llm_manager/
config.rs

1mod model_config;
2mod presets;
3mod profiles;
4mod store;
5
6use std::collections::HashSet;
7use std::path::PathBuf;
8
9use chrono::Local;
10use serde::{Deserialize, Serialize};
11
12pub use model_config::ModelConfigStore;
13
14pub use profiles::ProfileStore;
15
16use crate::models::{
17    Backend, CacheType, CacheTypeK, CacheTypeV, Mirostat, NumMode, RopeScaling, Samplers, SplitMode,
18};
19use crate::tui::app::ActivePanel;
20pub use presets::PresetStore;
21
22/// Resolve the base config directory with a safe fallback chain.
23///
24/// Prefers `dirs::config_dir()` (XDG on Linux, ~/Library/Application Support on macOS,
25/// etc.), falls back to `~/.config`, and lastly `./.llm-manager` if both fail.
26pub fn config_base_dir() -> PathBuf {
27    if let Some(d) = dirs::config_dir() {
28        return d;
29    }
30    if let Some(home) = dirs::home_dir() {
31        return home.join(".config");
32    }
33    PathBuf::from(".").join(".llm-manager")
34}
35
36/// Count physical CPU cores on Linux (ignores hyperthreading).
37/// Falls back to 1 if the file can't be read or parsing fails.
38pub fn physical_cores() -> u32 {
39    let content = match std::fs::read_to_string("/proc/cpuinfo") {
40        Ok(c) => c,
41        Err(_) => {
42            return std::thread::available_parallelism()
43                .map(|p| p.get() as u32)
44                .unwrap_or(1);
45        }
46    };
47    let mut seen = HashSet::new();
48    let mut cur_phys: Option<&str> = None;
49    let mut cur_core: Option<&str> = None;
50    for line in content.lines() {
51        if let Some((key, val)) = line.split_once(':') {
52            let key = key.trim();
53            let val = val.trim();
54            match key {
55                "physical id" => cur_phys = Some(val),
56                "core id" => cur_core = Some(val),
57                _ => {}
58            }
59            if let (Some(phys), Some(core)) = (cur_phys, cur_core) {
60                seen.insert((phys, core));
61            }
62        }
63    }
64    seen.len() as u32
65}
66
67/// A remote RPC worker for distributed inference.
68#[derive(Debug, Clone, Serialize, Deserialize)]
69pub struct RpcWorker {
70    #[serde(default)]
71    pub selected: bool,
72    #[serde(default)]
73    pub name: String,
74    pub ip: String,
75    #[serde(default = "default_rpc_port")]
76    pub port: u16,
77}
78
79fn default_rpc_port() -> u16 {
80    50052
81}
82
83/// Global configuration.
84#[derive(Debug, Clone, Serialize, Deserialize)]
85pub struct Config {
86    pub models_dirs: Vec<PathBuf>,
87    pub llama_server: PathBuf,
88    pub default: DefaultParams,
89    /// Per-model overrides (keyed by model file name, stored as YAML in models/).
90    #[serde(default, skip)]
91    pub model_overrides: ModelConfigStore,
92    /// Named profiles of settings presets (stored as YAML in profiles/).
93    #[serde(default, skip)]
94    pub profiles: ProfileStore,
95    /// System prompt presets (stored as YAML in presets/).
96    #[serde(default, skip)]
97    pub system_prompt_presets: PresetStore,
98    /// RPC Workers for distributed inference.
99    #[serde(default)]
100    pub rpc_workers: Vec<RpcWorker>,
101    /// Number of results per HuggingFace search query.
102    #[serde(default = "default_search_limit")]
103    pub search_limit: u32,
104    /// The last focused panel position (for restoring on next launch).
105    #[serde(default)]
106    pub active_panel: crate::tui::app::ActivePanel,
107    /// Left panel width percentage (20-80).
108    #[serde(default = "default_left_pct")]
109    pub left_pct: u16,
110}
111
112fn default_left_pct() -> u16 {
113    55
114}
115
116fn default_search_limit() -> u32 {
117    50
118}
119
120/// A named profile of settings.
121#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
122pub struct Profile {
123    pub name: String,
124    /// Brief description shown in the profile list.
125    pub description: String,
126    /// The settings for this profile.
127    #[serde(default)]
128    pub settings: ModelOverride,
129}
130
131impl Profile {
132    /// Apply this profile's settings to a base ModelSettings.
133    pub fn apply(&self, mut base: crate::models::ModelSettings) -> crate::models::ModelSettings {
134        self.settings.apply(&mut base);
135        base
136    }
137}
138
139/// A named system prompt preset.
140#[derive(Debug, Clone, Serialize, Deserialize)]
141pub struct SystemPromptPreset {
142    pub name: String,
143    pub description: String,
144    pub content: String,
145}
146
147/// Built-in system prompt presets.
148pub fn builtin_system_prompt_presets() -> Vec<SystemPromptPreset> {
149    vec![
150        SystemPromptPreset {
151            name: "General".into(),
152            description: "General-purpose assistant".into(),
153            content: "You are a helpful assistant.".into(),
154        },
155        SystemPromptPreset {
156            name: "Coder".into(),
157            description: "Expert software developer".into(),
158            content: "You are an expert software developer. Write clean, well-documented code. Explain your reasoning and suggest improvements.".into(),
159        },
160        SystemPromptPreset {
161            name: "Thinker".into(),
162            description: "Analytical and thoughtful".into(),
163            content: "You are a thoughtful and analytical AI assistant. Think carefully before answering. Provide well-reasoned responses with clear explanations.".into(),
164        },
165        SystemPromptPreset {
166            name: "Mathematician".into(),
167            description: "Expert in mathematics".into(),
168            content: "You are an expert in mathematics. Provide clear, step-by-step solutions to mathematical problems. Show your reasoning and explain key concepts.".into(),
169        },
170    ]
171}
172
173#[derive(Debug, Clone, Serialize, Deserialize, Default, PartialEq)]
174pub struct ModelOverride {
175    // Loading
176    pub context_length: Option<u32>,
177    pub batch_size: Option<u32>,
178    pub ubatch_size: Option<u32>,
179    pub cache_type_k: Option<CacheTypeK>,
180    pub cache_type_v: Option<CacheTypeV>,
181    pub keep: Option<i32>,
182    pub swa_full: Option<bool>,
183    pub mlock: Option<bool>,
184    pub mmap: Option<bool>,
185    pub numa: Option<NumMode>,
186    pub uniform_cache: Option<bool>,
187    pub system_prompt: Option<String>,
188    pub system_prompt_preset_name: Option<String>,
189    pub max_concurrent_predictions: Option<u32>,
190    pub threads: Option<u32>,
191    pub threads_batch: Option<u32>,
192    pub parallel: Option<u32>,
193
194    // GPU
195    pub gpu_layers: Option<i32>,
196    pub split_mode: Option<SplitMode>,
197    pub tensor_split: Option<String>,
198    pub main_gpu: Option<i32>,
199    pub fit: Option<bool>,
200    pub lora: Option<PathBuf>,
201    pub lora_scaled: Option<(PathBuf, f32)>,
202    pub rpc: Option<String>,
203    pub embedding: Option<bool>,
204    pub kv_cache_offload: Option<bool>,
205    pub flash_attn: Option<bool>,
206    pub jinja: Option<bool>,
207    pub chat_template: Option<String>,
208    pub chat_template_kwargs: Option<String>,
209    pub expert_count: Option<i32>,
210    pub gpu_layers_mode: Option<crate::models::GpuLayersMode>,
211
212    // Sampling
213    pub seed: Option<i32>,
214    pub temperature: Option<f32>,
215    pub top_k: Option<i32>,
216    pub top_p: Option<f32>,
217    pub min_p: Option<f32>,
218    pub typical_p: Option<f32>,
219    pub mirostat: Option<Mirostat>,
220    pub mirostat_lr: Option<f32>,
221    pub mirostat_ent: Option<f32>,
222    pub ignore_eos: Option<bool>,
223    pub samplers: Option<Samplers>,
224
225    // Repetition
226    pub repeat_penalty: Option<f32>,
227    pub repeat_last_n: Option<i32>,
228    pub presence_penalty: Option<f32>,
229    pub frequency_penalty: Option<f32>,
230    pub dry_multiplier: Option<f32>,
231    pub dry_base: Option<f32>,
232    pub dry_allowed_length: Option<i32>,
233    pub dry_penalty_last_n: Option<i32>,
234
235    // RoPE
236    pub rope_scaling: Option<RopeScaling>,
237    pub rope_scale: Option<f32>,
238    pub rope_freq_base: Option<f32>,
239    pub rope_freq_scale: Option<f32>,
240    pub rope_yarn_enabled: Option<bool>,
241
242    // Server
243    pub cache_prompt: Option<bool>,
244    pub cache_reuse: Option<u32>,
245    pub webui: Option<bool>,
246
247    // Other
248    pub max_tokens: Option<u32>,
249    pub cache_type: Option<CacheType>,
250    pub llama_cpp_version_cpu: Option<String>,
251    pub llama_cpp_version_vulkan: Option<String>,
252    pub llama_cpp_version_rocm: Option<String>,
253    pub llama_cpp_version_rocm_lemonade: Option<String>,
254    pub llama_cpp_version_cuda: Option<String>,
255    pub spec_type: Option<String>,
256    pub draft_tokens: Option<u32>,
257    pub tags: Option<Vec<String>>,
258}
259
260/// Apply a scalar Copy field from override: `base.f = self.f.unwrap_or(base.f)`.
261macro_rules! apply_scalar {
262    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
263        $(
264            $base.$field = $self.$field.unwrap_or($base.$field);
265        )+
266    };
267}
268
269/// Apply a Clone field from override: `if let Some(v) = &self.f { base.f = v.clone(); }`.
270macro_rules! apply_clone {
271    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
272        $(
273            if let Some(v) = &$self.$field {
274                $base.$field = v.clone();
275            }
276        )+
277    };
278}
279
280/// Apply an Option<T> field from override: `if let Some(v) = &self.f { base.f = Some(v.clone()); }`.
281macro_rules! apply_option {
282    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
283        $(
284            if let Some(v) = &$self.$field {
285                $base.$field = Some(v.clone());
286            }
287        )+
288    };
289}
290
291impl ModelOverride {
292    pub fn from_settings(s: &crate::models::ModelSettings) -> Self {
293        Self {
294            context_length: Some(s.context_length),
295            batch_size: Some(s.batch_size),
296            ubatch_size: Some(s.ubatch_size),
297            cache_type_k: s.cache_type_k,
298            cache_type_v: s.cache_type_v,
299            keep: Some(s.keep),
300            swa_full: Some(s.swa_full),
301            mlock: Some(s.mlock),
302            mmap: Some(s.mmap),
303            numa: Some(s.numa),
304            uniform_cache: Some(s.uniform_cache),
305            system_prompt: Some(s.system_prompt.clone()),
306            system_prompt_preset_name: Some(s.system_prompt_preset_name.clone()),
307            max_concurrent_predictions: s.max_concurrent_predictions,
308            threads: Some(s.threads),
309            threads_batch: Some(s.threads_batch),
310            parallel: Some(s.parallel),
311            gpu_layers: Some(match s.gpu_layers_mode {
312                crate::models::GpuLayersMode::Auto => 0,
313                crate::models::GpuLayersMode::Specific(n) => n as i32,
314                crate::models::GpuLayersMode::All => -1,
315            }),
316            gpu_layers_mode: Some(s.gpu_layers_mode),
317            split_mode: Some(s.split_mode),
318            tensor_split: Some(s.tensor_split.clone()),
319            main_gpu: Some(s.main_gpu),
320            fit: Some(s.fit),
321            lora: s.lora.clone(),
322            lora_scaled: s.lora_scaled.clone(),
323            rpc: Some(s.rpc.clone()),
324            embedding: Some(s.embedding),
325            kv_cache_offload: Some(s.kv_cache_offload),
326            flash_attn: Some(s.flash_attn),
327            jinja: Some(s.jinja),
328            chat_template: s.chat_template.clone(),
329            chat_template_kwargs: s.chat_template_kwargs.clone(),
330            expert_count: Some(s.expert_count),
331            seed: Some(s.seed),
332            temperature: Some(s.temperature),
333            top_k: Some(s.top_k),
334            top_p: Some(s.top_p),
335            min_p: Some(s.min_p),
336            typical_p: Some(s.typical_p),
337            mirostat: Some(s.mirostat),
338            mirostat_lr: Some(s.mirostat_lr),
339            mirostat_ent: Some(s.mirostat_ent),
340            ignore_eos: Some(s.ignore_eos),
341            samplers: Some(s.samplers.clone()),
342            repeat_penalty: Some(s.repeat_penalty),
343            repeat_last_n: Some(s.repeat_last_n),
344            presence_penalty: s.presence_penalty,
345            frequency_penalty: s.frequency_penalty,
346            dry_multiplier: Some(s.dry_multiplier),
347            dry_base: Some(s.dry_base),
348            dry_allowed_length: Some(s.dry_allowed_length),
349            dry_penalty_last_n: Some(s.dry_penalty_last_n),
350            rope_scaling: Some(s.rope_scaling),
351            rope_scale: Some(s.rope_scale),
352            rope_freq_base: Some(s.rope_freq_base),
353            rope_freq_scale: Some(s.rope_freq_scale),
354            rope_yarn_enabled: Some(s.rope_yarn_enabled),
355            cache_prompt: Some(s.cache_prompt),
356            cache_reuse: Some(s.cache_reuse),
357            webui: Some(s.webui),
358            max_tokens: s.max_tokens,
359            cache_type: Some(s.cache_type),
360            llama_cpp_version_cpu: s.llama_cpp_version_cpu.clone(),
361            llama_cpp_version_vulkan: s.llama_cpp_version_vulkan.clone(),
362            llama_cpp_version_rocm: s.llama_cpp_version_rocm.clone(),
363            llama_cpp_version_rocm_lemonade: s.llama_cpp_version_rocm_lemonade.clone(),
364            llama_cpp_version_cuda: s.llama_cpp_version_cuda.clone(),
365            spec_type: Some(s.spec_type.clone()),
366            draft_tokens: Some(s.draft_tokens),
367           tags: Some(s.tags.clone()),
368        }
369    }
370
371    /// Merge override into a base ModelSettings (in-place).
372    pub fn apply(&self, base: &mut crate::models::ModelSettings) {
373        // Override values always take precedence. For Option<T> fields,
374        // the override value (even None) is explicitly set by the user.
375
376        // Scalar Copy fields: base.f = self.f.unwrap_or(base.f)
377        apply_scalar!(self, base,
378            context_length, batch_size, ubatch_size, keep, swa_full, mlock, mmap,
379            numa, uniform_cache, kv_cache_offload, threads, threads_batch, parallel,
380            split_mode, main_gpu, fit, embedding, flash_attn, jinja, expert_count,
381            seed, temperature, top_k, top_p, min_p, typical_p,
382            mirostat, mirostat_lr, mirostat_ent, ignore_eos,
383            repeat_penalty, repeat_last_n,
384            dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
385            rope_scaling, rope_scale, rope_freq_base, rope_freq_scale, rope_yarn_enabled,
386            cache_prompt, cache_reuse, webui, cache_type,
387            draft_tokens, gpu_layers_mode,
388        );
389
390        // Cloneable fields: if let Some(v) = &self.f { base.f = v.clone(); }
391        apply_clone!(self, base,
392            system_prompt, system_prompt_preset_name, tensor_split, rpc,
393            samplers, spec_type, tags,
394        );
395
396        // Option<T> fields: if let Some(v) = &self.f { base.f = Some(v.clone()); }
397        apply_option!(self, base,
398            lora, lora_scaled, chat_template, chat_template_kwargs,
399            llama_cpp_version_cpu, llama_cpp_version_vulkan,
400            llama_cpp_version_rocm, llama_cpp_version_rocm_lemonade,
401            llama_cpp_version_cuda,
402        );
403
404        // Direct Option<T> assignment (same type in both structs)
405        base.cache_type_k = self.cache_type_k;
406        base.cache_type_v = self.cache_type_v;
407        base.presence_penalty = self.presence_penalty;
408        base.frequency_penalty = self.frequency_penalty;
409        base.max_tokens = self.max_tokens;
410
411        // Special: max_concurrent_predictions uses or() for Option chaining
412        base.max_concurrent_predictions = self
413            .max_concurrent_predictions
414            .or(base.max_concurrent_predictions);
415
416        // Special: gpu_layers converts i32 legacy field to GpuLayersMode enum
417        // Only applies when gpu_layers is explicitly set in the override.
418        if let Some(n) = self.gpu_layers {
419            base.gpu_layers_mode = match n {
420                n if n < 0 => crate::models::GpuLayersMode::All,
421                n => crate::models::GpuLayersMode::Specific(n as u32),
422            };
423        }
424
425        // FIELD ACCOUNTING (ModelOverride: 87 fields):
426        // - apply_scalar: 53 fields
427        // - apply_clone: 7 fields
428        // - apply_option: 10 fields
429        // - direct Option assign: 5 fields (cache_type_k, cache_type_v, presence_penalty,
430        //   frequency_penalty, max_tokens)
431        // - special: 1 field (max_concurrent_predictions)
432        // - conditional: gpu_layers overrides gpu_layers_mode only when Some
433        // - NOT in ModelSettings: 0 (all ModelOverride fields mapped above)
434        //
435        // ModelSettings fields NOT in ModelOverride (not overridable):
436        // host, port, timeout, backend, platform, router_max_models, server_mode,
437        // api_endpoint_enabled, api_endpoint_port
438        //
439        // When adding a field: ensure it appears in exactly one category above.
440    }
441}
442
443/// Built-in profiles with sensible defaults for popular model families.
444pub fn builtin_profiles() -> Vec<Profile> {
445    vec![
446        Profile {
447            name: "Qwen".into(),
448            description: "Optimized for Qwen models (dense)".into(),
449            settings: ModelOverride {
450                context_length: Some(131072),
451                temperature: Some(0.7),
452                top_k: Some(20),
453                top_p: Some(0.95),
454                max_tokens: Some(4096),
455                presence_penalty: Some(0.0),
456                uniform_cache: Some(true),
457                jinja: Some(true),
458                ..Default::default()
459            },
460        },
461        Profile {
462            name: "Qwen-MoE".into(),
463            description: "Optimized for Qwen MoE models (35B-A3B)".into(),
464            settings: ModelOverride {
465                context_length: Some(131072),
466                temperature: Some(0.8),
467                top_k: Some(20),
468                top_p: Some(0.95),
469                max_tokens: Some(4096),
470                presence_penalty: Some(1.5),
471                uniform_cache: Some(true),
472                jinja: Some(true),
473                ..Default::default()
474            },
475        },
476        Profile {
477            name: "Qwen-Coding".into(),
478            description: "Optimized for Qwen models in coding mode".into(),
479            settings: ModelOverride {
480                context_length: Some(131072),
481                temperature: Some(0.6),
482                top_k: Some(20),
483                top_p: Some(0.95),
484                max_tokens: Some(4096),
485                presence_penalty: Some(0.0),
486                uniform_cache: Some(true),
487                jinja: Some(true),
488                ..Default::default()
489            },
490        },
491        Profile {
492            name: "Gemma".into(),
493            description: "Optimized for Gemma 2/4 models".into(),
494            settings: ModelOverride {
495                context_length: Some(131072),
496                min_p: Some(0.1),
497                temperature: Some(1.0),
498                top_k: Some(65),
499                top_p: Some(0.95),
500                max_tokens: Some(4096),
501                uniform_cache: Some(true),
502                jinja: Some(true),
503                ..Default::default()
504            },
505        },
506        Profile {
507            name: "Llama".into(),
508            description: "Optimized for Llama 3.1/3.3 models".into(),
509            settings: ModelOverride {
510                context_length: Some(131072),
511                temperature: Some(0.7),
512                top_p: Some(0.9),
513                repeat_penalty: Some(1.1),
514                max_tokens: Some(4096),
515                uniform_cache: Some(true),
516                jinja: Some(true),
517                ..Default::default()
518            },
519        },
520        Profile {
521            name: "Mistral".into(),
522            description: "Optimized for Mistral 7B/NeMo models".into(),
523            settings: ModelOverride {
524                context_length: Some(131072),
525                temperature: Some(0.7),
526                top_k: Some(50),
527                top_p: Some(0.9),
528                max_tokens: Some(4096),
529                uniform_cache: Some(true),
530                jinja: Some(true),
531                ..Default::default()
532            },
533        },
534        Profile {
535            name: "Phi".into(),
536            description: "Optimized for Phi 3.5 Mini models".into(),
537            settings: ModelOverride {
538                context_length: Some(131072),
539                temperature: Some(0.7),
540                top_k: Some(50),
541                top_p: Some(0.9),
542                repeat_penalty: Some(1.1),
543                max_tokens: Some(4096),
544                uniform_cache: Some(true),
545                ..Default::default()
546            },
547        },
548    ]
549}
550
551#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
552#[serde(default)]
553pub struct DefaultParams {
554    // Loading
555    #[serde(default)]
556    pub context_length: u32,
557    #[serde(default)]
558    pub threads: u32,
559    #[serde(default)]
560    pub threads_batch: u32,
561    #[serde(default)]
562    pub batch_size: u32,
563    #[serde(default)]
564    pub ubatch_size: u32,
565    #[serde(default = "default_cache_type_k")]
566    pub cache_type_k: Option<CacheTypeK>,
567    #[serde(default = "default_cache_type_v")]
568    pub cache_type_v: Option<CacheTypeV>,
569    #[serde(default)]
570    pub keep: i32,
571    #[serde(default)]
572    pub swa_full: bool,
573    #[serde(default)]
574    pub mlock: bool,
575    #[serde(default)]
576    pub mmap: bool,
577    #[serde(default)]
578    pub numa: NumMode,
579    #[serde(default)]
580    pub uniform_cache: bool,
581    #[serde(default)]
582    pub kv_cache_offload: bool,
583    #[serde(default)]
584    pub parallel: u32,
585    #[serde(default)]
586    pub max_concurrent_predictions: Option<u32>,
587    #[serde(default)]
588    pub system_prompt: String,
589    #[serde(default = "default_system_prompt_preset_name")]
590    pub system_prompt_preset_name: String,
591    // GPU
592    #[serde(default)]
593    pub gpu_layers: i32,
594    #[serde(default = "default_gpu_layers_mode")]
595    pub gpu_layers_mode: crate::models::GpuLayersMode,
596    #[serde(default)]
597    pub split_mode: SplitMode,
598    #[serde(default)]
599    pub tensor_split: String,
600    #[serde(default)]
601    pub main_gpu: i32,
602    #[serde(default)]
603    pub fit: bool,
604    #[serde(default)]
605    pub lora: Option<PathBuf>,
606    #[serde(default)]
607    pub lora_scaled: Option<(PathBuf, f32)>,
608    #[serde(default)]
609    pub rpc: String,
610    #[serde(default)]
611    pub embedding: bool,
612    #[serde(default)]
613    pub flash_attn: bool,
614    #[serde(default)]
615    pub jinja: bool,
616    #[serde(default)]
617    pub chat_template: Option<String>,
618    #[serde(default)]
619    pub chat_template_kwargs: Option<String>,
620    #[serde(default)]
621    pub expert_count: i32,
622
623    // Sampling
624    #[serde(default)]
625    pub seed: i32,
626    #[serde(default)]
627    pub temperature: f32,
628    #[serde(default)]
629    pub top_k: i32,
630    #[serde(default)]
631    pub top_p: f32,
632    #[serde(default)]
633    pub min_p: f32,
634    #[serde(default)]
635    pub typical_p: f32,
636    #[serde(default)]
637    pub mirostat: Mirostat,
638    #[serde(default)]
639    pub mirostat_lr: f32,
640    #[serde(default)]
641    pub mirostat_ent: f32,
642    #[serde(default)]
643    pub ignore_eos: bool,
644    #[serde(default)]
645    pub samplers: Samplers,
646
647    // Repetition
648    #[serde(default)]
649    pub repeat_penalty: f32,
650    #[serde(default)]
651    pub repeat_last_n: i32,
652    #[serde(default = "default_presence_penalty")]
653    pub presence_penalty: Option<f32>,
654    #[serde(default = "default_frequency_penalty")]
655    pub frequency_penalty: Option<f32>,
656    #[serde(default)]
657    pub dry_multiplier: f32,
658    #[serde(default)]
659    pub dry_base: f32,
660    #[serde(default)]
661    pub dry_allowed_length: i32,
662    #[serde(default)]
663    pub dry_penalty_last_n: i32,
664
665    // RoPE
666    #[serde(default)]
667    pub rope_scaling: RopeScaling,
668    #[serde(default)]
669    pub rope_scale: f32,
670    #[serde(default)]
671    pub rope_freq_base: f32,
672    #[serde(default)]
673    pub rope_freq_scale: f32,
674    #[serde(default)]
675    pub rope_yarn_enabled: bool,
676
677    // Server
678    #[serde(default)]
679    pub host: String,
680    #[serde(default)]
681    pub port: u16,
682    #[serde(default)]
683    pub timeout: u32,
684    #[serde(default = "default_cache_prompt")]
685    pub cache_prompt: bool,
686    #[serde(default)]
687    pub cache_reuse: u32,
688    #[serde(default)]
689    pub webui: bool,
690    #[serde(default)]
691    pub ws_server_enabled: bool,
692    #[serde(default = "default_ws_server_port")]
693    pub ws_server_port: u16,
694    #[serde(default)]
695    pub ws_server_auth_key: Option<String>,
696    #[serde(default = "default_ws_server_tls_enabled")]
697    pub ws_server_tls_enabled: bool,
698    #[serde(default)]
699    pub ws_server_tls_cert: Option<String>,
700    #[serde(default)]
701    pub ws_server_tls_key: Option<String>,
702    #[serde(default)]
703    pub router_max_models: u32,
704    #[serde(default)]
705    pub server_mode: crate::models::ServerMode,
706
707    // Other
708    #[serde(default = "default_max_tokens")]
709    pub max_tokens: Option<u32>,
710    #[serde(default)]
711    pub cache_type: CacheType,
712    #[serde(default)]
713    pub backend: Backend,
714    /// Platform override: "linux", "windows", or "macos". If None, auto-detected.
715    #[serde(default)]
716    pub platform: Option<String>,
717    #[serde(default)]
718    pub llama_cpp_version_cpu: Option<String>,
719    #[serde(default)]
720    pub llama_cpp_version_vulkan: Option<String>,
721    #[serde(default)]
722    pub llama_cpp_version_rocm: Option<String>,
723    #[serde(default)]
724    pub llama_cpp_version_rocm_lemonade: Option<String>,
725    #[serde(default)]
726    pub llama_cpp_version_cuda: Option<String>,
727
728    // API
729    #[serde(default)]
730    pub api_endpoint_enabled: bool,
731    #[serde(default = "default_api_endpoint_port")]
732    pub api_endpoint_port: u16,
733    #[serde(default)]
734    pub spec_type: String,
735    #[serde(default)]
736    pub draft_tokens: u32,
737    #[serde(default)]
738    pub tags: Vec<String>,
739}
740
741fn default_api_endpoint_port() -> u16 {
742    49222
743}
744
745fn default_system_prompt_preset_name() -> String {
746    "General".to_string()
747}
748
749fn default_cache_type_k() -> Option<CacheTypeK> {
750    None
751}
752fn default_cache_type_v() -> Option<CacheTypeV> {
753    None
754}
755fn default_presence_penalty() -> Option<f32> {
756    None
757}
758fn default_frequency_penalty() -> Option<f32> {
759    None
760}
761fn default_max_tokens() -> Option<u32> {
762    None
763}
764fn default_cache_prompt() -> bool {
765    true
766}
767fn default_ws_server_port() -> u16 {
768     49223
769 }
770 fn default_ws_server_tls_enabled() -> bool {
771     true
772 }
773 fn default_gpu_layers_mode() -> crate::models::GpuLayersMode {
774    crate::models::GpuLayersMode::Auto
775}
776
777impl Default for DefaultParams {
778    fn default() -> Self {
779        Self {
780            // Loading
781            context_length: 131072,
782            threads: physical_cores(),
783            threads_batch: 8,
784            batch_size: 512,
785            ubatch_size: 512,
786            cache_type_k: None,
787            cache_type_v: None,
788            keep: 0,
789            swa_full: false,
790            mlock: false,
791            mmap: true,
792            numa: NumMode::None,
793            uniform_cache: true,
794            kv_cache_offload: true,
795            parallel: 1,
796            max_concurrent_predictions: None,
797            system_prompt: "You are a helpful assistant.".to_string(),
798            system_prompt_preset_name: "General".to_string(),
799
800            // GPU
801            gpu_layers: -1,
802            gpu_layers_mode: crate::models::GpuLayersMode::Auto,
803            split_mode: SplitMode::Layer,
804            tensor_split: String::new(),
805            main_gpu: 0,
806            fit: true,
807            lora: None,
808            lora_scaled: None,
809            rpc: String::new(),
810            embedding: false,
811            flash_attn: true,
812            jinja: true,
813            chat_template: None,
814            chat_template_kwargs: None,
815            expert_count: -1,
816
817            // Sampling
818            seed: -1,
819            temperature: 0.8,
820            top_k: 40,
821            top_p: 0.95,
822            min_p: 0.0,
823            typical_p: 1.0,
824            mirostat: Mirostat::Off,
825            mirostat_lr: 0.1,
826            mirostat_ent: 5.0,
827            ignore_eos: false,
828            samplers: Samplers::default(),
829
830            // Repetition
831            repeat_penalty: 1.1,
832            repeat_last_n: 64,
833            presence_penalty: None,
834            frequency_penalty: None,
835            dry_multiplier: 0.0,
836            dry_base: 1.75,
837            dry_allowed_length: 2,
838            dry_penalty_last_n: -1,
839
840            // RoPE
841            rope_scaling: RopeScaling::None,
842            rope_scale: 1.0,
843            rope_freq_base: 0.0,
844            rope_freq_scale: 1.0,
845            rope_yarn_enabled: false,
846
847            // Server
848            host: "127.0.0.1".to_string(),
849            port: 8080,
850            timeout: 600,
851            cache_prompt: true,
852            cache_reuse: 0,
853            webui: false,
854            ws_server_enabled: false,
855            ws_server_port: 49223,
856            ws_server_auth_key: None,
857            ws_server_tls_enabled: true,
858            ws_server_tls_cert: None,
859            ws_server_tls_key: None,
860            router_max_models: 4,
861            server_mode: crate::models::ServerMode::Normal,
862
863            // Other
864            max_tokens: None,
865            cache_type: CacheType::F16,
866            backend: {
867                use crate::backend::hardware::{GpuVendor, detect_gpu_vendors};
868                let vendors = detect_gpu_vendors();
869                let mut result = Backend::Cpu;
870                for v in &vendors {
871                    if matches!(v, GpuVendor::Nvidia) {
872                        result = Backend::Cuda;
873                        break;
874                    }
875                    if matches!(v, GpuVendor::Amd) {
876                        result = Backend::Rocm;
877                        break;
878                    }
879                    if matches!(v, GpuVendor::Intel) {
880                        result = Backend::Vulkan;
881                        break;
882                    }
883                }
884                result
885            },
886            platform: None,
887            llama_cpp_version_cpu: None,
888            llama_cpp_version_vulkan: None,
889            llama_cpp_version_rocm: None,
890            llama_cpp_version_rocm_lemonade: None,
891            llama_cpp_version_cuda: None,
892            api_endpoint_enabled: false,
893            api_endpoint_port: 49222,
894            spec_type: String::new(),
895            draft_tokens: 0,
896            tags: Vec::new(),
897        }
898    }
899}
900
901impl Default for Config {
902    fn default() -> Self {
903        Self {
904            models_dirs: vec![
905                dirs::data_dir()
906                    .unwrap_or_default()
907                    .join("llm-manager")
908                    .join("models"),
909            ],
910            llama_server: "llama-server".into(),
911            default: DefaultParams::default(),
912            model_overrides: Default::default(),
913            profiles: Default::default(),
914            system_prompt_presets: Default::default(),
915            rpc_workers: Vec::new(),
916            search_limit: default_search_limit(),
917            active_panel: ActivePanel::Models,
918            left_pct: 55,
919        }
920    }
921}
922
923impl Config {
924    pub fn config_path() -> PathBuf {
925        config_base_dir()
926            .join("llm-manager")
927            .join("config.yaml")
928    }
929
930    /// Validate config values and return a list of warnings for invalid entries.
931    pub fn validate(&self) -> Vec<String> {
932        let mut warnings = Vec::new();
933        let default = &self.default;
934
935        // Numeric range checks
936        if default.context_length < 512 || default.context_length > 131072 {
937            warnings.push(format!(
938                "context_length {} is outside recommended range 512-131072",
939                default.context_length
940            ));
941        }
942        if default.temperature < 0.0 || default.temperature > 2.0 {
943            warnings.push(format!(
944                "temperature {} is outside recommended range 0.0-2.0",
945                default.temperature
946            ));
947        }
948        if (default.top_p < 0.0 || default.top_p > 1.0) && default.top_p != 0.0 {
949            warnings.push(format!(
950                "top_p {} is outside recommended range 0.0-1.0",
951                default.top_p
952            ));
953        }
954        if (default.repeat_penalty < 0.0 || default.repeat_penalty > 3.0)
955            && default.repeat_penalty != 1.0
956        {
957            warnings.push(format!(
958                "repeat_penalty {} is outside recommended range 0.0-3.0",
959                default.repeat_penalty
960            ));
961        }
962        if default.mirostat_lr < 0.0 || default.mirostat_lr > 1.0 {
963            warnings.push(format!(
964                "mirostat_lr {} is outside recommended range 0.0-1.0",
965                default.mirostat_lr
966            ));
967        }
968        if default.mirostat_ent < 0.0 || default.mirostat_ent > 10.0 {
969            warnings.push(format!(
970                "mirostat_ent {} is outside recommended range 0.0-10.0",
971                default.mirostat_ent
972            ));
973        }
974
975        if default.timeout < 1 {
976            warnings.push(format!(
977                "timeout {} must be at least 1 second",
978                default.timeout
979            ));
980        }
981
982        // Path validation
983        if let Some(lora) = &default.lora
984            && !lora.exists() {
985                warnings.push(format!("lora path {} does not exist", lora.display()));
986            }
987        if let Some((lora, _)) = &default.lora_scaled
988            && !lora.exists() {
989                warnings.push(format!("lora path {} does not exist", lora.display()));
990            }
991
992        // Model override validation
993        for model_name in self.model_overrides.keys() {
994            if let Some(override_settings) = self.model_overrides.get(model_name.as_str()) {
995                if let Some(lora) = &override_settings.lora
996                    && !lora.exists() {
997                        warnings.push(format!(
998                            "model '{}' lora path {} does not exist",
999                            model_name,
1000                            lora.display()
1001                        ));
1002                    }
1003                if let Some((lora, _)) = &override_settings.lora_scaled
1004                    && !lora.exists() {
1005                        warnings.push(format!(
1006                            "model '{}' lora path {} does not exist",
1007                            model_name,
1008                            lora.display()
1009                        ));
1010                    }
1011            }
1012        }
1013
1014        warnings
1015    }
1016
1017    /// Resolve settings for a specific model and profile.
1018    pub fn resolve_settings(
1019        &self,
1020        model_name: Option<&str>,
1021        profile_name: Option<&str>,
1022    ) -> crate::models::ModelSettings {
1023        let mut settings = crate::models::ModelSettings::from_config(self);
1024
1025        // Apply model-specific override
1026        if let Some(name) = model_name
1027            && let Some(override_settings) = self.model_overrides.get(name)
1028        {
1029            override_settings.apply(&mut settings);
1030        }
1031
1032        // Apply profile override if specified
1033        if let Some(p_name) = profile_name {
1034            if let Some(profile) = self.profiles.get(p_name) {
1035                profile.settings.apply(&mut settings);
1036            } else if let Some(profile) = builtin_profiles().iter().find(|p| p.name == p_name) {
1037                profile.settings.apply(&mut settings);
1038            }
1039        }
1040
1041        settings
1042    }
1043
1044    /// Get a system prompt preset content by name.
1045    pub fn get_preset_content(&self, name: &str) -> Option<String> {
1046        self.system_prompt_presets
1047            .get(name)
1048            .map(|p| p.content.clone())
1049    }
1050
1051    fn normalize_config(mut config: Config) -> Config {
1052        // normalize models_dirs
1053        for path in &mut config.models_dirs {
1054            let path_str = path.to_string_lossy();
1055            if let Some(stripped) = path_str.strip_prefix("~/") {
1056                let home = dirs::home_dir().unwrap_or_default();
1057                *path = home.join(stripped);
1058            } else if !path.is_absolute() {
1059                let home = dirs::home_dir().unwrap_or_default();
1060                *path = home.join(path_str.as_ref());
1061            }
1062        }
1063
1064        // Merge built-in profiles into in-memory cache (do not persist to disk)
1065        for p in builtin_profiles() {
1066            if config.profiles.get(&p.name).is_none() {
1067                config.profiles.insert_builtin(p);
1068            }
1069        }
1070
1071        // Merge built-in system prompt presets into in-memory cache (do not persist to disk)
1072        for p in builtin_system_prompt_presets() {
1073            if config.system_prompt_presets.get(&p.name).is_none() {
1074                config.system_prompt_presets.insert_builtin(p);
1075            }
1076        }
1077        config
1078    }
1079
1080    fn load_impl(path: &PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
1081        let content = std::fs::read_to_string(path)?;
1082        let config: Config = serde_yaml::from_str(&content)
1083            .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
1084        let config = Self::normalize_config(config);
1085        let config = config.auto_detect_platform();
1086        let warnings = config.validate();
1087        if !warnings.is_empty() {
1088            eprintln!("Config validation warnings:");
1089            for warning in &warnings {
1090                eprintln!("  - {}", warning);
1091            }
1092        }
1093        Ok(config)
1094    }
1095
1096    pub fn load() -> Result<Self, Box<dyn std::error::Error>> {
1097        let path = Self::config_path();
1098        if path.exists() {
1099            Self::load_impl(&path)
1100        } else {
1101            let mut config = Config::default();
1102            config.save()?;
1103            Ok(config)
1104        }
1105    }
1106
1107    pub fn load_from(path: PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
1108        if path.exists() {
1109            Self::load_impl(&path)
1110        } else {
1111            Err(format!("Config file not found: {}", path.display()).into())
1112        }
1113    }
1114
1115    /// Auto-detect the platform if not explicitly set in config.
1116    fn auto_detect_platform(mut self) -> Self {
1117        if self.default.platform.is_none() {
1118            self.default.platform =
1119                Some(
1120                    crate::backend::hardware::platform_name(
1121                        crate::backend::hardware::detect_platform(),
1122                    )
1123                    .to_string(),
1124                );
1125        }
1126        self
1127    }
1128
1129    pub fn save(&mut self) -> Result<(), Box<dyn std::error::Error>> {
1130        let path = Self::config_path();
1131        if let Some(parent) = path.parent() {
1132            std::fs::create_dir_all(parent)?;
1133        }
1134        let content = serde_yaml::to_string(self)?;
1135        std::fs::write(&path, content)?;
1136        // Persist model configs to individual YAML files
1137        let entries: Vec<(String, ModelOverride)> = self
1138            .model_overrides
1139            .keys()
1140            .iter()
1141            .filter_map(|k| self.model_overrides.get(k).map(|v| (k.clone(), v.clone())))
1142            .collect();
1143        for (name, cfg) in entries {
1144            self.model_overrides.save(&name, &cfg);
1145        }
1146        // Persist user profiles to individual YAML files (skip built-ins)
1147        for profile in self.profiles.user_profiles() {
1148            self.profiles.save(&profile);
1149        }
1150        // Persist user presets to individual YAML files (skip built-ins)
1151        for preset in self.system_prompt_presets.user_presets() {
1152            self.system_prompt_presets.save(&preset);
1153        }
1154        Ok(())
1155    }
1156
1157    pub fn merged_profiles(&self) -> Vec<Profile> {
1158        self.profiles.all()
1159    }
1160
1161    pub fn merged_presets(&self) -> Vec<SystemPromptPreset> {
1162        self.system_prompt_presets.all()
1163    }
1164}
1165
1166#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1167pub enum LogLevel {
1168    Info,
1169    Warning,
1170    Error,
1171}
1172
1173impl LogLevel {
1174    pub fn label(&self) -> &'static str {
1175        match self {
1176            LogLevel::Info => "INFO",
1177            LogLevel::Warning => "WARNING",
1178            LogLevel::Error => "ERROR",
1179        }
1180    }
1181}
1182
1183#[derive(Debug, Clone)]
1184pub struct LogEntry {
1185    pub timestamp: String,
1186    pub level: LogLevel,
1187    pub message: String,
1188}
1189
1190impl LogEntry {
1191    pub fn new(message: impl Into<String>, level: LogLevel) -> Self {
1192        let timestamp = Local::now().format("%H:%M:%S").to_string();
1193        let message = sanitize_log(&message.into());
1194        Self {
1195            timestamp,
1196            level,
1197            message,
1198        }
1199    }
1200}
1201
1202/// Sanitize log messages to prevent TUI layout breakages.
1203/// Strips non-printable characters and control sequences, and limits length.
1204fn sanitize_log(input: &str) -> String {
1205    // Limit length to avoid layout/perf issues with massive lines
1206    let max_len = 2000;
1207    let chars: Vec<char> = input.chars().collect();
1208    let truncated = chars.len() > max_len;
1209    let chars = if truncated {
1210        chars[..max_len].to_vec()
1211    } else {
1212        chars
1213    };
1214
1215    let mut output = String::with_capacity(chars.len());
1216    for c in chars {
1217        // Strip ALL control characters except newline and tab.
1218        // Critically: strip \r (carriage return) as it breaks TUI rendering.
1219        if c.is_control() && c != '\n' && c != '\t' {
1220            continue;
1221        }
1222        output.push(c);
1223    }
1224
1225    // Replace tabs with spaces for consistent rendering
1226    let output = output.replace('\t', "    ");
1227
1228    // Final trim to remove trailing junk
1229    let mut result = output.trim_end().to_string();
1230    if truncated {
1231        result.push_str("... (truncated)");
1232    }
1233    result
1234}
llm_manager/config.rs

llm_manager/
config.rs
