(l);
        }
        let mut ref_stack = vec![0f32; N_LAYERS * nq * batch * 4];
        for (li, r) in intermediate_ref_boxes.iter().enumerate() {
            ref_stack[li * nq * batch * 4..(li + 1) * nq * batch * 4].copy_from_slice(r);
        }
        let mut presence_stack = vec![0f32; N_LAYERS * batch];
        for (li, p) in presence_logits.iter().enumerate() {
            for b in 0..batch {
                presence_stack[li * batch + b] = p[b];
            }
        }
        let _ = nh;
        let _ = lq;
        Ok((int_stack, ref_stack, presence_stack, presence))
    }
}

/// IR-compiled detector decoder on the requested device (6 layer graphs).
#[allow(clippy::too_many_arguments)]
pub fn forward_decoder_ir_on(
    weights: &Sam3DecoderWeights,
    memory: &[f32],
    memory_pos: &[f32],
    memory_text: &[f32],
    text_attention_mask: &[u8],
    batch: usize,
    h: usize,
    w: usize,
    seq_len: usize,
    device: Device,
) -> Result<Sam3DecoderOutput> {
    forward_decoder_ir_on_with_profile(
        weights,
        memory,
        memory_pos,
        memory_text,
        text_attention_mask,
        batch,
        h,
        w,
        seq_len,
        device,
        &CompileProfile::sam3(),
        None,
    )
}

/// Same as [`forward_decoder_ir_on`] with an explicit tier-1 profile.
#[allow(clippy::too_many_arguments)]
pub fn forward_decoder_ir_on_with_profile(
    weights: &Sam3DecoderWeights,
    memory: &[f32],
    memory_pos: &[f32],
    memory_text: &[f32],
    text_attention_mask: &[u8],
    batch: usize,
    h: usize,
    w: usize,
    seq_len: usize,
    device: Device,
    profile: &CompileProfile,
    gguf_packed: Option<&GgufPackedParams>,
) -> Result<Sam3DecoderOutput> {
    ensure!(weights.loaded, "decoder weights not loaded");
    ensure!(batch == 1, "decoder IR forward requires batch=1 for boxRPB");
    let hw = h * w;
    let mut dec = Sam3CompiledDecoder::new_with_profile_and_gguf(
        weights,
        batch,
        hw,
        seq_len,
        device,
        profile,
        gguf_packed,
    )?;
    let (intermediate, intermediate_ref_boxes, presence_logits, presence_feats) =
        dec.run(memory, memory_pos, memory_text, text_attention_mask, h, w)?;
    Ok(Sam3DecoderOutput {
        intermediate,
        intermediate_ref_boxes,
        presence_logits,
        presence_feats,
        num_layers: N_LAYERS,
        num_queries: NUM_QUERIES,
        batch,
        d_model: D_MODEL,
    })
}

// ── Host helpers (sineembed, boxRPB, mlp, sigmoid) ─────────────────────

fn sigmoid(x: f32) -> f32 {
    1.0 / (1.0 + (-x).exp())
}

fn inv_sigmoid(x: f32) -> f32 {
    let eps = 1e-3f32;
    let x = x.clamp(0.0, 1.0).max(eps).min(1.0 - eps);
    (x / (1.0 - x)).ln()
}

fn layer_norm_host(x: &[f32], gamma: &[f32], beta: &[f32], dim: usize) -> Vec<f32> {
    let rows = x.len() / dim;
    let mut out = vec![0f32; x.len()];
    for r in 0..rows {
        let row = &x[r * dim..(r + 1) * dim];
        let mean = row.iter().sum::<f32>() / dim as f32;
        let var = row.iter().map(|v| (*v - mean).powi(2)).sum::<f32>() / dim as f32;
        let inv = 1.0 / (var + 1e-5).sqrt();
        for c in 0..dim {
            out[r * dim + c] = (row[c] - mean) * inv * gamma[c] + beta[c];
        }
    }
    out
}

#[allow(dead_code)]
fn host_mlp2_forward(mlp: &Mlp2, x: &[f32], rows: usize) -> Result<Vec<f32>> {
    let h = matmul_bias_relu(x, &mlp.w0_t, &mlp.b0, rows, mlp.in_dim, mlp.hidden);
    Ok(matmul_bias(
        &h,
        &mlp.w1_t,
        &mlp.b1,
        rows,
        mlp.hidden,
        mlp.out_dim,
    ))
}

/// In-place mlp2: `out = w1·relu(w0·x + b0) + b1`. Caller provides the
/// hidden scratch buffer and output buffer — no allocation in the hot
/// path. First layer uses fused matmul+bias+relu epilogue.
#[allow(dead_code)]
fn host_mlp2_forward_into(mlp: &Mlp2, x: &[f32], rows: usize, hidden: &mut [f32], out: &mut [f32]) {
    rlx_cpu::blas::sgemm_bias_epilogue(
        x,
        &mlp.w0_t,
        &mlp.b0,
        hidden,
        rows,
        mlp.in_dim,
        mlp.hidden,
        |v| if v < 0.0 { 0.0 } else { v },
    );
    rlx_cpu::blas::sgemm_bias(
        hidden,
        &mlp.w1_t,
        &mlp.b1,
        out,
        rows,
        mlp.hidden,
        mlp.out_dim,
    );
}

#[allow(dead_code)]
fn host_mlp3_forward(mlp: &Mlp3, x: &[f32], rows: usize) -> Result<Vec<f32>> {
    let h = matmul_bias_relu(x, &mlp.w0_t, &mlp.b0, rows, mlp.in_dim, mlp.hidden);
    let h = matmul_bias_relu(&h, &mlp.w1_t, &mlp.b1, rows, mlp.hidden, mlp.hidden);
    Ok(matmul_bias(
        &h,
        &mlp.w2_t,
        &mlp.b2,
        rows,
        mlp.hidden,
        mlp.out_dim,
    ))
}

#[allow(dead_code)]
fn host_mlp3_forward_into(
    mlp: &Mlp3,
    x: &[f32],
    rows: usize,
    h0: &mut [f32],
    h1: &mut [f32],
    out: &mut [f32],
) {
    let relu = |v: f32| if v < 0.0 { 0.0 } else { v };
    rlx_cpu::blas::sgemm_bias_epilogue(
        x, &mlp.w0_t, &mlp.b0, h0, rows, mlp.in_dim, mlp.hidden, relu,
    );
    rlx_cpu::blas::sgemm_bias_epilogue(
        h0, &mlp.w1_t, &mlp.b1, h1, rows, mlp.hidden, mlp.hidden, relu,
    );
    rlx_cpu::blas::sgemm_bias(h1, &mlp.w2_t, &mlp.b2, out, rows, mlp.hidden, mlp.out_dim);
}

#[allow(dead_code)]
fn matmul_bias(x: &[f32], w_t: &[f32], b: &[f32], rows: usize, k: usize, n: usize) -> Vec<f32> {
    let mut out = vec![0f32; rows * n];
    rlx_cpu::blas::sgemm_bias(x, w_t, b, &mut out, rows, k, n);
    out
}

#[allow(dead_code)]
fn matmul_bias_relu(
    x: &[f32],
    w_t: &[f32],
    b: &[f32],
    rows: usize,
    k: usize,
    n: usize,
) -> Vec<f32> {
    let mut out = matmul_bias(x, w_t, b, rows, k, n);
    for v in out.iter_mut() {
        if *v < 0.0 {
            *v = 0.0;
        }
    }
    out
}

fn sineembed_4d(pos: &[f32], batch: usize, nq: usize, d_model: usize) -> Vec<f32> {
    let mut out = vec![0.0f32; batch * nq * 2 * d_model];
    sineembed_4d_into(pos, batch, nq, d_model, &mut out);
    out
}

fn sineembed_4d_into(pos: &[f32], batch: usize, nq: usize, d_model: usize, out: &mut [f32]) {
    let half = d_model / 2;
    let scale = 2.0 * std::f32::consts::PI;
    let mut dim_t = vec![0.0f32; half];
    for i in 0..half {
        let exp = 2.0 * ((i / 2) as f32) / half as f32;
        dim_t[i] = 10000.0f32.powf(exp);
    }
    debug_assert_eq!(out.len(), batch * nq * 2 * d_model);
    for b in 0..batch {
        for q in 0..nq {
            let p = &pos[(b * nq + q) * 4..(b * nq + q + 1) * 4];
            let vals = [p[1] * scale, p[0] * scale, p[2] * scale, p[3] * scale];
            let base = (b * nq + q) * 2 * d_model;
            for axis in 0..4 {
                let slot = base + axis * half;
                for i in 0..half {
                    let theta = vals[axis] / dim_t[i];
                    out[slot + i] = if i % 2 == 0 { theta.sin() } else { theta.cos() };
                }
            }
        }
    }
}

/// Owning version that allocates; kept for the construct-time cache
/// where we run it once. Hot path uses `boxrpb_log_full_into` with a
/// pre-allocated scratch buffer.
fn boxrpb_log_full(
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    reference_boxes: &[f32],
    batch: usize,
    nq: usize,
    h: usize,
    w: usize,
    gguf_packed: Option<&GgufPackedParams>,
) -> Result<Vec<f32>> {
    let nh = N_HEADS;
    let lq = nq + 1;
    let mut out = vec![0f32; batch * nh * lq * h * w];
    let mut dx_thq = vec![0f32; nh * nq * w];
    let mut dy_thq = vec![0f32; nh * nq * h];
    let mut deltas_x = vec![0f32; nq * w * 2];
    let mut deltas_y = vec![0f32; nq * h * 2];
    let mut hidden_x = vec![0f32; nq * w * boxrpb_x.hidden];
    let mut hidden_y = vec![0f32; nq * h * boxrpb_y.hidden];
    let mut feats_x = vec![0f32; nq * w * boxrpb_x.out_dim];
    let mut feats_y = vec![0f32; nq * h * boxrpb_y.out_dim];
    boxrpb_log_full_into(
        boxrpb_x,
        boxrpb_y,
        reference_boxes,
        batch,
        nq,
        h,
        w,
        &mut out,
        &mut dx_thq,
        &mut dy_thq,
        &mut deltas_x,
        &mut deltas_y,
        &mut hidden_x,
        &mut hidden_y,
        &mut feats_x,
        &mut feats_y,
        gguf_packed,
    )?;
    Ok(out)
}

#[allow(clippy::too_many_arguments)]
fn boxrpb_log_full_into(
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    reference_boxes: &[f32],
    batch: usize,
    nq: usize,
    h: usize,
    w: usize,
    out: &mut [f32],
    dx_thq: &mut [f32],
    dy_thq: &mut [f32],
    deltas_x: &mut [f32],
    deltas_y: &mut [f32],
    hidden_x: &mut [f32],
    hidden_y: &mut [f32],
    feats_x: &mut [f32],
    feats_y: &mut [f32],
    gguf_packed: Option<&GgufPackedParams>,
) -> Result<()> {
    let nh = N_HEADS;
    let lq = nq + 1;
    debug_assert_eq!(out.len(), batch * nh * lq * h * w);
    debug_assert_eq!(dx_thq.len(), nh * nq * w);
    debug_assert_eq!(dy_thq.len(), nh * nq * h);
    debug_assert_eq!(deltas_x.len(), nq * w * 2);
    debug_assert_eq!(deltas_y.len(), nq * h * 2);
    debug_assert_eq!(hidden_x.len(), nq * w * boxrpb_x.hidden);
    debug_assert_eq!(hidden_y.len(), nq * h * boxrpb_y.hidden);
    debug_assert_eq!(feats_x.len(), nq * w * boxrpb_x.out_dim);
    debug_assert_eq!(feats_y.len(), nq * h * boxrpb_y.out_dim);
    // Zero the presence rows once — non-presence rows get overwritten.
    for head in 0..nh {
        for b in 0..batch {
            let off = b * nh * lq * h * w + head * lq * h * w;
            // Presence row at lq=0
            for i in 0..h * w {
                out[off + i] = 0.0;
            }
        }
    }
    let coords_h: Vec<f32> = (0..h).map(|y| y as f32 / h as f32).collect();
    let coords_w: Vec<f32> = (0..w).map(|x| x as f32 / w as f32).collect();

    for b in 0..batch {
        for q in 0..nq {
            let p = &reference_boxes[(b * nq + q) * 4..(b * nq + q + 1) * 4];
            let (cx, cy, bw, bh) = (p[0], p[1], p[2], p[3]);
            let x0 = cx - 0.5 * bw;
            let x1 = cx + 0.5 * bw;
            let y0 = cy - 0.5 * bh;
            let y1 = cy + 0.5 * bh;
            for xi in 0..w {
                let dx0 = (coords_w[xi] - x0) * 8.0;
                let dx1 = (coords_w[xi] - x1) * 8.0;
                deltas_x[(q * w + xi) * 2] = log_norm(dx0);
                deltas_x[(q * w + xi) * 2 + 1] = log_norm(dx1);
            }
            for yi in 0..h {
                let dy0 = (coords_h[yi] - y0) * 8.0;
                let dy1 = (coords_h[yi] - y1) * 8.0;
                deltas_y[(q * h + yi) * 2] = log_norm(dy0);
                deltas_y[(q * h + yi) * 2 + 1] = log_norm(dy1);
            }
        }
        mlp2_forward_into(boxrpb_x, deltas_x, nq * w, hidden_x, feats_x, gguf_packed)?;
        mlp2_forward_into(boxrpb_y, deltas_y, nq * h, hidden_y, feats_y, gguf_packed)?;
        let dx_feats: &[f32] = feats_x;
        let dy_feats: &[f32] = feats_y;
        // Transpose dx/dy from [pos, head] to [head, q, pos] so the
        // outer-add per (head, q) reads contiguous slices.
        for q in 0..nq {
            for xi in 0..w {
                let src_base = (q * w + xi) * nh;
                for head in 0..nh {
                    dx_thq[(head * nq + q) * w + xi] = dx_feats[src_base + head];
                }
            }
            for yi in 0..h {
                let src_base = (q * h + yi) * nh;
                for head in 0..nh {
                    dy_thq[(head * nq + q) * h + yi] = dy_feats[src_base + head];
                }
            }
        }
        let base = b * nh * lq * h * w;
        let total = nh * nq;
        let out_ptr = out.as_mut_ptr() as usize;
        let dx_ptr = dx_thq.as_ptr() as usize;
        let dy_ptr = dy_thq.as_ptr() as usize;
        rlx_cpu::pool::par_for(total, 8, &|off, cnt| unsafe {
            for idx in off..off + cnt {
                let head = idx / nq;
                let q = idx % nq;
                let dst = (out_ptr as *mut f32).add(base + (head * lq + 1 + q) * h * w);
                let dx_row =
                    std::slice::from_raw_parts((dx_ptr as *const f32).add((head * nq + q) * w), w);
                let dy_row =
                    std::slice::from_raw_parts((dy_ptr as *const f32).add((head * nq + q) * h), h);
                for y in 0..h {
                    let dy = dy_row[y];
                    let row_dst = dst.add(y * w);
                    for x in 0..w {
                        *row_dst.add(x) = dy + dx_row[x];
                    }
                }
            }
        });
    }
    Ok(())
}

fn log_norm(v: f32) -> f32 {
    let s = if v < 0.0 { -1.0 } else { 1.0 };
    s * (v.abs() + 1.0).log2() / 8.0f32.log2()
}

/// Geometry-only delta computation. Output layout matches the IR
/// `deltas_x [B, nq, w, 2]` / `deltas_y [B, nq, h, 2]` inputs that
/// feed the boxRPB subgraph: per (batch, query, spatial-coord), pair
/// of `log_norm((coord - left)*8)` and `log_norm((coord - right)*8)`.
fn compute_deltas_into(
    reference_boxes: &[f32],
    batch: usize,
    nq: usize,
    h: usize,
    w: usize,
    deltas_x: &mut [f32],
    deltas_y: &mut [f32],
) {
    debug_assert_eq!(deltas_x.len(), batch * nq * w * 2);
    debug_assert_eq!(deltas_y.len(), batch * nq * h * 2);
    let coords_h: Vec<f32> = (0..h).map(|y| y as f32 / h as f32).collect();
    let coords_w: Vec<f32> = (0..w).map(|x| x as f32 / w as f32).collect();
    for b in 0..batch {
        for q in 0..nq {
            let p = &reference_boxes[(b * nq + q) * 4..(b * nq + q + 1) * 4];
            let (cx, cy, bw, bh) = (p[0], p[1], p[2], p[3]);
            let x0 = cx - 0.5 * bw;
            let x1 = cx + 0.5 * bw;
            let y0 = cy - 0.5 * bh;
            let y1 = cy + 0.5 * bh;
            let dx_off = ((b * nq + q) * w) * 2;
            for xi in 0..w {
                let dx0 = (coords_w[xi] - x0) * 8.0;
                let dx1 = (coords_w[xi] - x1) * 8.0;
                deltas_x[dx_off + xi * 2] = log_norm(dx0);
                deltas_x[dx_off + xi * 2 + 1] = log_norm(dx1);
            }
            let dy_off = ((b * nq + q) * h) * 2;
            for yi in 0..h {
                let dy0 = (coords_h[yi] - y0) * 8.0;
                let dy1 = (coords_h[yi] - y1) * 8.0;
                deltas_y[dy_off + yi * 2] = log_norm(dy0);
                deltas_y[dy_off + yi * 2 + 1] = log_norm(dy1);
            }
        }
    }
}

/// Standalone boxRPB subgraph for CPU parity checks (`build_boxrpb_subgraph`).
#[allow(dead_code)]
fn build_boxrpb_check_hir(
    boxrpb_x: &Mlp2,
    boxrpb_y: &Mlp2,
    batch: usize,
    nq: usize,
    h: usize,
    w: usize,
) -> Result<(HirModule, HashMap<String, Vec<f32>>)> {
    let nh = N_HEADS;
    let mut hir = HirModule::new("sam3_boxrpb_check");
    let mut params = HashMap::new();
    let mut typed = Vec::new();
    let mut gguf_cache = HashMap::new();
    {
        let mut g = HirMut::new(&mut hir);
        let f = DType::F32;
        let deltas_x = g.input("deltas_x", Shape::new(&[batch, nq, w, 2], f));
        let deltas_y = g.input("deltas_y", Shape::new(&[batch, nq, h, 2], f));
        let out = build_boxrpb_subgraph(
            &mut g,
            &mut params,
            &mut typed,
            &mut gguf_cache,
            None,
            boxrpb_x,
            boxrpb_y,
            deltas_x,
            deltas_y,
            batch,
            nq,
            nh,
            h,
            w,
        )?;
        g.set_outputs(vec![out]);
    }
    Ok((hir, params))
}

#[cfg(test)]
mod tests {
    use super::*;

    fn synth_mlp2(in_d: usize, hidden: usize, out_d: usize) -> Mlp2 {
        Mlp2 {
            w0_t: vec![0.01; in_d * hidden],
            b0: vec![0.0; hidden],
            w1_t: vec![0.02; hidden * out_d],
            b1: vec![0.0; out_d],
            in_dim: in_d,
            hidden,
            out_dim: out_d,
            w0_gguf_key: None,
            w1_gguf_key: None,
        }
    }

    #[test]
    fn sam3_boxrpb_ir_matches_host_cpu() -> Result<()> {
        let batch = 1usize;
        let nq = 2usize;
        let h = 4usize;
        let w = 4usize;
        let nh = N_HEADS;
        let boxrpb_x = synth_mlp2(2, 16, nh);
        let boxrpb_y = synth_mlp2(2, 16, nh);
        let ref_boxes = vec![
            0.5, 0.5, 0.4, 0.4, //
            0.3, 0.7, 0.2, 0.3,
        ];
        let host = boxrpb_log_full(&boxrpb_x, &boxrpb_y, &ref_boxes, batch, nq, h, w, None)?;

        let mut deltas_x = vec![0f32; batch * nq * w * 2];
        let mut deltas_y = vec![0f32; batch * nq * h * 2];
        compute_deltas_into(&ref_boxes, batch, nq, h, w, &mut deltas_x, &mut deltas_y);

        let (hir, params) = build_boxrpb_check_hir(&boxrpb_x, &boxrpb_y, batch, nq, h, w)?;
        let mut compiled = rlx_core::flow_bridge::compile_hir_sam(Device::Cpu, hir)?;
        for (name, data) in &params {
            compiled.set_param(name, data);
        }
        let ir = compiled
            .run(&[("deltas_x", &deltas_x), ("deltas_y", &deltas_y)])
            .into_iter()
            .next()
            .unwrap();

        let fd = host
            .iter()
            .zip(&ir)
            .map(|(a, b)| (a - b).abs())
            .fold(0f32, f32::max);
        assert!(fd < 5e-2, "sam3 boxRPB IR vs host max |Δ| = {fd:.3e}");
        Ok(())
    }
}