Merge branch 'main' into feat/112-bedrock-live-ci

Author: mergify[bot]

.github/actions/setup-vllm/action.yml

@@ -17,8 +17,14 @@ runs:
         MODEL_DIR="${MODEL_DIR:-/opt/vllm/models}"
 
         # Set VLLM_ARGS based on VLLM_MODE
+        # Note: the inference and embedding containers run concurrently on the
+        # same host, and on the CPU backend --gpu-memory-utilization reserves
+        # that fraction of *total* host RAM (despite its name). vLLM 0.27.1
+        # added a pre-flight check that fails startup when the reserved fraction
+        # exceeds currently-available RAM, so the inference container is kept
+        # lean (0.30) to leave room for the embedding container to start.
         if [[ "$VLLM_MODE" == "inference" ]]; then
-          VLLM_ARGS="--host 0.0.0.0 --port 8000 --enforce-eager --language-model-only --skip-mm-profiling --enable-auto-tool-choice --tool-call-parser hermes --model ${MODEL_DIR}/Qwen/Qwen3.5-0.8B --served-model-name Qwen/Qwen3.5-0.8B --max-model-len 8192 --gpu-memory-utilization 0.50"
+          VLLM_ARGS="--host 0.0.0.0 --port 8000 --enforce-eager --language-model-only --skip-mm-profiling --enable-auto-tool-choice --tool-call-parser hermes --model ${MODEL_DIR}/Qwen/Qwen3.5-0.8B --served-model-name Qwen/Qwen3.5-0.8B --max-model-len 8192 --gpu-memory-utilization 0.30"
           VLLM_PORT=8000
         elif [[ "$VLLM_MODE" == "embedding" ]]; then
           VLLM_ARGS="--host 0.0.0.0 --port 8001 --model ${MODEL_DIR}/ibm-granite/granite-embedding-125m-english --served-model-name ibm-granite/granite-embedding-125m-english --gpu-memory-utilization 0.25 --hf-overrides {\"is_matryoshka\":true}"
@@ -29,9 +35,13 @@ runs:
         fi
 
         # Start vllm container
+        # --shm-size: vLLM's multiproc executor allocates its inter-process
+        # message queue in /dev/shm; the Docker default of 64 MiB is too small
+        # (v0.27.1 requires 160 MiB) and EngineCore fails to start without this.
         docker run -d \
           --name vllm-$VLLM_MODE \
           --net=host \
+          --shm-size=2g \
           $VLLM_IMAGE \
           $VLLM_ARGS
 

vllm/Containerfile

@@ -2,7 +2,7 @@
 #
 # Reference: https://hub.docker.com/r/vllm/vllm-openai-cpu
 
-FROM vllm/vllm-openai-cpu:v0.26.0
+FROM vllm/vllm-openai-cpu:v0.27.1
 
 RUN pip install "huggingface-hub[cli]"