Merge branch 'main' into patch-1

Author: mishig25

packages/gguf/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@huggingface/gguf",
 	"packageManager": "[email protected]",
-	"version": "0.1.10",
+	"version": "0.1.11",
 	"description": "a GGUF parser that works on remotely hosted files",
 	"repository": "https://github.com/huggingface/huggingface.js.git",
 	"publishConfig": {

packages/gguf/src/gguf.spec.ts

@@ -1,6 +1,13 @@
 import { beforeAll, describe, expect, it } from "vitest";
 import type { GGUFParseOutput } from "./gguf";
-import { GGMLFileQuantizationType, GGMLQuantizationType, gguf, ggufAllShards, parseGgufShardFilename } from "./gguf";
+import {
+	GGMLFileQuantizationType,
+	GGMLQuantizationType,
+	gguf,
+	ggufAllShards,
+	parseGgufShardFilename,
+	parseGGUFQuantLabel,
+} from "./gguf";
 import fs from "node:fs";
 
 const URL_LLAMA = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/resolve/191239b/llama-2-7b-chat.Q2_K.gguf";
@@ -266,4 +273,14 @@ describe("gguf", () => {
 		const { parameterCount } = await ggufAllShards(URL_SHARDED_GROK);
 		expect(parameterCount).toEqual(316_490_127_360); // 316B
 	});
+
+	it("parse quant label", async () => {
+		expect(parseGGUFQuantLabel("Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf")).toEqual("Q4_K_M");
+		expect(parseGGUFQuantLabel("subdir/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf")).toEqual("Q4_K_M");
+		expect(parseGGUFQuantLabel("Codestral-22B-v0.1-Q2_K.gguf")).toEqual("Q2_K");
+		expect(parseGGUFQuantLabel("Codestral-22B-v0.1.gguf")).toEqual(undefined);
+		expect(parseGGUFQuantLabel("Codestral-22B-v0.1-F32-Q2_K.gguf")).toEqual("Q2_K"); // gguf name with two quant labels [F32, Q2_K]
+		expect(parseGGUFQuantLabel("Codestral-22B-v0.1-IQ3_XS.gguf")).toEqual(undefined); // TODO: investigate IQ3_XS
+		expect(parseGGUFQuantLabel("Codestral-22B-v0.1-Q4_0_4_4.gguf")).toEqual("Q4_0"); // TODO: investigate Q4_0_4_4
+	});
 });

packages/gguf/src/gguf.ts

@@ -1,5 +1,5 @@
 import type { MetadataValue, Version, GGUFMetadata, GGUFTensorInfo, GGUFParseOutput } from "./types";
-import { GGUFValueType } from "./types";
+import { GGMLQuantizationType, GGUFValueType } from "./types";
 import { isBackend } from "./utils/isBackend";
 import { promisesQueue } from "./utils/promisesQueue";
 
@@ -29,6 +29,15 @@ export function parseGgufShardFilename(filename: string): GgufShardFileInfo | nu
 	return null;
 }
 
+const ggufQuants = Object.values(GGMLQuantizationType).filter((v): v is string => typeof v === "string");
+export const GGUF_QUANT_RE = new RegExp(`(?<quant>${ggufQuants.join("|")})` + "(_(?<sizeVariation>[A-Z]+))?");
+export const GGUF_QUANT_RE_GLOBAL = new RegExp(GGUF_QUANT_RE, "g");
+
+export function parseGGUFQuantLabel(fname: string): string | undefined {
+	const quantLabel = fname.toUpperCase().match(GGUF_QUANT_RE_GLOBAL)?.at(-1); // if there is multiple quant substrings in a name, we prefer the last one
+	return quantLabel;
+}
+
 const isVersion = (version: number): version is Version => version === 1 || version === 2 || version === 3;
 
 /**