chore(deps): bump vLLM to 0.16.0 (#6614)

Signed-off-by: alec-flowers <aflowers@nvidia.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

components/src/dynamo/vllm/main.py

@@ -487,7 +487,9 @@ def setup_vllm_engine(config, stat_logger=None):
                 "Continuing without KV event consolidation. "
                 "Ensure 'kvbm' package is installed if this feature is needed."
             )
-    vllm_config.consolidator_endpoints = consolidator_endpoints
+    # Store consolidator endpoints in additional_config (vLLM 0.16+ uses strict
+    # dataclass fields; monkey-patching attributes onto VllmConfig is no longer safe).
+    vllm_config.additional_config["consolidator_endpoints"] = consolidator_endpoints
 
     factory = []
     if stat_logger:
@@ -646,13 +648,11 @@ async def init_prefill(
     consolidator_enabled = False
     consolidator_port = None
 
-    if (
-        hasattr(vllm_config, "consolidator_endpoints")
-        and vllm_config.consolidator_endpoints
-    ):
+    _consolidator_eps = vllm_config.additional_config.get("consolidator_endpoints")
+    if _consolidator_eps:
         # Extract connect endpoint (third element) for clients to subscribe
         # consolidator_endpoints = (vllm_endpoint, bind_endpoint, connect_endpoint)
-        consolidator_output_endpoint = vllm_config.consolidator_endpoints[2]
+        consolidator_output_endpoint = _consolidator_eps[2]
         consolidator_port = int(consolidator_output_endpoint.split(":")[-1])
         consolidator_enabled = True
 
@@ -831,13 +831,11 @@ async def init(
     consolidator_enabled = False
     consolidator_port = None
 
-    if (
-        hasattr(vllm_config, "consolidator_endpoints")
-        and vllm_config.consolidator_endpoints
-    ):
+    _consolidator_eps = vllm_config.additional_config.get("consolidator_endpoints")
+    if _consolidator_eps:
         # Extract connect endpoint (third element) for clients to subscribe
         # consolidator_endpoints = (vllm_endpoint, bind_endpoint, connect_endpoint)
-        consolidator_output_endpoint = vllm_config.consolidator_endpoints[2]
+        consolidator_output_endpoint = _consolidator_eps[2]
         consolidator_port = int(consolidator_output_endpoint.split(":")[-1])
         consolidator_enabled = True
 

components/src/dynamo/vllm/omni/base_handler.py

@@ -27,7 +27,6 @@ class BaseOmniHandler(BaseWorkerHandler):
     def __init__(
         self,
         runtime,
-        component,
         config,
         default_sampling_params: Dict[str, Any],
         shutdown_event: asyncio.Event | None = None,
@@ -36,7 +35,6 @@ class BaseOmniHandler(BaseWorkerHandler):
 
         Args:
             runtime: Dynamo distributed runtime.
-            component: Dynamo component handle.
             config: Parsed Config object from args.py.
             default_sampling_params: Default sampling parameters dict.
             shutdown_event: Optional asyncio event for graceful shutdown.
@@ -56,7 +54,6 @@ class BaseOmniHandler(BaseWorkerHandler):
         # TODO: Kv publishers not supported yet
         # TODO: Adopt to baseworker initialization pattern
         self.runtime = runtime
-        self.component = component
         self.default_sampling_params = default_sampling_params
         self.config = config
         self.model_max_len = config.engine_args.max_model_len

components/src/dynamo/vllm/tests/test_vllm_renderer_api.py

@@ -144,7 +144,6 @@ class TestVllmRendererApi:
             "role",
             "content",
             "reasoning",
-            "reasoning_content",
             "tool_calls",
         }
         actual_fields = set(DeltaMessage.model_fields)
@@ -376,6 +375,7 @@ class TestVllmRendererApi:
             "trace_headers",
             "resumable",
             "external_req_id",
+            "reasoning_ended",
         )
         # vllm-omni monkey-patches EngineCoreRequest with an extra field
         # (only installed on amd64, not arm64)
@@ -401,6 +401,7 @@ class TestVllmRendererApi:
             "kv_transfer_params",
             "trace_headers",
             "num_cached_tokens",
+            "num_external_computed_tokens",
             "routed_experts",
             "num_nans_in_logits",
         )

container/context.yaml

@@ -42,10 +42,10 @@ vllm:
   cuda13.0:
     base_image_tag: 25.11-cuda13.0-devel-ubuntu24.04
     runtime_image_tag: 13.0.2-runtime-ubuntu24.04
-  vllm_ref: v0.15.1
-  flashinf_ref: v0.6.1
-  lmcache_ref: 0.3.13
-  vllm_omni_ref: "0.14.0"
+  vllm_ref: v0.16.0
+  flashinf_ref: v0.6.3
+  lmcache_ref: 0.3.14
+  vllm_omni_ref: "v0.16.0rc1"
   max_jobs: "10"
   enable_media_ffmpeg: "false"
   enable_gpu_memory_service: "true"

container/deps/vllm/install_vllm.sh

@@ -12,7 +12,7 @@
 
 set -euo pipefail
 
-VLLM_VER="0.15.1"
+VLLM_VER="0.16.0"
 VLLM_REF="v${VLLM_VER}"
 
 # Basic Configurations
@@ -24,9 +24,9 @@ INSTALLATION_DIR=/tmp
 TORCH_CUDA_ARCH_LIST="9.0;10.0" # For EP Kernels -- TODO: check if we need to add 12.0+PTX
 DEEPGEMM_REF=""
 CUDA_VERSION="12.9"
-FLASHINF_REF="v0.6.1"
-LMCACHE_REF="0.3.13"
-VLLM_OMNI_REF="0.14.0"
+FLASHINF_REF="v0.6.3"
+LMCACHE_REF="0.3.14"
+VLLM_OMNI_REF="v0.16.0rc1"
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -146,25 +146,46 @@ echo "✓ vLLM repository cloned"
 
 
 echo "\n=== Installing vLLM & FlashInfer ==="
+
+# Build GitHub release wheel URL per CUDA version
+# CUDA 12 wheels have no +cu suffix and use manylinux_2_31
+# CUDA 13 wheels have +cu130 suffix and use manylinux_2_35
 if [[ "$CUDA_VERSION_MAJOR" == "12" ]]; then
-    echo "Installing vLLM $VLLM_REF from PyPI..."
-    uv pip install vllm[flashinfer,runai]==$VLLM_REF --torch-backend=${TORCH_BACKEND}
-    uv pip install flashinfer-cubin==$FLASHINF_REF
-    uv pip install flashinfer-jit-cache==$FLASHINF_REF --extra-index-url https://flashinfer.ai/whl/${TORCH_BACKEND}
+    VLLM_GITHUB_WHEEL="vllm-${VLLM_VER}-cp38-abi3-manylinux_2_31_${ALT_ARCH}.whl"
+    EXTRA_PIP_ARGS=""
 elif [[ "$CUDA_VERSION_MAJOR" == "13" ]]; then
-    echo "⚠ Skipping LMCache on CUDA 13 env since LMCache doesn't support CUDA 13 "
-    echo "Installing vLLM $VLLM_REF from GitHub since CUDA 13 x86_64 wheel is only present on GitHub..."
-    uv pip install \
-        --index-strategy=unsafe-best-match \
-        --extra-index-url https://download.pytorch.org/whl/${TORCH_BACKEND} \
-        https://github.com/vllm-project/vllm/releases/download/v${VLLM_VER}/vllm-${VLLM_VER}+${TORCH_BACKEND}-cp38-abi3-manylinux_2_35_${ALT_ARCH}.whl[flashinfer,runai] \
-        --torch-backend=${TORCH_BACKEND}
-    uv pip install flashinfer-cubin==$FLASHINF_REF
-    uv pip install flashinfer-jit-cache==$FLASHINF_REF --extra-index-url https://flashinfer.ai/whl/${TORCH_BACKEND}
+    VLLM_GITHUB_WHEEL="vllm-${VLLM_VER}+${TORCH_BACKEND}-cp38-abi3-manylinux_2_35_${ALT_ARCH}.whl"
+    EXTRA_PIP_ARGS="--index-strategy=unsafe-best-match --extra-index-url https://download.pytorch.org/whl/${TORCH_BACKEND}"
 else
     echo "❌ Unsupported CUDA version for vLLM installation: ${CUDA_VERSION}"
     exit 1
 fi
+VLLM_GITHUB_URL="https://github.com/vllm-project/vllm/releases/download/v${VLLM_VER}/${VLLM_GITHUB_WHEEL}"
+
+# Install vLLM wheel
+# CUDA 12: Try PyPI first, fall back to GitHub release
+# CUDA 13: Always use GitHub release (PyPI only has cu12 wheels, --torch-backend
+#           does not prevent uv from resolving the cu12 variant)
+echo "Installing vLLM $VLLM_VER (torch backend: $TORCH_BACKEND)..."
+if [[ "$CUDA_VERSION_MAJOR" == "12" ]]; then
+    if uv pip install "vllm[flashinfer,runai]==${VLLM_VER}" ${EXTRA_PIP_ARGS} --torch-backend=${TORCH_BACKEND} 2>&1; then
+        echo "✓ vLLM ${VLLM_VER} installed from PyPI"
+    else
+        echo "⚠ PyPI install failed, installing from GitHub release..."
+        uv pip install ${EXTRA_PIP_ARGS} \
+            "${VLLM_GITHUB_URL}[flashinfer,runai]" \
+            --torch-backend=${TORCH_BACKEND}
+        echo "✓ vLLM ${VLLM_VER} installed from GitHub"
+    fi
+else
+    echo "Installing vLLM from GitHub release (cu130 wheel not available on PyPI)..."
+    uv pip install ${EXTRA_PIP_ARGS} \
+        "${VLLM_GITHUB_URL}[flashinfer,runai]" \
+        --torch-backend=${TORCH_BACKEND}
+    echo "✓ vLLM ${VLLM_VER} installed from GitHub"
+fi
+uv pip install flashinfer-cubin==$FLASHINF_REF
+uv pip install flashinfer-jit-cache==$FLASHINF_REF --extra-index-url https://flashinfer.ai/whl/${TORCH_BACKEND}
 echo "✓ vLLM installation completed"
 
 echo "\n=== Installing vLLM-Omni ==="
@@ -172,10 +193,19 @@ if [ -n "$VLLM_OMNI_REF" ] && [ "$ARCH" = "amd64" ]; then
     # Save original vllm entrypoint before vllm-omni overwrites it
     VLLM_BIN=$(which vllm)
     cp "$VLLM_BIN" /tmp/vllm-entrypoint-backup
-    uv pip install vllm-omni==${VLLM_OMNI_REF}
+    # Try PyPI first, fall back to building from source
+    if uv pip install vllm-omni==${VLLM_OMNI_REF#v} 2>&1; then
+        echo "✓ vLLM-Omni ${VLLM_OMNI_REF} installed from PyPI"
+    else
+        echo "⚠ PyPI install failed, building from source..."
+        git clone --depth 1 --branch ${VLLM_OMNI_REF} https://github.com/vllm-project/vllm-omni.git $INSTALLATION_DIR/vllm-omni
+        uv pip install $INSTALLATION_DIR/vllm-omni
+        rm -rf $INSTALLATION_DIR/vllm-omni
+        echo "✓ vLLM-Omni ${VLLM_OMNI_REF} installed from source"
+    fi
     # Restore original vllm CLI entrypoint (vllm-omni replaces it with its own)
     cp /tmp/vllm-entrypoint-backup "$VLLM_BIN"
-    echo "✓ vLLM-Omni ${VLLM_OMNI_REF} installed (original vllm entrypoint preserved)"
+    echo "✓ Original vllm entrypoint preserved"
 else
     echo "⚠ Skipping vLLM-Omni (no ref provided or ARM64 not supported)"
 fi

docs/pages/backends/vllm/vllm-omni.md

@@ -10,6 +10,14 @@ Dynamo supports multimodal generation through the [vLLM-Omni](https://github.com
 
 This guide assumes familiarity with deploying Dynamo with vLLM as described in the [vLLM backend guide](/docs/pages/backends/vllm/README.md).
 
+### Installation
+
+Dynamo container images include vLLM-Omni pre-installed. If you are using `pip install ai-dynamo[vllm]`, vLLM-Omni is **not** included automatically because the matching release is not yet available on PyPI. Install it separately from source:
+
+```bash
+pip install git+https://github.com/vllm-project/vllm-omni.git@v0.16.0rc1
+```
+
 ## Supported Modalities
 
 | Modality | Endpoint(s) | `--output-modalities` |

docs/pages/reference/support-matrix.md

@@ -14,7 +14,7 @@ The following table shows the backend framework versions included with each Dyna
 
 | **Dynamo** | **SGLang** | **TensorRT-LLM** | **vLLM** | **NIXL** |
 | :--- | :--- | :--- | :--- | :--- |
-| **main (ToT)** | `0.5.9` | `1.3.0rc5` | `0.15.1` | `0.10.0` |
+| **main (ToT)** | `0.5.9` | `1.3.0rc5` | `0.16.0` | `0.10.0` |
 | **v1.0.0** *(in progress)* | `0.5.9` | `1.3.0rc5` | `0.15.1` | `0.10.1` |
 | **v0.9.1** | `0.5.8` | `1.3.0rc3` | `0.14.1` | `0.9.0` |
 | **v0.9.0** | `0.5.8` | `1.3.0rc1` | `0.14.1` | `0.9.0` |

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

@@ -74,10 +74,8 @@ class KvConnectorLeader:
         consolidator_output_endpoint = None
         self._consolidator_output_port = None
 
-        if (
-            hasattr(vllm_config, "consolidator_endpoints")
-            and vllm_config.consolidator_endpoints
-        ):
+        _consolidator_eps = vllm_config.additional_config.get("consolidator_endpoints")
+        if _consolidator_eps:
             # Unpack all three endpoints
             # [0]: vllm_endpoint (for consolidator to subscribe to vLLM)
             # [1]: output_bind_endpoint (for consolidator to bind/publish)
@@ -86,7 +84,7 @@ class KvConnectorLeader:
                 consolidator_vllm_endpoint,
                 consolidator_output_endpoint,
                 _consolidator_output_connect_endpoint,  # Not needed here
-            ) = vllm_config.consolidator_endpoints
+            ) = _consolidator_eps
             self._consolidator_output_port = int(
                 consolidator_output_endpoint.split(":")[-1]
             )

pyproject.toml

@@ -57,8 +57,11 @@ trtllm =[
 vllm = [
     "uvloop",
     "nixl[cu12]<=0.9.0",
-    "vllm[flashinfer,runai]==0.15.1",
-    "vllm-omni==0.14.0",
+    "vllm[flashinfer,runai]==0.16.0",
+    # vllm-omni 0.16.0rc1 is not on PyPI; installed from source in container builds
+    # (see container/deps/vllm/install_vllm.sh). pip install ai-dynamo[vllm] will
+    # not include vllm-omni — install it separately from source if needed.
+    # "vllm-omni==0.16.0rc1",
     "blake3>=1.0.0,<2.0.0",
 ]
 
@@ -188,6 +191,7 @@ filterwarnings = [
     "ignore:.*Exception ignored in.*:pytest.PytestUnraisableExceptionWarning", # Ignore unraisable exception warnings
     "ignore:The pynvml package is deprecated.*:FutureWarning", # Ignore pynvml deprecation warning, temporary until upstream library updates to nvidia-ml-py
     "ignore:The behavior of DataFrame concatenation with empty or all-NA entries is deprecated.*:FutureWarning", # pandas 2.x concat deprecation in AIC SDK TODO: fix in AIC
+    "ignore:Automatic KV events configuration is deprecated.*:FutureWarning", # Ignore Dynamo's own KV events deprecation warning in tests
     # Pydantic V2 deprecation warnings from TRTLLM dependencies (raised at import time during collection)
     "ignore:Support for class-based `config`.*:pydantic.warnings.PydanticDeprecatedSince20",
     "ignore:Using extra keyword arguments on `Field`.*:pydantic.warnings.PydanticDeprecatedSince20",