-
Notifications
You must be signed in to change notification settings - Fork 2.1k
/
hf.ts
199 lines (167 loc) Β· 5.72 KB
/
hf.ts
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
import { CallbackManagerForLLMRun } from "@langchain/core/callbacks/manager";
import { LLM, type BaseLLMParams } from "@langchain/core/language_models/llms";
import { GenerationChunk } from "@langchain/core/outputs";
import { getEnvironmentVariable } from "@langchain/core/utils/env";
/**
* Interface defining the parameters for configuring the Hugging Face
* model for text generation.
*/
export interface HFInput {
/** Model to use */
model: string;
/** Custom inference endpoint URL to use */
endpointUrl?: string;
/** Sampling temperature to use */
temperature?: number;
/**
* Maximum number of tokens to generate in the completion.
*/
maxTokens?: number;
/**
* The model will stop generating text when one of the strings in the list is generated.
*/
stopSequences?: string[];
/** Total probability mass of tokens to consider at each step */
topP?: number;
/** Integer to define the top tokens considered within the sample operation to create new text. */
topK?: number;
/** Penalizes repeated tokens according to frequency */
frequencyPenalty?: number;
/** API key to use. */
apiKey?: string;
/**
* Credentials to use for the request. If this is a string, it will be passed straight on. If it's a boolean, true will be "include" and false will not send credentials at all.
*/
includeCredentials?: string | boolean;
}
/**
* Class implementing the Large Language Model (LLM) interface using the
* Hugging Face Inference API for text generation.
* @example
* ```typescript
* const model = new HuggingFaceInference({
* model: "gpt2",
* temperature: 0.7,
* maxTokens: 50,
* });
*
* const res = await model.invoke(
* "Question: What would be a good company name for a company that makes colorful socks?\nAnswer:"
* );
* console.log({ res });
* ```
*/
export class HuggingFaceInference extends LLM implements HFInput {
lc_serializable = true;
get lc_secrets(): { [key: string]: string } | undefined {
return {
apiKey: "HUGGINGFACEHUB_API_KEY",
};
}
model = "gpt2";
temperature: number | undefined = undefined;
maxTokens: number | undefined = undefined;
stopSequences: string[] | undefined = undefined;
topP: number | undefined = undefined;
topK: number | undefined = undefined;
frequencyPenalty: number | undefined = undefined;
apiKey: string | undefined = undefined;
endpointUrl: string | undefined = undefined;
includeCredentials: string | boolean | undefined = undefined;
constructor(fields?: Partial<HFInput> & BaseLLMParams) {
super(fields ?? {});
this.model = fields?.model ?? this.model;
this.temperature = fields?.temperature ?? this.temperature;
this.maxTokens = fields?.maxTokens ?? this.maxTokens;
this.stopSequences = fields?.stopSequences ?? this.stopSequences;
this.topP = fields?.topP ?? this.topP;
this.topK = fields?.topK ?? this.topK;
this.frequencyPenalty = fields?.frequencyPenalty ?? this.frequencyPenalty;
this.apiKey =
fields?.apiKey ?? getEnvironmentVariable("HUGGINGFACEHUB_API_KEY");
this.endpointUrl = fields?.endpointUrl;
this.includeCredentials = fields?.includeCredentials;
if (!this.apiKey) {
throw new Error(
`Please set an API key for HuggingFace Hub in the environment variable "HUGGINGFACEHUB_API_KEY" or in the apiKey field of the HuggingFaceInference constructor.`
);
}
}
_llmType() {
return "hf";
}
invocationParams(options?: this["ParsedCallOptions"]) {
return {
model: this.model,
parameters: {
// make it behave similar to openai, returning only the generated text
return_full_text: false,
temperature: this.temperature,
max_new_tokens: this.maxTokens,
stop: options?.stop ?? this.stopSequences,
top_p: this.topP,
top_k: this.topK,
repetition_penalty: this.frequencyPenalty,
},
};
}
async *_streamResponseChunks(
prompt: string,
options: this["ParsedCallOptions"],
runManager?: CallbackManagerForLLMRun
): AsyncGenerator<GenerationChunk> {
const hfi = await this._prepareHFInference();
const stream = await this.caller.call(async () =>
hfi.textGenerationStream({
...this.invocationParams(options),
inputs: prompt,
})
);
for await (const chunk of stream) {
const token = chunk.token.text;
yield new GenerationChunk({ text: token, generationInfo: chunk });
await runManager?.handleLLMNewToken(token ?? "");
// stream is done
if (chunk.generated_text)
yield new GenerationChunk({
text: "",
generationInfo: { finished: true },
});
}
}
/** @ignore */
async _call(
prompt: string,
options: this["ParsedCallOptions"]
): Promise<string> {
const hfi = await this._prepareHFInference();
const args = { ...this.invocationParams(options), inputs: prompt };
const res = await this.caller.callWithOptions(
{ signal: options.signal },
hfi.textGeneration.bind(hfi),
args
);
return res.generated_text;
}
/** @ignore */
private async _prepareHFInference() {
const { HfInference } = await HuggingFaceInference.imports();
const hfi = new HfInference(this.apiKey, {
includeCredentials: this.includeCredentials,
});
return this.endpointUrl ? hfi.endpoint(this.endpointUrl) : hfi;
}
/** @ignore */
static async imports(): Promise<{
HfInference: typeof import("@huggingface/inference").HfInference;
}> {
try {
const { HfInference } = await import("@huggingface/inference");
return { HfInference };
} catch (e) {
throw new Error(
"Please install huggingface as a dependency with, e.g. `yarn add @huggingface/inference`"
);
}
}
}