Skip to content

Support multiple chat templates per model #134

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 9 commits into from
Oct 3, 2024
Merged

Support multiple chat templates per model #134

Changes from 1 commit
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
48f9167
Improve chat template parsing
DePasqualeOrg Oct 1, 2024
ac91113
Clean up
DePasqualeOrg Oct 2, 2024
852ea26
Improve chat template selection
DePasqualeOrg Oct 2, 2024
b916247
Add tests for chat templates
DePasqualeOrg Oct 2, 2024
cb47ba4
Update Sources/Tokenizers/Tokenizer.swift
DePasqualeOrg Oct 2, 2024
19a5da7
Improve template selection
DePasqualeOrg Oct 2, 2024
de740f8
More elegant solution for chatTemplate argument
DePasqualeOrg Oct 3, 2024
62ccb41
Update Sources/Tokenizers/Tokenizer.swift
DePasqualeOrg Oct 3, 2024
c0355dd
Add overload with `chatTemplate` argument of type `String`
DePasqualeOrg Oct 3, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
32 changes: 28 additions & 4 deletions Sources/Tokenizers/Tokenizer.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,8 @@ public protocol Tokenizer {
chatTemplate: String?,
addGenerationPrompt: Bool,
truncation: Bool,
maxLength: Int?
maxLength: Int?,
tools: [[String: Any]]?
) throws -> [Int]
}

Expand Down Expand Up @@ -323,12 +324,35 @@ public class PreTrainedTokenizer: Tokenizer {

public func applyChatTemplate(
messages: [[String: String]],
chatTemplate: String?,
chatTemplate: String? = nil,
addGenerationPrompt: Bool = false,
Copy link
Contributor Author

@DePasqualeOrg DePasqualeOrg Oct 2, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

addGenerationPrompt is false by default in this method (which mirrors the Python implementation), but it is currently set to true in the overload method. Should we make the behavior consistent in all applyChatTemplate methods? Should it be true or false by default?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it should be true for the overloads as I expect that to be the most common use of the API, but I'm not fully sure.

truncation: Bool = false,
maxLength: Int?
maxLength: Int? = nil,
tools: [[String: Any]]? = nil
) throws -> [Int] {
let template = try Template(chatTemplate ?? tokenizerConfig.chatTemplate?.stringValue ?? defaultChatTemplate)
var chatTemplateFromConfig: String?
if let chatTemplateValue = tokenizerConfig.chatTemplate {
if let chatTemplateStringValue = chatTemplateValue.stringValue {
chatTemplateFromConfig = chatTemplateStringValue
} else if let chatTemplateArrayValue = chatTemplateValue.arrayValue {
// If a list of chat templates is specified, convert them to a dict
let templateDict = Dictionary<String, String>(uniqueKeysWithValues: chatTemplateArrayValue.compactMap { template in
guard let name = template[dynamicMember: "name"]?.stringValue,
let templateString = template[dynamicMember: "template"]?.stringValue else {
return nil
}
return (name, templateString)
})
// Choose the appropriate template
if let tools = tools, !tools.isEmpty, let toolUseTemplate = templateDict["tool_use"] {
chatTemplateFromConfig = toolUseTemplate
} else {
chatTemplateFromConfig = templateDict["default"]
}
}
}

let template = try Template(chatTemplate ?? chatTemplateFromConfig ?? defaultChatTemplate)
var context: [String: Any] = [
"messages": messages,
"add_generation_prompt": addGenerationPrompt
Expand Down