agent: Allow customizing temperature by provider/model (#30033)
Adds a new `agent.model_parameters` setting that allows the user to
specify a custom temperature for a provider AND/OR model:
```json5
"model_parameters": [
// To set parameters for all requests to OpenAI models:
{
"provider": "openai",
"temperature": 0.5
},
// To set parameters for all requests in general:
{
"temperature": 0
},
// To set parameters for a specific provider and model:
{
"provider": "zed.dev",
"model": "claude-3-7-sonnet-latest",
"temperature": 1.0
}
],
```
Release Notes:
- agent: Allow customizing temperature by provider/model
---------
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
This commit is contained in:
Agus Zubiaga
GitHub
parent
0055a20512
commit
3cdf5ce947
22 changed files with 348 additions and 106 deletions
|
|
@ -1417,7 +1417,10 @@ impl ActiveThread {
|
|||
messages: vec![request_message],
|
||||
tools: vec![],
|
||||
stop: vec![],
|
||||
temperature: None,
|
||||
temperature: AssistantSettings::temperature_for_model(
|
||||
&configured_model.model,
|
||||
cx,
|
||||
),
|
||||
};
|
||||
|
||||
Some(configured_model.model.count_tokens(request, cx))
|
||||
|
|
|
|||
|
|
@ -2,6 +2,7 @@ use crate::context::ContextLoadResult;
|
|||
use crate::inline_prompt_editor::CodegenStatus;
|
||||
use crate::{context::load_context, context_store::ContextStore};
|
||||
use anyhow::Result;
|
||||
use assistant_settings::AssistantSettings;
|
||||
use client::telemetry::Telemetry;
|
||||
use collections::HashSet;
|
||||
use editor::{Anchor, AnchorRangeExt, MultiBuffer, MultiBufferSnapshot, ToOffset as _, ToPoint};
|
||||
|
|
@ -383,7 +384,7 @@ impl CodegenAlternative {
|
|||
if user_prompt.trim().to_lowercase() == "delete" {
|
||||
async { Ok(LanguageModelTextStream::default()) }.boxed_local()
|
||||
} else {
|
||||
let request = self.build_request(user_prompt, cx)?;
|
||||
let request = self.build_request(&model, user_prompt, cx)?;
|
||||
cx.spawn(async move |_, cx| model.stream_completion_text(request.await, &cx).await)
|
||||
.boxed_local()
|
||||
};
|
||||
|
|
@ -393,6 +394,7 @@ impl CodegenAlternative {
|
|||
|
||||
fn build_request(
|
||||
&self,
|
||||
model: &Arc<dyn LanguageModel>,
|
||||
user_prompt: String,
|
||||
cx: &mut App,
|
||||
) -> Result<Task<LanguageModelRequest>> {
|
||||
|
|
@ -441,6 +443,8 @@ impl CodegenAlternative {
|
|||
}
|
||||
});
|
||||
|
||||
let temperature = AssistantSettings::temperature_for_model(&model, cx);
|
||||
|
||||
Ok(cx.spawn(async move |_cx| {
|
||||
let mut request_message = LanguageModelRequestMessage {
|
||||
role: Role::User,
|
||||
|
|
@ -463,7 +467,7 @@ impl CodegenAlternative {
|
|||
mode: None,
|
||||
tools: Vec::new(),
|
||||
stop: Vec::new(),
|
||||
temperature: None,
|
||||
temperature,
|
||||
messages: vec![request_message],
|
||||
}
|
||||
}))
|
||||
|
|
|
|||
|
|
@ -8,7 +8,7 @@ use crate::ui::{
|
|||
AnimatedLabel, MaxModeTooltip,
|
||||
preview::{AgentPreview, UsageCallout},
|
||||
};
|
||||
use assistant_settings::CompletionMode;
|
||||
use assistant_settings::{AssistantSettings, CompletionMode};
|
||||
use buffer_diff::BufferDiff;
|
||||
use client::UserStore;
|
||||
use collections::{HashMap, HashSet};
|
||||
|
|
@ -1273,7 +1273,7 @@ impl MessageEditor {
|
|||
messages: vec![request_message],
|
||||
tools: vec![],
|
||||
stop: vec![],
|
||||
temperature: None,
|
||||
temperature: AssistantSettings::temperature_for_model(&model.model, cx),
|
||||
};
|
||||
|
||||
Some(model.model.count_tokens(request, cx))
|
||||
|
|
|
|||
|
|
@ -6,6 +6,7 @@ use crate::inline_prompt_editor::{
|
|||
use crate::terminal_codegen::{CLEAR_INPUT, CodegenEvent, TerminalCodegen};
|
||||
use crate::thread_store::{TextThreadStore, ThreadStore};
|
||||
use anyhow::{Context as _, Result};
|
||||
use assistant_settings::AssistantSettings;
|
||||
use client::telemetry::Telemetry;
|
||||
use collections::{HashMap, VecDeque};
|
||||
use editor::{MultiBuffer, actions::SelectAll};
|
||||
|
|
@ -266,6 +267,12 @@ impl TerminalInlineAssistant {
|
|||
load_context(contexts, project, &assist.prompt_store, cx)
|
||||
})?;
|
||||
|
||||
let ConfiguredModel { model, .. } = LanguageModelRegistry::read_global(cx)
|
||||
.inline_assistant_model()
|
||||
.context("No inline assistant model")?;
|
||||
|
||||
let temperature = AssistantSettings::temperature_for_model(&model, cx);
|
||||
|
||||
Ok(cx.background_spawn(async move {
|
||||
let mut request_message = LanguageModelRequestMessage {
|
||||
role: Role::User,
|
||||
|
|
@ -287,7 +294,7 @@ impl TerminalInlineAssistant {
|
|||
messages: vec![request_message],
|
||||
tools: Vec::new(),
|
||||
stop: Vec::new(),
|
||||
temperature: None,
|
||||
temperature,
|
||||
}
|
||||
}))
|
||||
}
|
||||
|
|
|
|||
|
|
@ -1145,7 +1145,7 @@ impl Thread {
|
|||
messages: vec![],
|
||||
tools: Vec::new(),
|
||||
stop: Vec::new(),
|
||||
temperature: None,
|
||||
temperature: AssistantSettings::temperature_for_model(&model, cx),
|
||||
};
|
||||
|
||||
let available_tools = self.available_tools(cx, model.clone());
|
||||
|
|
@ -1251,7 +1251,12 @@ impl Thread {
|
|||
request
|
||||
}
|
||||
|
||||
fn to_summarize_request(&self, added_user_message: String) -> LanguageModelRequest {
|
||||
fn to_summarize_request(
|
||||
&self,
|
||||
model: &Arc<dyn LanguageModel>,
|
||||
added_user_message: String,
|
||||
cx: &App,
|
||||
) -> LanguageModelRequest {
|
||||
let mut request = LanguageModelRequest {
|
||||
thread_id: None,
|
||||
prompt_id: None,
|
||||
|
|
@ -1259,7 +1264,7 @@ impl Thread {
|
|||
messages: vec![],
|
||||
tools: Vec::new(),
|
||||
stop: Vec::new(),
|
||||
temperature: None,
|
||||
temperature: AssistantSettings::temperature_for_model(model, cx),
|
||||
};
|
||||
|
||||
for message in &self.messages {
|
||||
|
|
@ -1696,7 +1701,7 @@ impl Thread {
|
|||
If the conversation is about a specific subject, include it in the title. \
|
||||
Be descriptive. DO NOT speak in the first person.";
|
||||
|
||||
let request = self.to_summarize_request(added_user_message.into());
|
||||
let request = self.to_summarize_request(&model.model, added_user_message.into(), cx);
|
||||
|
||||
self.pending_summary = cx.spawn(async move |this, cx| {
|
||||
async move {
|
||||
|
|
@ -1782,7 +1787,7 @@ impl Thread {
|
|||
4. Any action items or next steps if any\n\
|
||||
Format it in Markdown with headings and bullet points.";
|
||||
|
||||
let request = self.to_summarize_request(added_user_message.into());
|
||||
let request = self.to_summarize_request(&model, added_user_message.into(), cx);
|
||||
|
||||
*self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
|
||||
message_id: last_message_id,
|
||||
|
|
@ -2655,7 +2660,7 @@ struct PendingCompletion {
|
|||
mod tests {
|
||||
use super::*;
|
||||
use crate::{ThreadStore, context::load_context, context_store::ContextStore, thread_store};
|
||||
use assistant_settings::AssistantSettings;
|
||||
use assistant_settings::{AssistantSettings, LanguageModelParameters};
|
||||
use assistant_tool::ToolRegistry;
|
||||
use editor::EditorSettings;
|
||||
use gpui::TestAppContext;
|
||||
|
|
@ -3066,6 +3071,100 @@ fn main() {{
|
|||
);
|
||||
}
|
||||
|
||||
#[gpui::test]
|
||||
async fn test_temperature_setting(cx: &mut TestAppContext) {
|
||||
init_test_settings(cx);
|
||||
|
||||
let project = create_test_project(
|
||||
cx,
|
||||
json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
|
||||
)
|
||||
.await;
|
||||
|
||||
let (_workspace, _thread_store, thread, _context_store, model) =
|
||||
setup_test_environment(cx, project.clone()).await;
|
||||
|
||||
// Both model and provider
|
||||
cx.update(|cx| {
|
||||
AssistantSettings::override_global(
|
||||
AssistantSettings {
|
||||
model_parameters: vec![LanguageModelParameters {
|
||||
provider: Some(model.provider_id().0.to_string().into()),
|
||||
model: Some(model.id().0.clone()),
|
||||
temperature: Some(0.66),
|
||||
}],
|
||||
..AssistantSettings::get_global(cx).clone()
|
||||
},
|
||||
cx,
|
||||
);
|
||||
});
|
||||
|
||||
let request = thread.update(cx, |thread, cx| {
|
||||
thread.to_completion_request(model.clone(), cx)
|
||||
});
|
||||
assert_eq!(request.temperature, Some(0.66));
|
||||
|
||||
// Only model
|
||||
cx.update(|cx| {
|
||||
AssistantSettings::override_global(
|
||||
AssistantSettings {
|
||||
model_parameters: vec![LanguageModelParameters {
|
||||
provider: None,
|
||||
model: Some(model.id().0.clone()),
|
||||
temperature: Some(0.66),
|
||||
}],
|
||||
..AssistantSettings::get_global(cx).clone()
|
||||
},
|
||||
cx,
|
||||
);
|
||||
});
|
||||
|
||||
let request = thread.update(cx, |thread, cx| {
|
||||
thread.to_completion_request(model.clone(), cx)
|
||||
});
|
||||
assert_eq!(request.temperature, Some(0.66));
|
||||
|
||||
// Only provider
|
||||
cx.update(|cx| {
|
||||
AssistantSettings::override_global(
|
||||
AssistantSettings {
|
||||
model_parameters: vec![LanguageModelParameters {
|
||||
provider: Some(model.provider_id().0.to_string().into()),
|
||||
model: None,
|
||||
temperature: Some(0.66),
|
||||
}],
|
||||
..AssistantSettings::get_global(cx).clone()
|
||||
},
|
||||
cx,
|
||||
);
|
||||
});
|
||||
|
||||
let request = thread.update(cx, |thread, cx| {
|
||||
thread.to_completion_request(model.clone(), cx)
|
||||
});
|
||||
assert_eq!(request.temperature, Some(0.66));
|
||||
|
||||
// Same model name, different provider
|
||||
cx.update(|cx| {
|
||||
AssistantSettings::override_global(
|
||||
AssistantSettings {
|
||||
model_parameters: vec![LanguageModelParameters {
|
||||
provider: Some("anthropic".into()),
|
||||
model: Some(model.id().0.clone()),
|
||||
temperature: Some(0.66),
|
||||
}],
|
||||
..AssistantSettings::get_global(cx).clone()
|
||||
},
|
||||
cx,
|
||||
);
|
||||
});
|
||||
|
||||
let request = thread.update(cx, |thread, cx| {
|
||||
thread.to_completion_request(model.clone(), cx)
|
||||
});
|
||||
assert_eq!(request.temperature, None);
|
||||
}
|
||||
|
||||
fn init_test_settings(cx: &mut TestAppContext) {
|
||||
cx.update(|cx| {
|
||||
let settings_store = SettingsStore::test(cx);
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue