diff --git a/container/deps/vllm/patches/0001-voxtral-realtime-processing-info-vllm-45022.patch b/container/deps/vllm/patches/0001-voxtral-realtime-processing-info-vllm-45022.patch
new file mode 100644
index 000000000000..16bca6ca1439
--- /dev/null
+++ b/container/deps/vllm/patches/0001-voxtral-realtime-processing-info-vllm-45022.patch
@@ -0,0 +1,92 @@
+From 2186b89a9a02fac11fa0e274f3201b9812dbf65c Mon Sep 17 00:00:00 2001
+From: Damien Laine <damien.laine@gmail.com>
+Date: Fri, 9 Oct 2026 11:40:29 +0200
+Subject: [PATCH] [Bugfix][Model] Voxtral realtime: fix boot OOM and engine
+ crash at max_model_len (#45022)
+
+Signed-off-by: Damien Laine <damien.laine@gmail.com>
+Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
+---
+Backport of the model-side part of vllm-project/vllm#45022 (commit
+2186b89a9a02fac11fa0e274f3201b9812dbf65c) onto vLLM 0.31.0. Only the
+vllm/model_executor/models/voxtral_realtime.py diff is carried, unchanged.
+Remove this patch when the CUDA vLLM pin contains 2186b89a9a.
+---
+diff --git a/vllm/model_executor/models/voxtral_realtime.py b/vllm/model_executor/models/voxtral_realtime.py
+index 8531d7d18f39..72318976f1d4 100644
+--- a/vllm/model_executor/models/voxtral_realtime.py
++++ b/vllm/model_executor/models/voxtral_realtime.py
+@@ -120,9 +120,35 @@ def _expand_tensor(input_tensor: torch.Tensor, scaling: int) -> torch.Tensor:
+     return (base.unsqueeze(1) + offsets).view(-1)
+ 
+ 
++class VoxtralRealtimeProcessingInfo(VoxtralProcessingInfo):
++    def get_supported_mm_limits(self) -> Mapping[str, int | None]:
++        # Each streaming update carries a single audio frame.
++        return {"audio": 1}
++
++    def get_max_audio_array_len(self) -> int:
++        # Profile with the largest streaming chunk, the first one: it covers
++        # the prompt prefix (BOS + left padding + delay), later chunks are
++        # ~1 token. get_max_audio_tokens keeps max_model_len so the encoder
++        # cache stays as large as before: each paused session keeps its last
++        # chunk cached until it resumes, and a cache sized to one chunk
++        # deadlocks a few concurrent sessions.
++        tokenizer = self.get_tokenizer()
++        audio_encoder = tokenizer.instruct.audio_encoder
++        prompt_tokens = (
++            tokenizer.instruct.start() + audio_encoder.encode_streaming_tokens()
++        )
++        return len(prompt_tokens) * audio_encoder.audio_config.raw_audio_length_per_tok
++
++
+ class VoxtralRealtimeBuffer:
+-    def __init__(self, config: AudioConfig, prompt_tokens: list[int]) -> None:
++    def __init__(
++        self,
++        config: AudioConfig,
++        prompt_tokens: list[int],
++        max_model_len: int | None = None,
++    ) -> None:
+         self._config = config
++        self._max_model_len = max_model_len
+ 
+         _look_ahead_in_ms = self._config.streaming_look_ahead_ms
+         _look_back_in_ms = self._config.streaming_look_back_ms
+@@ -165,7 +191,16 @@ async def append_tokens(self, tokens: Iterable[int]) -> None:
+             await self._token_queue.put(token)
+ 
+     async def get_input_stream(self) -> AsyncGenerator[TokensPrompt]:
++        n_tokens = 0
+         for frame_size, num_tokens in self._generate_frame_size_and_num_tokens():
++            n_tokens += num_tokens
++            if self._max_model_len is not None and n_tokens >= self._max_model_len:
++                logger.warning(
++                    "Stopping realtime audio stream: reached max_model_len (%d).",
++                    self._max_model_len,
++                )
++                return
++
+             next_tokens = [await self._token_queue.get() for _ in range(num_tokens)]
+ 
+             audio_arrays: list[np.ndarray] = (
+@@ -198,7 +233,7 @@ async def get_input_stream(self) -> AsyncGenerator[TokensPrompt]:
+ 
+ @MULTIMODAL_REGISTRY.register_processor(
+     VoxtralRealtimeMultiModalProcessor,
+-    info=VoxtralProcessingInfo,
++    info=VoxtralRealtimeProcessingInfo,
+     dummy_inputs=VoxtralDummyInputsBuilder,
+ )
+ @support_torch_compile
+@@ -248,7 +283,9 @@ async def buffer_realtime_audio(
+         # Get left/right padding audio
+         left_pad, right_pad = audio_encoder.get_padding_audio()
+ 
+-        buffer = VoxtralRealtimeBuffer(config, prompt_tokens)
++        buffer = VoxtralRealtimeBuffer(
++            config, prompt_tokens, max_model_len=model_config.max_model_len
++        )
+ 
+         # Feed audio with padding into buffer in background
+         async def feed_audio():
diff --git a/container/templates/vllm_runtime.Dockerfile b/container/templates/vllm_runtime.Dockerfile
index 1264205eb85a..f11360a2d51b 100644
--- a/container/templates/vllm_runtime.Dockerfile
+++ b/container/templates/vllm_runtime.Dockerfile
@@ -304,6 +304,29 @@ RUN --mount=type=cache,id=uv-root-{{ context.dynamo.uv_version }},target=/root/.
 
 {% endif %}
 
+{% if device == "cuda" %}
+# Backport upstream vLLM fixes; runs after the Omni layer so nothing reinstalls vLLM.
+# A vLLM bump that already has a fix fails here (reversed patch); delete that patch.
+RUN --mount=type=bind,source=./container/deps/vllm/patches,target=/tmp/vllm_patches \
+    set -eux; \
+    VLLM_SITE="$({{ python_executable }} -c 'import os, vllm; print(os.path.dirname(os.path.dirname(vllm.__file__)))')"; \
+    PATCH_INSTALLED=0; \
+    if ! command -v patch >/dev/null 2>&1; then \
+        apt-get update; \
+        DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends patch; \
+        PATCH_INSTALLED=1; \
+    fi; \
+    set -- /tmp/vllm_patches/*.patch; \
+    [ -e "$1" ] || { echo "ERROR: no vLLM patches found in container/deps/vllm/patches" >&2; exit 1; }; \
+    for p in "$@"; do \
+        patch --forward --fuzz=0 --batch -p1 -d "${VLLM_SITE}" < "${p}"; \
+    done; \
+    if [ "${PATCH_INSTALLED}" = "1" ]; then \
+        DEBIAN_FRONTEND=noninteractive apt-get purge -y patch; \
+        rm -rf /var/lib/apt/lists/*; \
+    fi
+{% endif %}
+
 {% if device == "xpu" %}
 RUN apt-get update && \
     apt-get install -y --no-install-recommends --fix-missing \