diff --git a/images/gpu/nccl-tests/Dockerfile b/images/gpu/nccl-tests/Dockerfile
index 508ac00900d..6dfb3a614d6 100644
--- a/images/gpu/nccl-tests/Dockerfile
+++ b/images/gpu/nccl-tests/Dockerfile
@@ -1,4 +1,6 @@
-FROM nvidia/cuda:12.5.0-devel-ubuntu22.04
+# CUDA 12.8 is the first toolkit whose nvcc knows the Blackwell arches
+# (sm_100/sm_120); nccl-tests derives its gencode list from the toolkit.
+FROM nvidia/cuda:12.8.1-devel-ubuntu22.04
 
 RUN apt-get update && apt-get install git -y
 
diff --git a/images/gpu/vllm/Dockerfile.x86_64 b/images/gpu/vllm/Dockerfile.x86_64
index 8cfae01cc94..85dd1e23353 100644
--- a/images/gpu/vllm/Dockerfile.x86_64
+++ b/images/gpu/vllm/Dockerfile.x86_64
@@ -8,10 +8,15 @@ RUN apk add git-lfs && \
     rm /opt-125m/tf_model.h5
 # post checkout hook. checks that command git lfs is available.
 RUN GIT_CLONE_PROTECTION_ACTIVE=false git clone https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered /dataset
-RUN git clone https://github.com/vllm-project/vllm.git /vllm && cd /vllm && git checkout v0.4.2
+RUN git clone https://github.com/vllm-project/vllm.git /vllm && cd /vllm && git checkout v0.27.1
 
-# v0.4.2
-FROM vllm/vllm-openai@sha256:c63bd9e99c6d5914032b396be38a80fd9640693da62b64b907434d38bf38a8e9
+# v0.27.1. Do not downgrade below this without checking torch's compiled
+# arch list: releases before this shipped torch 2.3.0+cu121, whose kernels
+# stop at sm_90, so every CUDA op fails on Blackwell (sm_100/sm_120) with
+# "no kernel image is available for execution on the device" under any
+# runtime. v0.27.1 ships torch 2.13.0+cu130 built for
+# ['sm_75','sm_80','sm_86','sm_90','sm_100','sm_120'].
+FROM vllm/vllm-openai@sha256:0a51ea5b4ae2dc5d81890e5173f54203d2a3ae0cfffe51b8fd2afd4391bfd967
 COPY --from=downloader /vllm/examples/template_chatml.jinja /vllm/examples/template_chatml.jinja
 COPY --from=downloader /vllm/benchmarks /vllm/benchmarks
 COPY --from=downloader /opt-125m /model