1use anyhow::{Result, anyhow};
2use collections::HashMap;
3use futures::Stream;
4use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};
5use gpui::{AnyView, App, AsyncApp, Context, Subscription, Task};
6use http_client::HttpClient;
7use language_model::{
8 AuthenticateError, LanguageModelCompletionError, LanguageModelCompletionEvent,
9 LanguageModelToolChoice, LanguageModelToolResultContent, LanguageModelToolUse, MessageContent,
10 StopReason, TokenUsage,
11};
12use language_model::{
13 LanguageModel, LanguageModelId, LanguageModelName, LanguageModelProvider,
14 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
15 LanguageModelRequest, RateLimiter, Role,
16};
17use lmstudio::{ModelType, get_models};
18use schemars::JsonSchema;
19use serde::{Deserialize, Serialize};
20use settings::{Settings, SettingsStore};
21use std::pin::Pin;
22use std::str::FromStr;
23use std::{collections::BTreeMap, sync::Arc};
24use ui::{ButtonLike, Indicator, List, prelude::*};
25use util::ResultExt;
26
27use crate::AllLanguageModelSettings;
28use crate::ui::InstructionListItem;
29
30const LMSTUDIO_DOWNLOAD_URL: &str = "https://lmstudio.ai/download";
31const LMSTUDIO_CATALOG_URL: &str = "https://lmstudio.ai/models";
32const LMSTUDIO_SITE: &str = "https://lmstudio.ai/";
33
34const PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("lmstudio");
35const PROVIDER_NAME: LanguageModelProviderName = LanguageModelProviderName::new("LM Studio");
36
37#[derive(Default, Debug, Clone, PartialEq)]
38pub struct LmStudioSettings {
39 pub api_url: String,
40 pub available_models: Vec<AvailableModel>,
41}
42
43#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
44pub struct AvailableModel {
45 pub name: String,
46 pub display_name: Option<String>,
47 pub max_tokens: u64,
48 pub supports_tool_calls: bool,
49 pub supports_images: bool,
50}
51
52pub struct LmStudioLanguageModelProvider {
53 http_client: Arc<dyn HttpClient>,
54 state: gpui::Entity<State>,
55}
56
57pub struct State {
58 http_client: Arc<dyn HttpClient>,
59 available_models: Vec<lmstudio::Model>,
60 fetch_model_task: Option<Task<Result<()>>>,
61 _subscription: Subscription,
62}
63
64impl State {
65 fn is_authenticated(&self) -> bool {
66 !self.available_models.is_empty()
67 }
68
69 fn fetch_models(&mut self, cx: &mut Context<Self>) -> Task<Result<()>> {
70 let settings = &AllLanguageModelSettings::get_global(cx).lmstudio;
71 let http_client = self.http_client.clone();
72 let api_url = settings.api_url.clone();
73
74 // As a proxy for the server being "authenticated", we'll check if its up by fetching the models
75 cx.spawn(async move |this, cx| {
76 let models = get_models(http_client.as_ref(), &api_url, None).await?;
77
78 let mut models: Vec<lmstudio::Model> = models
79 .into_iter()
80 .filter(|model| model.r#type != ModelType::Embeddings)
81 .map(|model| {
82 lmstudio::Model::new(
83 &model.id,
84 None,
85 model
86 .loaded_context_length
87 .or_else(|| model.max_context_length),
88 model.capabilities.supports_tool_calls(),
89 model.capabilities.supports_images() || model.r#type == ModelType::Vlm,
90 )
91 })
92 .collect();
93
94 models.sort_by(|a, b| a.name.cmp(&b.name));
95
96 this.update(cx, |this, cx| {
97 this.available_models = models;
98 cx.notify();
99 })
100 })
101 }
102
103 fn restart_fetch_models_task(&mut self, cx: &mut Context<Self>) {
104 let task = self.fetch_models(cx);
105 self.fetch_model_task.replace(task);
106 }
107
108 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
109 if self.is_authenticated() {
110 return Task::ready(Ok(()));
111 }
112
113 let fetch_models_task = self.fetch_models(cx);
114 cx.spawn(async move |_this, _cx| Ok(fetch_models_task.await?))
115 }
116}
117
118impl LmStudioLanguageModelProvider {
119 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
120 let this = Self {
121 http_client: http_client.clone(),
122 state: cx.new(|cx| {
123 let subscription = cx.observe_global::<SettingsStore>({
124 let mut settings = AllLanguageModelSettings::get_global(cx).lmstudio.clone();
125 move |this: &mut State, cx| {
126 let new_settings = &AllLanguageModelSettings::get_global(cx).lmstudio;
127 if &settings != new_settings {
128 settings = new_settings.clone();
129 this.restart_fetch_models_task(cx);
130 cx.notify();
131 }
132 }
133 });
134
135 State {
136 http_client,
137 available_models: Default::default(),
138 fetch_model_task: None,
139 _subscription: subscription,
140 }
141 }),
142 };
143 this.state
144 .update(cx, |state, cx| state.restart_fetch_models_task(cx));
145 this
146 }
147}
148
149impl LanguageModelProviderState for LmStudioLanguageModelProvider {
150 type ObservableEntity = State;
151
152 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
153 Some(self.state.clone())
154 }
155}
156
157impl LanguageModelProvider for LmStudioLanguageModelProvider {
158 fn id(&self) -> LanguageModelProviderId {
159 PROVIDER_ID
160 }
161
162 fn name(&self) -> LanguageModelProviderName {
163 PROVIDER_NAME
164 }
165
166 fn icon(&self) -> IconName {
167 IconName::AiLmStudio
168 }
169
170 fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
171 // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.
172 // In a constrained environment where user might not have enough resources it'll be a bad UX to select something
173 // to load by default.
174 None
175 }
176
177 fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
178 // See explanation for default_model.
179 None
180 }
181
182 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
183 let mut models: BTreeMap<String, lmstudio::Model> = BTreeMap::default();
184
185 // Add models from the LM Studio API
186 for model in self.state.read(cx).available_models.iter() {
187 models.insert(model.name.clone(), model.clone());
188 }
189
190 // Override with available models from settings
191 for model in AllLanguageModelSettings::get_global(cx)
192 .lmstudio
193 .available_models
194 .iter()
195 {
196 models.insert(
197 model.name.clone(),
198 lmstudio::Model {
199 name: model.name.clone(),
200 display_name: model.display_name.clone(),
201 max_tokens: model.max_tokens,
202 supports_tool_calls: model.supports_tool_calls,
203 supports_images: model.supports_images,
204 },
205 );
206 }
207
208 models
209 .into_values()
210 .map(|model| {
211 Arc::new(LmStudioLanguageModel {
212 id: LanguageModelId::from(model.name.clone()),
213 model: model.clone(),
214 http_client: self.http_client.clone(),
215 request_limiter: RateLimiter::new(4),
216 }) as Arc<dyn LanguageModel>
217 })
218 .collect()
219 }
220
221 fn is_authenticated(&self, cx: &App) -> bool {
222 self.state.read(cx).is_authenticated()
223 }
224
225 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
226 self.state.update(cx, |state, cx| state.authenticate(cx))
227 }
228
229 fn configuration_view(&self, _window: &mut Window, cx: &mut App) -> AnyView {
230 let state = self.state.clone();
231 cx.new(|cx| ConfigurationView::new(state, cx)).into()
232 }
233
234 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
235 self.state.update(cx, |state, cx| state.fetch_models(cx))
236 }
237}
238
239pub struct LmStudioLanguageModel {
240 id: LanguageModelId,
241 model: lmstudio::Model,
242 http_client: Arc<dyn HttpClient>,
243 request_limiter: RateLimiter,
244}
245
246impl LmStudioLanguageModel {
247 fn to_lmstudio_request(
248 &self,
249 request: LanguageModelRequest,
250 ) -> lmstudio::ChatCompletionRequest {
251 let mut messages = Vec::new();
252
253 for message in request.messages {
254 for content in message.content {
255 match content {
256 MessageContent::Text(text) => add_message_content_part(
257 lmstudio::MessagePart::Text { text },
258 message.role,
259 &mut messages,
260 ),
261 MessageContent::Thinking { .. } => {}
262 MessageContent::RedactedThinking(_) => {}
263 MessageContent::Image(image) => {
264 add_message_content_part(
265 lmstudio::MessagePart::Image {
266 image_url: lmstudio::ImageUrl {
267 url: image.to_base64_url(),
268 detail: None,
269 },
270 },
271 message.role,
272 &mut messages,
273 );
274 }
275 MessageContent::ToolUse(tool_use) => {
276 let tool_call = lmstudio::ToolCall {
277 id: tool_use.id.to_string(),
278 content: lmstudio::ToolCallContent::Function {
279 function: lmstudio::FunctionContent {
280 name: tool_use.name.to_string(),
281 arguments: serde_json::to_string(&tool_use.input)
282 .unwrap_or_default(),
283 },
284 },
285 };
286
287 if let Some(lmstudio::ChatMessage::Assistant { tool_calls, .. }) =
288 messages.last_mut()
289 {
290 tool_calls.push(tool_call);
291 } else {
292 messages.push(lmstudio::ChatMessage::Assistant {
293 content: None,
294 tool_calls: vec![tool_call],
295 });
296 }
297 }
298 MessageContent::ToolResult(tool_result) => {
299 let content = match &tool_result.content {
300 LanguageModelToolResultContent::Text(text) => {
301 vec![lmstudio::MessagePart::Text {
302 text: text.to_string(),
303 }]
304 }
305 LanguageModelToolResultContent::Image(image) => {
306 vec![lmstudio::MessagePart::Image {
307 image_url: lmstudio::ImageUrl {
308 url: image.to_base64_url(),
309 detail: None,
310 },
311 }]
312 }
313 };
314
315 messages.push(lmstudio::ChatMessage::Tool {
316 content: content.into(),
317 tool_call_id: tool_result.tool_use_id.to_string(),
318 });
319 }
320 }
321 }
322 }
323
324 lmstudio::ChatCompletionRequest {
325 model: self.model.name.clone(),
326 messages,
327 stream: true,
328 max_tokens: Some(-1),
329 stop: Some(request.stop),
330 // In LM Studio you can configure specific settings you'd like to use for your model.
331 // For example Qwen3 is recommended to be used with 0.7 temperature.
332 // It would be a bad UX to silently override these settings from Zed, so we pass no temperature as a default.
333 temperature: request.temperature.or(None),
334 tools: request
335 .tools
336 .into_iter()
337 .map(|tool| lmstudio::ToolDefinition::Function {
338 function: lmstudio::FunctionDefinition {
339 name: tool.name,
340 description: Some(tool.description),
341 parameters: Some(tool.input_schema),
342 },
343 })
344 .collect(),
345 tool_choice: request.tool_choice.map(|choice| match choice {
346 LanguageModelToolChoice::Auto => lmstudio::ToolChoice::Auto,
347 LanguageModelToolChoice::Any => lmstudio::ToolChoice::Required,
348 LanguageModelToolChoice::None => lmstudio::ToolChoice::None,
349 }),
350 }
351 }
352
353 fn stream_completion(
354 &self,
355 request: lmstudio::ChatCompletionRequest,
356 cx: &AsyncApp,
357 ) -> BoxFuture<
358 'static,
359 Result<futures::stream::BoxStream<'static, Result<lmstudio::ResponseStreamEvent>>>,
360 > {
361 let http_client = self.http_client.clone();
362 let Ok(api_url) = cx.update(|cx| {
363 let settings = &AllLanguageModelSettings::get_global(cx).lmstudio;
364 settings.api_url.clone()
365 }) else {
366 return futures::future::ready(Err(anyhow!("App state dropped"))).boxed();
367 };
368
369 let future = self.request_limiter.stream(async move {
370 let request = lmstudio::stream_chat_completion(http_client.as_ref(), &api_url, request);
371 let response = request.await?;
372 Ok(response)
373 });
374
375 async move { Ok(future.await?.boxed()) }.boxed()
376 }
377}
378
379impl LanguageModel for LmStudioLanguageModel {
380 fn id(&self) -> LanguageModelId {
381 self.id.clone()
382 }
383
384 fn name(&self) -> LanguageModelName {
385 LanguageModelName::from(self.model.display_name().to_string())
386 }
387
388 fn provider_id(&self) -> LanguageModelProviderId {
389 PROVIDER_ID
390 }
391
392 fn provider_name(&self) -> LanguageModelProviderName {
393 PROVIDER_NAME
394 }
395
396 fn supports_tools(&self) -> bool {
397 self.model.supports_tool_calls()
398 }
399
400 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
401 self.supports_tools()
402 && match choice {
403 LanguageModelToolChoice::Auto => true,
404 LanguageModelToolChoice::Any => true,
405 LanguageModelToolChoice::None => true,
406 }
407 }
408
409 fn supports_images(&self) -> bool {
410 self.model.supports_images
411 }
412
413 fn telemetry_id(&self) -> String {
414 format!("lmstudio/{}", self.model.id())
415 }
416
417 fn max_token_count(&self) -> u64 {
418 self.model.max_token_count()
419 }
420
421 fn count_tokens(
422 &self,
423 request: LanguageModelRequest,
424 _cx: &App,
425 ) -> BoxFuture<'static, Result<u64>> {
426 // Endpoint for this is coming soon. In the meantime, hacky estimation
427 let token_count = request
428 .messages
429 .iter()
430 .map(|msg| msg.string_contents().split_whitespace().count())
431 .sum::<usize>();
432
433 let estimated_tokens = (token_count as f64 * 0.75) as u64;
434 async move { Ok(estimated_tokens) }.boxed()
435 }
436
437 fn stream_completion(
438 &self,
439 request: LanguageModelRequest,
440 cx: &AsyncApp,
441 ) -> BoxFuture<
442 'static,
443 Result<
444 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
445 LanguageModelCompletionError,
446 >,
447 > {
448 let request = self.to_lmstudio_request(request);
449 let completions = self.stream_completion(request, cx);
450 async move {
451 let mapper = LmStudioEventMapper::new();
452 Ok(mapper.map_stream(completions.await?).boxed())
453 }
454 .boxed()
455 }
456}
457
458struct LmStudioEventMapper {
459 tool_calls_by_index: HashMap<usize, RawToolCall>,
460}
461
462impl LmStudioEventMapper {
463 fn new() -> Self {
464 Self {
465 tool_calls_by_index: HashMap::default(),
466 }
467 }
468
469 pub fn map_stream(
470 mut self,
471 events: Pin<Box<dyn Send + Stream<Item = Result<lmstudio::ResponseStreamEvent>>>>,
472 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
473 {
474 events.flat_map(move |event| {
475 futures::stream::iter(match event {
476 Ok(event) => self.map_event(event),
477 Err(error) => vec![Err(LanguageModelCompletionError::from(error))],
478 })
479 })
480 }
481
482 pub fn map_event(
483 &mut self,
484 event: lmstudio::ResponseStreamEvent,
485 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
486 let Some(choice) = event.choices.into_iter().next() else {
487 return vec![Err(LanguageModelCompletionError::from(anyhow!(
488 "Response contained no choices"
489 )))];
490 };
491
492 let mut events = Vec::new();
493 if let Some(content) = choice.delta.content {
494 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
495 }
496
497 if let Some(reasoning_content) = choice.delta.reasoning_content {
498 events.push(Ok(LanguageModelCompletionEvent::Thinking {
499 text: reasoning_content,
500 signature: None,
501 }));
502 }
503
504 if let Some(tool_calls) = choice.delta.tool_calls {
505 for tool_call in tool_calls {
506 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
507
508 if let Some(tool_id) = tool_call.id {
509 entry.id = tool_id;
510 }
511
512 if let Some(function) = tool_call.function {
513 if let Some(name) = function.name {
514 // At the time of writing this code LM Studio (0.3.15) is incompatible with the OpenAI API:
515 // 1. It sends function name in the first chunk
516 // 2. It sends empty string in the function name field in all subsequent chunks for arguments
517 // According to https://platform.openai.com/docs/guides/function-calling?api-mode=responses#streaming
518 // function name field should be sent only inside the first chunk.
519 if !name.is_empty() {
520 entry.name = name;
521 }
522 }
523
524 if let Some(arguments) = function.arguments {
525 entry.arguments.push_str(&arguments);
526 }
527 }
528 }
529 }
530
531 if let Some(usage) = event.usage {
532 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
533 input_tokens: usage.prompt_tokens,
534 output_tokens: usage.completion_tokens,
535 cache_creation_input_tokens: 0,
536 cache_read_input_tokens: 0,
537 })));
538 }
539
540 match choice.finish_reason.as_deref() {
541 Some("stop") => {
542 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
543 }
544 Some("tool_calls") => {
545 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
546 match serde_json::Value::from_str(&tool_call.arguments) {
547 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
548 LanguageModelToolUse {
549 id: tool_call.id.into(),
550 name: tool_call.name.into(),
551 is_input_complete: true,
552 input,
553 raw_input: tool_call.arguments,
554 },
555 )),
556 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
557 id: tool_call.id.into(),
558 tool_name: tool_call.name.into(),
559 raw_input: tool_call.arguments.into(),
560 json_parse_error: error.to_string(),
561 }),
562 }
563 }));
564
565 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
566 }
567 Some(stop_reason) => {
568 log::error!("Unexpected LMStudio stop_reason: {stop_reason:?}",);
569 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
570 }
571 None => {}
572 }
573
574 events
575 }
576}
577
578#[derive(Default)]
579struct RawToolCall {
580 id: String,
581 name: String,
582 arguments: String,
583}
584
585fn add_message_content_part(
586 new_part: lmstudio::MessagePart,
587 role: Role,
588 messages: &mut Vec<lmstudio::ChatMessage>,
589) {
590 match (role, messages.last_mut()) {
591 (Role::User, Some(lmstudio::ChatMessage::User { content }))
592 | (
593 Role::Assistant,
594 Some(lmstudio::ChatMessage::Assistant {
595 content: Some(content),
596 ..
597 }),
598 )
599 | (Role::System, Some(lmstudio::ChatMessage::System { content })) => {
600 content.push_part(new_part);
601 }
602 _ => {
603 messages.push(match role {
604 Role::User => lmstudio::ChatMessage::User {
605 content: lmstudio::MessageContent::from(vec![new_part]),
606 },
607 Role::Assistant => lmstudio::ChatMessage::Assistant {
608 content: Some(lmstudio::MessageContent::from(vec![new_part])),
609 tool_calls: Vec::new(),
610 },
611 Role::System => lmstudio::ChatMessage::System {
612 content: lmstudio::MessageContent::from(vec![new_part]),
613 },
614 });
615 }
616 }
617}
618
619struct ConfigurationView {
620 state: gpui::Entity<State>,
621 loading_models_task: Option<Task<()>>,
622}
623
624impl ConfigurationView {
625 pub fn new(state: gpui::Entity<State>, cx: &mut Context<Self>) -> Self {
626 let loading_models_task = Some(cx.spawn({
627 let state = state.clone();
628 async move |this, cx| {
629 if let Some(task) = state
630 .update(cx, |state, cx| state.authenticate(cx))
631 .log_err()
632 {
633 task.await.log_err();
634 }
635 this.update(cx, |this, cx| {
636 this.loading_models_task = None;
637 cx.notify();
638 })
639 .log_err();
640 }
641 }));
642
643 Self {
644 state,
645 loading_models_task,
646 }
647 }
648
649 fn retry_connection(&self, cx: &mut App) {
650 self.state
651 .update(cx, |state, cx| state.fetch_models(cx))
652 .detach_and_log_err(cx);
653 }
654}
655
656impl Render for ConfigurationView {
657 fn render(&mut self, _window: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
658 let is_authenticated = self.state.read(cx).is_authenticated();
659
660 let lmstudio_intro = "Run local LLMs like Llama, Phi, and Qwen.";
661
662 if self.loading_models_task.is_some() {
663 div().child(Label::new("Loading models...")).into_any()
664 } else {
665 v_flex()
666 .gap_2()
667 .child(
668 v_flex().gap_1().child(Label::new(lmstudio_intro)).child(
669 List::new()
670 .child(InstructionListItem::text_only(
671 "LM Studio needs to be running with at least one model downloaded.",
672 ))
673 .child(InstructionListItem::text_only(
674 "To get your first model, try running `lms get qwen2.5-coder-7b`",
675 )),
676 ),
677 )
678 .child(
679 h_flex()
680 .w_full()
681 .justify_between()
682 .gap_2()
683 .child(
684 h_flex()
685 .w_full()
686 .gap_2()
687 .map(|this| {
688 if is_authenticated {
689 this.child(
690 Button::new("lmstudio-site", "LM Studio")
691 .style(ButtonStyle::Subtle)
692 .icon(IconName::ArrowUpRight)
693 .icon_size(IconSize::Small)
694 .icon_color(Color::Muted)
695 .on_click(move |_, _window, cx| {
696 cx.open_url(LMSTUDIO_SITE)
697 })
698 .into_any_element(),
699 )
700 } else {
701 this.child(
702 Button::new(
703 "download_lmstudio_button",
704 "Download LM Studio",
705 )
706 .style(ButtonStyle::Subtle)
707 .icon(IconName::ArrowUpRight)
708 .icon_size(IconSize::Small)
709 .icon_color(Color::Muted)
710 .on_click(move |_, _window, cx| {
711 cx.open_url(LMSTUDIO_DOWNLOAD_URL)
712 })
713 .into_any_element(),
714 )
715 }
716 })
717 .child(
718 Button::new("view-models", "Model Catalog")
719 .style(ButtonStyle::Subtle)
720 .icon(IconName::ArrowUpRight)
721 .icon_size(IconSize::Small)
722 .icon_color(Color::Muted)
723 .on_click(move |_, _window, cx| {
724 cx.open_url(LMSTUDIO_CATALOG_URL)
725 }),
726 ),
727 )
728 .map(|this| {
729 if is_authenticated {
730 this.child(
731 ButtonLike::new("connected")
732 .disabled(true)
733 .cursor_style(gpui::CursorStyle::Arrow)
734 .child(
735 h_flex()
736 .gap_2()
737 .child(Indicator::dot().color(Color::Success))
738 .child(Label::new("Connected"))
739 .into_any_element(),
740 ),
741 )
742 } else {
743 this.child(
744 Button::new("retry_lmstudio_models", "Connect")
745 .icon_position(IconPosition::Start)
746 .icon_size(IconSize::XSmall)
747 .icon(IconName::PlayFilled)
748 .on_click(cx.listener(move |this, _, _window, cx| {
749 this.retry_connection(cx)
750 })),
751 )
752 }
753 }),
754 )
755 .into_any()
756 }
757 }
758}