Bump to composer 0.17

setup.py

@@ -47,7 +47,7 @@
 ]
 
 install_requires = [
-    'mosaicml[libcloud,wandb,mlflow,oci,gcs]>=0.16.4,<0.17',
+    'mosaicml[libcloud,wandb,mlflow,oci,gcs]>=0.17,<0.18',
     'accelerate>=0.20,<0.21',  # for HF inference `device_map`
     'transformers>=4.34.1,<4.35',
     'mosaicml-streaming>=0.7.1,<0.8',
@@ -84,11 +84,11 @@
 ]
 
 extra_deps['databricks'] = [
-    'mosaicml[databricks]',
+    'mosaicml[databricks]>=0.17,<0.18',
 ]
 
 extra_deps['tensorboard'] = [
-    'mosaicml[tensorboard]>=0.16.1,<0.17',
+    'mosaicml[tensorboard]>=0.17,<0.18',
 ]
 
 extra_deps['gpu'] = [

tests/test_mpt_gen.py

@@ -95,17 +95,15 @@ def test_mpt_generate_multi_gpu(attn_impl: str, use_alibi: bool,
 @pytest.mark.gpu
 @pytest.mark.parametrize('attn_impl', ['triton', 'torch'])
 @pytest.mark.parametrize('use_alibi', [True, False])
-@pytest.mark.parametrize('tie_word_embeddings', [True, False])
 def test_mpt_generate_callback(attn_impl: str, use_alibi: bool,
-                               tie_word_embeddings: bool,
                                build_tiny_mpt: Callable[...,
                                                         ComposerMPTCausalLM],
                                tiny_ft_dataloader: DataLoader):
     device = get_device('gpu')
 
     # build mpt model
     model = build_tiny_mpt(
-        tie_word_embeddings=tie_word_embeddings,
+        tie_word_embeddings=True,
         attn_config={
             'attn_impl': attn_impl,
             'attn_uses_sequence_id': False,
@@ -143,3 +141,54 @@ def test_mpt_generate_callback(attn_impl: str, use_alibi: bool,
 
     generate.generate.assert_called_once()
     trainer.logger.log_table.assert_called_once()
+
+
+@pytest.mark.gpu
+@pytest.mark.parametrize('attn_impl', ['triton', 'torch'])
+@pytest.mark.parametrize('use_alibi', [True, False])
+def test_mpt_generate_callback_not_tied(
+        use_alibi: bool, attn_impl: str,
+        build_tiny_mpt: Callable[..., ComposerMPTCausalLM],
+        tiny_ft_dataloader: DataLoader):
+    device = get_device('gpu')
+
+    # build mpt model
+    model = build_tiny_mpt(
+        tie_word_embeddings=False,
+        attn_config={
+            'attn_impl': attn_impl,
+            'attn_uses_sequence_id': False,
+            'alibi': use_alibi,
+        },
+    )
+    model = device.module_to_device(model)
+
+    # generate callback
+    prompts = [
+        'The best banana bread recipe is',
+        '2+2=',
+        'how much wood could a woodchuck chuck',
+    ]
+    gen_interval = 1
+    generate = ComposerGenerate(
+        prompts,
+        interval=f'{gen_interval}ba',
+        max_new_tokens=5,
+        batch_size=len(prompts),
+        use_cache=True,
+    )
+    generate.generate = Mock(wraps=generate.generate, autospec=True)
+
+    # build trainer
+    trainer = Trainer(
+        model=model,
+        train_dataloader=tiny_ft_dataloader,
+        device=device,
+        max_duration=f'{gen_interval}ba',
+        callbacks=[generate],
+    )
+    trainer.logger.log_table = Mock()
+    trainer.fit()
+
+    generate.generate.assert_called_once()
+    trainer.logger.log_table.assert_called_once()