[WIP] Add support for Glm

Author: xenova

src/configs.js

@@ -117,6 +117,7 @@ function getNormalizedConfig(config) {
             break;
         case 'gemma':
         case 'gemma2':
+        case 'glm':
             mapping['num_heads'] = 'num_key_value_heads';
             mapping['num_layers'] = 'num_hidden_layers';
             mapping['dim_kv'] = 'head_dim';

src/models.js

@@ -4037,6 +4037,23 @@ export class Gemma2Model extends Gemma2PreTrainedModel { }
 export class Gemma2ForCausalLM extends Gemma2PreTrainedModel { }
 //////////////////////////////////////////////////
 
+
+//////////////////////////////////////////////////
+// Glm models
+
+/**
+ * The bare Glm Model outputting raw hidden-states without any specific head on top.
+ */
+export class GlmPreTrainedModel extends PreTrainedModel { }
+/**
+ * The bare Glm Model outputting raw hidden-states without any specific head on top.
+ */
+export class GlmModel extends GlmPreTrainedModel { }
+
+export class GlmForCausalLM extends GlmPreTrainedModel { }
+//////////////////////////////////////////////////
+
+
 //////////////////////////////////////////////////
 export class OpenELMPreTrainedModel extends PreTrainedModel { }
 export class OpenELMModel extends OpenELMPreTrainedModel { }
@@ -6765,6 +6782,7 @@ const MODEL_MAPPING_NAMES_DECODER_ONLY = new Map([
     ['cohere', ['CohereModel', CohereModel]],
     ['gemma', ['GemmaModel', GemmaModel]],
     ['gemma2', ['Gemma2Model', Gemma2Model]],
+    ['glm', ['GlmModel', GlmModel]],
     ['openelm', ['OpenELMModel', OpenELMModel]],
     ['qwen2', ['Qwen2Model', Qwen2Model]],
     ['phi', ['PhiModel', PhiModel]],
@@ -6856,6 +6874,7 @@ const MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = new Map([
     ['cohere', ['CohereForCausalLM', CohereForCausalLM]],
     ['gemma', ['GemmaForCausalLM', GemmaForCausalLM]],
     ['gemma2', ['Gemma2ForCausalLM', Gemma2ForCausalLM]],
+    ['glm', ['GlmForCausalLM', GlmForCausalLM]],
     ['openelm', ['OpenELMForCausalLM', OpenELMForCausalLM]],
     ['qwen2', ['Qwen2ForCausalLM', Qwen2ForCausalLM]],
     ['phi', ['PhiForCausalLM', PhiForCausalLM]],

tests/tiny_random.test.js

@@ -10,7 +10,6 @@ import {
   BertTokenizer,
   T5Tokenizer,
   WhisperTokenizer,
-  BartTokenizer,
   MarianTokenizer,
   PreTrainedTokenizer,
   AutoTokenizer,
@@ -29,6 +28,7 @@ import {
   CohereForCausalLM,
   GemmaForCausalLM,
   Gemma2ForCausalLM,
+  GlmForCausalLM,
   OPTForCausalLM,
   GPTNeoXForCausalLM,
   GPTJForCausalLM,
@@ -1366,7 +1366,7 @@ describe("Tiny random models", () => {
     });
   });
 
-  describe("gemma", () => {
+  describe("gemma2", () => {
     describe("Gemma2ForCausalLM", () => {
       const model_id = "hf-internal-testing/tiny-random-Gemma2ForCausalLM";
       /** @type {Gemma2ForCausalLM} */
@@ -1417,6 +1417,57 @@ describe("Tiny random models", () => {
     });
   });
 
+  describe("glm", () => {
+    describe("GlmForCausalLM", () => {
+      const model_id = "hf-internal-testing/tiny-random-GlmForCausalLM";
+      /** @type {GlmForCausalLM} */
+      let model;
+      /** @type {PreTrainedTokenizer} */
+      let tokenizer;
+      beforeAll(async () => {
+        model = await GlmForCausalLM.from_pretrained(model_id, {
+          // TODO move to config
+          ...DEFAULT_MODEL_OPTIONS,
+        });
+        tokenizer = await PreTrainedTokenizer.from_pretrained(model_id);
+        // tokenizer.padding_side = "left";
+      }, MAX_MODEL_LOAD_TIME);
+
+      it(
+        "batch_size=1",
+        async () => {
+          const inputs = tokenizer("hello");
+          const outputs = await model.generate({
+            ...inputs,
+            max_length: 10,
+          });
+          expect(outputs.tolist()).toEqual([[23582n,  5797n, 38238n, 24486n, 36539n, 34489n,  6948n, 34489n,  6948n, 34489n]]);
+        },
+        MAX_TEST_EXECUTION_TIME,
+      );
+
+      it(
+        "batch_size>1",
+        async () => {
+          const inputs = tokenizer(["hello", "hello world"], { padding: true });
+          const outputs = await model.generate({
+            ...inputs,
+            max_length: 10,
+          });
+          expect(outputs.tolist()).toEqual([
+            [59246n, 23582n,  5797n, 38238n, 24486n, 36539n, 34489n,  6948n, 34489n,  6948n],
+            [23582n,  2901n, 39936n, 25036n, 55411n, 10337n,  3424n, 39183n, 30430n, 37285n]
+          ]);
+        },
+        MAX_TEST_EXECUTION_TIME,
+      );
+
+      afterAll(async () => {
+        await model?.dispose();
+      }, MAX_MODEL_DISPOSE_TIME);
+    });
+  });
+
   describe("gpt_neo", () => {
     describe("GPTNeoForCausalLM", () => {
       const model_id = "hf-internal-testing/tiny-random-GPTNeoForCausalLM";