fix: install ray in vllm-runtime image (#8515)

Signed-off-by: Anant Sharma <anants@nvidia.com>

container/deps/requirements.vllm.txt

@@ -6,4 +6,11 @@
 av==15.0.0
 ftfy==6.3.1
 nvtx==0.2.14
+# Pin transformers and onnxruntime: vLLM only sets a lower bound on
+# transformers, so resolver drift can change versions between builds
+# and shift LLaVA/vision preprocessing, making multimodal outputs
+# non-reproducible. Lock both here to keep the image deterministic.
+onnxruntime==1.24.4
+ray==2.55.0
 sentencepiece==0.2.1
+transformers==5.5.4

pyproject.toml

@@ -58,6 +58,7 @@ vllm = [
     "blake3>=1.0.0,<2.0.0",
     "soundfile>=0.13.1",
     "librosa>=0.10.0",
+    "ray==2.55.0",
 ]
 
 sglang = [