llm_manager/tui/

settings.rs

use crate::config::Profile;
use crate::models::{CacheQuantType, GpuLayersMode, Mirostat, ModelSettings, NumMode, SplitMode};
use crate::tui::colors::*;
use crate::tui::format_context_k;
use ratatui::{
    style::{Modifier, Style},
    text::{Line, Span},
};

// ── Function pointer types (zero-cost, no dynamic dispatch) ──────────────────

pub type DisplayFn = fn(&ModelSettings) -> String;
pub type DirtyFn = fn(&ModelSettings, &ModelSettings) -> bool;
pub type AdjustFn = fn(&mut ModelSettings, i32, u32); // u32 = context_limit (0 = no limit)
pub type ApplyEditFn = fn(&mut ModelSettings, &str);
pub type CtrlEToggleFn = fn(&mut ModelSettings);

// ── SettingField ─────────────────────────────────────────────────────────────

pub struct SettingField {
    pub id: &'static str,
    pub name: &'static str,
    pub section: &'static str,
    pub display: DisplayFn,
    pub dirty: DirtyFn,
    pub adjust: AdjustFn,
    pub apply_edit: ApplyEditFn,
    pub ctrl_e_toggle: Option<CtrlEToggleFn>,
    pub is_expert: bool,
    pub is_ultra: bool,
    pub is_enabled: Option<fn(&ModelSettings) -> bool>,
}

impl SettingField {
    pub fn name(&self) -> &str {
        self.name
    }

    pub fn display(&self, settings: &ModelSettings) -> String {
        (self.display)(settings)
    }

    pub fn is_dirty(&self, settings: &ModelSettings, cached: &ModelSettings) -> bool {
        (self.dirty)(settings, cached)
    }

    pub fn adjust(&self, settings: &mut ModelSettings, delta: i32, context_limit: u32) {
        (self.adjust)(settings, delta, context_limit);
    }

    pub fn apply_edit(&self, settings: &mut ModelSettings, buf: &str) {
        (self.apply_edit)(settings, buf);
    }

    pub fn ctrl_e_toggle(&self, settings: &mut ModelSettings) {
        if let Some(toggle) = self.ctrl_e_toggle {
            toggle(settings);
        }
    }

    /// Returns true if this field starts a new section (different from the previous field's section).
    pub fn is_new_section(&self, prev_section: Option<&str>) -> bool {
        Some(self.section) != prev_section
    }
}

// ── Helper constructors (generated from macro) ───────────────────────────────

/// Generate a field constructor function.
/// Variants:
///   - `field` / `expert_field` / `ultra_field` — no ctrl_e_toggle
///   - `field_with_toggle` / `expert_field_with_toggle` / `ultra_field_with_toggle` — with ctrl_e_toggle
macro_rules! make_field_fn {
    ($fn:ident, $expert:expr, $ultra:expr, toggle) => {
        fn $fn(
            id: &'static str,
            name: &'static str,
            section: &'static str,
            display: DisplayFn,
            dirty: DirtyFn,
            adjust: AdjustFn,
            apply_edit: ApplyEditFn,
            ctrl_e_toggle: CtrlEToggleFn,
            _help_text: &'static str,
        ) -> SettingField {
            SettingField {
                id,
                name,
                section,
                display,
                dirty,
                adjust,
                apply_edit,
                ctrl_e_toggle: Some(ctrl_e_toggle),
                is_expert: $expert,
                is_ultra: $ultra,
                is_enabled: None,
            }
        }
    };
    ($fn:ident, $expert:expr, $ultra:expr, @none) => {
        fn $fn(
            id: &'static str,
            name: &'static str,
            section: &'static str,
            display: DisplayFn,
            dirty: DirtyFn,
            adjust: AdjustFn,
            apply_edit: ApplyEditFn,
            _help_text: &'static str,
        ) -> SettingField {
            SettingField {
                id,
                name,
                section,
                display,
                dirty,
                adjust,
                apply_edit,
                ctrl_e_toggle: None,
                is_expert: $expert,
                is_ultra: $ultra,
                is_enabled: None,
            }
        }
    };
}

make_field_fn!(field, false, false, @none);
make_field_fn!(expert_field, true, false, @none);
make_field_fn!(ultra_field, true, true, @none);
make_field_fn!(field_with_toggle, false, false, toggle);
make_field_fn!(expert_field_with_toggle, true, false, toggle);
make_field_fn!(ultra_field_with_toggle, true, true, toggle);

// ── Shared adjustment and toggle logic ───────────────────────────────────────

fn gpu_layers_adjust(settings: &mut ModelSettings, delta: i32, _context_limit: u32) {
    settings.gpu_layers_mode = match (delta, &settings.gpu_layers_mode) {
        (1, GpuLayersMode::Auto) => GpuLayersMode::Specific(1),
        (1, GpuLayersMode::Specific(n)) => GpuLayersMode::Specific(n + 1),
        (1, GpuLayersMode::All) => GpuLayersMode::Auto,
        (-1, GpuLayersMode::Auto) => GpuLayersMode::Auto,
        (-1, GpuLayersMode::Specific(n)) if *n == 0 => GpuLayersMode::Auto,
        (-1, GpuLayersMode::Specific(n)) if *n == 1 => GpuLayersMode::Specific(0),
        (-1, GpuLayersMode::Specific(n)) => GpuLayersMode::Specific(n - 1),
        (-1, GpuLayersMode::All) => GpuLayersMode::All,
        _ => settings.gpu_layers_mode,
    };
}

fn gpu_layers_apply(settings: &mut ModelSettings, buf: &str) {
    if let Ok(v) = buf.parse::<i32>() {
        settings.gpu_layers_mode = if v < 0 {
            GpuLayersMode::All
        } else {
            GpuLayersMode::Specific(v as u32)
        };
    }
}

fn toggle_mlock(settings: &mut ModelSettings) {
    settings.mlock = !settings.mlock;
}
fn toggle_flash_attn(settings: &mut ModelSettings) {
    settings.flash_attn = !settings.flash_attn;
}
fn toggle_auto_chat_template(settings: &mut ModelSettings) {
    settings.auto_chat_template = !settings.auto_chat_template;
}

fn toggle_fit(settings: &mut ModelSettings) {
    settings.fit = !settings.fit;
}
fn toggle_kv_cache_offload(settings: &mut ModelSettings) {
    settings.kv_cache_offload = !settings.kv_cache_offload;
}
fn toggle_uniform_cache(settings: &mut ModelSettings) {
    settings.uniform_cache = !settings.uniform_cache;
}
fn toggle_swa_full(settings: &mut ModelSettings) {
    settings.swa_full = !settings.swa_full;
}
fn toggle_mtp(settings: &mut ModelSettings) {
    if settings.spec_type.is_empty() {
        settings.spec_type = "draft-mtp".to_string();
    } else {
        settings.spec_type = String::new();
    }
}

fn toggle_rope_yarn_enabled(settings: &mut ModelSettings) {
    settings.rope_yarn_enabled = !settings.rope_yarn_enabled;
}
fn toggle_ignore_eos(settings: &mut ModelSettings) {
    settings.ignore_eos = !settings.ignore_eos;
}
fn toggle_max_tokens(settings: &mut ModelSettings) {
    settings.max_tokens = settings.max_tokens.map_or(Some(2048), |_| None);
}
fn toggle_max_concurrent_predictions(settings: &mut ModelSettings) {
    settings.max_concurrent_predictions = settings
        .max_concurrent_predictions
        .map_or(Some(1), |_| None);
}
fn toggle_cache_type_k(settings: &mut ModelSettings) {
    settings.cache_type_k = settings
        .cache_type_k
        .map_or(Some(CacheQuantType::F16), |_| None);
}
fn toggle_cache_type_v(settings: &mut ModelSettings) {
    settings.cache_type_v = settings
        .cache_type_v
        .map_or(Some(CacheQuantType::F16), |_| None);
}
fn toggle_expert_count(settings: &mut ModelSettings) {
    settings.expert_count = match settings.expert_count {
        0 => 1,
        -1 => 0,
        _ => -1,
    };
}
fn toggle_presence_penalty(settings: &mut ModelSettings) {
    settings.presence_penalty = settings.presence_penalty.map_or(Some(0.0), |_| None);
}
fn toggle_frequency_penalty(settings: &mut ModelSettings) {
    settings.frequency_penalty = settings.frequency_penalty.map_or(Some(0.0), |_| None);
}

// ── Diff macros for profile settings comparison ──────────────────────────────

macro_rules! diff_int {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(v) = $s.$field
            && v != $c.$field
        {
            $parts.push(format!("{}={}", $label, v));
        }
    };
}
macro_rules! diff_float {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(v) = $s.$field
            && (v - $c.$field).abs() > 0.001
        {
            $parts.push(format!("{}={:.2}", $label, v));
        }
    };
}
macro_rules! diff_bool {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(v) = $s.$field
            && v != $c.$field
        {
            $parts.push(format!("{}={}", $label, v));
        }
    };
}
macro_rules! diff_string {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(v) = &$s.$field
            && v != &$c.$field
        {
            $parts.push(format!("{}={}", $label, v));
        }
    };
}
macro_rules! diff_enum {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(ref v) = $s.$field
            && *v != $c.$field
        {
            $parts.push(format!("{}={}", $label, v));
        }
    };
}
macro_rules! diff_option {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if $s.$field != $c.$field {
            if let Some(ref v) = $s.$field {
                $parts.push(format!("{}={}", $label, v));
            }
        }
    };
}
macro_rules! diff_option_float {
    ($parts:expr, $s:expr, $c:expr, $field:ident, $label:literal) => {
        if let Some(v) = $s.$field {
            let current_val = $c.$field.unwrap_or(0.0);
            if (v - current_val).abs() > 0.001 {
                $parts.push(format!("{}={:.2}", $label, v));
            }
        }
    };
}

// ── All Fields (Interleaved for context-aware expert mode) ────────────────────

pub fn all_fields() -> Vec<SettingField> {
    vec![
        // ── Loading ───────────────────────────────────────────────────────────
        field(
            "system_prompt_preset_name",
            "Prompt",
            "Loading",
            |s| s.system_prompt_preset_name.clone(),
            |s, c| s.system_prompt_preset_name != c.system_prompt_preset_name,
            |_, _, _| {},
            |_, _| {},
            "System prompt preset. Pre-configured prompts that shape how the model behaves (e.g., 'coder', 'assistant', 'creative'). Affects the model's personality and output style.",
        ),
        field(
            "context_length",
            "Context",
            "Loading",
            |s| {
                format_context_k(s.context_length, s.rope_yarn_enabled, s.rope_scale)
            },
            |s, c| s.context_length != c.context_length,
            |s, delta, ctx_limit| {
                let mut val = (s.context_length as i32 + delta * 128).max(128) as u32;
                if ctx_limit > 0 {
                    val = val.min(ctx_limit);
                }
                s.context_length = val;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.context_length = v.max(128);
                }
            },
            "Context window size in tokens. Determines how much of the conversation history is kept in memory. A larger context allows longer conversations but uses more RAM. Typical: 32k-256k depending on model and RAM.",
        ),
        expert_field_with_toggle(
            "rope_yarn_enabled",
            "Yarn RoPE",
            "Loading",
            |s| s.rope_yarn_enabled.to_string(),
            |s, c| s.rope_yarn_enabled != c.rope_yarn_enabled,
            |_, _, _| {},
            |_, _| {},
            toggle_rope_yarn_enabled,
            "Enable YaRN (Yet another RoPE extensioN) for scaling context beyond training limits. YaRN uses a frequency rescaling technique to handle longer contexts. Toggle on/off with Enter.",
        ),
        {
            let mut f = expert_field(
                "yarn_params",
                "Yarn Params",
                "Loading",
                |s| {
                    format!(
                        "scale={:.2} base={:.2} scale_f={:.2}",
                        s.rope_scale, s.rope_freq_base, s.rope_freq_scale
                    )
                },
                |s, c| {
                    s.rope_scale != c.rope_scale
                        || s.rope_freq_base != c.rope_freq_base
                        || s.rope_freq_scale != c.rope_freq_scale
                },
                |_, _, _| {},
                |_, _| {},
                "YaRN configuration: rope_scale (context multiplier), rope_freq_base (frequency base), rope_freq_scale (frequency scaling). Press Enter to open the YaRN parameter editor.",
            );
            f.is_enabled = Some(|s| s.rope_yarn_enabled);
            f
        },
        ultra_field(
            "threads_batch",
            "Threads Batch",
            "Loading",
            |s| s.threads_batch.to_string(),
            |s, c| s.threads_batch != c.threads_batch,
            |s, delta, _| {
                s.threads_batch = (s.threads_batch as i32 + delta).max(1) as u32;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.threads_batch = v.max(1);
                }
            },
            "CPU threads for batch processing (1 to 32). Separate from Threads (inference threads). Keep equal for most workloads, or reduce batch threads to lower CPU usage during batch operations.",
        ),
        ultra_field(
            "ubatch_size",
            "UBatch Size",
            "Loading",
            |s| s.ubatch_size.to_string(),
            |s, c| s.ubatch_size != c.ubatch_size,
            |s, delta, _| {
                s.ubatch_size = (s.ubatch_size as i32 + delta * 64).max(1) as u32;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.ubatch_size = v.max(1);
                }
            },
            "Unlimited batch size for prompt processing. Larger values improve prompt evaluation throughput but use more RAM. Typical: 512-2048. Set to 0 to match context_length.",
        ),
        ultra_field(
            "keep",
            "Keep",
            "Loading",
            |s| s.keep.to_string(),
            |s, c| s.keep != c.keep,
            |s, delta, _| {
                s.keep = (s.keep + delta).max(0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.keep = v;
                }
            },
            "Number of layers to keep in memory when swapping (negative = all). Useful for fast reloading of the same model. Typical: -1 (all) or 0 (none).",
        ),
        field_with_toggle(
            "mlock",
            "Keep in memory (mlock)",
            "Loading",
            |s| s.mlock.to_string(),
            |s, c| s.mlock != c.mlock,
            |_, _, _| {},
            |_, _| {},
            toggle_mlock,
            "Lock model weights in RAM (mlock). Prevents the OS from swapping model weights to disk. Slows model load time but ensures faster inference once loaded. Useful for repeated use.",
        ),
        expert_field_with_toggle(
            "chat_template",
            "Chat Template",
            "Loading",
            |s| {
                if s.auto_chat_template {
                    "Auto (detect)".to_string()
                } else if let Some(ref t) = s.chat_template {
                    if t.ends_with(".jinja") {
                        let filename = std::path::Path::new(t)
                            .file_name()
                            .and_then(|n| n.to_str())
                            .unwrap_or(t);
                        filename.to_string()
                    } else {
                        t.clone()
                    }
                } else {
                    "None".to_string()
                }
            },
            |s, c| {
                s.auto_chat_template != c.auto_chat_template || s.chat_template != c.chat_template
            },
            |_, _, _| {},
            |_, _| {},
            toggle_auto_chat_template,
            "Select chat template: Auto (detect) uses GGUF architecture metadata, specific template names use llama.cpp built-in templates, Select a Template file picks a .jinja file from any directory, None disables template. Press Enter to open picker.",
        ),
        expert_field(
            "numa",
            "NUMA",
            "Loading",
            |s| s.numa.to_string(),
            |s, c| s.numa != c.numa,
            |s, delta, _| {
                let mut val = s.numa;
                val = match (delta, val) {
                    (1, NumMode::None) => NumMode::Distribute,
                    (1, NumMode::Distribute) => NumMode::Isolate,
                    (1, NumMode::Isolate) => NumMode::Numactl,
                    (1, NumMode::Numactl) => NumMode::None,
                    (-1, NumMode::None) => NumMode::Numactl,
                    (-1, NumMode::Distribute) => NumMode::None,
                    (-1, NumMode::Isolate) => NumMode::Distribute,
                    (-1, NumMode::Numactl) => NumMode::Isolate,
                    _ => val,
                };
                s.numa = val;
            },
            |_, _| {},
            "NUMA (Non-Uniform Memory Access) strategy: None, Distribute, Isolate, or Numactl. Affects CPU thread affinity on multi-socket systems. None = default.",
        ),
        // ── GPU Offload ───────────────────────────────────────────────────────
        field(
            "gpu_layers_mode",
            "GPU Layers",
            "GPU Offload",
            |s| match s.gpu_layers_mode {
                GpuLayersMode::Auto => "Auto".to_string(),
                GpuLayersMode::Specific(n) => n.to_string(),
                GpuLayersMode::All => "All".to_string(),
            },
            |s, c| s.gpu_layers_mode != c.gpu_layers_mode,
            gpu_layers_adjust,
            gpu_layers_apply,
            "How many model layers to offload to GPU. Arrow keys cycle: Auto → 1 → 2 → ... → N → All → Auto. Auto lets llama.cpp decide based on VRAM. All loads every layer (999). Specific number sets exact offload count.",
        ),
        ultra_field(
            "split_mode",
            "Split Mode",
            "GPU Offload",
            |s| s.split_mode.to_string(),
            |s, c| s.split_mode != c.split_mode,
            |s, delta, _| {
                let mut val = s.split_mode;
                val = match (delta, val) {
                    (1, SplitMode::None) => SplitMode::Layer,
                    (1, SplitMode::Layer) => SplitMode::Row,
                    (1, SplitMode::Row) => SplitMode::Tensor,
                    (1, SplitMode::Tensor) => SplitMode::None,
                    (-1, SplitMode::None) => SplitMode::Tensor,
                    (-1, SplitMode::Layer) => SplitMode::None,
                    (-1, SplitMode::Row) => SplitMode::Layer,
                    (-1, SplitMode::Tensor) => SplitMode::Row,
                    _ => val,
                };
                s.split_mode = val;
            },
            |_, _| {},
            "GPU split strategy: None, Layer (default), Row, or Tensor. Controls how model layers are distributed across multiple GPUs. Layer splits by layer count, Row/Tensor split by matrix dimensions for multi-GPU setups.",
        ),
        ultra_field(
            "tensor_split",
            "Tensor Split",
            "GPU Offload",
            |s| s.tensor_split.clone(),
            |s, c| s.tensor_split != c.tensor_split,
            |_, _, _| {},
            |_, _| {},
            "Fraction of model weights to load on each GPU (colon-separated for multi-GPU, e.g., '0.5:0.5'). For single GPU, leave empty. Press Enter to edit.",
        ),
        expert_field(
            "main_gpu",
            "Main GPU",
            "GPU Offload",
            |s| s.main_gpu.to_string(),
            |s, c| s.main_gpu != c.main_gpu,
            |s, delta, _| {
                s.main_gpu = (s.main_gpu + delta).max(0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.main_gpu = v;
                }
            },
            "Index of the main GPU (0-based). Handles initial model loading and some computations. Typical: 0 for single GPU, 0 for primary in multi-GPU setups.",
        ),
        field_with_toggle(
            "fit",
            "Fit",
            "GPU Offload",
            |s| s.fit.to_string(),
            |s, c| s.fit != c.fit,
            |_, _, _| {},
            |_, _| {},
            toggle_fit,
            "Automatically adjust arguments to fit device memory. Toggle on/off with Enter.",
        ),
        field_with_toggle(
            "flash_attn",
            "Flash Attention",
            "GPU Offload",
            |s| s.flash_attn.to_string(),
            |s, c| s.flash_attn != c.flash_attn,
            |_, _, _| {},
            |_, _| {},
            toggle_flash_attn,
            "Enable Flash Attention (flash-attn) for faster inference. Requires compatible GPU (Ampere+ / Ada). Significantly speeds up long-context inference. Only works with certain GGUF formats.",
        ),
        field_with_toggle(
            "kv_cache_offload",
            "KV Cache Offload",
            "GPU Offload",
            |s| s.kv_cache_offload.to_string(),
            |s, c| s.kv_cache_offload != c.kv_cache_offload,
            |_, _, _| {},
            |_, _| {},
            toggle_kv_cache_offload,
            "Offload KV cache to RAM when GPU memory is full. Allows larger batch sizes and contexts at the cost of some speed. Useful when VRAM is limited but you still want longer conversations.",
        ),
        // ── Cache type fields ──────────────────────────────────────────────────
        {
            let mut f = expert_field(
                "cache_type_k",
                "Cache Type K",
                "GPU Offload",
                |s| {
                    s.cache_type_k
                        .map(|v| v.to_string())
                        .unwrap_or_else(|| "Disabled".to_string())
                },
                |s, c| s.cache_type_k != c.cache_type_k,
                |s, delta, _| {
                    let mut val = s.cache_type_k.unwrap_or(CacheQuantType::F16);
                    val = if delta > 0 { val.next() } else { val.prev() };
                    s.cache_type_k = Some(val);
                },
                |s, buf| {
                    if let Ok(n) = buf.parse::<u8>() {
                        s.cache_type_k = Some(CacheQuantType::from_u8(n));
                    }
                },
                "Quantization precision for KV cache keys. Lower precision (e.g., Q4, Q8) saves VRAM but may slightly reduce quality. Default is usually FP16. Use lower values if running out of VRAM.",
            );
            f.ctrl_e_toggle = Some(toggle_cache_type_k);
            f
        },
        {
            let mut f = expert_field(
                "cache_type_v",
                "Cache Type V",
                "GPU Offload",
                |s| {
                    s.cache_type_v
                        .map(|v| v.to_string())
                        .unwrap_or_else(|| "Disabled".to_string())
                },
                |s, c| s.cache_type_v != c.cache_type_v,
                |s, delta, _| {
                    let mut val = s.cache_type_v.unwrap_or(CacheQuantType::F16);
                    val = if delta > 0 { val.next() } else { val.prev() };
                    s.cache_type_v = Some(val);
                },
                |s, buf| {
                    if let Ok(n) = buf.parse::<u8>() {
                        s.cache_type_v = Some(CacheQuantType::from_u8(n));
                    }
                },
                "Quantization precision for KV cache values. Lower precision (e.g., Q4, Q8) saves VRAM but may slightly reduce quality. Default is usually FP16. Use lower values if running out of VRAM.",
            );
            f.ctrl_e_toggle = Some(toggle_cache_type_v);
            f
        },
        expert_field_with_toggle(
            "expert_count",
            "Active Experts",
            "GPU Offload",
            |s| {
                if s.expert_count > 0 {
                    s.expert_count.to_string()
                } else if s.expert_count == -1 {
                    "Auto".to_string()
                } else {
                    "Disabled".to_string()
                }
            },
            |s, c| s.expert_count != c.expert_count,
            |s, delta, _| {
                s.expert_count = (s.expert_count + delta).clamp(-1, 99);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.expert_count = v.clamp(-1, 99);
                }
            },
            toggle_expert_count,
            "Number of MoE (Mixture of Experts) experts to activate per token. -1 = auto (all active). Reducing this speeds up inference for MoE models like Mixtral but may reduce quality. Typical: 2-8 for Mixtral.",
        ),
        // ── Evaluation ────────────────────────────────────────────────────────
        field(
            "batch_size",
            "Eval Batch",
            "Evaluation",
            |s| s.batch_size.to_string(),
            |s, c| s.batch_size != c.batch_size,
            |s, delta, _| {
                s.batch_size = (s.batch_size as i32 + delta * 64).max(1) as u32;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.batch_size = v.max(1);
                }
            },
            "Batch size for evaluation (inference). Larger batches use more VRAM but can improve throughput via parallelism. Small values (1-8) for low VRAM, larger (16-128) for high VRAM setups.",
        ),
        field_with_toggle(
            "uniform_cache",
            "Unified KV",
            "Evaluation",
            |s| s.uniform_cache.to_string(),
            |s, c| s.uniform_cache != c.uniform_cache,
            |_, _, _| {},
            |_, _| {},
            toggle_uniform_cache,
            "Share KV cache across sequences. Reduces VRAM usage when running multiple requests by reusing allocated cache. May slightly reduce performance but enables more concurrent users.",
        ),
        expert_field(
            "cache_reuse",
            "Cache Reuse",
            "Evaluation",
            |s| s.cache_reuse.to_string(),
            |s, c| s.cache_reuse != c.cache_reuse,
            |s, delta, _| {
                s.cache_reuse = (s.cache_reuse as i32 + delta * 16).max(0) as u32;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.cache_reuse = v;
                }
            },
            "Minimum chunk size (in tokens) before KV cache is reused across requests. Higher values (e.g., 128, 256) only cache large shared prefixes, reducing disk write churn. Lower values (0-32) cache more aggressively. Adjust with Left/Right arrows (step 16).",
        ),
        expert_field_with_toggle(
            "swa_full",
            "SWA Full Cache",
            "Evaluation",
            |s| s.swa_full.to_string(),
            |s, c| s.swa_full != c.swa_full,
            |_, _, _| {},
            |_, _| {},
            toggle_swa_full,
            "Enable full-size sliding window attention cache. Stores complete KV entries for SWA layers instead of compressed representation. Uses more VRAM but preserves quality on very long contexts. Toggle with Enter.",
        ),
        expert_field_with_toggle(
            "max_concurrent_predictions",
            "Max Concurrent Pred",
            "Evaluation",
            |s| {
                s.max_concurrent_predictions
                    .map(|v| v.to_string())
                    .unwrap_or_else(|| "Off".to_string())
            },
            |s, c| s.max_concurrent_predictions != c.max_concurrent_predictions,
            |s, delta, _| match s.max_concurrent_predictions {
                Some(n) => {
                    s.max_concurrent_predictions = Some(((n as i32) + delta).clamp(1, 10) as u32)
                }
                None => s.max_concurrent_predictions = Some(1),
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.max_concurrent_predictions = Some(v.clamp(1, 10));
                }
            },
            toggle_max_concurrent_predictions,
            "Maximum number of models that can run simultaneously. Press Enter to open a picker that shows how context length divides per model. Each model needs its own VRAM/CPU resources.",
        ),
        // ── Sampling ──────────────────────────────────────────────────────────
        field(
            "seed",
            "Seed",
            "Sampling",
            |s| s.seed.to_string(),
            |s, c| s.seed != c.seed,
            |s, delta, _| {
                s.seed = (s.seed + delta).max(-1);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.seed = v;
                }
            },
            "Random seed for reproducible outputs. -1 = random (default). Set to a fixed value for deterministic, repeatable responses — useful for debugging or testing prompts.",
        ),
        field(
            "temperature",
            "Temp",
            "Sampling",
            |s| format!("{:.2}", s.temperature),
            |s, c| (s.temperature - c.temperature).abs() > 0.001,
            |s, delta, _| {
                s.temperature =
                    ((s.temperature * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 2.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.temperature = v.clamp(0.0, 2.0);
                }
            },
            "Sampling temperature. Controls creativity: 0 = deterministic (most predictable), 0.7 = balanced, 1.0+ = creative. Lower values produce more focused, factual outputs. Typical: 0.7-0.9 for general use.",
        ),
        field(
            "top_k",
            "Top-k",
            "Sampling",
            |s| s.top_k.to_string(),
            |s, c| s.top_k != c.top_k,
            |s, delta, _| {
                s.top_k = (s.top_k + delta).max(1);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.top_k = v.max(0);
                }
            },
            "Only consider the top k most likely tokens at each step. Smaller top-k (e.g., 10-40) makes output more deterministic. Larger values allow more variety. Typical: 40-50. Set to 0 to disable.",
        ),
        field(
            "top_p",
            "Top-p",
            "Sampling",
            |s| format!("{:.2}", s.top_p),
            |s, c| (s.top_p - c.top_p).abs() > 0.001,
            |s, delta, _| {
                s.top_p = ((s.top_p * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 1.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.top_p = v.clamp(0.0, 1.0);
                }
            },
            "Nucleus sampling: only consider tokens whose cumulative probability reaches p. Smaller top-p (e.g., 0.9) is more conservative, larger (e.g., 0.95-0.99) allows more variety. Often preferred over top-k. Typical: 0.9-0.95.",
        ),
        field(
            "min_p",
            "Min P",
            "Sampling",
            |s| format!("{:.2}", s.min_p),
            |s, c| (s.min_p - c.min_p).abs() > 0.001,
            |s, delta, _| {
                s.min_p = ((s.min_p * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 1.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.min_p = v.clamp(0.0, 1.0);
                }
            },
            "Minimum probability threshold relative to the most likely token. Tokens below min_p * max_prob are excluded. A filter that's more principled than top-k/top-p for controlling diversity. Typical: 0.01-0.1.",
        ),
        ultra_field(
            "typical_p",
            "Typical P",
            "Sampling",
            |s| format!("{:.2}", s.typical_p),
            |s, c| (s.typical_p - c.typical_p).abs() > 0.001,
            |s, delta, _| {
                s.typical_p = ((s.typical_p * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 1.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.typical_p = v.clamp(0.0, 1.0);
                }
            },
            "Locally typical sampling (typ_p). Controls diversity by keeping tokens with typical probability mass. Values near 1.0 = no effect, 0.1-0.5 = moderate diversity. Typical: 1.0 (off).",
        ),
        ultra_field(
            "mirostat",
            "Mirostat",
            "Sampling",
            |s| s.mirostat.to_string(),
            |s, c| s.mirostat != c.mirostat,
            |s, delta, _| {
                let mut val = s.mirostat;
                val = match (delta, val) {
                    (1, Mirostat::Off) => Mirostat::V1,
                    (1, Mirostat::V1) => Mirostat::Mirostat2,
                    (1, Mirostat::Mirostat2) => Mirostat::Off,
                    (-1, Mirostat::Off) => Mirostat::Mirostat2,
                    (-1, Mirostat::V1) => Mirostat::Off,
                    (-1, Mirostat::Mirostat2) => Mirostat::V1,
                    _ => val,
                };
                s.mirostat = val;
            },
            |_, _| {},
            "Mirostat sampling mode: Off (default), Mirostat, or Mirostat2. Adaptive temperature control that maintains target perplexity. Mirostat2 is more aggressive. Useful for consistent output quality.",
        ),
        ultra_field(
            "mirostat_lr",
            "Mirostat LR",
            "Sampling",
            |s| format!("{:.2}", s.mirostat_lr),
            |s, c| (s.mirostat_lr - c.mirostat_lr).abs() > 0.001,
            |s, delta, _| {
                s.mirostat_lr =
                    ((s.mirostat_lr * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 1.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.mirostat_lr = v.clamp(0.0, 1.0);
                }
            },
            "Mirostat learning rate (eta). Controls how quickly the temperature adapts. Smaller = smoother adjustments. Typical: 0.1.",
        ),
        ultra_field(
            "mirostat_ent",
            "Mirostat Ent",
            "Sampling",
            |s| format!("{:.2}", s.mirostat_ent),
            |s, c| (s.mirostat_ent - c.mirostat_ent).abs() > 0.001,
            |s, delta, _| {
                s.mirostat_ent =
                    ((s.mirostat_ent * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 10.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.mirostat_ent = v.clamp(0.0, 10.0);
                }
            },
            "Mirostat target entropy. Controls the diversity of output. Higher = more diverse. Typical: 5.0.",
        ),
        ultra_field_with_toggle(
            "ignore_eos",
            "Ignore EOS",
            "Sampling",
            |s| s.ignore_eos.to_string(),
            |s, c| s.ignore_eos != c.ignore_eos,
            |_, _, _| {},
            |_, _| {},
            toggle_ignore_eos,
            "Ignore end-of-sequence tokens during generation. Toggle on/off with Enter. Useful when you want to force the model to continue generating.",
        ),
        ultra_field(
            "samplers",
            "Samplers",
            "Sampling",
            |s| s.samplers.0.clone(),
            |s, c| s.samplers.0 != c.samplers.0,
            |_, _, _| {},
            |_, _| {},
            "Semicolon-separated sampler order string (e.g., 'mirostat;temperature;top_k;top_p'). Controls which samplers are applied and in what order. Press Enter to edit.",
        ),
        field_with_toggle(
            "max_tokens",
            "Max Tokens",
            "Sampling",
            |s| {
                s.max_tokens
                    .map(|v| v.to_string())
                    .unwrap_or_else(|| "Disabled".to_string())
            },
            |s, c| s.max_tokens != c.max_tokens,
            |s, delta, _| {
                let current = s.max_tokens.unwrap_or(0);
                s.max_tokens = Some((current as i32 + delta * 16).max(16) as u32);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.max_tokens = if v == 0 { None } else { Some(v as u32) };
                }
            },
            toggle_max_tokens,
            "Maximum number of tokens to generate in the response. Prevents runaway responses. Set to 0 or Disabled for no limit. Typical: 4096-8192 for chat, higher for code generation.",
        ),
        // ── Repetition ────────────────────────────────────────────────────────
        field(
            "repeat_penalty",
            "Repeat Penalty",
            "Repetition",
            |s| format!("{:.2}", s.repeat_penalty),
            |s, c| (s.repeat_penalty - c.repeat_penalty).abs() > 0.001,
            |s, delta, _| {
                s.repeat_penalty =
                    ((s.repeat_penalty * 100.0 + delta as f32 * 5.0) / 100.0).clamp(1.0, 2.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.repeat_penalty = v.clamp(0.0, 2.0);
                }
            },
            "Controls repetition penalty (1.0 = no penalty, 1.1 = mild, 1.2 = strong). Higher values discourage the model from repeating phrases. Typical: 1.05-1.15 for most use cases.",
        ),
        field(
            "repeat_last_n",
            "Repeat Last N",
            "Repetition",
            |s| s.repeat_last_n.to_string(),
            |s, c| s.repeat_last_n != c.repeat_last_n,
            |s, delta, _| {
                s.repeat_last_n = (s.repeat_last_n + delta).max(0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.repeat_last_n = v.max(0);
                }
            },
            "How many recent tokens to check for repetition (0 = all). Smaller values (32-64) focus on local repetition, larger values (128-256) catch longer patterns. Typical: 64.",
        ),
        expert_field_with_toggle(
            "presence_penalty",
            "Presence Penalty",
            "Repetition",
            |s| {
                s.presence_penalty
                    .map(|v| {
                        if (v - 0.0).abs() < 0.001 {
                            "Off".to_string()
                        } else {
                            format!("{:.2}", v)
                        }
                    })
                    .unwrap_or_else(|| "Off".to_string())
            },
            |s, c| match (s.presence_penalty, c.presence_penalty) {
                (Some(v1), Some(v2)) => (v1 - v2).abs() > 0.001,
                (None, None) => false,
                _ => true,
            },
            |s, delta, _| {
                let current = s.presence_penalty.unwrap_or(0.0);
                s.presence_penalty =
                    Some(((current * 100.0 + delta as f32 * 5.0) / 100.0).clamp(-2.0, 2.0));
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.presence_penalty = Some(v.clamp(-2.0, 2.0));
                }
            },
            toggle_presence_penalty,
            "Encourages the model to talk about new topics (+) or stay on topic (-). Positive values reduce topic repetition, negative values encourage deeper exploration. Typical: 0.0 (off).",
        ),
        expert_field_with_toggle(
            "frequency_penalty",
            "Freq Penalty",
            "Repetition",
            |s| {
                s.frequency_penalty
                    .map(|v| {
                        if (v - 0.0).abs() < 0.001 {
                            "Off".to_string()
                        } else {
                            format!("{:.2}", v)
                        }
                    })
                    .unwrap_or_else(|| "Off".to_string())
            },
            |s, c| match (s.frequency_penalty, c.frequency_penalty) {
                (Some(v1), Some(v2)) => (v1 - v2).abs() > 0.001,
                (None, None) => false,
                _ => true,
            },
            |s, delta, _| {
                let current = s.frequency_penalty.unwrap_or(0.0);
                s.frequency_penalty =
                    Some(((current * 100.0 + delta as f32 * 5.0) / 100.0).clamp(-2.0, 2.0));
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.frequency_penalty = Some(v.clamp(-2.0, 2.0));
                }
            },
            toggle_frequency_penalty,
            "Penalizes tokens based on how often they appear in the text (+) or rewards them (-). Positive values reduce word repetition, negative values encourage denser language. Typical: 0.0 (off).",
        ),
        // ── DRY ───────────────────────────────────────────────────────────────
        ultra_field(
            "dry_multiplier",
            "DRY Multiplier",
            "DRY",
            |s| format!("{:.2}", s.dry_multiplier),
            |s, c| (s.dry_multiplier - c.dry_multiplier).abs() > 0.001,
            |s, delta, _| {
                s.dry_multiplier =
                    ((s.dry_multiplier * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 10.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.dry_multiplier = v.clamp(0.0, 10.0);
                }
            },
            "DRY (Don't Repeat Yourself) multiplier. Scales the penalty for repetition. Higher values = stronger anti-repetition. Typical: 1.75.",
        ),
        ultra_field(
            "dry_base",
            "DRY Base",
            "DRY",
            |s| format!("{:.2}", s.dry_base),
            |s, c| (s.dry_base - c.dry_base).abs() > 0.001,
            |s, delta, _| {
                s.dry_base = ((s.dry_base * 100.0 + delta as f32 * 5.0) / 100.0).clamp(0.0, 10.0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<f32>() {
                    s.dry_base = v.clamp(0.0, 10.0);
                }
            },
            "DRY penalty base (log scale). Controls the strength of the repetition penalty. Typical: 1.0 (log2) or 0.0 (linear).",
        ),
        ultra_field(
            "dry_allowed_length",
            "DRY Allowed Length",
            "DRY",
            |s| s.dry_allowed_length.to_string(),
            |s, c| s.dry_allowed_length != c.dry_allowed_length,
            |s, delta, _| {
                s.dry_allowed_length = (s.dry_allowed_length + delta).max(0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.dry_allowed_length = v;
                }
            },
            "Number of recent tokens to check for repetition (penalty starts after this). Higher values check longer context. Typical: 2.",
        ),
        ultra_field(
            "dry_penalty_last_n",
            "DRY Penalty Last N",
            "DRY",
            |s| s.dry_penalty_last_n.to_string(),
            |s, c| s.dry_penalty_last_n != c.dry_penalty_last_n,
            |s, delta, _| {
                s.dry_penalty_last_n = (s.dry_penalty_last_n + delta).max(0);
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<i32>() {
                    s.dry_penalty_last_n = v;
                }
            },
            "How many tokens to consider for DRY penalty (0 = all). Larger values catch longer repetition patterns. Typical: -1 (all) or 128.",
        ),
        // ── Speculative Decoding ─────────────────────────────────────────────
        expert_field_with_toggle(
            "is_mtp",
            "MTP",
            "Speculative",
            |s| {
                if s.spec_type.is_empty() {
                    "Off".to_string()
                } else {
                    s.spec_type.clone()
                }
            },
            |s, c| s.spec_type != c.spec_type,
            |_, _, _| {},
            |_, _| {},
            toggle_mtp,
            "Speculative decoding method for faster inference. Options: Off, draft-mtp (MTP-based), draft-simple, draft-eagle3, ngram-simple, ngram-map-k, ngram-map-k4v, ngram-mod, ngram-cache. Draft-mtp requires a compatible model with MTP architecture.",
        ),
        expert_field(
            "spec_type",
            "Spec Type",
            "Speculative",
            |s| {
                if s.spec_type.is_empty() {
                    "Off".to_string()
                } else {
                    s.spec_type.clone()
                }
            },
            |s, c| s.spec_type != c.spec_type,
            |_, _, _| {},
            |_, _| {},
            "Speculative decoding method for faster inference. Options: Off, draft-mtp (MTP-based), draft-simple, draft-eagle3, ngram-simple, ngram-map-k, ngram-map-k4v, ngram-mod, ngram-cache. Draft-mtp requires a compatible model with MTP architecture.",
        ),
        expert_field(
            "draft_tokens",
            "Spec Draft N Max",
            "Speculative",
            |s| s.draft_tokens.to_string(),
            |s, c| s.draft_tokens != c.draft_tokens,
            |s, delta, _| {
                s.draft_tokens = (s.draft_tokens as i32 + delta).clamp(0, 16) as u32;
            },
            |s, buf| {
                if let Ok(v) = buf.parse::<u32>() {
                    s.draft_tokens = v.min(16);
                }
            },
            "Maximum number of draft tokens per step (0-16). More drafts = more potential speedup but also more wasted computation if drafts are rejected. Typical: 4-8 for draft-mtp.",
        ),
        // ── Tags ──────────────────────────────────────────────────────────────
        field(
            "tags",
            "Tags (Enter to edit)",
            "Tags",
            |s| {
                if s.tags.is_empty() {
                    "None".to_string()
                } else {
                    s.tags.join(", ")
                }
            },
            |s, c| s.tags != c.tags,
            |_, _, _| {},
            |_, _| {},
            "Comma-separated labels for the model (e.g., 'coding, chat, reasoning'). Used for filtering and organization. Press Enter to open a tag editor.",
        ),
        // ── Backend ───────────────────────────────────────────────────────────
        field(
            "backend_version",
            "LLama.cpp Version",
            "Backend",
            |s| s.get_active_backend_version_display().to_string(),
            |s, c| s.get_active_backend_version() != c.get_active_backend_version(),
            |_, _, _| {},
            |_, _| {},
            "Select the llama.cpp backend binary (CPU / Vulkan / ROCm / CUDA). Press Enter to open a version picker. Different backends support different GPU types and features.",
        ),
    ]
}

pub fn filtered_fields(expert_mode: bool) -> Vec<SettingField> {
    all_fields()
        .into_iter()
        .filter(|f| {
            if !expert_mode {
                !f.is_expert
            } else {
                !f.is_ultra // In expert mode, hide ultra experts
            }
        })
        .collect()
}

// ── Simple helper for the server settings panel (tabbed.rs) ──────────────────

/// Render a single setting line for the server settings panel.
#[allow(clippy::too_many_arguments)]
pub fn add_setting(
    lines: &mut Vec<Line<'static>>,
    total_count: &mut usize,
    _settings: &ModelSettings,
    _cached: &ModelSettings,
    selected_line_idx: &mut usize,
    selected_content_line: &mut usize,
    idx: usize,
    name: &str,
    val: &str,
    selected: usize,
    _edit_buf: &str,
    _editing: bool,
    disabled: bool,
) {
    let current_line = lines.len();
    let name_style = if disabled {
        Style::default().fg(GRAY)
    } else {
        Style::default().fg(YELLOW)
    };
    let val_style = if disabled {
        Style::default()
            .fg(GRAY)
    } else {
        Style::default().fg(WHITE)
    };
    if idx == selected {
        *selected_line_idx = current_line;
        *selected_content_line = current_line;
        lines.push(Line::from(vec![
            Span::styled(
                "> ",
                Style::default()
                    .fg(YELLOW)
                    .add_modifier(if disabled {
                        Modifier::DIM
                    } else {
                        Modifier::BOLD
                    }),
            ),
            Span::styled(format!("{name}: "), name_style),
            Span::styled(
                 val.to_string(),
                 Style::default()
                     .fg(BLACK)
                     .bg(GREEN)
                     .add_modifier(Modifier::BOLD),
             ),
        ]));
    } else {
        lines.push(Line::from(vec![
            Span::styled("  ", name_style),
            Span::styled(format!("{name}: "), name_style),
            Span::styled(val.to_string(), val_style),
        ]));
    }
    *total_count += 1;
}

/// Build a list of setting names that differ between a profile and the current settings.
pub fn profile_settings_parts(profile: &Profile, current: &ModelSettings) -> Vec<String> {
    let mut parts = Vec::new();
    let s = &profile.settings;

    // ── Integers ──────────────────────────────────────────────────────────
    diff_int!(parts, s, current, context_length, "ctx");
    diff_int!(parts, s, current, threads, "threads");
    diff_int!(parts, s, current, threads_batch, "threads_batch");
    diff_int!(parts, s, current, batch_size, "batch");
    diff_int!(parts, s, current, ubatch_size, "ubatch");
    diff_int!(parts, s, current, parallel, "parallel");
    diff_option!(parts, s, current, max_concurrent_predictions, "concurrent");
    diff_int!(parts, s, current, cache_reuse, "cache_reuse");
    diff_option!(parts, s, current, max_tokens, "max_tokens");
    diff_int!(parts, s, current, draft_tokens, "draft_tokens");

    diff_int!(parts, s, current, keep, "keep");
    diff_int!(parts, s, current, main_gpu, "main_gpu");
    diff_int!(parts, s, current, expert_count, "expert_count");
    diff_int!(parts, s, current, seed, "seed");
    diff_int!(parts, s, current, top_k, "top_k");
    diff_int!(parts, s, current, repeat_last_n, "repeat_last_n");
    diff_int!(parts, s, current, dry_allowed_length, "dry_allowed");
    diff_int!(parts, s, current, dry_penalty_last_n, "dry_penalty_last_n");

    // ── Floats ────────────────────────────────────────────────────────────
    diff_float!(parts, s, current, temperature, "temp");
    diff_float!(parts, s, current, top_p, "top_p");
    diff_float!(parts, s, current, min_p, "min_p");
    diff_float!(parts, s, current, typical_p, "typical_p");
    diff_float!(parts, s, current, mirostat_lr, "mirostat_lr");
    diff_float!(parts, s, current, mirostat_ent, "mirostat_ent");
    diff_float!(parts, s, current, repeat_penalty, "rep_pen");
    diff_option_float!(parts, s, current, presence_penalty, "pres_pen");
    diff_option_float!(parts, s, current, frequency_penalty, "freq_pen");
    diff_float!(parts, s, current, dry_multiplier, "dry_mult");
    diff_float!(parts, s, current, dry_base, "dry_base");
    diff_float!(parts, s, current, rope_scale, "rope_scale");
    diff_float!(parts, s, current, rope_freq_base, "rope_freq_base");
    diff_float!(parts, s, current, rope_freq_scale, "rope_freq_scale");

    // ── Bools ─────────────────────────────────────────────────────────────
    diff_bool!(parts, s, current, swa_full, "swa_full");
    diff_bool!(parts, s, current, mlock, "mlock");
    diff_bool!(parts, s, current, mmap, "mmap");
    diff_bool!(parts, s, current, uniform_cache, "uniform_cache");
    diff_bool!(parts, s, current, kv_cache_offload, "kv_cache_offload");
    diff_bool!(parts, s, current, fit, "fit");
    diff_bool!(parts, s, current, embedding, "embedding");
    diff_bool!(parts, s, current, flash_attn, "flash_attn");
    diff_bool!(parts, s, current, jinja, "jinja");
    diff_bool!(parts, s, current, auto_chat_template, "auto_chat_template");
    diff_bool!(parts, s, current, ignore_eos, "ignore_eos");
    diff_bool!(parts, s, current, rope_yarn_enabled, "yarn_enabled");
    diff_bool!(parts, s, current, cache_prompt, "cache_prompt");
    diff_bool!(parts, s, current, webui, "webui");
    // ── Strings ───────────────────────────────────────────────────────────
    diff_string!(parts, s, current, system_prompt_preset_name, "preset");
    diff_string!(parts, s, current, tensor_split, "tensor_split");
    diff_string!(parts, s, current, rpc, "rpc");
    if s.chat_template != current.chat_template {
        if let Some(ref v) = s.chat_template {
            let filename = std::path::Path::new(v)
                .file_name()
                .and_then(|n| n.to_str())
                .unwrap_or(v);
            parts.push(format!("chat_template={}", filename));
        }
    }
    diff_option!(
        parts,
        s,
        current,
        chat_template_kwargs,
        "chat_template_kwargs"
    );
    diff_option!(parts, s, current, llama_cpp_version_cpu, "llama_cpp_cpu");
    diff_option!(
        parts,
        s,
        current,
        llama_cpp_version_vulkan,
        "llama_cpp_vulkan"
    );
    diff_option!(parts, s, current, llama_cpp_version_rocm, "llama_cpp_rocm");
    diff_option!(
        parts,
        s,
        current,
        llama_cpp_version_rocm_lemonade,
        "llama_cpp_rocm_lemonade"
    );
    diff_option!(parts, s, current, llama_cpp_version_cuda, "llama_cpp_cuda");
    diff_string!(parts, s, current, spec_type, "spec_type");

    // ── Enums ─────────────────────────────────────────────────────────────
    diff_enum!(parts, s, current, numa, "numa");
    diff_enum!(parts, s, current, split_mode, "split_mode");
    diff_enum!(parts, s, current, mirostat, "mirostat");
    diff_enum!(parts, s, current, samplers, "samplers");
    diff_enum!(parts, s, current, rope_scaling, "rope_scaling");
    diff_enum!(parts, s, current, cache_type, "cache_type");
    diff_option!(parts, s, current, cache_type_k, "cache_type_k");
    diff_option!(parts, s, current, cache_type_v, "cache_type_v");

    // ── Special (custom display) ──────────────────────────────────────────
    if let Some(v) = s.gpu_layers_mode
        && v != current.gpu_layers_mode
    {
        let display = match v {
            crate::models::GpuLayersMode::Auto => "Auto".to_string(),
            crate::models::GpuLayersMode::Specific(n) => n.to_string(),
            crate::models::GpuLayersMode::All => "All".to_string(),
        };
        parts.push(format!("gpu_layers={}", display));
    }

    parts
}