Add MII v0.1 unit tests (#262)

.github/workflows/nv-torch-latest-v100.yaml

@@ -50,5 +50,5 @@ jobs:
         run: |
           unset TORCH_CUDA_ARCH_LIST # only jit compile for current arch
           if [[ -d ./torch-extensions ]]; then rm -rf ./torch-extensions; fi
-          cd tests
+          cd tests/legacy
           TRANSFORMERS_CACHE=/blob/transformers_cache/ TORCH_EXTENSIONS_DIR=./torch-extensions pytest --color=yes --durations=0 --verbose ./

mii/batching/ragged_batching.py

@@ -134,6 +134,7 @@ class RaggedRequest:
     sampler: BaseGenerationSampler
     stop_criterion: BaseGenerationStopCriterion
     stream: bool = False
+    ignore_eos: bool = False
 
     _next_token: Union[None, torch.Tensor] = None
     _is_done: bool = False
@@ -150,6 +151,8 @@ def next_token(self, next_token: Union[None, torch.Tensor]) -> None:
 
     @property
     def is_done(self) -> bool:
+        if self.ignore_eos:
+            return False
         return self._is_done
 
     @is_done.setter
@@ -536,6 +539,7 @@ def make_request(self,
         max_length = kwargs.pop("max_length", self.max_length)
         max_new_tokens = kwargs.pop("max_new_tokens", max_length - len(input_tokens))
         stream = kwargs.pop("stream", False)
+        ignore_eos = kwargs.pop("ignore_eos", False)
         # TODO: Add back this check
         # if self.policy.get_length(uid) + len(token_ids) >= max_length:
         #    raise ValueError(f"Session {uid} has reached max length {max_length}.")
@@ -578,6 +582,7 @@ def make_request(self,
                 sampler=sampler,
                 stop_criterion=stop_criterion,
                 stream=stream,
+                ignore_eos=ignore_eos,
             )
         ]
 
@@ -712,6 +717,8 @@ def put_request(self,
                     kwargs: Dict,
                     session_id: Union[str,
                                       None] = None) -> int:
+        if not self.is_rank_0:
+            return
         if self.stop_thread:
             raise RuntimeError("The request queue was shutdown.")
 

tests/conftest.py

@@ -4,36 +4,25 @@
 # DeepSpeed Team
 
 import pytest
+import time
+import torch
 import os
-import mii.legacy as mii
+import mii
 from types import SimpleNamespace
 
 
-@pytest.fixture(scope="function", params=["fp16"])
-def dtype(request):
-    return request.param
-
-
-@pytest.fixture(scope="function", params=[1])
+@pytest.fixture(scope="function", params=[None])
 def tensor_parallel(request):
-    return request.param
+    if request.param is not None:
+        return request.param
+    return int(os.getenv("WORLD_SIZE", "1"))
 
 
 @pytest.fixture(scope="function", params=[50050])
 def port_number(request):
     return request.param
 
 
-@pytest.fixture(scope="function", params=[False])
-def meta_tensor(request):
-    return request.param
-
-
-@pytest.fixture(scope="function", params=[False])
-def load_with_sys_mem(request):
-    return request.param
-
-
 @pytest.fixture(scope="function", params=[1])
 def replica_num(request):
     return request.param
@@ -49,66 +38,40 @@ def restful_api_port(request):
     return request.param
 
 
-@pytest.fixture(scope="function", params=["text-generation"])
+@pytest.fixture(scope="function", params=[mii.TaskType.TEXT_GENERATION])
 def task_name(request):
     return request.param
 
 
-@pytest.fixture(scope="function", params=["bigscience/bloom-560m"])
+@pytest.fixture(scope="function", params=["facebook/opt-1.3b"])
 def model_name(request):
     return request.param
 
 
-@pytest.fixture(scope="function")
-def deployment_name(model_name):
-    return model_name + "-deployment"
-
-
 @pytest.fixture(scope="function", params=[mii.DeploymentType.LOCAL])
 def deployment_type(request):
     return request.param
 
 
 @pytest.fixture(scope="function", params=[True])
-def enable_deepspeed(request):
-    return request.param
-
-
-@pytest.fixture(scope="function", params=[False])
-def enable_zero(request):
-    return request.param
-
-
-@pytest.fixture(scope="function", params=[{}])
-def ds_config(request):
+def all_rank_output(request):
     return request.param
 
 
 @pytest.fixture(scope="function")
 def model_config(
     task_name: str,
     model_name: str,
-    dtype: str,
     tensor_parallel: int,
-    meta_tensor: bool,
-    load_with_sys_mem: bool,
     replica_num: int,
-    enable_deepspeed: bool,
-    enable_zero: bool,
-    ds_config: dict,
+    all_rank_output: bool,
 ):
     config = SimpleNamespace(
+        model_name_or_path=model_name,
         task=task_name,
-        model=model_name,
-        dtype=dtype,
         tensor_parallel=tensor_parallel,
-        model_path=os.getenv("TRANSFORMERS_CACHE",
-                             ""),
-        meta_tensor=meta_tensor,
         replica_num=replica_num,
-        enable_deepspeed=enable_deepspeed,
-        enable_zero=enable_zero,
-        ds_config=ds_config,
+        all_rank_output=all_rank_output,
     )
     return config.__dict__
 
@@ -129,31 +92,38 @@ def mii_config(
     return config.__dict__
 
 
-@pytest.fixture(scope="function", params=[None])
+@pytest.fixture(scope="function", params=[None], ids=["nofail"])
 def expected_failure(request):
     return request.param
 
 
 @pytest.fixture(scope="function")
-def deployment(deployment_name, mii_config, model_config, expected_failure):
+def pipeline(model_config, expected_failure):
+    if expected_failure is not None:
+        with pytest.raises(expected_failure) as excinfo:
+            mii.pipeline(model_config=model_config)
+        yield excinfo
+    else:
+        pipe = mii.pipeline(model_config=model_config)
+        yield pipe
+        del pipe.inference_engine
+        del pipe
+        torch.cuda.empty_cache()
+
+
+@pytest.fixture(scope="function")
+def deployment(mii_config, model_config, expected_failure):
     if expected_failure is not None:
         with pytest.raises(expected_failure) as excinfo:
-            mii.deploy(
-                deployment_name=deployment_name,
-                mii_config=mii_config,
-                model_config=model_config,
-            )
+            mii.serve(model_config=model_config, mii_config=mii_config)
         yield excinfo
     else:
-        mii.deploy(
-            deployment_name=deployment_name,
-            mii_config=mii_config,
-            model_config=model_config,
-        )
-        yield deployment_name
-        mii.terminate(deployment_name)
+        client = mii.serve(model_config=model_config, mii_config=mii_config)
+        yield client
+        client.terminate_server()
+        time.sleep(1)
 
 
-@pytest.fixture(scope="function", params=[{"query": "DeepSpeed is the greatest"}])
+@pytest.fixture(scope="function", params=["DeepSpeed is the greatest"], ids=["query0"])
 def query(request):
     return request.param

tests/legacy/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team

tests/legacy/conftest.py

@@ -0,0 +1,159 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+import pytest
+import os
+import mii.legacy as mii
+from types import SimpleNamespace
+
+
+@pytest.fixture(scope="function", params=["fp16"])
+def dtype(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[1])
+def tensor_parallel(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[50050])
+def port_number(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[False])
+def meta_tensor(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[False])
+def load_with_sys_mem(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[1])
+def replica_num(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[False])
+def enable_restful_api(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[28080])
+def restful_api_port(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=["text-generation"])
+def task_name(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=["bigscience/bloom-560m"])
+def model_name(request):
+    return request.param
+
+
+@pytest.fixture(scope="function")
+def deployment_name(model_name):
+    return model_name + "-deployment"
+
+
+@pytest.fixture(scope="function", params=[mii.DeploymentType.LOCAL])
+def deployment_type(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[True])
+def enable_deepspeed(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[False])
+def enable_zero(request):
+    return request.param
+
+
+@pytest.fixture(scope="function", params=[{}])
+def ds_config(request):
+    return request.param
+
+
+@pytest.fixture(scope="function")
+def model_config(
+    task_name: str,
+    model_name: str,
+    dtype: str,
+    tensor_parallel: int,
+    meta_tensor: bool,
+    load_with_sys_mem: bool,
+    replica_num: int,
+    enable_deepspeed: bool,
+    enable_zero: bool,
+    ds_config: dict,
+):
+    config = SimpleNamespace(
+        task=task_name,
+        model=model_name,
+        dtype=dtype,
+        tensor_parallel=tensor_parallel,
+        model_path=os.getenv("TRANSFORMERS_CACHE",
+                             ""),
+        meta_tensor=meta_tensor,
+        replica_num=replica_num,
+        enable_deepspeed=enable_deepspeed,
+        enable_zero=enable_zero,
+        ds_config=ds_config,
+    )
+    return config.__dict__
+
+
+@pytest.fixture(scope="function")
+def mii_config(
+    deployment_type: str,
+    port_number: int,
+    enable_restful_api: bool,
+    restful_api_port: int,
+):
+    config = SimpleNamespace(
+        deployment_type=deployment_type,
+        port_number=port_number,
+        enable_restful_api=enable_restful_api,
+        restful_api_port=restful_api_port,
+    )
+    return config.__dict__
+
+
+@pytest.fixture(scope="function", params=[None])
+def expected_failure(request):
+    return request.param
+
+
+@pytest.fixture(scope="function")
+def deployment(deployment_name, mii_config, model_config, expected_failure):
+    if expected_failure is not None:
+        with pytest.raises(expected_failure) as excinfo:
+            mii.deploy(
+                deployment_name=deployment_name,
+                mii_config=mii_config,
+                model_config=model_config,
+            )
+        yield excinfo
+    else:
+        mii.deploy(
+            deployment_name=deployment_name,
+            mii_config=mii_config,
+            model_config=model_config,
+        )
+        yield deployment_name
+        mii.terminate(deployment_name)
+
+
+@pytest.fixture(scope="function", params=[{"query": "DeepSpeed is the greatest"}])
+def query(request):
+    return request.param

tests/legacy/pytest.ini

@@ -0,0 +1,3 @@
+[pytest]
+markers =
+    deepspeed:Run test for deepspeed CI

tests/pytest.ini

@@ -1,3 +1,2 @@
 [pytest]
-markers =
-    deepspeed:Run test for deepspeed CI
+addopts = --ignore=legacy