fuel-transformers 0.19.1

// SPDX-License-Identifier: MIT OR Apache-2.0
//! Persimmon decoder ported to the lazy-graph API.
//!
//! Phase D LLM port. Persimmon (Adept AI) combines several quirks:
//! - **LayerNorm with bias** (not RmsNorm) on pre-attention +
//!   pre-FFN paths.
//! - **QK-LayerNorm** — separate LayerNorm applied to projected Q
//!   and K BEFORE head reshape (`q_norm` shape `[hidden_size]`,
//!   `k_norm` shape `[num_kv_heads * head_dim]`).
//! - **Partial rotary** (factor 0.5 for the 8B base model). Reuses
//!   `Tensor::rope_partial`.
//! - **ReLU MLP** — `down(relu(up(x)))`, no gate path.
//! - **Q/K/V/O biases** always present (no gating flag).
//!
//! Custom `PersimmonLayerWeights` because the LN+bias surface and
//! QK-LN structure don't fit `LayerWeights`.

use fuel_ir::Shape;
use fuel_ir::error::Result;
use fuel_tensor::Device;
use fuel_tensor::{lazy::LayerNormPair, lazy::Tensor, lazy::WeightStorage};
use std::sync::Arc;

#[derive(Debug, Clone, PartialEq)]
pub struct PersimmonConfig {
    pub vocab_size: usize,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub num_hidden_layers: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: usize,
    pub head_dim: usize,
    pub layer_norm_eps: f64,
    pub rope_theta: f64,
    pub max_position_embeddings: usize,
    pub partial_rotary_factor: f64,
    pub qk_layernorm: bool,
}

impl PersimmonConfig {
    pub fn rope_dim(&self) -> usize {
        let rd = (self.head_dim as f64 * self.partial_rotary_factor) as usize;
        (rd / 2) * 2
    }
}

// ROADMAP item 8 (II): config-from-path on the #57 template. A `serde` raw with
// HF field names + Persimmon's own constant defaults, then `resolve` routes kv
// heads + head_dim through the shared `fuel_loaders::hf_config` rules. Persimmon uses
// LayerNorm (`layer_norm_eps`), partial rotary (`partial_rotary_factor`, HF
// default 0.5) and QK-LayerNorm (`qk_layernorm`, HF default true).
#[derive(Debug, Clone, serde::Deserialize)]
struct PersimmonConfigRaw {
    vocab_size: usize,
    hidden_size: usize,
    intermediate_size: usize,
    num_hidden_layers: usize,
    num_attention_heads: usize,
    #[serde(default)]
    num_key_value_heads: Option<usize>,
    #[serde(default)]
    head_dim: Option<usize>,
    #[serde(default = "default_persimmon_layer_norm_eps")]
    layer_norm_eps: f64,
    #[serde(default = "default_persimmon_rope_theta")]
    rope_theta: f64,
    max_position_embeddings: usize,
    #[serde(default = "default_persimmon_partial_rotary_factor")]
    partial_rotary_factor: f64,
    #[serde(default = "default_persimmon_qk_layernorm")]
    qk_layernorm: bool,
}

fn default_persimmon_layer_norm_eps() -> f64 {
    1e-5
}
fn default_persimmon_rope_theta() -> f64 {
    25_000.0
}
fn default_persimmon_partial_rotary_factor() -> f64 {
    0.5
}
fn default_persimmon_qk_layernorm() -> bool {
    true
}

impl PersimmonConfigRaw {
    fn from_json_str(json: &str) -> fuel_ir::error::Result<Self> {
        serde_json::from_str(json)
            .map_err(|e| fuel_ir::error::Error::Msg(format!("parsing Persimmon config.json: {e}")))
    }

    fn resolve(self) -> Result<PersimmonConfig> {
        Ok(PersimmonConfig {
            vocab_size: self.vocab_size,
            hidden_size: self.hidden_size,
            intermediate_size: self.intermediate_size,
            num_hidden_layers: self.num_hidden_layers,
            num_attention_heads: self.num_attention_heads,
            num_key_value_heads: fuel_loaders::hf_config::num_key_value_heads(
                self.num_key_value_heads,
                self.num_attention_heads,
            )?,
            head_dim: fuel_loaders::hf_config::head_dim(
                self.head_dim,
                self.hidden_size,
                self.num_attention_heads,
            )?,
            layer_norm_eps: self.layer_norm_eps,
            rope_theta: self.rope_theta,
            max_position_embeddings: self.max_position_embeddings,
            partial_rotary_factor: self.partial_rotary_factor,
            qk_layernorm: self.qk_layernorm,
        })
    }
}

impl PersimmonConfig {
    /// Parse a HuggingFace `config.json` string into a [`PersimmonConfig`].
    ///
    /// ROADMAP item 8 (II): reads the artifact rather than returning a preset —
    /// see the born-red `persimmon_config_from_hf_json_parses_the_artifact`.
    pub fn from_hf_json_str(json: &str) -> fuel_ir::error::Result<Self> {
        PersimmonConfigRaw::from_json_str(json)?.resolve()
    }
}

#[derive(Debug, Clone)]
pub struct PersimmonLayerWeights {
    pub input_ln_gain: Arc<[f32]>,
    pub input_ln_bias: Arc<[f32]>,
    pub post_attn_ln_gain: Arc<[f32]>,
    pub post_attn_ln_bias: Arc<[f32]>,
    pub attn_q: WeightStorage,
    pub attn_q_bias: Arc<[f32]>,
    pub attn_k: WeightStorage,
    pub attn_k_bias: Arc<[f32]>,
    pub attn_v: WeightStorage,
    pub attn_v_bias: Arc<[f32]>,
    pub attn_o: WeightStorage,
    pub attn_o_bias: Arc<[f32]>,
    /// QK-LN gain + bias (present iff `cfg.qk_layernorm`).
    pub q_norm: Option<LayerNormPair>,
    pub k_norm: Option<LayerNormPair>,
    /// `[hidden_size, intermediate_size]`.
    pub mlp_up: WeightStorage,
    pub mlp_up_bias: Arc<[f32]>,
    /// `[intermediate_size, hidden_size]`.
    pub mlp_down: WeightStorage,
    pub mlp_down_bias: Arc<[f32]>,
}

#[derive(Debug, Clone)]
pub struct PersimmonWeights {
    pub token_embedding: Arc<[f32]>,
    pub layers: Vec<PersimmonLayerWeights>,
    pub final_ln_gain: Arc<[f32]>,
    pub final_ln_bias: Arc<[f32]>,
    pub output: WeightStorage,
}

#[derive(Debug, Clone)]
pub struct PersimmonModel {
    pub config: PersimmonConfig,
    pub weights: PersimmonWeights,
}

impl PersimmonModel {
    pub fn forward(&self, tokens: &[u32], start_pos: usize) -> Result<Tensor> {
        let h_norm = self.run_backbone(tokens, start_pos)?;
        self.apply_lm_head(&h_norm)
    }

    /// Run the decoder forward up to the final LayerNorm and
    /// return per-token hidden states `(1, seq, hidden_size)`.
    /// Persimmon uses partial RoPE + LayerNorm (gain+bias).
    pub fn forward_hidden(&self, tokens: &[u32], start_pos: usize) -> Result<Tensor> {
        self.run_backbone(tokens, start_pos)
    }

    /// Multimodal entry point. Skips token embedding; runs the decoder
    /// over pre-embedded inputs. Persimmon does NOT scale embeddings.
    pub fn forward_embeds(&self, embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        let h_norm = self.run_backbone_embeds(embeds, start_pos)?;
        self.apply_lm_head(&h_norm)
    }

    /// Hidden-state variant of [`Self::forward_embeds`].
    pub fn forward_hidden_embeds(&self, embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        self.run_backbone_embeds(embeds, start_pos)
    }

    /// Build per-token embeddings without running the decoder.
    pub fn embed_tokens_anchored(&self, anchor: &Tensor, tokens: &[u32]) -> Result<Tensor> {
        let cfg = &self.config;
        anchor.embed_tokens_anchored(
            self.weights.token_embedding.clone(),
            cfg.vocab_size,
            cfg.hidden_size,
            tokens,
        )
    }

    fn apply_lm_head(&self, h_norm: &Tensor) -> Result<Tensor> {
        let cfg = &self.config;
        self.weights
            .output
            .apply_linear(h_norm, cfg.hidden_size, cfg.vocab_size)
    }

    fn run_backbone(&self, tokens: &[u32], start_pos: usize) -> Result<Tensor> {
        let cfg = &self.config;
        let weights = &self.weights;
        let seq = tokens.len();
        assert!(seq > 0);

        let h = Tensor::embed_tokens(
            weights.token_embedding.clone(),
            cfg.vocab_size,
            cfg.hidden_size,
            tokens,
            &Device::cpu(),
        )?;
        self.run_backbone_embeds(&h, start_pos)
    }

    fn run_backbone_embeds(&self, embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        let cfg = &self.config;
        let weights = &self.weights;
        let dims = embeds.shape();
        let dims = dims.dims();
        if dims.len() != 3 || dims[0] != 1 || dims[2] != cfg.hidden_size {
            return Err(fuel_ir::error::Error::Msg(format!(
                "PersimmonModel::forward_embeds: expected embeds shape (1, seq, hidden_size={}), got {:?}",
                cfg.hidden_size, dims,
            )).bt());
        }
        let seq = dims[1];
        if seq == 0 {
            return Err(fuel_ir::error::Error::Msg(
                "PersimmonModel::forward_embeds: seq must be > 0".into(),
            )
            .bt());
        }
        if cfg.num_attention_heads * cfg.head_dim != cfg.hidden_size {
            return Err(fuel_ir::error::Error::Msg(
                "PersimmonConfig: num_attention_heads * head_dim must equal hidden_size".into(),
            )
            .bt());
        }
        let mut h = embeds.clone();

        let rope_dim = cfg.rope_dim();
        let (rope_cos, rope_sin) = h.rope_tables_const(cfg.rope_theta, start_pos, seq, rope_dim);

        for layer in &weights.layers {
            h = self.apply_layer(&h, layer, &rope_cos, &rope_sin)?;
        }
        h.layer_norm_affine(
            std::sync::Arc::clone(&weights.final_ln_gain),
            std::sync::Arc::clone(&weights.final_ln_bias),
            cfg.layer_norm_eps,
        )
    }

    fn apply_layer(
        &self,
        x: &Tensor,
        layer: &PersimmonLayerWeights,
        rope_cos: &Tensor,
        rope_sin: &Tensor,
    ) -> Result<Tensor> {
        let cfg = &self.config;
        let head_dim = cfg.head_dim;
        let x_shape = x.shape();
        let dims = x_shape.dims();
        let batch = dims[0];
        let seq = dims[1];
        let kv_dim = cfg.num_key_value_heads * head_dim;

        let x_norm = x.layer_norm_affine(
            std::sync::Arc::clone(&layer.input_ln_gain),
            std::sync::Arc::clone(&layer.input_ln_bias),
            cfg.layer_norm_eps,
        )?;

        // Q/K/V projections — always have biases on Persimmon.
        let q = layer.attn_q.apply_linear_with_bias(
            &x_norm,
            cfg.hidden_size,
            cfg.hidden_size,
            std::sync::Arc::clone(&layer.attn_q_bias),
        )?;
        let k = layer.attn_k.apply_linear_with_bias(
            &x_norm,
            cfg.hidden_size,
            kv_dim,
            std::sync::Arc::clone(&layer.attn_k_bias),
        )?;
        let v = layer.attn_v.apply_linear_with_bias(
            &x_norm,
            cfg.hidden_size,
            kv_dim,
            std::sync::Arc::clone(&layer.attn_v_bias),
        )?;

        // QK-LayerNorm BEFORE head reshape.
        let (q, k) = match (&layer.q_norm, &layer.k_norm) {
            (Some((qg, qb)), Some((kg, kb))) => {
                let q = q.layer_norm_affine(
                    std::sync::Arc::clone(qg),
                    std::sync::Arc::clone(qb),
                    cfg.layer_norm_eps,
                )?;
                let k = k.layer_norm_affine(
                    std::sync::Arc::clone(kg),
                    std::sync::Arc::clone(kb),
                    cfg.layer_norm_eps,
                )?;
                (q, k)
            }
            _ => (q, k),
        };

        let _ = (batch, seq);
        let q = q.split_heads(cfg.num_attention_heads, head_dim)?;
        let k = k.split_heads(cfg.num_key_value_heads, head_dim)?;
        let v = v.split_heads(cfg.num_key_value_heads, head_dim)?;

        let q_r = q.rope_partial(rope_cos, rope_sin, cfg.rope_dim())?;
        let k_r = k.rope_partial(rope_cos, rope_sin, cfg.rope_dim())?;

        let n_rep = cfg.num_attention_heads / cfg.num_key_value_heads;
        let k_full = k_r.repeat_interleave(1_usize, n_rep)?;
        let v_full = v.repeat_interleave(1_usize, n_rep)?;

        let k_t = k_full.transpose()?;
        let scale = 1.0_f64 / (head_dim as f64).sqrt();
        let scores = q_r.matmul(&k_t)?;
        let scores_scaled = scores.mul_scalar(scale);
        let mask = Tensor::additive_causal_mask_like(x, seq)
            .reshape(Shape::from_dims(&[1, 1, seq, seq]))?;
        let scores_masked = scores_scaled.broadcast_add(&mask)?;
        let attn = scores_masked.softmax_last_dim()?;
        let attn_v = attn.matmul(&v_full)?;

        let merged = attn_v.merge_heads()?;
        let attn_out = layer.attn_o.apply_linear_with_bias(
            &merged,
            cfg.hidden_size,
            cfg.hidden_size,
            std::sync::Arc::clone(&layer.attn_o_bias),
        )?;

        let h1 = x.add(&attn_out)?;
        let h1_norm = h1.layer_norm_affine(
            std::sync::Arc::clone(&layer.post_attn_ln_gain),
            std::sync::Arc::clone(&layer.post_attn_ln_bias),
            cfg.layer_norm_eps,
        )?;
        // MLP: simple `down(relu(up(x)))`.
        let up = layer.mlp_up.apply_linear_with_bias(
            &h1_norm,
            cfg.hidden_size,
            cfg.intermediate_size,
            std::sync::Arc::clone(&layer.mlp_up_bias),
        )?;
        let up_act = up.relu();
        let ffn_out = layer.mlp_down.apply_linear_with_bias(
            &up_act,
            cfg.intermediate_size,
            cfg.hidden_size,
            std::sync::Arc::clone(&layer.mlp_down_bias),
        )?;
        h1.add(&ffn_out)
    }
}

// ---- HuggingFace safetensors loader ----------------------------------------

impl PersimmonWeights {
    /// Load Persimmon (adept/persimmon-8b-{chat,base}) weights from HF safetensors.
    pub fn load_from_mmapped(
        st: &fuel_loaders::safetensors::MmapedSafetensors,
        cfg: &PersimmonConfig,
    ) -> Result<Self> {
        use fuel_tensor::{
            lazy::load_tensor_as_f32, lazy::load_transposed_matrix_preserve_dtype as ltm,
        };
        let h = cfg.hidden_size;
        let inter = cfg.intermediate_size;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        let kv_dim = cfg.num_key_value_heads * cfg.head_dim;

        let token_embedding = Arc::from(load_tensor_as_f32(st, "model.embed_tokens.weight")?);

        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
        for i in 0..cfg.num_hidden_layers {
            let p = format!("model.layers.{i}");
            let input_ln_gain = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.input_layernorm.weight"),
            )?);
            let input_ln_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.input_layernorm.bias"),
            )?);
            let post_attn_ln_gain = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.post_attention_layernorm.weight"),
            )?);
            let post_attn_ln_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.post_attention_layernorm.bias"),
            )?);
            let attn_q = ltm(st, &format!("{p}.self_attn.q_proj.weight"), q_dim, h)?;
            let attn_q_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.self_attn.q_proj.bias"),
            )?);
            let attn_k = ltm(st, &format!("{p}.self_attn.k_proj.weight"), kv_dim, h)?;
            let attn_k_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.self_attn.k_proj.bias"),
            )?);
            let attn_v = ltm(st, &format!("{p}.self_attn.v_proj.weight"), kv_dim, h)?;
            let attn_v_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.self_attn.v_proj.bias"),
            )?);
            let attn_o = ltm(st, &format!("{p}.self_attn.dense.weight"), h, q_dim)?;
            let attn_o_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.self_attn.dense.bias"),
            )?);

            let q_norm = if cfg.qk_layernorm {
                let g = Arc::from(load_tensor_as_f32(
                    st,
                    &format!("{p}.self_attn.q_layernorm.weight"),
                )?);
                let b = Arc::from(load_tensor_as_f32(
                    st,
                    &format!("{p}.self_attn.q_layernorm.bias"),
                )?);
                Some((g, b))
            } else {
                None
            };
            let k_norm = if cfg.qk_layernorm {
                let g = Arc::from(load_tensor_as_f32(
                    st,
                    &format!("{p}.self_attn.k_layernorm.weight"),
                )?);
                let b = Arc::from(load_tensor_as_f32(
                    st,
                    &format!("{p}.self_attn.k_layernorm.bias"),
                )?);
                Some((g, b))
            } else {
                None
            };

            let mlp_up = ltm(st, &format!("{p}.mlp.dense_h_to_4h.weight"), inter, h)?;
            let mlp_up_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.mlp.dense_h_to_4h.bias"),
            )?);
            let mlp_down = ltm(st, &format!("{p}.mlp.dense_4h_to_h.weight"), h, inter)?;
            let mlp_down_bias = Arc::from(load_tensor_as_f32(
                st,
                &format!("{p}.mlp.dense_4h_to_h.bias"),
            )?);

            layers.push(PersimmonLayerWeights {
                input_ln_gain,
                input_ln_bias,
                post_attn_ln_gain,
                post_attn_ln_bias,
                attn_q,
                attn_q_bias,
                attn_k,
                attn_k_bias,
                attn_v,
                attn_v_bias,
                attn_o,
                attn_o_bias,
                q_norm,
                k_norm,
                mlp_up,
                mlp_up_bias,
                mlp_down,
                mlp_down_bias,
            });
        }

        let final_ln_gain = Arc::from(load_tensor_as_f32(st, "model.final_layernorm.weight")?);
        let final_ln_bias = Arc::from(load_tensor_as_f32(st, "model.final_layernorm.bias")?);
        let output = ltm(st, "lm_head.weight", cfg.vocab_size, h)?;

        Ok(Self {
            token_embedding,
            layers,
            final_ln_gain,
            final_ln_bias,
            output,
        })
    }
}

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

    // ROADMAP item 8 (II). Golden values from adept/persimmon-8b-base's real
    // config.json. Persimmon has no size preset, so a second distinct config
    // provides the constant-parser discrimination. `partial_rotary_factor` is
    // absent from the artifact → defaulted to 0.5; `qk_layernorm` is present.
    const PERSIMMON_8B_CONFIG_JSON: &str = r#"{
        "architectures": ["PersimmonForCausalLM"],
        "model_type": "persimmon",
        "vocab_size": 262144,
        "hidden_size": 4096,
        "intermediate_size": 16384,
        "num_hidden_layers": 36,
        "num_attention_heads": 64,
        "num_key_value_heads": 64,
        "max_position_embeddings": 16384,
        "layer_norm_eps": 1e-05,
        "rms_norm_eps": 1e-06,
        "rope_theta": 25000.0,
        "qk_layernorm": true
    }"#;

    #[test]
    fn persimmon_config_from_hf_json_parses_the_artifact() {
        let cfg = PersimmonConfig::from_hf_json_str(PERSIMMON_8B_CONFIG_JSON).unwrap();
        assert_eq!(cfg.hidden_size, 4096);
        assert_eq!(cfg.num_hidden_layers, 36);
        assert_eq!(cfg.num_attention_heads, 64);
        assert_eq!(cfg.vocab_size, 262_144);
        assert_eq!(cfg.intermediate_size, 16_384);
        assert_eq!(cfg.num_key_value_heads, 64);
        // head_dim absent → derived 4096/64 = 64.
        assert_eq!(cfg.head_dim, 64);
        assert_eq!(cfg.layer_norm_eps, 1e-5);
        assert_eq!(cfg.rope_theta, 25_000.0);
        assert!(cfg.qk_layernorm); // present in the artifact
        // partial_rotary_factor absent → defaulted to 0.5.
        assert_eq!(cfg.partial_rotary_factor, 0.5);
    }

    /// A SECOND distinct config, exercising the default path (layer_norm_eps/
    /// rope_theta/partial_rotary_factor/qk_layernorm omitted) AND the take-if-
    /// present head_dim branch (explicit 80 ≠ 1024/16 = 64).
    #[test]
    fn persimmon_config_reads_a_second_distinct_config_with_explicit_head_dim() {
        let json = r#"{
            "model_type": "persimmon",
            "vocab_size": 50000,
            "hidden_size": 1024,
            "intermediate_size": 4096,
            "num_hidden_layers": 8,
            "num_attention_heads": 16,
            "num_key_value_heads": 4,
            "head_dim": 80,
            "max_position_embeddings": 2048
        }"#;
        let cfg = PersimmonConfig::from_hf_json_str(json).unwrap();
        assert_eq!(cfg.num_key_value_heads, 4);
        // explicit head_dim WINS over the quotient (1024/16 = 64).
        assert_eq!(cfg.head_dim, 80);
        assert_ne!(cfg.head_dim, 1024 / 16);
        // omitted → resolve defaults (Persimmon's own consts)
        assert_eq!(cfg.layer_norm_eps, 1e-5);
        assert_eq!(cfg.rope_theta, 25_000.0);
        assert_eq!(cfg.partial_rotary_factor, 0.5);
        assert!(cfg.qk_layernorm); // default true
        assert_ne!(cfg.hidden_size, 4096);
    }

    /// `num_key_value_heads` ABSENT → defaults to `num_attention_heads`.
    #[test]
    fn persimmon_config_gqa_defaults_to_num_heads_when_absent() {
        let json = r#"{
            "model_type": "persimmon",
            "vocab_size": 1000,
            "hidden_size": 64,
            "intermediate_size": 128,
            "num_hidden_layers": 2,
            "num_attention_heads": 8,
            "max_position_embeddings": 128
        }"#;
        let cfg = PersimmonConfig::from_hf_json_str(json).unwrap();
        assert_eq!(cfg.num_key_value_heads, 8);
    }

    /// TRUE MQA (`num_key_value_heads = 1`) survives, not collapsed.
    #[test]
    fn persimmon_config_preserves_true_mqa() {
        let json = r#"{
            "model_type": "persimmon",
            "vocab_size": 1000,
            "hidden_size": 64,
            "intermediate_size": 128,
            "num_hidden_layers": 2,
            "num_attention_heads": 8,
            "num_key_value_heads": 1,
            "max_position_embeddings": 128
        }"#;
        let cfg = PersimmonConfig::from_hf_json_str(json).unwrap();
        assert_eq!(cfg.num_key_value_heads, 1);
    }

    fn tiny_weights(cfg: &PersimmonConfig) -> PersimmonWeights {
        let mut s: u32 = 77777;
        let next = || -> f32 {
            s = s.wrapping_mul(1103515245).wrapping_add(12345);
            ((s >> 16) as u16 as f32 / 65535.0 - 0.5) * 0.05
        };
        let vec_of = |n: usize, next: &mut dyn FnMut() -> f32| -> Arc<[f32]> {
            Arc::from((0..n).map(|_| next()).collect::<Vec<_>>())
        };
        let h = cfg.hidden_size;
        let i = cfg.intermediate_size;
        let kv = cfg.num_key_value_heads * cfg.head_dim;
        let mut nb: Box<dyn FnMut() -> f32> = Box::new(next);
        let token_embedding = vec_of(cfg.vocab_size * h, &mut *nb);
        let layers: Vec<PersimmonLayerWeights> = (0..cfg.num_hidden_layers)
            .map(|_| PersimmonLayerWeights {
                input_ln_gain: Arc::from(vec![1.0_f32; h]),
                input_ln_bias: Arc::from(vec![0.0_f32; h]),
                post_attn_ln_gain: Arc::from(vec![1.0_f32; h]),
                post_attn_ln_bias: Arc::from(vec![0.0_f32; h]),
                attn_q: WeightStorage::F32(vec_of(h * h, &mut *nb)),
                attn_q_bias: vec_of(h, &mut *nb),
                attn_k: WeightStorage::F32(vec_of(h * kv, &mut *nb)),
                attn_k_bias: vec_of(kv, &mut *nb),
                attn_v: WeightStorage::F32(vec_of(h * kv, &mut *nb)),
                attn_v_bias: vec_of(kv, &mut *nb),
                attn_o: WeightStorage::F32(vec_of(h * h, &mut *nb)),
                attn_o_bias: vec_of(h, &mut *nb),
                q_norm: if cfg.qk_layernorm {
                    Some((Arc::from(vec![1.0_f32; h]), Arc::from(vec![0.0_f32; h])))
                } else {
                    None
                },
                k_norm: if cfg.qk_layernorm {
                    Some((Arc::from(vec![1.0_f32; kv]), Arc::from(vec![0.0_f32; kv])))
                } else {
                    None
                },
                mlp_up: WeightStorage::F32(vec_of(h * i, &mut *nb)),
                mlp_up_bias: vec_of(i, &mut *nb),
                mlp_down: WeightStorage::F32(vec_of(i * h, &mut *nb)),
                mlp_down_bias: vec_of(h, &mut *nb),
            })
            .collect();
        let final_ln_gain = Arc::from(vec![1.0_f32; h]);
        let final_ln_bias = Arc::from(vec![0.0_f32; h]);
        let output = WeightStorage::F32(vec_of(h * cfg.vocab_size, &mut *nb));
        PersimmonWeights {
            token_embedding,
            layers,
            final_ln_gain,
            final_ln_bias,
            output,
        }
    }

    #[test]
    fn forward_with_qk_layernorm_and_partial_rotary() {
        let cfg = PersimmonConfig {
            vocab_size: 32,
            hidden_size: 16,
            intermediate_size: 32,
            num_hidden_layers: 2,
            num_attention_heads: 4,
            num_key_value_heads: 4,
            head_dim: 4,
            layer_norm_eps: 1e-5,
            rope_theta: 25_000.0,
            max_position_embeddings: 64,
            partial_rotary_factor: 1.0,
            qk_layernorm: true,
        };
        let model = PersimmonModel {
            config: cfg.clone(),
            weights: tiny_weights(&cfg),
        };
        let logits = model.forward(&[1, 2, 3], 0).unwrap();
        assert_eq!(logits.shape().dims(), &[1, 3, cfg.vocab_size]);
        for &v in &logits.realize_f32() {
            assert!(v.is_finite());
        }
    }

    #[test]
    fn forward_hidden_shape_and_finite() {
        let cfg = PersimmonConfig {
            vocab_size: 32,
            hidden_size: 16,
            intermediate_size: 32,
            num_hidden_layers: 2,
            num_attention_heads: 4,
            num_key_value_heads: 4,
            head_dim: 4,
            layer_norm_eps: 1e-5,
            rope_theta: 25_000.0,
            max_position_embeddings: 64,
            partial_rotary_factor: 1.0,
            qk_layernorm: true,
        };
        let model = PersimmonModel {
            config: cfg.clone(),
            weights: tiny_weights(&cfg),
        };
        let tokens: Vec<u32> = vec![1, 2, 3, 4];
        let hidden = model.forward_hidden(&tokens, 0).unwrap();
        assert_eq!(hidden.shape().dims(), &[1, tokens.len(), cfg.hidden_size]);
        for &v in &hidden.realize_f32() {
            assert!(v.is_finite(), "non-finite hidden: {v}");
        }
    }

    fn forward_embeds_test_cfg() -> PersimmonConfig {
        PersimmonConfig {
            vocab_size: 32,
            hidden_size: 16,
            intermediate_size: 32,
            num_hidden_layers: 2,
            num_attention_heads: 4,
            num_key_value_heads: 4,
            head_dim: 4,
            layer_norm_eps: 1e-5,
            rope_theta: 25_000.0,
            max_position_embeddings: 64,
            partial_rotary_factor: 1.0,
            qk_layernorm: true,
        }
    }

    #[test]
    fn forward_embeds_matches_forward_after_token_lookup() {
        let cfg = forward_embeds_test_cfg();
        let model = PersimmonModel {
            config: cfg.clone(),
            weights: tiny_weights(&cfg),
        };
        let tokens: Vec<u32> = vec![1, 2, 3];
        let logits_ref = model.forward(&tokens, 0).unwrap().realize_f32();
        let anchor =
            Tensor::from_f32(vec![0.0_f32], Shape::from_dims(&[1]), &Device::cpu()).unwrap();
        let embeds = model.embed_tokens_anchored(&anchor, &tokens).unwrap();
        let logits_via_embeds = model.forward_embeds(&embeds, 0).unwrap().realize_f32();
        let max_diff = logits_ref
            .iter()
            .zip(logits_via_embeds.iter())
            .map(|(a, b)| (a - b).abs())
            .fold(0.0_f32, f32::max);
        assert!(
            max_diff < 1e-5,
            "Persimmon forward vs forward_embeds must agree (max diff {max_diff})"
        );
    }

    #[test]
    fn forward_embeds_rejects_bad_shape() {
        let cfg = forward_embeds_test_cfg();
        let model = PersimmonModel {
            config: cfg.clone(),
            weights: tiny_weights(&cfg),
        };
        let bad = Tensor::from_f32(
            vec![0.0_f32; 3 * (cfg.hidden_size + 1)],
            Shape::from_dims(&[1, 3, cfg.hidden_size + 1]),
            &Device::cpu(),
        )
        .unwrap();
        assert!(model.forward_embeds(&bad, 0).is_err());
    }

    #[test]
    fn forward_hidden_embeds_matches_forward_hidden() {
        let cfg = forward_embeds_test_cfg();
        let model = PersimmonModel {
            config: cfg.clone(),
            weights: tiny_weights(&cfg),
        };
        let tokens: Vec<u32> = vec![5, 7];
        let h_ref = model.forward_hidden(&tokens, 0).unwrap().realize_f32();
        let anchor =
            Tensor::from_f32(vec![0.0_f32], Shape::from_dims(&[1]), &Device::cpu()).unwrap();
        let embeds = model.embed_tokens_anchored(&anchor, &tokens).unwrap();
        let h_via_embeds = model
            .forward_hidden_embeds(&embeds, 0)
            .unwrap()
            .realize_f32();
        let max_diff = h_ref
            .iter()
            .zip(h_via_embeds.iter())
            .map(|(a, b)| (a - b).abs())
            .fold(0.0_f32, f32::max);
        assert!(
            max_diff < 1e-5,
            "Persimmon forward_hidden vs forward_hidden_embeds must agree (max diff {max_diff})"
        );
    }
}