chore: bump vLLM to 0.11.2

.github/actions/pytest/action.yml

@@ -54,12 +54,13 @@ runs:
         # Run pytest with detailed output and JUnit XML
         set +e  # Don't exit on test failures
 
+        # -W ignore::DeprecationWarning suppresses DeprecationWarnings from SWIG/SciPy (vLLM dependency). Remove once vLLM updates to a SciPy version which uses SWIG 4.4.x.
         docker run --runtime=nvidia --gpus all -w /workspace \
           --cpus=${NUM_CPUS} \
           --network host \
           --name ${{ env.CONTAINER_ID }}_pytest \
           ${{ inputs.image_tag }} \
-          bash -c "mkdir -p /workspace/test-results && pytest -v --tb=short --basetemp=/tmp -o cache_dir=/tmp/.pytest_cache --junitxml=/workspace/test-results/${{ env.PYTEST_XML_FILE }} --durations=10 -m \"${{ inputs.pytest_marks }}\""
+          bash -c "mkdir -p /workspace/test-results && pytest -v --tb=short --basetemp=/tmp -o cache_dir=/tmp/.pytest_cache --junitxml=/workspace/test-results/${{ env.PYTEST_XML_FILE }} --durations=10 -W ignore::DeprecationWarning -m \"${{ inputs.pytest_marks }}\""
 
         TEST_EXIT_CODE=$?
         echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> $GITHUB_ENV

container/Dockerfile.vllm

@@ -3,21 +3,22 @@
 # SPDX-License-Identifier: Apache-2.0
 
 ARG BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
-# TODO OPS-612: NCCL will hang with 25.03, so use 25.01 for now
+# Note: CUDA 12.9 requires NGC 25.04+ (25.01 only has CUDA 12.8)
+# TODO OPS-612: NCCL hang issue with 25.03 - verify if resolved in 25.04
 # Please check https://github.com/ai-dynamo/dynamo/pull/1065
 # for details and reproducer to manually test if the image
 # can be updated to later versions.
-ARG BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
+ARG BASE_IMAGE_TAG="25.04-cuda12.9-devel-ubuntu24.04"
 ARG ENABLE_KVBM=false
 ARG RUNTIME_IMAGE="nvcr.io/nvidia/cuda"
-ARG RUNTIME_IMAGE_TAG="12.8.1-runtime-ubuntu24.04"
-ARG CUDA_VERSION="12.8"
+ARG RUNTIME_IMAGE_TAG="12.9.0-runtime-ubuntu24.04"
+ARG CUDA_VERSION="12.9"
 
 # Make sure to update the dependency version in pyproject.toml when updating this
-ARG VLLM_REF="v0.11.0"
+ARG VLLM_REF="v0.11.2"
 # FlashInfer only respected when building vLLM from source, ie when VLLM_REF does not start with 'v' or for arm64 builds
-ARG FLASHINF_REF="v0.3.1"
-ARG TORCH_BACKEND="cu128"
+ARG FLASHINF_REF="v0.5.2"
+ARG TORCH_BACKEND="cu129"
 
 # If left blank, then we will fallback to vLLM defaults
 ARG DEEPGEMM_REF=""
@@ -206,7 +207,7 @@ RUN apt-get update && \
         # prometheus dependencies
         ca-certificates \
         # DeepGemm uses 'cuobjdump' which does not come with CUDA image
-        cuda-command-line-tools-12-8 && \
+        cuda-command-line-tools-12-9 && \
     rm -rf /var/lib/apt/lists/*
 
 # Copy CUDA development tools (nvcc, headers, dependencies, etc.) from base devel image

container/build.sh

@@ -102,11 +102,12 @@ DEFAULT_TENSORRTLLM_PIP_WHEEL="tensorrt-llm==1.2.0rc2"
 TENSORRTLLM_PIP_WHEEL=""
 
 VLLM_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
-# FIXME: NCCL will hang with 25.03, so use 25.01 for now
+# Note: CUDA 12.9 requires NGC 25.04+ (25.01 only has CUDA 12.8)
+# FIXME: NCCL hang issue with 25.03 - verify if resolved in 25.04
 # Please check https://github.com/ai-dynamo/dynamo/pull/1065
 # for details and reproducer to manually test if the image
 # can be updated to later versions.
-VLLM_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
+VLLM_BASE_IMAGE_TAG="25.04-cuda12.9-devel-ubuntu24.04"
 
 NONE_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
 NONE_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"

container/deps/vllm/install_vllm.sh

@@ -13,23 +13,23 @@
 
 set -euo pipefail
 
-VLLM_REF="v0.11.0"
+VLLM_REF="v0.11.2"
 
 # Basic Configurations
 ARCH=$(uname -m)
 MAX_JOBS=16
 INSTALLATION_DIR=/tmp
 
 # VLLM and Dependency Configurations
-TORCH_BACKEND="cu128"
+TORCH_BACKEND="cu129"
 TORCH_CUDA_ARCH_LIST="9.0;10.0" # For EP Kernels
 DEEPGEMM_REF=""
-CUDA_VERSION="12.8" # For DEEPGEMM
+CUDA_VERSION="12.9" # For DEEPGEMM
 
 # These flags are applicable when installing vLLM from source code
 EDITABLE=true
 VLLM_GIT_URL="https://github.com/vllm-project/vllm.git"
-FLASHINF_REF="v0.3.1"
+FLASHINF_REF="v0.5.2"
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -141,6 +141,8 @@ if [[ $VLLM_REF =~ ^v ]] && { [ "$ARCH" = "amd64" ] || { [ "$ARCH" = "arm64" ] &
     echo "Installing vLLM $VLLM_REF from PyPI... (ARCH=$ARCH, TORCH_BACKEND=$TORCH_BACKEND)"
 
     uv pip install vllm[flashinfer]==$VLLM_REF --torch-backend=$TORCH_BACKEND
+    uv pip install flashinfer-cubin==$FLASHINF_REF
+    uv pip install flashinfer-jit-cache==$FLASHINF_REF --extra-index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
 else
     # VLLM_REF does not start with 'v' or amd64 - use git checkout path

examples/backends/sglang/slurm_jobs/scripts/vllm/benchmark_serving.py

@@ -8,7 +8,6 @@
     vLLM OpenAI API server
     vllm serve <your_model> \
         --swap-space 16 \
-        --disable-log-requests
 
     (TGI backend)
     ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>

examples/backends/vllm/deploy/agg_kvbm.yaml

@@ -42,7 +42,6 @@ spec:
             - Qwen/Qwen3-8B
             - --gpu-memory-utilization
             - "0.45"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager

examples/backends/vllm/deploy/disagg_kvbm.yaml

@@ -35,7 +35,6 @@ spec:
             - Qwen/Qwen3-8B
             - --gpu-memory-utilization
             - "0.3"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager
@@ -68,7 +67,6 @@ spec:
             - --is-prefill-worker
             - --gpu-memory-utilization
             - "0.3"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager

examples/backends/vllm/deploy/disagg_kvbm_2p2d.yaml

@@ -35,7 +35,6 @@ spec:
             - Qwen/Qwen3-8B
             - --gpu-memory-utilization
             - "0.3"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager
@@ -68,7 +67,6 @@ spec:
             - --is-prefill-worker
             - --gpu-memory-utilization
             - "0.3"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager

examples/backends/vllm/deploy/disagg_kvbm_tp2.yaml

@@ -37,7 +37,6 @@ spec:
             - Qwen/Qwen3-8B
             - --gpu-memory-utilization
             - "0.23"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager
@@ -72,7 +71,6 @@ spec:
             - --is-prefill-worker
             - --gpu-memory-utilization
             - "0.23"
-            - --disable-log-requests
             - --max-model-len
             - "32000"
             - --enforce-eager

lib/bindings/kvbm/python/kvbm/vllm_integration/connector/dynamo_connector.py

@@ -23,6 +23,7 @@
     from vllm.config import VllmConfig
     from vllm.forward_context import ForwardContext
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.kv_cache_interface import KVCacheConfig
     from vllm.v1.request import Request
 
 
@@ -40,8 +41,15 @@ def __init__(self, metadata: bytes):
 
 
 class DynamoConnector(KVConnectorBase_V1):
-    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
-        super().__init__(vllm_config=vllm_config, role=role)
+    def __init__(
+        self,
+        vllm_config: "VllmConfig",
+        role: KVConnectorRole,
+        kv_cache_config: Optional["KVCacheConfig"] = None,
+    ):
+        super().__init__(
+            vllm_config=vllm_config, role=role, kv_cache_config=kv_cache_config
+        )
 
         assert vllm_config.kv_transfer_config is not None
         assert vllm_config.kv_transfer_config.engine_id is not None

lib/bindings/kvbm/python/kvbm/vllm_integration/connector/pd_connector.py

@@ -29,6 +29,7 @@
         LMCacheConnectorV1,
     )
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.kv_cache_interface import KVCacheConfig
     from vllm.v1.request import Request
 
 
@@ -46,8 +47,15 @@ class PdConnector(MultiConnector):
     - The second connector must be NIXL and will be used by decode worker to get KV blocks from prefill worker.
     """
 
-    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
-        super().__init__(vllm_config=vllm_config, role=role)
+    def __init__(
+        self,
+        vllm_config: "VllmConfig",
+        role: KVConnectorRole,
+        kv_cache_config: "KVCacheConfig",
+    ):
+        super().__init__(
+            vllm_config=vllm_config, role=role, kv_cache_config=kv_cache_config
+        )
         if len(self._connectors) != 2:
             raise ValueError(
                 f"PdConnector requires exactly two connectors (got {len(self._connectors)})"

lib/bindings/kvbm/python/kvbm/vllm_integration/connector_worker.py

@@ -14,7 +14,7 @@
 from vllm.config import VllmConfig
 from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
 from vllm.model_executor.models.utils import extract_layer_index
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
+from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
 
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionMetadata

pyproject.toml

@@ -56,7 +56,7 @@ trtllm =[
 vllm = [
     "uvloop",
     "nixl[cu12]<=0.7.1",
-    "vllm[flashinfer]==0.11.0",
+    "vllm[flashinfer]==0.11.2",
 ]
 
 sglang = [

recipes/llama-3-70b/vllm/agg/deploy.yaml

@@ -43,7 +43,7 @@ spec:
             - name: HF_HOME
               value: /opt/models
           args:
-          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 4 --data-parallel-size 1 --disable-log-requests --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 4 --data-parallel-size 1 --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
           command:
           - /bin/sh
           - -c

recipes/llama-3-70b/vllm/disagg-multi-node/deploy.yaml

@@ -43,7 +43,7 @@ spec:
             - name: HF_HOME
               value: /opt/models
           args:
-          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 8 --data-parallel-size 1 --disable-log-requests --is-prefill-worker --gpu-memory-utilization 0.95 --no-enable-prefix-caching --block-size 128"
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 8 --data-parallel-size 1 --is-prefill-worker --gpu-memory-utilization 0.95 --no-enable-prefix-caching --block-size 128"
           command:
           - /bin/sh
           - -c
@@ -74,7 +74,7 @@ spec:
             - name: HF_HOME
               value: /opt/models
           args:
-          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 8 --data-parallel-size 1 --disable-log-requests --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 8 --data-parallel-size 1 --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
           command:
           - /bin/sh
           - -c

recipes/llama-3-70b/vllm/disagg-single-node/deploy.yaml

@@ -55,7 +55,7 @@ spec:
             - name: HF_HOME
               value: /opt/models
           args:
-          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 2 --data-parallel-size 1 --disable-log-requests --is-prefill-worker --gpu-memory-utilization 0.95 --no-enable-prefix-caching --block-size 128"
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 2 --data-parallel-size 1 --is-prefill-worker --gpu-memory-utilization 0.95 --no-enable-prefix-caching --block-size 128"
           command:
           - /bin/sh
           - -c
@@ -98,7 +98,7 @@ spec:
             - name: HF_HOME
               value: /opt/models
           args:
-          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 4 --data-parallel-size 1 --disable-log-requests --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 4 --data-parallel-size 1 --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
           command:
           - /bin/sh
           - -c

tests/dependencies/test_vllm_imports.py

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unit tests to sanity check that required dependencies can be imported."""
+
+import pytest
+
+
+@pytest.mark.vllm
+@pytest.mark.unit
+@pytest.mark.gpu_1
+def test_import_deep_ep():
+    """Test that deep_ep module can be imported."""
+    try:
+        import deep_ep
+
+        assert deep_ep is not None
+    except ImportError as e:
+        pytest.fail(f"Failed to import deep_ep: {e}")
+
+
+@pytest.mark.vllm
+@pytest.mark.unit
+@pytest.mark.gpu_1
+def test_import_pplx_kernels():
+    """Test that pplx_kernels module can be imported."""
+    try:
+        import pplx_kernels
+
+        assert pplx_kernels is not None
+    except ImportError as e:
+        pytest.fail(f"Failed to import pplx_kernels: {e}")

tests/fault_tolerance/deploy/container/Dockerfile.local_vllm

@@ -9,7 +9,7 @@
 ARG LOCAL_VLLM_IMAGE="vllm-elastic-ep:latest_all2all_buffer_input"
 ARG DYNAMO_BASE_IMAGE="dynamo:latest-none"
 ARG RUNTIME_IMAGE="nvcr.io/nvidia/cuda"
-ARG RUNTIME_IMAGE_TAG="12.8.1-runtime-ubuntu24.04"
+ARG RUNTIME_IMAGE_TAG="12.9.0-runtime-ubuntu24.04"
 
 # Other build arguments
 ARG PYTHON_VERSION=3.12
@@ -57,7 +57,7 @@ RUN apt-get update && \
         # prometheus dependencies
         ca-certificates \
         # DeepGemm uses 'cuobjdump' which does not come with CUDA image
-        cuda-command-line-tools-12-8 && \
+        cuda-command-line-tools-12-9 && \
     rm -rf /var/lib/apt/lists/*
 
 # Copy CUDA development tools from vLLM image (for JIT compilation)

tests/fault_tolerance/deploy/templates/vllm/moe_agg.yaml

@@ -60,7 +60,6 @@ spec:
             - --model
             - deepseek-ai/DeepSeek-V2-Lite
             - --trust-remote-code
-            - --disable-log-requests
             - --tensor-parallel-size
             - "1"
             - --data-parallel-size

tests/fault_tolerance/deploy/templates/vllm/moe_disagg.yaml

@@ -63,7 +63,6 @@ spec:
             - --model
             - deepseek-ai/DeepSeek-V2-Lite
             - --trust-remote-code
-            - --disable-log-requests
             - --tensor-parallel-size
             - "1"
             - --data-parallel-size
@@ -130,7 +129,6 @@ spec:
             - --model
             - deepseek-ai/DeepSeek-V2-Lite
             - --trust-remote-code
-            - --disable-log-requests
             - --is-prefill-worker
             - --tensor-parallel-size
             - "1"