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/// Default system prompt used when no preset is selected.
23pub const DEFAULT_SYSTEM_PROMPT: &str =
24    "You are an expert software developer. Write clean, well-documented code. Explain your reasoning and suggest improvements.";
25
26/// Resolve the base config directory with a safe fallback chain.
27///
28/// Prefers `dirs::config_dir()` (XDG on Linux, ~/Library/Application Support on macOS,
29/// etc.), falls back to `~/.config`, and lastly `./.llm-manager` if both fail.
30pub fn config_base_dir() -> PathBuf {
31    if let Some(d) = dirs::config_dir() {
32        return d;
33    }
34    if let Some(home) = dirs::home_dir() {
35        return home.join(".config");
36    }
37    PathBuf::from(".").join(".llm-manager")
38}
39
40/// Count physical CPU cores on Linux (ignores hyperthreading).
41/// Falls back to 1 if the file can't be read or parsing fails.
42pub fn physical_cores() -> u32 {
43    let content = match std::fs::read_to_string("/proc/cpuinfo") {
44        Ok(c) => c,
45        Err(_) => {
46            return std::thread::available_parallelism()
47                .map(|p| p.get() as u32)
48                .unwrap_or(1);
49        }
50    };
51    let mut seen = HashSet::new();
52    let mut cur_phys: Option<&str> = None;
53    let mut cur_core: Option<&str> = None;
54    for line in content.lines() {
55        if let Some((key, val)) = line.split_once(':') {
56            let key = key.trim();
57            let val = val.trim();
58            match key {
59                "physical id" => cur_phys = Some(val),
60                "core id" => cur_core = Some(val),
61                _ => {}
62            }
63            if let (Some(phys), Some(core)) = (cur_phys, cur_core) {
64                seen.insert((phys, core));
65            }
66        }
67    }
68    seen.len() as u32
69}
70
71/// A remote RPC worker for distributed inference.
72#[derive(Debug, Clone, Serialize, Deserialize)]
73pub struct RpcWorker {
74    #[serde(default)]
75    pub selected: bool,
76    #[serde(default)]
77    pub name: String,
78    pub ip: String,
79    #[serde(default = "default_rpc_port")]
80    pub port: u16,
81}
82
83fn default_rpc_port() -> u16 {
84    50052
85}
86
87/// Global configuration.
88#[derive(Debug, Clone, Serialize, Deserialize)]
89pub struct Config {
90    pub models_dirs: Vec<PathBuf>,
91    pub llama_server: PathBuf,
92    pub default: DefaultParams,
93    /// Per-model overrides (keyed by model file name, stored as YAML in models/).
94    #[serde(default, skip)]
95    pub model_overrides: ModelConfigStore,
96    /// Named profiles of settings presets (stored as YAML in profiles/).
97    #[serde(default, skip)]
98    pub profiles: ProfileStore,
99    /// System prompt presets (stored as YAML in presets/).
100    #[serde(default, skip)]
101    pub system_prompt_presets: PresetStore,
102    /// RPC Workers for distributed inference.
103    #[serde(default)]
104    pub rpc_workers: Vec<RpcWorker>,
105    /// Number of results per HuggingFace search query.
106    #[serde(default = "default_search_limit")]
107    pub search_limit: u32,
108    /// The last focused panel position (for restoring on next launch).
109    #[serde(default)]
110    pub active_panel: crate::tui::app::ActivePanel,
111    /// Left panel width percentage (20-80).
112    #[serde(default = "default_left_pct")]
113    pub left_pct: u16,
114}
115
116fn default_left_pct() -> u16 {
117    55
118}
119
120fn default_search_limit() -> u32 {
121    50
122}
123
124/// A named profile of settings.
125#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
126pub struct Profile {
127    pub name: String,
128    /// Brief description shown in the profile list.
129    pub description: String,
130    /// The settings for this profile.
131    #[serde(default)]
132    pub settings: ModelOverride,
133}
134
135impl Profile {
136    /// Apply this profile's settings to a base ModelSettings.
137    pub fn apply(&self, mut base: crate::models::ModelSettings) -> crate::models::ModelSettings {
138        self.settings.apply(&mut base);
139        base
140    }
141}
142
143/// A named system prompt preset.
144#[derive(Debug, Clone, Serialize, Deserialize)]
145pub struct SystemPromptPreset {
146    pub name: String,
147    pub description: String,
148    pub content: String,
149}
150
151/// Built-in system prompt presets.
152pub fn builtin_system_prompt_presets() -> Vec<SystemPromptPreset> {
153    vec![
154        SystemPromptPreset {
155            name: "General".into(),
156            description: "General-purpose assistant".into(),
157            content: "You are a helpful assistant.".into(),
158        },
159        SystemPromptPreset {
160            name: "Coder".into(),
161            description: "Expert software developer".into(),
162            content: "You are an expert software developer. Write clean, well-documented code. Explain your reasoning and suggest improvements.".into(),
163        },
164        SystemPromptPreset {
165            name: "Thinker".into(),
166            description: "Analytical and thoughtful".into(),
167            content: "You are a thoughtful and analytical AI assistant. Think carefully before answering. Provide well-reasoned responses with clear explanations.".into(),
168        },
169        SystemPromptPreset {
170            name: "Mathematician".into(),
171            description: "Expert in mathematics".into(),
172            content: "You are an expert in mathematics. Provide clear, step-by-step solutions to mathematical problems. Show your reasoning and explain key concepts.".into(),
173        },
174    ]
175}
176
177#[derive(Debug, Clone, Serialize, Deserialize, Default, PartialEq)]
178pub struct ModelOverride {
179    // Loading
180    pub context_length: Option<u32>,
181    pub batch_size: Option<u32>,
182    pub ubatch_size: Option<u32>,
183    pub cache_type_k: Option<CacheTypeK>,
184    pub cache_type_v: Option<CacheTypeV>,
185    pub keep: Option<i32>,
186    pub swa_full: Option<bool>,
187    pub mlock: Option<bool>,
188    pub mmap: Option<bool>,
189    pub numa: Option<NumMode>,
190    pub uniform_cache: Option<bool>,
191    pub system_prompt: Option<String>,
192    pub system_prompt_preset_name: Option<String>,
193    pub max_concurrent_predictions: Option<u32>,
194    pub threads: Option<u32>,
195    pub threads_batch: Option<u32>,
196    pub parallel: Option<u32>,
197
198    // GPU
199    pub gpu_layers: Option<i32>,
200    pub split_mode: Option<SplitMode>,
201    pub tensor_split: Option<String>,
202    pub main_gpu: Option<i32>,
203    pub fit: Option<bool>,
204    pub lora: Option<PathBuf>,
205    pub lora_scaled: Option<(PathBuf, f32)>,
206    pub rpc: Option<String>,
207    pub embedding: Option<bool>,
208    pub kv_cache_offload: Option<bool>,
209    pub flash_attn: Option<bool>,
210    pub jinja: Option<bool>,
211    pub chat_template: Option<String>,
212    pub chat_template_kwargs: Option<String>,
213    pub expert_count: Option<i32>,
214    pub gpu_layers_mode: Option<crate::models::GpuLayersMode>,
215
216    // Sampling
217    pub seed: Option<i32>,
218    pub temperature: Option<f32>,
219    pub top_k: Option<i32>,
220    pub top_p: Option<f32>,
221    pub min_p: Option<f32>,
222    pub typical_p: Option<f32>,
223    pub mirostat: Option<Mirostat>,
224    pub mirostat_lr: Option<f32>,
225    pub mirostat_ent: Option<f32>,
226    pub ignore_eos: Option<bool>,
227    pub samplers: Option<Samplers>,
228
229    // Repetition
230    pub repeat_penalty: Option<f32>,
231    pub repeat_last_n: Option<i32>,
232    pub presence_penalty: Option<f32>,
233    pub frequency_penalty: Option<f32>,
234    pub dry_multiplier: Option<f32>,
235    pub dry_base: Option<f32>,
236    pub dry_allowed_length: Option<i32>,
237    pub dry_penalty_last_n: Option<i32>,
238
239    // RoPE
240    pub rope_scaling: Option<RopeScaling>,
241    pub rope_scale: Option<f32>,
242    pub rope_freq_base: Option<f32>,
243    pub rope_freq_scale: Option<f32>,
244    pub rope_yarn_enabled: Option<bool>,
245
246    // Server
247    pub cache_prompt: Option<bool>,
248    pub cache_reuse: Option<u32>,
249    pub webui: Option<bool>,
250
251    // Other
252    pub max_tokens: Option<u32>,
253    pub cache_type: Option<CacheType>,
254    pub llama_cpp_version_cpu: Option<String>,
255    pub llama_cpp_version_vulkan: Option<String>,
256    pub llama_cpp_version_rocm: Option<String>,
257    pub llama_cpp_version_rocm_lemonade: Option<String>,
258    pub llama_cpp_version_cuda: Option<String>,
259    pub spec_type: Option<String>,
260    pub draft_tokens: Option<u32>,
261    pub tags: Option<Vec<String>>,
262}
263
264/// Apply a scalar Copy field from override: `base.f = self.f.unwrap_or(base.f)`.
265macro_rules! apply_scalar {
266    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
267        $(
268            $base.$field = $self.$field.unwrap_or($base.$field);
269        )+
270    };
271}
272
273/// Apply a Clone field from override: `if let Some(v) = &self.f { base.f = v.clone(); }`.
274macro_rules! apply_clone {
275    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
276        $(
277            if let Some(v) = &$self.$field {
278                $base.$field = v.clone();
279            }
280        )+
281    };
282}
283
284/// Apply an Option<T> field from override: `if let Some(v) = &self.f { base.f = Some(v.clone()); }`.
285macro_rules! apply_option {
286    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
287        $(
288            if let Some(v) = &$self.$field {
289                $base.$field = Some(v.clone());
290            }
291        )+
292    };
293}
294
295impl ModelOverride {
296    pub fn from_settings(s: &crate::models::ModelSettings) -> Self {
297        Self {
298            context_length: Some(s.context_length),
299            batch_size: Some(s.batch_size),
300            ubatch_size: Some(s.ubatch_size),
301            cache_type_k: s.cache_type_k,
302            cache_type_v: s.cache_type_v,
303            keep: Some(s.keep),
304            swa_full: Some(s.swa_full),
305            mlock: Some(s.mlock),
306            mmap: Some(s.mmap),
307            numa: Some(s.numa),
308            uniform_cache: Some(s.uniform_cache),
309            system_prompt: Some(s.system_prompt.clone()),
310            system_prompt_preset_name: Some(s.system_prompt_preset_name.clone()),
311            max_concurrent_predictions: s.max_concurrent_predictions,
312            threads: Some(s.threads),
313            threads_batch: Some(s.threads_batch),
314            parallel: Some(s.parallel),
315            gpu_layers: Some(match s.gpu_layers_mode {
316                crate::models::GpuLayersMode::Auto => 0,
317                crate::models::GpuLayersMode::Specific(n) => n as i32,
318                crate::models::GpuLayersMode::All => -1,
319            }),
320            gpu_layers_mode: Some(s.gpu_layers_mode),
321            split_mode: Some(s.split_mode),
322            tensor_split: Some(s.tensor_split.clone()),
323            main_gpu: Some(s.main_gpu),
324            fit: Some(s.fit),
325            lora: s.lora.clone(),
326            lora_scaled: s.lora_scaled.clone(),
327            rpc: Some(s.rpc.clone()),
328            embedding: Some(s.embedding),
329            kv_cache_offload: Some(s.kv_cache_offload),
330            flash_attn: Some(s.flash_attn),
331            jinja: Some(s.jinja),
332            chat_template: s.chat_template.clone(),
333            chat_template_kwargs: s.chat_template_kwargs.clone(),
334            expert_count: Some(s.expert_count),
335            seed: Some(s.seed),
336            temperature: Some(s.temperature),
337            top_k: Some(s.top_k),
338            top_p: Some(s.top_p),
339            min_p: Some(s.min_p),
340            typical_p: Some(s.typical_p),
341            mirostat: Some(s.mirostat),
342            mirostat_lr: Some(s.mirostat_lr),
343            mirostat_ent: Some(s.mirostat_ent),
344            ignore_eos: Some(s.ignore_eos),
345            samplers: Some(s.samplers.clone()),
346            repeat_penalty: Some(s.repeat_penalty),
347            repeat_last_n: Some(s.repeat_last_n),
348            presence_penalty: s.presence_penalty,
349            frequency_penalty: s.frequency_penalty,
350            dry_multiplier: Some(s.dry_multiplier),
351            dry_base: Some(s.dry_base),
352            dry_allowed_length: Some(s.dry_allowed_length),
353            dry_penalty_last_n: Some(s.dry_penalty_last_n),
354            rope_scaling: Some(s.rope_scaling),
355            rope_scale: Some(s.rope_scale),
356            rope_freq_base: Some(s.rope_freq_base),
357            rope_freq_scale: Some(s.rope_freq_scale),
358            rope_yarn_enabled: Some(s.rope_yarn_enabled),
359            cache_prompt: Some(s.cache_prompt),
360            cache_reuse: Some(s.cache_reuse),
361            webui: Some(s.webui),
362            max_tokens: s.max_tokens,
363            cache_type: Some(s.cache_type),
364            llama_cpp_version_cpu: s.llama_cpp_version_cpu.clone(),
365            llama_cpp_version_vulkan: s.llama_cpp_version_vulkan.clone(),
366            llama_cpp_version_rocm: s.llama_cpp_version_rocm.clone(),
367            llama_cpp_version_rocm_lemonade: s.llama_cpp_version_rocm_lemonade.clone(),
368            llama_cpp_version_cuda: s.llama_cpp_version_cuda.clone(),
369            spec_type: Some(s.spec_type.clone()),
370            draft_tokens: Some(s.draft_tokens),
371            tags: Some(s.tags.clone()),
372        }
373    }
374
375    /// Merge override into a base ModelSettings (in-place).
376    pub fn apply(&self, base: &mut crate::models::ModelSettings) {
377        // Override values always take precedence. For Option<T> fields,
378        // the override value (even None) is explicitly set by the user.
379
380        // Scalar Copy fields: base.f = self.f.unwrap_or(base.f)
381        apply_scalar!(self, base,
382            context_length, batch_size, ubatch_size, keep, swa_full, mlock, mmap,
383            numa, uniform_cache, kv_cache_offload, threads, threads_batch, parallel,
384            split_mode, main_gpu, fit, embedding, flash_attn, jinja, expert_count,
385            seed, temperature, top_k, top_p, min_p, typical_p,
386            mirostat, mirostat_lr, mirostat_ent, ignore_eos,
387            repeat_penalty, repeat_last_n,
388            dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
389            rope_scaling, rope_scale, rope_freq_base, rope_freq_scale, rope_yarn_enabled,
390            cache_prompt, cache_reuse, webui, cache_type,
391            draft_tokens, gpu_layers_mode,
392        );
393
394        // Cloneable fields: if let Some(v) = &self.f { base.f = v.clone(); }
395        apply_clone!(self, base,
396            system_prompt, system_prompt_preset_name, tensor_split, rpc,
397            samplers, spec_type, tags,
398        );
399
400        // Option<T> fields: if let Some(v) = &self.f { base.f = Some(v.clone()); }
401        apply_option!(self, base,
402            lora, lora_scaled, chat_template, chat_template_kwargs,
403            llama_cpp_version_cpu, llama_cpp_version_vulkan,
404            llama_cpp_version_rocm, llama_cpp_version_rocm_lemonade,
405            llama_cpp_version_cuda,
406        );
407
408        // Direct Option<T> assignment (same type in both structs)
409        base.cache_type_k = self.cache_type_k;
410        base.cache_type_v = self.cache_type_v;
411        base.presence_penalty = self.presence_penalty;
412        base.frequency_penalty = self.frequency_penalty;
413        base.max_tokens = self.max_tokens;
414
415        // Special: max_concurrent_predictions uses or() for Option chaining
416        base.max_concurrent_predictions = self
417            .max_concurrent_predictions
418            .or(base.max_concurrent_predictions);
419
420        // Special: gpu_layers converts i32 legacy field to GpuLayersMode enum
421        // Only applies when gpu_layers is explicitly set in the override.
422        if let Some(n) = self.gpu_layers {
423            base.gpu_layers_mode = match n {
424                n if n < 0 => crate::models::GpuLayersMode::All,
425                n => crate::models::GpuLayersMode::Specific(n as u32),
426            };
427        }
428
429        // FIELD ACCOUNTING (ModelOverride: 87 fields):
430        // - apply_scalar: 53 fields
431        // - apply_clone: 7 fields
432        // - apply_option: 10 fields
433        // - direct Option assign: 5 fields (cache_type_k, cache_type_v, presence_penalty,
434        //   frequency_penalty, max_tokens)
435        // - special: 1 field (max_concurrent_predictions)
436        // - conditional: gpu_layers overrides gpu_layers_mode only when Some
437        // - NOT in ModelSettings: 0 (all ModelOverride fields mapped above)
438        //
439        // ModelSettings fields NOT in ModelOverride (not overridable):
440        // host, port, timeout, backend, platform, router_max_models, server_mode,
441        // api_endpoint_enabled, api_endpoint_port
442        //
443        // When adding a field: ensure it appears in exactly one category above.
444    }
445}
446
447/// Built-in profiles with sensible defaults for popular model families.
448pub fn builtin_profiles() -> Vec<Profile> {
449    vec![
450        Profile {
451            name: "Qwen".into(),
452            description: "Optimized for Qwen models (dense)".into(),
453            settings: ModelOverride {
454                context_length: Some(131072),
455                temperature: Some(0.7),
456                top_k: Some(20),
457                top_p: Some(0.95),
458                max_tokens: Some(4096),
459                presence_penalty: Some(0.0),
460                uniform_cache: Some(true),
461                jinja: Some(true),
462                ..Default::default()
463            },
464        },
465        Profile {
466            name: "Qwen-MoE".into(),
467            description: "Optimized for Qwen MoE models (35B-A3B)".into(),
468            settings: ModelOverride {
469                context_length: Some(131072),
470                temperature: Some(0.8),
471                top_k: Some(20),
472                top_p: Some(0.95),
473                max_tokens: Some(4096),
474                presence_penalty: Some(1.5),
475                uniform_cache: Some(true),
476                jinja: Some(true),
477                ..Default::default()
478            },
479        },
480        Profile {
481            name: "Qwen-Coding".into(),
482            description: "Optimized for Qwen models in coding mode".into(),
483            settings: ModelOverride {
484                context_length: Some(131072),
485                temperature: Some(0.6),
486                top_k: Some(20),
487                top_p: Some(0.95),
488                max_tokens: Some(4096),
489                presence_penalty: Some(0.0),
490                uniform_cache: Some(true),
491                jinja: Some(true),
492                ..Default::default()
493            },
494        },
495        Profile {
496            name: "Gemma".into(),
497            description: "Optimized for Gemma 2/4 models".into(),
498            settings: ModelOverride {
499                context_length: Some(131072),
500                min_p: Some(0.1),
501                temperature: Some(1.0),
502                top_k: Some(65),
503                top_p: Some(0.95),
504                max_tokens: Some(4096),
505                uniform_cache: Some(true),
506                jinja: Some(true),
507                ..Default::default()
508            },
509        },
510        Profile {
511            name: "Llama".into(),
512            description: "Optimized for Llama 3.1/3.3 models".into(),
513            settings: ModelOverride {
514                context_length: Some(131072),
515                temperature: Some(0.7),
516                top_p: Some(0.9),
517                repeat_penalty: Some(1.1),
518                max_tokens: Some(4096),
519                uniform_cache: Some(true),
520                jinja: Some(true),
521                ..Default::default()
522            },
523        },
524        Profile {
525            name: "Mistral".into(),
526            description: "Optimized for Mistral 7B/NeMo models".into(),
527            settings: ModelOverride {
528                context_length: Some(131072),
529                temperature: Some(0.7),
530                top_k: Some(50),
531                top_p: Some(0.9),
532                max_tokens: Some(4096),
533                uniform_cache: Some(true),
534                jinja: Some(true),
535                ..Default::default()
536            },
537        },
538        Profile {
539            name: "Phi".into(),
540            description: "Optimized for Phi 3.5 Mini models".into(),
541            settings: ModelOverride {
542                context_length: Some(131072),
543                temperature: Some(0.7),
544                top_k: Some(50),
545                top_p: Some(0.9),
546                repeat_penalty: Some(1.1),
547                max_tokens: Some(4096),
548                uniform_cache: Some(true),
549                ..Default::default()
550            },
551        },
552    ]
553}
554
555#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
556#[serde(default)]
557pub struct DefaultParams {
558    // Loading
559    #[serde(default)]
560    pub context_length: u32,
561    #[serde(default)]
562    pub threads: u32,
563    #[serde(default)]
564    pub threads_batch: u32,
565    #[serde(default)]
566    pub batch_size: u32,
567    #[serde(default)]
568    pub ubatch_size: u32,
569    #[serde(default = "default_cache_type_k")]
570    pub cache_type_k: Option<CacheTypeK>,
571    #[serde(default = "default_cache_type_v")]
572    pub cache_type_v: Option<CacheTypeV>,
573    #[serde(default)]
574    pub keep: i32,
575    #[serde(default)]
576    pub swa_full: bool,
577    #[serde(default)]
578    pub mlock: bool,
579    #[serde(default)]
580    pub mmap: bool,
581    #[serde(default)]
582    pub numa: NumMode,
583    #[serde(default)]
584    pub uniform_cache: bool,
585    #[serde(default)]
586    pub kv_cache_offload: bool,
587    #[serde(default)]
588    pub parallel: u32,
589    #[serde(default)]
590    pub max_concurrent_predictions: Option<u32>,
591    #[serde(default)]
592    pub system_prompt: String,
593    #[serde(default = "default_system_prompt_preset_name")]
594    pub system_prompt_preset_name: String,
595    // GPU
596    #[serde(default)]
597    pub gpu_layers: i32,
598    #[serde(default = "default_gpu_layers_mode")]
599    pub gpu_layers_mode: crate::models::GpuLayersMode,
600    #[serde(default)]
601    pub split_mode: SplitMode,
602    #[serde(default)]
603    pub tensor_split: String,
604    #[serde(default)]
605    pub main_gpu: i32,
606    #[serde(default)]
607    pub fit: bool,
608    #[serde(default)]
609    pub lora: Option<PathBuf>,
610    #[serde(default)]
611    pub lora_scaled: Option<(PathBuf, f32)>,
612    #[serde(default)]
613    pub rpc: String,
614    #[serde(default)]
615    pub embedding: bool,
616    #[serde(default)]
617    pub flash_attn: bool,
618    #[serde(default)]
619    pub jinja: bool,
620    #[serde(default)]
621    pub chat_template: Option<String>,
622    #[serde(default)]
623    pub chat_template_kwargs: Option<String>,
624    #[serde(default)]
625    pub expert_count: i32,
626
627    // Sampling
628    #[serde(default)]
629    pub seed: i32,
630    #[serde(default)]
631    pub temperature: f32,
632    #[serde(default)]
633    pub top_k: i32,
634    #[serde(default)]
635    pub top_p: f32,
636    #[serde(default)]
637    pub min_p: f32,
638    #[serde(default)]
639    pub typical_p: f32,
640    #[serde(default)]
641    pub mirostat: Mirostat,
642    #[serde(default)]
643    pub mirostat_lr: f32,
644    #[serde(default)]
645    pub mirostat_ent: f32,
646    #[serde(default)]
647    pub ignore_eos: bool,
648    #[serde(default)]
649    pub samplers: Samplers,
650
651    // Repetition
652    #[serde(default)]
653    pub repeat_penalty: f32,
654    #[serde(default)]
655    pub repeat_last_n: i32,
656    #[serde(default = "default_presence_penalty")]
657    pub presence_penalty: Option<f32>,
658    #[serde(default = "default_frequency_penalty")]
659    pub frequency_penalty: Option<f32>,
660    #[serde(default)]
661    pub dry_multiplier: f32,
662    #[serde(default)]
663    pub dry_base: f32,
664    #[serde(default)]
665    pub dry_allowed_length: i32,
666    #[serde(default)]
667    pub dry_penalty_last_n: i32,
668
669    // RoPE
670    #[serde(default)]
671    pub rope_scaling: RopeScaling,
672    #[serde(default)]
673    pub rope_scale: f32,
674    #[serde(default)]
675    pub rope_freq_base: f32,
676    #[serde(default)]
677    pub rope_freq_scale: f32,
678    #[serde(default)]
679    pub rope_yarn_enabled: bool,
680
681    // Server
682    #[serde(default)]
683    pub host: String,
684    #[serde(default)]
685    pub port: u16,
686    #[serde(default)]
687    pub timeout: u32,
688    #[serde(default = "default_cache_prompt")]
689    pub cache_prompt: bool,
690    #[serde(default)]
691    pub cache_reuse: u32,
692    #[serde(default)]
693    pub webui: bool,
694    #[serde(default)]
695    pub ws_server_enabled: bool,
696    #[serde(default = "default_ws_server_port")]
697    pub ws_server_port: u16,
698    #[serde(default)]
699    pub ws_server_auth_key: Option<String>,
700    #[serde(default = "default_ws_server_tls_enabled")]
701    pub ws_server_tls_enabled: bool,
702    #[serde(default)]
703    pub ws_server_tls_cert: Option<String>,
704    #[serde(default)]
705    pub ws_server_tls_key: Option<String>,
706    #[serde(default)]
707    pub router_max_models: u32,
708    #[serde(default)]
709    pub server_mode: crate::models::ServerMode,
710
711    // Other
712    #[serde(default = "default_max_tokens")]
713    pub max_tokens: Option<u32>,
714    #[serde(default)]
715    pub cache_type: CacheType,
716    #[serde(default)]
717    pub backend: Backend,
718    /// Platform override: "linux", "windows", or "macos". If None, auto-detected.
719    #[serde(default)]
720    pub platform: Option<String>,
721    #[serde(default)]
722    pub llama_cpp_version_cpu: Option<String>,
723    #[serde(default)]
724    pub llama_cpp_version_vulkan: Option<String>,
725    #[serde(default)]
726    pub llama_cpp_version_rocm: Option<String>,
727    #[serde(default)]
728    pub llama_cpp_version_rocm_lemonade: Option<String>,
729    #[serde(default)]
730    pub llama_cpp_version_cuda: Option<String>,
731
732    // API
733    #[serde(default)]
734    pub api_endpoint_enabled: bool,
735    #[serde(default = "default_api_endpoint_port")]
736    pub api_endpoint_port: u16,
737    #[serde(default)]
738    pub spec_type: String,
739    #[serde(default)]
740    pub draft_tokens: u32,
741    #[serde(default)]
742    pub tags: Vec<String>,
743}
744
745fn default_api_endpoint_port() -> u16 {
746    49222
747}
748
749fn default_system_prompt_preset_name() -> String {
750    "General".to_string()
751}
752
753fn default_cache_type_k() -> Option<CacheTypeK> {
754    None
755}
756fn default_cache_type_v() -> Option<CacheTypeV> {
757    None
758}
759fn default_presence_penalty() -> Option<f32> {
760    None
761}
762fn default_frequency_penalty() -> Option<f32> {
763    None
764}
765fn default_max_tokens() -> Option<u32> {
766    None
767}
768fn default_cache_prompt() -> bool {
769    true
770}
771fn default_ws_server_port() -> u16 {
772    49223
773}
774
775fn default_ws_server_tls_enabled() -> bool {
776    true
777}
778
779fn default_gpu_layers_mode() -> crate::models::GpuLayersMode {
780    crate::models::GpuLayersMode::Auto
781}
782
783impl Default for DefaultParams {
784    fn default() -> Self {
785        Self {
786            // Loading
787            context_length: 131072,
788            threads: physical_cores(),
789            threads_batch: 8,
790            batch_size: 512,
791            ubatch_size: 512,
792            cache_type_k: None,
793            cache_type_v: None,
794            keep: 0,
795            swa_full: false,
796            mlock: false,
797            mmap: true,
798            numa: NumMode::None,
799            uniform_cache: true,
800            kv_cache_offload: true,
801            parallel: 1,
802            max_concurrent_predictions: None,
803            system_prompt: DEFAULT_SYSTEM_PROMPT.to_string(),
804            system_prompt_preset_name: "Coder".to_string(),
805
806            // GPU
807            gpu_layers: -1,
808            gpu_layers_mode: crate::models::GpuLayersMode::Auto,
809            split_mode: SplitMode::Layer,
810            tensor_split: String::new(),
811            main_gpu: 0,
812            fit: true,
813            lora: None,
814            lora_scaled: None,
815            rpc: String::new(),
816            embedding: false,
817            flash_attn: true,
818            jinja: true,
819            chat_template: None,
820            chat_template_kwargs: None,
821            expert_count: -1,
822
823            // Sampling
824            seed: -1,
825            temperature: 0.8,
826            top_k: 40,
827            top_p: 0.95,
828            min_p: 0.0,
829            typical_p: 1.0,
830            mirostat: Mirostat::Off,
831            mirostat_lr: 0.1,
832            mirostat_ent: 5.0,
833            ignore_eos: false,
834            samplers: Samplers::default(),
835
836            // Repetition
837            repeat_penalty: 1.1,
838            repeat_last_n: 64,
839            presence_penalty: None,
840            frequency_penalty: None,
841            dry_multiplier: 0.0,
842            dry_base: 1.75,
843            dry_allowed_length: 2,
844            dry_penalty_last_n: -1,
845
846            // RoPE
847            rope_scaling: RopeScaling::None,
848            rope_scale: 1.0,
849            rope_freq_base: 0.0,
850            rope_freq_scale: 1.0,
851            rope_yarn_enabled: false,
852
853            // Server
854            host: "127.0.0.1".to_string(),
855            port: 8080,
856            timeout: 600,
857            cache_prompt: true,
858            cache_reuse: 0,
859            webui: false,
860            ws_server_enabled: false,
861            ws_server_port: 49223,
862            ws_server_auth_key: None,
863            ws_server_tls_enabled: true,
864            ws_server_tls_cert: None,
865            ws_server_tls_key: None,
866            router_max_models: 4,
867            server_mode: crate::models::ServerMode::Normal,
868
869            // Other
870            max_tokens: None,
871            cache_type: CacheType::F16,
872            backend: {
873                use crate::backend::hardware::{GpuVendor, detect_gpu_vendors};
874                let vendors = detect_gpu_vendors();
875                let mut result = Backend::Cpu;
876                for v in &vendors {
877                    if matches!(v, GpuVendor::Nvidia) {
878                        result = Backend::Cuda;
879                        break;
880                    }
881                    if matches!(v, GpuVendor::Amd) {
882                        result = Backend::Rocm;
883                        break;
884                    }
885                    if matches!(v, GpuVendor::Intel) {
886                        result = Backend::Vulkan;
887                        break;
888                    }
889                }
890                result
891            },
892            platform: None,
893            llama_cpp_version_cpu: None,
894            llama_cpp_version_vulkan: None,
895            llama_cpp_version_rocm: None,
896            llama_cpp_version_rocm_lemonade: None,
897            llama_cpp_version_cuda: None,
898            api_endpoint_enabled: false,
899            api_endpoint_port: 49222,
900            spec_type: String::new(),
901            draft_tokens: 0,
902            tags: Vec::new(),
903        }
904    }
905}
906
907impl Default for Config {
908    fn default() -> Self {
909        Self {
910            models_dirs: vec![
911                dirs::data_dir()
912                    .unwrap_or_default()
913                    .join("llm-manager")
914                    .join("models"),
915            ],
916            llama_server: "llama-server".into(),
917            default: DefaultParams::default(),
918            model_overrides: Default::default(),
919            profiles: Default::default(),
920            system_prompt_presets: Default::default(),
921            rpc_workers: Vec::new(),
922            search_limit: default_search_limit(),
923            active_panel: ActivePanel::Models,
924            left_pct: 55,
925        }
926    }
927}
928
929impl Config {
930    pub fn config_path() -> PathBuf {
931        config_base_dir()
932            .join("llm-manager")
933            .join("config.yaml")
934    }
935
936    /// Validate config values and return a list of warnings for invalid entries.
937    pub fn validate(&self) -> Vec<String> {
938        let mut warnings = Vec::new();
939        let default = &self.default;
940
941        // Numeric range checks
942        if default.context_length < 512 || default.context_length > 131072 {
943            warnings.push(format!(
944                "context_length {} is outside recommended range 512-131072",
945                default.context_length
946            ));
947        }
948        if default.temperature < 0.0 || default.temperature > 2.0 {
949            warnings.push(format!(
950                "temperature {} is outside recommended range 0.0-2.0",
951                default.temperature
952            ));
953        }
954        if (default.top_p < 0.0 || default.top_p > 1.0) && default.top_p != 0.0 {
955            warnings.push(format!(
956                "top_p {} is outside recommended range 0.0-1.0",
957                default.top_p
958            ));
959        }
960        if (default.repeat_penalty < 0.0 || default.repeat_penalty > 3.0)
961            && default.repeat_penalty != 1.0
962        {
963            warnings.push(format!(
964                "repeat_penalty {} is outside recommended range 0.0-3.0",
965                default.repeat_penalty
966            ));
967        }
968        if default.mirostat_lr < 0.0 || default.mirostat_lr > 1.0 {
969            warnings.push(format!(
970                "mirostat_lr {} is outside recommended range 0.0-1.0",
971                default.mirostat_lr
972            ));
973        }
974        if default.mirostat_ent < 0.0 || default.mirostat_ent > 10.0 {
975            warnings.push(format!(
976                "mirostat_ent {} is outside recommended range 0.0-10.0",
977                default.mirostat_ent
978            ));
979        }
980
981        if default.timeout < 1 {
982            warnings.push(format!(
983                "timeout {} must be at least 1 second",
984                default.timeout
985            ));
986        }
987
988        // Path validation
989        if let Some(lora) = &default.lora
990            && !lora.exists() {
991                warnings.push(format!("lora path {} does not exist", lora.display()));
992            }
993        if let Some((lora, _)) = &default.lora_scaled
994            && !lora.exists() {
995                warnings.push(format!("lora path {} does not exist", lora.display()));
996            }
997
998        // Model override validation
999        for model_name in self.model_overrides.keys() {
1000            if let Some(override_settings) = self.model_overrides.get(model_name.as_str()) {
1001                if let Some(lora) = &override_settings.lora
1002                    && !lora.exists() {
1003                        warnings.push(format!(
1004                            "model '{}' lora path {} does not exist",
1005                            model_name,
1006                            lora.display()
1007                        ));
1008                    }
1009                if let Some((lora, _)) = &override_settings.lora_scaled
1010                    && !lora.exists() {
1011                        warnings.push(format!(
1012                            "model '{}' lora path {} does not exist",
1013                            model_name,
1014                            lora.display()
1015                        ));
1016                    }
1017            }
1018        }
1019
1020        warnings
1021    }
1022
1023    /// Resolve settings for a specific model and profile.
1024    pub fn resolve_settings(
1025        &self,
1026        model_name: Option<&str>,
1027        profile_name: Option<&str>,
1028    ) -> crate::models::ModelSettings {
1029        let mut settings = crate::models::ModelSettings::from_config(self);
1030
1031        // Apply model-specific override
1032        if let Some(name) = model_name
1033            && let Some(override_settings) = self.model_overrides.get(name)
1034        {
1035            override_settings.apply(&mut settings);
1036        }
1037
1038        // Apply profile override if specified
1039        if let Some(p_name) = profile_name {
1040            if let Some(profile) = self.profiles.get(p_name) {
1041                profile.settings.apply(&mut settings);
1042            } else if let Some(profile) = builtin_profiles().iter().find(|p| p.name == p_name) {
1043                profile.settings.apply(&mut settings);
1044            }
1045        }
1046
1047        settings
1048    }
1049
1050    /// Get a system prompt preset content by name.
1051    pub fn get_preset_content(&self, name: &str) -> Option<String> {
1052        self.system_prompt_presets
1053            .get(name)
1054            .map(|p| p.content.clone())
1055    }
1056
1057    fn normalize_config(mut config: Config) -> Config {
1058        // normalize models_dirs
1059        for path in &mut config.models_dirs {
1060            let path_str = path.to_string_lossy();
1061            if let Some(stripped) = path_str.strip_prefix("~/") {
1062                let home = dirs::home_dir().unwrap_or_default();
1063                *path = home.join(stripped);
1064            } else if !path.is_absolute() {
1065                let home = dirs::home_dir().unwrap_or_default();
1066                *path = home.join(path_str.as_ref());
1067            }
1068        }
1069
1070        // Merge built-in profiles into in-memory cache (do not persist to disk)
1071        for p in builtin_profiles() {
1072            if config.profiles.get(&p.name).is_none() {
1073                config.profiles.insert_builtin(p);
1074            }
1075        }
1076
1077        // Merge built-in system prompt presets into in-memory cache (do not persist to disk)
1078        for p in builtin_system_prompt_presets() {
1079            if config.system_prompt_presets.get(&p.name).is_none() {
1080                config.system_prompt_presets.insert_builtin(p);
1081            }
1082        }
1083        config
1084    }
1085
1086    fn load_impl(path: &PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
1087        let content = std::fs::read_to_string(path)?;
1088        let config: Config = serde_yaml::from_str(&content)
1089            .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
1090        let config = Self::normalize_config(config);
1091        let config = config.auto_detect_platform();
1092        let warnings = config.validate();
1093        if !warnings.is_empty() {
1094            eprintln!("Config validation warnings:");
1095            for warning in &warnings {
1096                eprintln!("  - {}", warning);
1097            }
1098        }
1099        Ok(config)
1100    }
1101
1102    pub fn load() -> Result<Self, Box<dyn std::error::Error>> {
1103        let path = Self::config_path();
1104        if path.exists() {
1105            Self::load_impl(&path)
1106        } else {
1107            let mut config = Config::default();
1108            config.save()?;
1109            Ok(config)
1110        }
1111    }
1112
1113    pub fn load_from(path: PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
1114        if path.exists() {
1115            Self::load_impl(&path)
1116        } else {
1117            Err(format!("Config file not found: {}", path.display()).into())
1118        }
1119    }
1120
1121    /// Auto-detect the platform if not explicitly set in config.
1122    fn auto_detect_platform(mut self) -> Self {
1123        if self.default.platform.is_none() {
1124            self.default.platform =
1125                Some(
1126                    crate::backend::hardware::platform_name(
1127                        crate::backend::hardware::detect_platform(),
1128                    )
1129                    .to_string(),
1130                );
1131        }
1132        self
1133    }
1134
1135    pub fn save(&mut self) -> Result<(), Box<dyn std::error::Error>> {
1136        let path = Self::config_path();
1137        if let Some(parent) = path.parent() {
1138            std::fs::create_dir_all(parent)?;
1139        }
1140        let content = serde_yaml::to_string(self)?;
1141        std::fs::write(&path, content)?;
1142        // Persist model configs to individual YAML files
1143        let entries: Vec<(String, ModelOverride)> = self
1144            .model_overrides
1145            .keys()
1146            .iter()
1147            .filter_map(|k| self.model_overrides.get(k).map(|v| (k.clone(), v.clone())))
1148            .collect();
1149        for (name, cfg) in entries {
1150            self.model_overrides.save(&name, &cfg);
1151        }
1152        // Persist user profiles to individual YAML files (skip built-ins)
1153        for profile in self.profiles.user_profiles() {
1154            self.profiles.save(&profile);
1155        }
1156        // Persist user presets to individual YAML files (skip built-ins)
1157        for preset in self.system_prompt_presets.user_presets() {
1158            self.system_prompt_presets.save(&preset);
1159        }
1160        Ok(())
1161    }
1162
1163    pub fn merged_profiles(&self) -> Vec<Profile> {
1164        self.profiles.all()
1165    }
1166
1167    pub fn merged_presets(&self) -> Vec<SystemPromptPreset> {
1168        self.system_prompt_presets.all()
1169    }
1170}
1171
1172#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1173pub enum LogLevel {
1174    Info,
1175    Warning,
1176    Error,
1177}
1178
1179impl LogLevel {
1180    pub fn label(&self) -> &'static str {
1181        match self {
1182            LogLevel::Info => "INFO",
1183            LogLevel::Warning => "WARNING",
1184            LogLevel::Error => "ERROR",
1185        }
1186    }
1187}
1188
1189#[derive(Debug, Clone)]
1190pub struct LogEntry {
1191    pub timestamp: String,
1192    pub level: LogLevel,
1193    pub message: String,
1194}
1195
1196impl LogEntry {
1197    pub fn new(message: impl Into<String>, level: LogLevel) -> Self {
1198        let timestamp = Local::now().format("%H:%M:%S").to_string();
1199        let message = sanitize_log(&message.into());
1200        Self {
1201            timestamp,
1202            level,
1203            message,
1204        }
1205    }
1206}
1207
1208/// Sanitize log messages to prevent TUI layout breakages.
1209/// Strips non-printable characters and control sequences, and limits length.
1210fn sanitize_log(input: &str) -> String {
1211    // Limit length to avoid layout/perf issues with massive lines
1212    let max_len = 2000;
1213    let chars: Vec<char> = input.chars().collect();
1214    let truncated = chars.len() > max_len;
1215    let chars = if truncated {
1216        chars[..max_len].to_vec()
1217    } else {
1218        chars
1219    };
1220
1221    let mut output = String::with_capacity(chars.len());
1222    for c in chars {
1223        // Strip ALL control characters except newline and tab.
1224        // Critically: strip \r (carriage return) as it breaks TUI rendering.
1225        if c.is_control() && c != '\n' && c != '\t' {
1226            continue;
1227        }
1228        output.push(c);
1229    }
1230
1231    // Replace tabs with spaces for consistent rendering
1232    let output = output.replace('\t', "    ");
1233
1234    // Final trim to remove trailing junk
1235    let mut result = output.trim_end().to_string();
1236    if truncated {
1237        result.push_str("... (truncated)");
1238    }
1239    result
1240}
llm_manager/config.rs

llm_manager/
config.rs
