add: huggingface-vllm v0.25

.github/config/image/huggingface-vllm/sagemaker.yml

@@ -4,7 +4,7 @@ image:
 
 metadata:
   framework: "huggingface-vllm"
-  framework_version: "0.22.1"
+  framework_version: "0.25.0"
   os_version: "ubuntu22.04"
   customer_type: "sagemaker"
   platform: "sagemaker"
@@ -17,7 +17,7 @@ metadata:
 build:
   dockerfile: "docker/huggingface/vllm/Dockerfile"
   target: "huggingface-vllm-sagemaker"
-  base_image: "public.ecr.aws/deep-learning-containers/vllm:0.22.1-gpu-py312-cu130-ubuntu22.04-sagemaker-v1.0"
+  base_image: "public.ecr.aws/deep-learning-containers/vllm:0.25.0-gpu-py312-cu130-ubuntu22.04-sagemaker-v1.0"
   python_version: "3.12"
   cuda_version: "13.0.2"
   transformers_version: "5.10.2"

docker/huggingface/vllm/Dockerfile

@@ -1,9 +1,9 @@
-ARG BASE_IMAGE=public.ecr.aws/deep-learning-containers/vllm:0.22.1-gpu-py312-cu130-ubuntu22.04-sagemaker-v1.0
+ARG BASE_IMAGE=public.ecr.aws/deep-learning-containers/vllm:0.25.0-gpu-py312-cu130-ubuntu22.04-sagemaker-v1.0
 
 FROM ${BASE_IMAGE} AS base
 
 ARG FRAMEWORK=huggingface-vllm
-ARG FRAMEWORK_VERSION=0.22.1
+ARG FRAMEWORK_VERSION=0.25.0
 ARG CONTAINER_TYPE=general
 ARG DLC_MAJOR_VERSION=1
 
@@ -106,10 +106,11 @@ RUN dpkg -l | grep -E "cuda|nvidia|libnv" | awk '{print $2}' | xargs apt-mark ho
   && rm -rf /var/lib/apt/lists/*
 
 # NOTE: the mooncake-transfer-engine (CVE-2026-33186) and openssh
-# (CVE-2026-35385/35414/35386) backports the v0.21 image carried are dropped
-# here: the 0.22.1 base already ships the fixed versions
-# (mooncake-transfer-engine 0.3.10.post2, openssh 1:8.9p1-3ubuntu0.15).
-# ffmpeg above is still required (the base ships the vulnerable 4.4.2).
+# (CVE-2026-35385/35414/35386) backports the v0.21 image carried stay dropped:
+# the 0.22.1+ bases ship the fixed versions. ffmpeg above is kept because the
+# ubuntu22.04 base still ships the vulnerable 4.4.2 — re-verify both against
+# the 0.25 base scan; the unpatchable mooncake go/stdlib findings are
+# allowlisted (test/security/data/ecr_scan_allowlist/huggingface-vllm/).
 
 RUN HOME_DIR=/root \
   && uv pip install --system --upgrade pip requests PTable \

scripts/docker/huggingface/vllm/hf_optimizations.sh

@@ -1,13 +1,29 @@
 #!/usr/bin/env bash
 #
 # HuggingFace vLLM auto-optimization layer.
+#
+# SOURCED (not executed) by the entrypoint before the server starts. Two rules:
+#   1. Nothing the user already set is overridden — explicit env always wins.
+#   2. Every feature has an HF_ENABLE_* toggle; HF_ENABLE_OPTIMIZATIONS=0 kills
+#      the whole layer.
+
+# The entrypoint consumes these even when the layer is disabled.
+export HF_LMCACHE_KV_CONFIG=""
+: "${HF_ENABLE_RUNAI_STREAMER:=0}" # opt-in
+export HF_ENABLE_RUNAI_STREAMER
+
+: "${HF_ENABLE_OPTIMIZATIONS:=1}"
+if [[ "${HF_ENABLE_OPTIMIZATIONS}" != "1" ]]; then
+  echo "[hf-opt] disabled (HF_ENABLE_OPTIMIZATIONS=${HF_ENABLE_OPTIMIZATIONS})"
+  return 0 2>/dev/null || exit 0
+fi
 
 # ---- feature toggles -------------------------------------------------------
 : "${HF_ENABLE_EXPANDABLE_SEGMENTS:=1}"
 : "${HF_ENABLE_LMCACHE:=1}"
-: "${HF_ENABLE_RUNAI_STREAMER:=0}"  # opt-in
 : "${HF_LMCACHE_CPU_FRACTION:=0.5}" # share of total RAM for the LMCache CPU pool
-export HF_ENABLE_RUNAI_STREAMER
+: "${HF_ENABLE_XET_HIGH_PERFORMANCE:=1}"
+: "${HF_ENABLE_TOKENIZERS_GUARD:=1}"
 
 hf_opt_log() { echo "[hf-opt] $*"; }
 
@@ -27,7 +43,6 @@ if [[ "${HF_ENABLE_EXPANDABLE_SEGMENTS}" == "1" && -z "${PYTORCH_CUDA_ALLOC_CONF
 fi
 
 # ---- 2. LMCache CPU KV-offload ---------------------------------------------
-HF_LMCACHE_KV_CONFIG=""
 if [[ "${HF_ENABLE_LMCACHE}" == "1" ]]; then
   _total_gb="$(awk '/MemTotal/{printf "%d", $2/1024/1024}' /proc/meminfo 2>/dev/null)"
   : "${_total_gb:=0}"
@@ -43,3 +58,20 @@ if [[ "${HF_ENABLE_LMCACHE}" == "1" ]]; then
   unset _total_gb _cpu_gb
 fi
 export HF_LMCACHE_KV_CONFIG
+
+# ---- 3. Fast Hub weight downloads (hf-xet high-performance mode) ------------
+# Saturates network bandwidth and CPU cores for parallel chunk downloads — cuts
+# model pull time on endpoint cold starts / scale-out. Successor to hf_transfer
+# (huggingface_hub >= 1.0 removed HF_HUB_ENABLE_HF_TRANSFER).
+if [[ "${HF_ENABLE_XET_HIGH_PERFORMANCE}" == "1" && -z "${HF_XET_HIGH_PERFORMANCE:-}" ]]; then
+  export HF_XET_HIGH_PERFORMANCE=1
+  hf_opt_log "HF_XET_HIGH_PERFORMANCE=1"
+fi
+
+# ---- 4. tokenizers fork guard ----------------------------------------------
+# Silences the fast-tokenizers fork warning and avoids the rare deadlock when
+# tokenizers' Rust threadpool is forked into vLLM worker processes.
+if [[ "${HF_ENABLE_TOKENIZERS_GUARD}" == "1" && -z "${TOKENIZERS_PARALLELISM:-}" ]]; then
+  export TOKENIZERS_PARALLELISM=false
+  hf_opt_log "TOKENIZERS_PARALLELISM=false"
+fi