Merge ae7f3aeba1 into 719907410b

.github/workflows/build.yaml

@@ -129,9 +129,9 @@ jobs:
                 export label_extension="-gaudi"
                 export docker_volume="/mnt/cache"
                 export docker_devices=""
-                export runs_on="ubuntu-latest"
+                export runs_on="itac-bm-emr-gaudi3-dell-8gaudi"
                 export platform=""
-                export extra_pytest=""
+                export extra_pytest="--gaudi"
                 export target=""
           esac
           echo $dockerfile

backends/gaudi/Makefile

@@ -50,11 +50,14 @@ local-dev-install: install-dependencies
 
 # In order to run the integration tests, you need to first build the image (make -C backends/gaudi image)
 run-integration-tests:
-	pip install -U pip uv
-	uv pip install -r ${root_dir}/backends/gaudi/server/integration-tests/requirements.txt
 	DOCKER_VOLUME=${root_dir}/data \
 	HF_TOKEN=`cat ${HOME}/.cache/huggingface/token` \
-	uv run pytest --durations=0 -sv ${root_dir}/backends/gaudi/server/integration-tests
+    pytest --durations=0 -s -vv ${root_dir}/integration-tests --gaudi
+
+run-integration-tests-with-all-models:
+	DOCKER_VOLUME=${root_dir}/data \
+	HF_TOKEN=`cat ${HOME}/.cache/huggingface/token` \
+	pytest --durations=0 -s -vv ${root_dir}/integration-tests --gaudi --gaudi-all-models
 
 # This is used to capture the expected outputs for the integration tests offering an easy way to add more models to the integration tests
 capture-expected-outputs-for-integration-tests:

backends/gaudi/README.md

@@ -99,16 +99,26 @@ curl 127.0.0.1:8080/generate \
 
 ### Integration tests
 
+Install the dependencies:
+```bash
+pip install -r integration-tests/requirements.txt
+```
+
 To run the integration tests, you need to first build the image:
 ```bash
 make -C backends/gaudi image
 ```
 
-Then run the following command to run the integration tests:
+Then run the following command to run the integration tests (CI tests):
 ```bash
 make -C backends/gaudi run-integration-tests
 ```
 
+To run the integration tests with all models, you can run the following command:
+```bash
+make -C backends/gaudi run-integration-tests-with-all-models
+```
+
 To capture the expected outputs for the integration tests, you can run the following command:
 ```bash
 make -C backends/gaudi capture-expected-outputs-for-integration-tests

backends/gaudi/server/integration-tests/pytest.ini

@@ -1,2 +0,0 @@
-[pytest]
-asyncio_mode = auto

backends/gaudi/server/integration-tests/requirements.txt

@@ -1,7 +0,0 @@
-pytest >= 8.3.5
-pytest-asyncio >= 0.26.0
-docker >= 7.1.0
-Levenshtein >= 0.27.1
-loguru >= 0.7.3
-aiohttp >= 3.11.14
-text-generation

integration-tests/conftest.py

@@ -1,4 +1,8 @@
-pytest_plugins = ["fixtures.neuron.service", "fixtures.neuron.export_models"]
+pytest_plugins = [
+    "fixtures.neuron.service",
+    "fixtures.neuron.export_models",
+    "fixtures.gaudi.service",
+]
 # ruff: noqa: E402
 from _pytest.fixtures import SubRequest
 from huggingface_hub.inference._generated.types.chat_completion import (
@@ -68,6 +72,15 @@ def pytest_addoption(parser):
     parser.addoption(
         "--neuron", action="store_true", default=False, help="run neuron tests"
     )
+    parser.addoption(
+        "--gaudi", action="store_true", default=False, help="run gaudi tests"
+    )
+    parser.addoption(
+        "--gaudi-all-models",
+        action="store_true",
+        default=False,
+        help="Run tests for all models instead of just the default subset",
+    )
 
 
 def pytest_configure(config):
@@ -84,6 +97,22 @@ def pytest_collection_modifyitems(config, items):
                 item.add_marker(pytest.mark.skip(reason="need --release option to run"))
 
         selectors.append(skip_release)
+
+    if config.getoption("--gaudi"):
+
+        def skip_not_gaudi(item):
+            if "gaudi" not in item.keywords:
+                item.add_marker(pytest.mark.skip(reason="requires --gaudi to run"))
+
+        selectors.append(skip_not_gaudi)
+    else:
+
+        def skip_gaudi(item):
+            if "gaudi" in item.keywords:
+                item.add_marker(pytest.mark.skip(reason="requires --gaudi to run"))
+
+        selectors.append(skip_gaudi)
+
     if config.getoption("--neuron"):
 
         def skip_not_neuron(item):
@@ -100,6 +129,7 @@ def pytest_collection_modifyitems(config, items):
                 item.add_marker(pytest.mark.skip(reason="requires --neuron to run"))
 
         selectors.append(skip_neuron)
+
     for item in items:
         for selector in selectors:
             selector(item)

integration-tests/fixtures/gaudi/service.py

@@ -14,15 +14,21 @@ import docker
 import pytest
 from aiohttp import ClientConnectorError, ClientOSError, ServerDisconnectedError
 from docker.errors import NotFound
-from loguru import logger
-from test_model import TEST_CONFIGS
-from text_generation import AsyncClient
-from text_generation.types import Response
+import logging
+from huggingface_hub import AsyncInferenceClient, TextGenerationOutput
+import huggingface_hub
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="<green>{time:YYYY-MM-DD HH:mm:ss}</green> | <level>{level: <8}</level> | <cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> - <level>{message}</level>",
+    stream=sys.stdout,
+)
+logger = logging.getLogger(__file__)
 
 # Use the latest image from the local docker build
 DOCKER_IMAGE = os.getenv("DOCKER_IMAGE", "tgi-gaudi")
 DOCKER_VOLUME = os.getenv("DOCKER_VOLUME", None)
-HF_TOKEN = os.getenv("HF_TOKEN", None)
+HF_TOKEN = huggingface_hub.get_token()
 
 assert (
     HF_TOKEN is not None
@@ -48,12 +54,6 @@ HABANA_RUN_ARGS = {
     "cap_add": ["sys_nice"],
 }
 
-logger.add(
-    sys.stderr,
-    format="<green>{time:YYYY-MM-DD HH:mm:ss}</green> | <level>{level: <8}</level> | <cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> - <level>{message}</level>",
-    level="INFO",
-)
-
 
 def stream_container_logs(container, test_name):
     """Stream container logs in a separate thread."""
@@ -69,9 +69,15 @@ def stream_container_logs(container, test_name):
         logger.error(f"Error streaming container logs: {str(e)}")
 
 
+class TestClient(AsyncInferenceClient):
+    def __init__(self, service_name: str, base_url: str):
+        super().__init__(model=base_url)
+        self.service_name = service_name
+
+
 class LauncherHandle:
-    def __init__(self, port: int):
-        self.client = AsyncClient(f"http://localhost:{port}", timeout=3600)
+    def __init__(self, service_name: str, port: int):
+        self.client = TestClient(service_name, f"http://localhost:{port}")
 
     def _inner_health(self):
         raise NotImplementedError
@@ -87,7 +93,7 @@ class LauncherHandle:
                 raise RuntimeError("Launcher crashed")
 
             try:
-                await self.client.generate("test")
+                await self.client.text_generation("test", max_new_tokens=1)
                 elapsed = time.time() - start_time
                 logger.info(f"Health check passed after {elapsed:.1f}s")
                 return
@@ -111,7 +117,8 @@ class LauncherHandle:
 
 class ContainerLauncherHandle(LauncherHandle):
     def __init__(self, docker_client, container_name, port: int):
-        super(ContainerLauncherHandle, self).__init__(port)
+        service_name = container_name  # Use container name as service name
+        super(ContainerLauncherHandle, self).__init__(service_name, port)
         self.docker_client = docker_client
         self.container_name = container_name
 
@@ -132,7 +139,8 @@ class ContainerLauncherHandle(LauncherHandle):
 
 class ProcessLauncherHandle(LauncherHandle):
     def __init__(self, process, port: int):
-        super(ProcessLauncherHandle, self).__init__(port)
+        service_name = "process"  # Use generic name for process launcher
+        super(ProcessLauncherHandle, self).__init__(service_name, port)
         self.process = process
 
     def _inner_health(self) -> bool:
@@ -151,11 +159,13 @@ def data_volume():
 
 
 @pytest.fixture(scope="module")
-def launcher(data_volume):
+def gaudi_launcher():
     @contextlib.contextmanager
     def docker_launcher(
         model_id: str,
         test_name: str,
+        tgi_args: List[str] = None,
+        env_config: dict = None,
     ):
         logger.info(
             f"Starting docker launcher for model {model_id} and test {test_name}"
@@ -183,32 +193,40 @@ def launcher(data_volume):
             )
             container.stop()
             container.wait()
+            container.remove()
+            logger.info(f"Removed existing container {container_name}")
         except NotFound:
             pass
         except Exception as e:
             logger.error(f"Error handling existing container: {str(e)}")
 
-        model_name = next(
-            name for name, cfg in TEST_CONFIGS.items() if cfg["model_id"] == model_id
-        )
-
-        tgi_args = TEST_CONFIGS[model_name]["args"].copy()
+        if tgi_args is None:
+            tgi_args = []
+        else:
+            tgi_args = tgi_args.copy()
 
         env = BASE_ENV.copy()
 
         # Add model_id to env
         env["MODEL_ID"] = model_id
 
-        # Add env config that is definied in the fixture parameter
-        if "env_config" in TEST_CONFIGS[model_name]:
-            env.update(TEST_CONFIGS[model_name]["env_config"].copy())
+        # Add env config that is defined in the fixture parameter
+        if env_config is not None:
+            env.update(env_config.copy())
 
-        volumes = [f"{DOCKER_VOLUME}:/data"]
+        volumes = []
+        if DOCKER_VOLUME:
+            volumes = [f"{DOCKER_VOLUME}:/data"]
         logger.debug(f"Using volume {volumes}")
 
         try:
+            logger.debug(f"Using command {tgi_args}")
             logger.info(f"Creating container with name {container_name}")
 
+            logger.debug(f"Using environment {env}")
+            logger.debug(f"Using volumes {volumes}")
+            logger.debug(f"HABANA_RUN_ARGS {HABANA_RUN_ARGS}")
+
             # Log equivalent docker run command for debugging, this is not actually executed
             container = client.containers.run(
                 DOCKER_IMAGE,
@@ -271,15 +289,16 @@ def launcher(data_volume):
 
 
 @pytest.fixture(scope="module")
-def generate_load():
+def gaudi_generate_load():
     async def generate_load_inner(
-        client: AsyncClient, prompt: str, max_new_tokens: int, n: int
-    ) -> List[Response]:
+        client: AsyncInferenceClient, prompt: str, max_new_tokens: int, n: int
+    ) -> List[TextGenerationOutput]:
         try:
             futures = [
-                client.generate(
+                client.text_generation(
                     prompt,
                     max_new_tokens=max_new_tokens,
+                    details=True,
                     decoder_input_details=True,
                 )
                 for _ in range(n)

integration-tests/gaudi/capture_expected_outputs.py

@@ -3,7 +3,7 @@ import os
 from typing import Dict, Any, Generator
 
 import pytest
-from test_model import TEST_CONFIGS
+from test_gaudi_generate import TEST_CONFIGS
 
 UNKNOWN_CONFIGS = {
     name: config

integration-tests/gaudi/test_gaudi_generate.py

@@ -1,10 +1,16 @@
-from typing import Any, Dict
-
-from text_generation import AsyncClient
+from typing import Any, Dict, Generator
+from _pytest.fixtures import SubRequest
+from huggingface_hub import AsyncInferenceClient
 import pytest
-from Levenshtein import distance as levenshtein_distance
 
-# The "args" config is not optimized for speed but only check that the inference is working for the different models architectures
+
+def pytest_configure(config):
+    config.addinivalue_line(
+        "markers", "gaudi_all_models: mark test to run with all models"
+    )
+
+
+# The "args" values in TEST_CONFIGS are not optimized for speed but only check that the inference is working for the different models architectures.
 TEST_CONFIGS = {
     "meta-llama/Llama-3.1-8B-Instruct-shared": {
         "model_id": "meta-llama/Llama-3.1-8B-Instruct",
@@ -25,6 +31,7 @@ TEST_CONFIGS = {
             "--max-batch-prefill-tokens",
             "2048",
         ],
+        "run_by_default": True,
     },
     "meta-llama/Llama-3.1-8B-Instruct": {
         "model_id": "meta-llama/Llama-3.1-8B-Instruct",
@@ -42,6 +49,7 @@ TEST_CONFIGS = {
             "--max-batch-prefill-tokens",
             "2048",
         ],
+        "run_by_default": True,
     },
     "meta-llama/Llama-2-7b-chat-hf": {
         "model_id": "meta-llama/Llama-2-7b-chat-hf",
@@ -181,72 +189,98 @@ TEST_CONFIGS = {
     },
 }
 
-print(f"Testing {len(TEST_CONFIGS)} models")
+
+def pytest_generate_tests(metafunc):
+    if "test_config" in metafunc.fixturenames:
+        if metafunc.config.getoption("--gaudi-all-models"):
+            models = list(TEST_CONFIGS.keys())
+        else:
+            models = [
+                name
+                for name, config in TEST_CONFIGS.items()
+                if config.get("run_by_default", False)
+            ]
+        print(f"Testing {len(models)} models")
+        metafunc.parametrize("test_config", models, indirect=True)
 
 
-@pytest.fixture(scope="module", params=TEST_CONFIGS.keys())
-def test_config(request) -> Dict[str, Any]:
+@pytest.fixture(scope="module")
+def test_config(request: SubRequest) -> Dict[str, Any]:
     """Fixture that provides model configurations for testing."""
-    test_config = TEST_CONFIGS[request.param]
-    test_config["test_name"] = request.param
+    model_name = request.param
+    test_config = TEST_CONFIGS[model_name]
+    test_config["test_name"] = model_name
     return test_config
 
 
 @pytest.fixture(scope="module")
-def model_id(test_config):
+def model_id(test_config: Dict[str, Any]) -> Generator[str, None, None]:
     yield test_config["model_id"]
 
 
 @pytest.fixture(scope="module")
-def test_name(test_config):
+def test_name(test_config: Dict[str, Any]) -> Generator[str, None, None]:
     yield test_config["test_name"]
 
 
 @pytest.fixture(scope="module")
-def expected_outputs(test_config):
+def expected_outputs(test_config: Dict[str, Any]) -> Dict[str, str]:
     return {
         "greedy": test_config["expected_greedy_output"],
-        # "sampling": model_config["expected_sampling_output"],
         "batch": test_config["expected_batch_output"],
     }
 
 
 @pytest.fixture(scope="module")
-def input(test_config):
+def input(test_config: Dict[str, Any]) -> str:
     return test_config["input"]
 
 
 @pytest.fixture(scope="module")
-def tgi_service(launcher, model_id, test_name):
-    with launcher(model_id, test_name) as tgi_service:
+def tgi_service(
+    gaudi_launcher, model_id: str, test_name: str, test_config: Dict[str, Any]
+):
+    with gaudi_launcher(
+        model_id,
+        test_name,
+        tgi_args=test_config.get("args", []),
+        env_config=test_config.get("env_config", {}),
+    ) as tgi_service:
         yield tgi_service
 
 
 @pytest.fixture(scope="module")
-async def tgi_client(tgi_service) -> AsyncClient:
+async def tgi_client(tgi_service) -> AsyncInferenceClient:
     await tgi_service.health(1000)
     return tgi_service.client
 
 
 @pytest.mark.asyncio
+@pytest.mark.all_models
 async def test_model_single_request(
-    tgi_client: AsyncClient, expected_outputs: Dict[str, Any], input: str
+    tgi_client: AsyncInferenceClient, expected_outputs: Dict[str, str], input: str
 ):
     # Bounded greedy decoding without input
-    response = await tgi_client.generate(
+    response = await tgi_client.text_generation(
         input,
         max_new_tokens=32,
+        details=True,
+        decoder_input_details=True,
     )
     assert response.details.generated_tokens == 32
     assert response.generated_text == expected_outputs["greedy"]
 
 
 @pytest.mark.asyncio
+@pytest.mark.all_models
 async def test_model_multiple_requests(
-    tgi_client, generate_load, expected_outputs, input
+    tgi_client: AsyncInferenceClient,
+    gaudi_generate_load,
+    expected_outputs: Dict[str, str],
+    input: str,
 ):
     num_requests = 4
-    responses = await generate_load(
+    responses = await gaudi_generate_load(
         tgi_client,
         input,
         max_new_tokens=32,
@@ -257,6 +291,4 @@ async def test_model_multiple_requests(
     expected = expected_outputs["batch"]
     for r in responses:
         assert r.details.generated_tokens == 32
-        # Compute the similarity with the expectation using the levenshtein distance
-        # We should not have more than two substitutions or additions
-        assert levenshtein_distance(r.generated_text, expected) < 3
+        assert r.generated_text == expected