feat: split monolithic requirements.txt and remove test deps from runtime image (#6656)

Signed-off-by: Anant Sharma <anants@nvidia.com>

container/templates/vllm_runtime.Dockerfile

@@ -311,14 +311,19 @@ RUN if [ "${ENABLE_MODELEXPRESS_P2P}" = "true" ]; then \
     fi
 {% endif %}
 
-# Install common and test dependencies. Cache uv downloads; uv handles its own locking for this cache.
-RUN --mount=type=bind,source=./container/deps/requirements.txt,target=/tmp/requirements.txt \
-    --mount=type=bind,source=./container/deps/requirements.test.txt,target=/tmp/requirements.test.txt \
+# Install runtime dependencies (common + vllm-specific + planner + benchmarks).
+# Test and dev dependencies are NOT installed here — they go in the test and dev images.
+RUN --mount=type=bind,source=./container/deps/requirements.common.txt,target=/tmp/requirements.common.txt \
+    --mount=type=bind,source=./container/deps/requirements.vllm.txt,target=/tmp/requirements.vllm.txt \
+    --mount=type=bind,source=./container/deps/requirements.planner.txt,target=/tmp/requirements.planner.txt \
+    --mount=type=bind,source=./container/deps/requirements.benchmark.txt,target=/tmp/requirements.benchmark.txt \
     --mount=type=cache,target=/home/dynamo/.cache/uv,uid=1000,gid=0,mode=0775 \
     export UV_CACHE_DIR=/home/dynamo/.cache/uv UV_GIT_LFS=1 UV_HTTP_TIMEOUT=300 UV_HTTP_RETRIES=5 && \
     uv pip install \
-        --requirement /tmp/requirements.txt \
-        --requirement /tmp/requirements.test.txt
+        --requirement /tmp/requirements.common.txt \
+        --requirement /tmp/requirements.vllm.txt \
+        --requirement /tmp/requirements.planner.txt \
+        --requirement /tmp/requirements.benchmark.txt
 
 # Copy tests, deploy and components for CI with correct ownership
 # Pattern: COPY --chmod=775 <path>; chmod g+w <path> done later as root because COPY --chmod only affects <path>/*, not <path>

lib/bindings/python/examples/kserve_grpc_service/test_client.py

@@ -10,9 +10,15 @@
 import argparse
 
 import numpy as np
-import tritonclient.grpc as triton_grpc
+
+try:
+    import tritonclient.grpc as triton_grpc
+    from tritonclient.utils import InferenceServerException
+except ImportError:
+    triton_grpc = None
+    InferenceServerException = None
+
 from google.protobuf.json_format import MessageToDict
-from tritonclient.utils import InferenceServerException
 
 
 def main() -> None:

lib/bindings/python/tests/test_kserve_grpc.py

@@ -7,8 +7,13 @@ from contextlib import asynccontextmanager
 from typing import Any, AsyncIterator, Optional, Tuple
 
 import pytest
-import tritonclient.grpc.model_config_pb2 as mc
-from tritonclient.utils import InferenceServerException
+
+try:
+    import tritonclient.grpc.model_config_pb2 as mc
+    from tritonclient.utils import InferenceServerException
+except ImportError:
+    mc = None
+    InferenceServerException = None
 
 from dynamo.llm import KserveGrpcService, ModelRuntimeConfig, PythonAsyncEngine
 

pyproject.toml

@@ -15,10 +15,6 @@ requires-python = ">=3.10"
 dependencies = [
     "ai-dynamo-runtime==1.0.0",
     "transformers>=4.56.0",
-    "pytest>=8.3.4",
-    "types-aiofiles>=24.1.0",
-    "types-psutil>=7.0.0.20250218",
-    "types-requests>=2.32.4.20260107",
     "kubernetes>=32.0.1,<33.0.0",
     "fastapi>=0.115.0",
     "distro",
@@ -28,6 +24,7 @@ dependencies = [
     "click<8.2.0",
     "setuptools",
     "prometheus_client>=0.23.1,<1.0",
+    "msgpack==1.1.2",
 ]
 
 classifiers = [
@@ -52,7 +49,6 @@ Repository = "https://github.com/ai-dynamo/dynamo.git"
 [project.optional-dependencies]
 trtllm =[
     "uvloop",
-    "msgpack==1.1.2",
     "tensorrt-llm==1.3.0rc5.post1",
 ]
 

tests/fault_tolerance/deploy/container/Dockerfile.local_vllm

@@ -139,11 +139,19 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     /opt/dynamo/wheelhouse/nixl/nixl*.whl \
     && rm -rf /opt/dynamo/wheelhouse
 
-# Install common and test dependencies
-COPY container/deps/requirements.txt /tmp/requirements.txt
+# Install runtime dependencies (common + vllm-specific + planner + benchmarks) and test dependencies
+COPY container/deps/requirements.common.txt /tmp/requirements.common.txt
+COPY container/deps/requirements.vllm.txt /tmp/requirements.vllm.txt
+COPY container/deps/requirements.planner.txt /tmp/requirements.planner.txt
+COPY container/deps/requirements.benchmark.txt /tmp/requirements.benchmark.txt
 COPY container/deps/requirements.test.txt /tmp/requirements.test.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --requirement /tmp/requirements.txt --requirement /tmp/requirements.test.txt && \
+    uv pip install \
+        --requirement /tmp/requirements.common.txt \
+        --requirement /tmp/requirements.vllm.txt \
+        --requirement /tmp/requirements.planner.txt \
+        --requirement /tmp/requirements.benchmark.txt \
+        --requirement /tmp/requirements.test.txt && \
     rm /tmp/requirements*.txt
 
 # Copy workspace files

tests/frontend/conftest.py

@@ -15,7 +15,6 @@ import time
 
 import pytest
 import requests
-import tritonclient.grpc as grpcclient
 
 from tests.utils.constants import QWEN
 from tests.utils.managed_process import DynamoFrontendProcess, ManagedProcess
@@ -80,6 +79,8 @@ def check_grpc_server_ready(
     Raises:
         Exception: If server is not ready after max_attempts
     """
+    import tritonclient.grpc as grpcclient
+
     for attempt in range(max_attempts):
         try:
             client = grpcclient.InferenceServerClient(f"localhost:{port}")

tests/frontend/grpc/test_tensor_mocker_engine.py

@@ -20,8 +20,16 @@ from functools import partial
 
 import numpy as np
 import pytest
-import triton_echo_client
-import tritonclient.grpc as grpcclient
+
+try:
+    import tritonclient.grpc as grpcclient
+except ImportError:
+    grpcclient = None
+
+try:
+    import triton_echo_client
+except ImportError:
+    triton_echo_client = None
 
 from tests.utils.constants import QWEN
 from tests.utils.managed_process import ManagedProcess

tests/frontend/grpc/test_tensor_parameters.py

@@ -16,7 +16,11 @@ import shutil
 
 import numpy as np
 import pytest
-import tritonclient.grpc as grpcclient
+
+try:
+    import tritonclient.grpc as grpcclient
+except ImportError:
+    grpcclient = None
 
 from tests.utils.managed_process import ManagedProcess
 

tests/planner/README.md

@@ -10,7 +10,7 @@ This directory contains comprehensive testing tools for validating the SLA plann
 The SLA planner monitors metrics every 60 seconds (default adjustment interval) and scales
 prefill/decode workers based on TTFT, ITL, and request patterns.
 
-To setup the environment, simply use the released docker images for any backends, or build your own docker image following the READMEs in `./examples/backends/<vllm/sglang/trtllm>/README.md`, or follow the `Developing Locally` section in [README.md](../../README.md) to setup the environment locally. If using the local environment, make sure to install dependencies by running `UV_GIT_LFS=1 uv pip install --no-cache -r container/deps/requirements.txt`
+To setup the environment, simply use the released docker images for any backends, or build your own docker image following the READMEs in `./examples/backends/<vllm/sglang/trtllm>/README.md`, or follow the `Developing Locally` section in [README.md](../../README.md) to setup the environment locally. If using the local environment, make sure to install dependencies by running `UV_GIT_LFS=1 uv pip install --no-cache -r container/deps/requirements.common.txt -r container/deps/requirements.planner.txt`
 
 ## Pre-Requisite: Pre-Deployment Profiling Data