First upload version 0.0.1
This commit is contained in:
Neyra
224
node_modules/node-llama-cpp/dist/chatWrappers/QwenChatWrapper.js
generated
vendored
Normal file
224
node_modules/node-llama-cpp/dist/chatWrappers/QwenChatWrapper.js
generated
vendored
Normal file
@@ -0,0 +1,224 @@
|
||||
import { ChatWrapper } from "../ChatWrapper.js";
|
||||
import { isChatModelResponseFunctionCall, isChatModelResponseSegment } from "../types.js";
|
||||
import { LlamaText, SpecialToken, SpecialTokensText } from "../utils/LlamaText.js";
|
||||
import { GgufArchitectureType } from "../gguf/types/GgufMetadataTypes.js";
|
||||
import { ChatModelFunctionsDocumentationGenerator } from "./utils/ChatModelFunctionsDocumentationGenerator.js";
|
||||
// source: https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-1M/blob/main/tokenizer_config.json#L197
|
||||
export class QwenChatWrapper extends ChatWrapper {
|
||||
wrapperName = "Qwen";
|
||||
keepOnlyLastThought;
|
||||
thoughts;
|
||||
/** @internal */ _flatFunctionResultString;
|
||||
settings;
|
||||
constructor(options = {}) {
|
||||
super();
|
||||
const { keepOnlyLastThought = true, thoughts = "auto", _lineBreakBeforeFunctionCallPrefix = false, _flatFunctionResultString = false } = options;
|
||||
this.keepOnlyLastThought = keepOnlyLastThought;
|
||||
this.thoughts = thoughts;
|
||||
this._flatFunctionResultString = _flatFunctionResultString;
|
||||
this.settings = {
|
||||
supportsSystemMessages: true,
|
||||
functions: {
|
||||
call: {
|
||||
optionalPrefixSpace: true,
|
||||
prefix: LlamaText([
|
||||
_lineBreakBeforeFunctionCallPrefix
|
||||
? "\n"
|
||||
: "",
|
||||
new SpecialTokensText("<tool_call>"), '\n{"name": "'
|
||||
]),
|
||||
paramsPrefix: '", "arguments": ',
|
||||
suffix: LlamaText("}\n", new SpecialTokensText("</tool_call>")),
|
||||
emptyCallParamsPlaceholder: {}
|
||||
},
|
||||
result: {
|
||||
prefix: LlamaText(new SpecialTokensText("\n<tool_response>\n")),
|
||||
suffix: LlamaText(new SpecialTokensText("\n</tool_response>"))
|
||||
},
|
||||
parallelism: {
|
||||
call: {
|
||||
sectionPrefix: "",
|
||||
betweenCalls: _lineBreakBeforeFunctionCallPrefix
|
||||
? ""
|
||||
: "\n",
|
||||
sectionSuffix: LlamaText(new SpecialTokensText("<|im_end|>\n"))
|
||||
},
|
||||
result: {
|
||||
sectionPrefix: LlamaText(new SpecialTokensText("<|im_start|>user")),
|
||||
sectionSuffix: LlamaText(new SpecialTokensText("<|im_end|>\n<|im_start|>assistant\n"))
|
||||
}
|
||||
}
|
||||
},
|
||||
segments: {
|
||||
reiterateStackAfterFunctionCalls: true,
|
||||
thought: {
|
||||
prefix: LlamaText(new SpecialTokensText("<think>\n")),
|
||||
suffix: LlamaText(new SpecialTokensText("\n</think>"))
|
||||
}
|
||||
}
|
||||
};
|
||||
}
|
||||
generateContextState({ chatHistory, availableFunctions, documentFunctionParams }) {
|
||||
const historyWithFunctions = this.addAvailableFunctionsSystemMessageToHistory(chatHistory, availableFunctions, {
|
||||
documentParams: documentFunctionParams
|
||||
});
|
||||
const resultItems = [];
|
||||
let systemTexts = [];
|
||||
let userTexts = [];
|
||||
let modelTexts = [];
|
||||
let currentAggregateFocus = null;
|
||||
function flush() {
|
||||
if (systemTexts.length > 0 || userTexts.length > 0 || modelTexts.length > 0)
|
||||
resultItems.push({
|
||||
system: LlamaText.joinValues("\n\n", systemTexts),
|
||||
user: LlamaText.joinValues("\n\n", userTexts),
|
||||
model: LlamaText.joinValues("\n\n", modelTexts)
|
||||
});
|
||||
systemTexts = [];
|
||||
userTexts = [];
|
||||
modelTexts = [];
|
||||
}
|
||||
for (let i = 0; i < historyWithFunctions.length; i++) {
|
||||
const item = historyWithFunctions[i];
|
||||
const isLastItem = i === historyWithFunctions.length - 1;
|
||||
if (item.type === "system") {
|
||||
if (currentAggregateFocus !== "system")
|
||||
flush();
|
||||
currentAggregateFocus = "system";
|
||||
systemTexts.push(LlamaText.fromJSON(item.text));
|
||||
}
|
||||
else if (item.type === "user") {
|
||||
flush();
|
||||
currentAggregateFocus = null;
|
||||
userTexts.push(LlamaText(item.text));
|
||||
}
|
||||
else if (item.type === "model") {
|
||||
flush();
|
||||
const transformedModelResponse = (this.thoughts === "discourage" && isLastItem)
|
||||
? discourageThoughtsInModelResponse(item.response)
|
||||
: item.response;
|
||||
currentAggregateFocus = null;
|
||||
modelTexts.push(this.generateModelResponseText((this.keepOnlyLastThought && !isLastItem)
|
||||
? transformedModelResponse.filter((response) => (!isChatModelResponseSegment(response) || response.segmentType !== "thought"))
|
||||
: transformedModelResponse));
|
||||
}
|
||||
else
|
||||
void item;
|
||||
}
|
||||
flush();
|
||||
const contextText = LlamaText(resultItems.map(({ system, user, model }, index) => {
|
||||
const isLastItem = index === resultItems.length - 1;
|
||||
return LlamaText([
|
||||
(system.values.length === 0)
|
||||
? LlamaText([])
|
||||
: LlamaText([
|
||||
new SpecialTokensText("<|im_start|>system\n"),
|
||||
system,
|
||||
new SpecialTokensText("<|im_end|>\n")
|
||||
]),
|
||||
(user.values.length === 0)
|
||||
? LlamaText([])
|
||||
: LlamaText([
|
||||
new SpecialTokensText("<|im_start|>user\n"),
|
||||
user,
|
||||
new SpecialTokensText("<|im_end|>\n")
|
||||
]),
|
||||
(model.values.length === 0 && !isLastItem)
|
||||
? LlamaText([])
|
||||
: LlamaText([
|
||||
new SpecialTokensText("<|im_start|>assistant\n"),
|
||||
model,
|
||||
isLastItem
|
||||
? LlamaText([])
|
||||
: new SpecialTokensText("<|im_end|>\n")
|
||||
])
|
||||
]);
|
||||
}));
|
||||
return {
|
||||
contextText,
|
||||
stopGenerationTriggers: [
|
||||
LlamaText(new SpecialToken("EOS")),
|
||||
LlamaText(new SpecialTokensText("<|im_end|>")),
|
||||
LlamaText("<|im_end|>")
|
||||
]
|
||||
};
|
||||
}
|
||||
generateFunctionCallResult(functionName, functionParams, result) {
|
||||
if (this._flatFunctionResultString && typeof result === "string")
|
||||
return super._generateFunctionCallResult(functionName, functionParams, result);
|
||||
return super.generateFunctionCallResult(functionName, functionParams, result);
|
||||
}
|
||||
generateAvailableFunctionsSystemText(availableFunctions, { documentParams = true }) {
|
||||
const functionsDocumentationGenerator = new ChatModelFunctionsDocumentationGenerator(availableFunctions);
|
||||
if (!functionsDocumentationGenerator.hasAnyFunctions)
|
||||
return LlamaText([]);
|
||||
return LlamaText.joinValues("\n", [
|
||||
"# Tools",
|
||||
"",
|
||||
"You may call one or more functions to assist with the user query.",
|
||||
"",
|
||||
LlamaText("You are provided with function signatures within ", new SpecialTokensText("<tools></tools>"), " XML tags:"),
|
||||
LlamaText(new SpecialTokensText("<tools>")),
|
||||
functionsDocumentationGenerator.getQwenFunctionSignatures({ documentParams }),
|
||||
LlamaText(new SpecialTokensText("</tools>")),
|
||||
"",
|
||||
LlamaText("For each function call, return a json object with function name and arguments within ", new SpecialTokensText("<tool_call></tool_call>"), " XML tags:"),
|
||||
LlamaText(new SpecialTokensText("<tool_call>")),
|
||||
'{"name": <function-name>, "arguments": <args-json-object>}',
|
||||
LlamaText(new SpecialTokensText("</tool_call>"))
|
||||
]);
|
||||
}
|
||||
/** @internal */
|
||||
static _checkModelCompatibility(options) {
|
||||
const architecture = options.fileInfo?.metadata.general.architecture;
|
||||
return (architecture == null ||
|
||||
architecture === GgufArchitectureType.qwen2 ||
|
||||
architecture === GgufArchitectureType.qwen2moe ||
|
||||
architecture === GgufArchitectureType.qwen2vl ||
|
||||
architecture === GgufArchitectureType.qwen3 ||
|
||||
architecture === GgufArchitectureType.qwen3moe ||
|
||||
architecture === GgufArchitectureType.qwen3vl ||
|
||||
architecture === GgufArchitectureType.qwen3vlmoe);
|
||||
}
|
||||
/** @internal */
|
||||
static _getOptionConfigurationsToTestIfCanSupersedeJinjaTemplate() {
|
||||
return [
|
||||
[{}, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[{ _lineBreakBeforeFunctionCallPrefix: true }, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[{ thoughts: "discourage" }, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[{ thoughts: "discourage", _lineBreakBeforeFunctionCallPrefix: true }, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[{ _flatFunctionResultString: true }, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[
|
||||
{ _flatFunctionResultString: true, _lineBreakBeforeFunctionCallPrefix: true },
|
||||
{},
|
||||
{ _requireFunctionCallSettingsExtraction: true }
|
||||
],
|
||||
[{ _flatFunctionResultString: true, thoughts: "discourage" }, {}, { _requireFunctionCallSettingsExtraction: true }],
|
||||
[
|
||||
{ _flatFunctionResultString: true, thoughts: "discourage", _lineBreakBeforeFunctionCallPrefix: true },
|
||||
{},
|
||||
{ _requireFunctionCallSettingsExtraction: true }
|
||||
]
|
||||
];
|
||||
}
|
||||
}
|
||||
function discourageThoughtsInModelResponse(response) {
|
||||
const emptyThought = {
|
||||
type: "segment",
|
||||
segmentType: "thought",
|
||||
ended: true,
|
||||
text: "\n\n",
|
||||
raw: LlamaText(new SpecialTokensText("<think>\n\n</think>\n\n")).toJSON()
|
||||
};
|
||||
const res = [...response];
|
||||
for (let i = res.length - 1; i >= 0; i--) {
|
||||
const item = res[i];
|
||||
if (isChatModelResponseFunctionCall(item)) {
|
||||
res.splice(i + 1, 0, emptyThought);
|
||||
return res;
|
||||
}
|
||||
}
|
||||
res.unshift(emptyThought);
|
||||
return res;
|
||||
}
|
||||
//# sourceMappingURL=QwenChatWrapper.js.map
|
||||
Reference in New Issue
Block a user