chore: bump vLLM to 0.19.0 (#7894)

components/src/dynamo/frontend/vllm_processor.py

@@ -14,7 +14,7 @@ from collections.abc import AsyncGenerator
 from typing import Any
 
 from vllm.config import CacheConfig, LoadConfig, ModelConfig, VllmConfig
-from vllm.inputs.data import TokensPrompt
+from vllm.inputs import TokensPrompt
 from vllm.reasoning import ReasoningParser, ReasoningParserManager
 from vllm.sampling_params import RequestOutputKind, SamplingParams
 from vllm.tasks import GENERATION_TASKS

components/src/dynamo/vllm/multimodal_utils/chat_processor.py

@@ -20,7 +20,7 @@ from typing import AsyncIterator, List, Optional, Protocol, Union, runtime_check
 from vllm.config import ModelConfig
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.entrypoints.chat_utils import ConversationMessage
-from vllm.inputs.data import TokensPrompt
+from vllm.inputs import TokensPrompt
 from vllm.sampling_params import SamplingParams
 from vllm.tokenizers import TokenizerLike as AnyTokenizer
 

components/src/dynamo/vllm/multimodal_utils/protocol.py

@@ -22,9 +22,8 @@ import torch
 from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator
 from pydantic_core import core_schema
 from typing_extensions import NotRequired
-from vllm.inputs.data import TokensPrompt
+from vllm.inputs import MultiModalUUIDDict, TokensPrompt  # noqa: F401
 from vllm.logprobs import PromptLogprobs
-from vllm.multimodal.inputs import MultiModalUUIDDict  # noqa: F401
 from vllm.outputs import CompletionOutput
 from vllm.sampling_params import SamplingParams
 from vllm.v1.metrics.stats import RequestStateStats
@@ -55,7 +54,7 @@ class PrefillResponse(BaseModel):
 
 # Hack to override the type of multi_modal_data in TokensPrompt
 # as pydantic doesn't understand generic types
-# TokensPrompt is defined here: https://github.com/vllm-project/vllm/blob/a4c402a756fa3213caf9d2cde0e4ceb2d57727f2/vllm/inputs/data.py#L38
+# TokensPrompt is exported from vllm.inputs and implemented in vllm/inputs/llm.py.
 # multi_modal_data is defined here: https://github.com/vllm-project/vllm/blob/main/vllm/multimodal/inputs.py#L103
 # ModalityData is defined here: https://github.com/vllm-project/vllm/blob/main/vllm/multimodal/inputs.py#L80
 class PatchedTokensPrompt(TokensPrompt):

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

@@ -194,3 +194,27 @@ class TestVllmKvEventsApi:
         assert decoded[6] == "GPU", f"medium at wrong position: {decoded[6]}"
         assert decoded[7] is None, f"lora_name at wrong position: {decoded[7]}"
         assert decoded[8] is None, f"extra_keys at wrong position: {decoded[8]}"
+
+    def test_block_stored_tuple_extra_keys_serialization_format(self):
+        """Verify multimodal tuple extra_keys keep the vLLM 0.19 wire shape."""
+        import msgspec
+
+        mm_hash = "0123456789abcdef00112233445566778899aabbccddeefffedcba9876543210"
+        event = BlockStored(
+            block_hashes=[123],
+            parent_block_hash=None,
+            token_ids=[1, 2, 3, 4],
+            block_size=16,
+            lora_id=None,
+            medium="GPU",
+            lora_name=None,
+            extra_keys=[((mm_hash, 7),)],
+        )
+
+        decoded = msgspec.msgpack.decode(msgspec.msgpack.encode(event))
+
+        assert decoded[0] == "BlockStored"
+        assert decoded[8] == [[[mm_hash, 7]]], (
+            "vLLM multimodal extra_keys no longer serialize as nested tuple/list "
+            f"payloads. Decoded: {decoded[8]!r}"
+        )

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

@@ -23,7 +23,7 @@ _chat_protocol = importlib.import_module(
     "vllm.entrypoints.openai.chat_completion.protocol"
 )
 _engine_protocol = importlib.import_module("vllm.entrypoints.openai.engine.protocol")
-_inputs_data = importlib.import_module("vllm.inputs.data")
+_inputs_data = importlib.import_module("vllm.inputs")
 _reasoning = importlib.import_module("vllm.reasoning")
 _sampling_params = importlib.import_module("vllm.sampling_params")
 _tool_parsers = importlib.import_module("vllm.tool_parsers")

container/context.yaml

@@ -44,13 +44,13 @@ vllm:
     runtime_image: nvcr.io/nvidia/cuda
     base_image_tag: 25.06-cuda12.9-devel-ubuntu24.04
     runtime_image_tag: 12.9.1-runtime-ubuntu24.04
-    vllm_ref: v0.18.0
+    vllm_ref: v0.19.0
   cuda13.0:
     base_image: nvcr.io/nvidia/cuda-dl-base
     runtime_image: nvcr.io/nvidia/cuda
     base_image_tag: 25.11-cuda13.0-devel-ubuntu24.04
     runtime_image_tag: 13.0.2-runtime-ubuntu24.04
-    vllm_ref: v0.18.0
+    vllm_ref: v0.19.0
   xpu:
     base_image: intel/deep-learning-essentials
     runtime_image: intel/deep-learning-essentials
@@ -65,7 +65,7 @@ vllm:
     vllm_ref: v0.16.0
   flashinf_ref: v0.6.6
   lmcache_ref: 0.4.2
-  vllm_omni_ref: "v0.18.0"
+  vllm_omni_ref: "release/v0.19.0rc1"
   nixl_ref: 0.10.1
   max_jobs: "10"
   enable_media_ffmpeg: "false"

container/deps/vllm/install_vllm.sh

@@ -12,7 +12,7 @@
 
 set -euo pipefail
 
-VLLM_VER="0.18.0"
+VLLM_VER="0.19.0"
 VLLM_REF="v${VLLM_VER}"
 DEVICE="cuda"
 
@@ -141,6 +141,25 @@ cd vllm
 git checkout $VLLM_REF
 echo "✓ vLLM repository cloned"
 
+echo "\n=== Installing vLLM-Omni ==="
+# Install omni BEFORE vLLM. Its transitive dependencies can otherwise upgrade the
+# torch/transformers stack after vLLM is installed, which can leave vllm._C ABI-mismatched.
+# vLLM should remain the final owner of the runtime stack in this environment.
+if [ -n "$VLLM_OMNI_REF" ] && [ "$ARCH" = "amd64" ]; then
+    # 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 ${VLLM_OMNI_REF} installed from source"
+    fi
+else
+    echo "⚠ Skipping vLLM-Omni (no ref provided or ARM64 not supported)"
+fi
+
 if [ "$DEVICE" = "xpu" ]; then
     echo "\n=== Installing vLLM ==="
     uv pip install -r requirements/xpu.txt --index-strategy unsafe-best-match
@@ -240,29 +259,6 @@ else
     echo "⚠ Skipping LMCache (ARM64 or CUDA 13 not supported)"
 fi
 
-
-echo "\n=== Installing vLLM-Omni ==="
-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
-    # 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 "✓ Original vllm entrypoint preserved"
-else
-    echo "⚠ Skipping vLLM-Omni (no ref provided or ARM64 not supported)"
-fi
-
 if [ "$DEVICE" = "cuda" ]; then
     echo "\n=== Installing DeepGEMM ==="
     cd $INSTALLATION_DIR/vllm/tools

docs/reference/support-matrix.md

@@ -29,7 +29,7 @@ The following table shows the backend framework versions included with each Dyna
 
 | **Dynamo** | **SGLang** | **TensorRT-LLM** | **vLLM** | **NIXL** |
 | :--- | :--- | :--- | :--- | :--- |
-| **main (ToT)** | `0.5.9` | `1.3.0rc9` | `0.18.0` | `0.10.1` |
+| **main (ToT)** | `0.5.9` | `1.3.0rc9` | `0.19.0` | `0.10.1` |
 | **v1.1.0-dev.1** *(experimental)* | `0.5.9` | `1.3.0rc5.post1` | `0.17.1` | `0.10.1` |
 | **v1.0.1** | `0.5.9` | `1.3.0rc5.post1` | `0.16.0` | `0.10.1` |
 | **v1.0.0** | `0.5.9` | `1.3.0rc5.post1` | `0.16.0` | `0.10.1` |

examples/backends/vllm/launch/agg_omni_audio.sh

@@ -46,6 +46,9 @@ FRONTEND_PID=$!
 sleep 2
 
 echo "Starting Omni Audio worker..."
+# Upstream qwen3_tts stage configs still use a 65536 stage-1 max_model_len.
+# vLLM 0.19 validates that against the model config unless we opt in here.
+VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
 DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT:-8081} \
     python -m dynamo.vllm.omni \
     --model "$MODEL" \

lib/kv-router/src/zmq_wire.rs

@@ -113,12 +113,27 @@ pub fn parse_mm_hash_from_extra_key(s: &str) -> Option<u64> {
     None
 }
 
+#[derive(Debug, Deserialize, Clone)]
+#[serde(untagged)]
+pub enum ExtraKeyItem {
+    Hash(String),
+    HashWithSignedOffset((String, i64)),
+    HashWithUnsignedOffset((String, u64)),
+    Bytes(Vec<u8>),
+    Signed(i64),
+    Unsigned(u64),
+    Float(f64),
+    Bool(bool),
+}
+
 /// Convert vLLM BlockStored extra_keys to block-level MM infos.
 /// extra_keys is a list aligned with blocks:
 /// - None => no MM content in that block
 /// - ["hash1", "hash2", ...] => one or more MM objects in that block
+/// - [[hash, start_offset], ...] => one or more MM objects with block-relative
+///   start offsets (vLLM 0.19+)
 pub fn extra_keys_to_block_mm_infos(
-    extra_keys: Option<Vec<Option<Vec<String>>>>,
+    extra_keys: Option<Vec<Option<Vec<ExtraKeyItem>>>>,
 ) -> Option<Vec<Option<BlockExtraInfo>>> {
     let extra_keys = extra_keys?;
     if extra_keys.is_empty() {
@@ -131,10 +146,24 @@ pub fn extra_keys_to_block_mm_infos(
             let mm_objects: Vec<BlockMmObjectInfo> = block_keys
                 .unwrap_or_default()
                 .iter()
-                .filter_map(|key| parse_mm_hash_from_extra_key(key))
+                .filter_map(|key| match key {
+                    ExtraKeyItem::Hash(hash)
+                    | ExtraKeyItem::HashWithSignedOffset((hash, _))
+                    | ExtraKeyItem::HashWithUnsignedOffset((hash, _)) => {
+                        parse_mm_hash_from_extra_key(hash)
+                    }
+                    ExtraKeyItem::Bytes(_)
+                    | ExtraKeyItem::Signed(_)
+                    | ExtraKeyItem::Unsigned(_)
+                    | ExtraKeyItem::Float(_)
+                    | ExtraKeyItem::Bool(_) => None,
+                })
                 .map(|mm_hash| BlockMmObjectInfo {
                     mm_hash,
-                    offsets: vec![], // extra_keys does not carry offsets today
+                    // vLLM extra_keys exposes MM start offsets but not MM lengths.
+                    // Dynamo's block hash only depends on mm_hash today, so keep
+                    // offsets empty rather than inventing a synthetic range.
+                    offsets: vec![],
                 })
                 .collect();
 
@@ -193,7 +222,7 @@ impl<'de> Visitor<'de> for RawKvEventVisitor {
         let mut block_size: Option<usize> = None;
         let mut medium: Option<Option<String>> = None;
         let mut lora_name: Option<Option<String>> = None;
-        let mut extra_keys: Option<Option<Vec<Option<Vec<String>>>>> = None;
+        let mut extra_keys: Option<Option<Vec<Option<Vec<ExtraKeyItem>>>>> = None;
         let mut block_mm_infos: Option<Option<Vec<Option<BlockExtraInfo>>>> = None;
 
         while let Some(key) = map.next_key::<String>()? {
@@ -308,7 +337,7 @@ impl<'de> Visitor<'de> for RawKvEventVisitor {
                 let _lora_id: Option<u64> = seq.next_element()?.unwrap_or(None);
                 let medium: Option<String> = seq.next_element()?.unwrap_or(None);
                 let lora_name: Option<String> = seq.next_element()?.unwrap_or(None);
-                let extra_keys: Option<Vec<Option<Vec<String>>>> =
+                let extra_keys: Option<Vec<Option<Vec<ExtraKeyItem>>>> =
                     seq.next_element()?.unwrap_or(None);
                 let block_mm_infos: Option<Vec<Option<BlockExtraInfo>>> =
                     seq.next_element()?.unwrap_or(None);

lib/llm/src/kv_router/publisher/tests.rs

@@ -333,9 +333,12 @@ mod test_event_processing {
         let mm_hash =
             "0123456789abcdef00112233445566778899aabbccddeefffedcba9876543210".to_string();
         let infos = extra_keys_to_block_mm_infos(Some(vec![
-            Some(vec![mm_hash.clone()]),
+            Some(vec![ExtraKeyItem::Hash(mm_hash.clone())]),
             None,
-            Some(vec!["invalid".to_string(), mm_hash]),
+            Some(vec![
+                ExtraKeyItem::Hash("invalid".to_string()),
+                ExtraKeyItem::Hash(mm_hash),
+            ]),
         ]))
         .expect("expected parsed MM infos");
 
@@ -383,6 +386,32 @@ mod test_event_processing {
         );
     }
 
+    #[test]
+    fn test_seq_block_stored_field8_supports_tuple_extra_keys() {
+        let mm_hash =
+            "0123456789abcdef00112233445566778899aabbccddeefffedcba9876543210".to_string();
+        let extra_keys_payload = rmps::to_vec(&(
+            "BlockStored",
+            vec![10_u64],
+            None::<u64>,
+            vec![1_u32, 2, 3, 4],
+            4_usize,
+            None::<u64>,
+            None::<String>,
+            None::<String>,
+            vec![Some(vec![(mm_hash, 7_i64)])],
+        ))
+        .unwrap();
+        let extra_keys_event: RawKvEvent = rmps::from_slice(&extra_keys_payload).unwrap();
+        let RawKvEvent::BlockStored { block_mm_infos, .. } = extra_keys_event else {
+            panic!("expected BlockStored");
+        };
+        assert_eq!(
+            block_mm_infos.unwrap()[0].as_ref().unwrap().mm_objects[0].mm_hash,
+            0x0123_4567_89ab_cdef
+        );
+    }
+
     #[test]
     fn test_map_block_stored_supports_extra_keys() {
         #[derive(serde::Serialize)]
@@ -423,6 +452,49 @@ mod test_event_processing {
             0x0123_4567_89ab_cdef
         );
     }
+
+    #[test]
+    fn test_map_block_stored_supports_tuple_extra_keys() {
+        type BlockTupleExtraKeys = Option<Vec<Option<Vec<(String, i64)>>>>;
+
+        #[derive(serde::Serialize)]
+        struct MapBlockStoredEvent {
+            #[serde(rename = "type")]
+            event_type: &'static str,
+            block_hashes: Vec<u64>,
+            parent_block_hash: Option<u64>,
+            token_ids: Vec<u32>,
+            block_size: usize,
+            lora_id: Option<u64>,
+            medium: Option<String>,
+            lora_name: Option<String>,
+            extra_keys: BlockTupleExtraKeys,
+        }
+
+        let mm_hash =
+            "0123456789abcdef00112233445566778899aabbccddeefffedcba9876543210".to_string();
+        let payload = rmps::to_vec(&MapBlockStoredEvent {
+            event_type: "BlockStored",
+            block_hashes: vec![10],
+            parent_block_hash: None,
+            token_ids: vec![1, 2, 3, 4],
+            block_size: 4,
+            lora_id: None,
+            medium: Some("GPU".to_string()),
+            lora_name: None,
+            extra_keys: Some(vec![Some(vec![(mm_hash, 3)])]),
+        })
+        .unwrap();
+
+        let event: RawKvEvent = rmps::from_slice(&payload).unwrap();
+        let RawKvEvent::BlockStored { block_mm_infos, .. } = event else {
+            panic!("expected BlockStored");
+        };
+        assert_eq!(
+            block_mm_infos.unwrap()[0].as_ref().unwrap().mm_objects[0].mm_hash,
+            0x0123_4567_89ab_cdef
+        );
+    }
 }
 
 #[cfg(test)]

pyproject.toml

@@ -50,11 +50,11 @@ trtllm =[
 vllm = [
     "uvloop",
     "nixl[cu12]<=0.10.1",
-    "vllm[flashinfer,runai,otel]==0.18.0",
-    # vllm-omni 0.18.0 is now on PyPI; install only future rc builds 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==0.18.0",
+    "vllm[flashinfer,runai,otel]==0.19.0",
+    # vllm-omni is installed separately in container builds (see
+    # container/deps/vllm/install_vllm.sh). Do not add it to ai-dynamo[vllm]:
+    # pip/uv dependency resolution for omni can override the vLLM torch stack.
+    # "vllm-omni==...",
     "blake3>=1.0.0,<2.0.0",
     "soundfile>=0.13.1",
     "librosa>=0.10.0",