added tests for gguf-v3

Author: biw

src/index.ts

@@ -28,7 +28,7 @@ type MetadataBaseValue = string | number | bigint | boolean
 type MetadataArray = MetadataBaseValue[]
 type MetadataValue = MetadataBaseValue | MetadataArray
 
-type Version = 1 | 2
+type Version = 1 | 2 | 3
 const isVersion = (version: number): version is Version =>
   version === 1 || version === 2 || version === 3
 
@@ -134,6 +134,7 @@ const readVersionedSize = async (
       value = BigInt(n.value)
       break
     }
+    case 3:
     case 2: {
       const n = await readUint64(fd)
       if (n.error) return n

test/index.test.ts

@@ -15,7 +15,7 @@ const fetchPartialFile = async (url: string, start: number, end: number) => {
 }
 
 describe('gguf', () => {
-  it(
+  test(
     'Llama-2-7b-Chat-GGUF ',
     async () => {
       const file = await fetchPartialFile(
@@ -68,4 +68,114 @@ describe('gguf', () => {
     },
     1000 * 30,
   )
+
+  test(
+    'Nous-Capybara-7B-v1.9-GGUF',
+    async () => {
+      const file = await fetchPartialFile(
+        'https://huggingface.co/TheBloke/Nous-Capybara-7B-v1.9-GGUF/resolve/main/nous-capybara-7b-v1.9.Q4_K_M.gguf',
+        0,
+        // 10mb
+        1024 * 1024 * 10,
+      )
+
+      const fileName = path.join(
+        __dirname,
+        'models',
+        'nous-capybara-7b-v1.9.Q4_K_M.gguf',
+      )
+
+      await writeFile(fileName, Buffer.from(file))
+
+      const { error, metadata } = await gguf(fileName)
+
+      expect(error).toBe(undefined)
+      expect(metadata).not.toBe(undefined)
+      if (!metadata) return // for types
+
+      expect(metadata.general.architecture).toBe('llama')
+      if (!isLlamaMetadata(metadata)) return // for types
+
+      expect(metadata.llama).toBeTruthy()
+
+      expect(metadata).toEqual({
+        general: {
+          architecture: 'llama',
+          file_type: 'MOSTLY_Q4_K_M',
+          name: 'nousresearch_nous-capybara-7b-v1.9',
+          quantization_version: 2,
+        },
+        llama: {
+          attention: {
+            head_count: 32,
+            head_count_kv: 8,
+            layer_norm_rms_epsilon: 0.000009999999747378752,
+          },
+          context_length: 32768,
+          embedding_length: 4096,
+          feed_forward_length: 14336,
+          rope: {
+            dimension_count: 128,
+            freq_base: 10000,
+          },
+        },
+      })
+    },
+    1000 * 30,
+  )
+
+  test(
+    'Xwin-MLewd-7B-V0.2-GGUF',
+    async () => {
+      const file = await fetchPartialFile(
+        'https://huggingface.co/TheBloke/Xwin-MLewd-7B-V0.2-GGUF/resolve/main/xwin-mlewd-7b-v0.2.Q4_K_M.gguf',
+        0,
+        // 10mb
+        1024 * 1024 * 10,
+      )
+
+      const fileName = path.join(
+        __dirname,
+        'models',
+        'xwin-mlewd-7b-v0.2.Q4_K_M.gguf',
+      )
+
+      await writeFile(fileName, Buffer.from(file))
+
+      const { error, metadata } = await gguf(fileName)
+
+      expect(error).toBe(undefined)
+      expect(metadata).not.toBe(undefined)
+      if (!metadata) return // for types
+
+      expect(metadata.general.architecture).toBe('llama')
+      if (!isLlamaMetadata(metadata)) return // for types
+
+      expect(metadata.llama).toBeTruthy()
+
+      expect(metadata).toEqual({
+        general: {
+          architecture: 'llama',
+          file_type: 'MOSTLY_Q4_K_M',
+          name: 'LLaMA v2',
+          quantization_version: 2,
+        },
+        llama: {
+          attention: {
+            head_count: 32,
+            head_count_kv: 32,
+            layer_norm_rms_epsilon: 0.000009999999747378752,
+          },
+          context_length: 4096,
+          embedding_length: 4096,
+          feed_forward_length: 11008,
+          rope: {
+            dimension_count: 128,
+            freq_base: 10000,
+          },
+        },
+      })
+    },
+    1000 * 30,
+  )
 })