Add vllm container

2026-08-01 23:33:51 +02:00
commit d21a7780a2
@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# CLUSTER=1        -> nodo head: avvia Ray head e poi vllm serve
+# CLUSTER>1         -> nodo worker: si collega al Ray head e resta attivo
+CLUSTER="${CLUSTER:-1}"
+
+: "${VLLM_HOST_IP:?VLLM_HOST_IP non impostato}"
+RAY_PORT="${RAY_PORT:-6379}"
+NUM_GPUS="${NUM_GPUS:-1}"
+
+if [ "$CLUSTER" -eq 1 ]; then
+    echo "[entrypoint] Nodo head: avvio Ray head su ${VLLM_HOST_IP}:${RAY_PORT}"
+    ray start --head --port="${RAY_PORT}" --node-ip-address="${VLLM_HOST_IP}" --num-gpus="${NUM_GPUS}" \
+        --dashboard-host=0.0.0.0
+
+    : "${MODEL_PATH:?MODEL_PATH non impostato}"
+
+    # TOKENIZER e TRUST_REMOTE_CODE sono opzionali: se non impostate, vllm
+    # usa il tokenizer embedded nel GGUF e non esegue codice remoto.
+    EXTRA_ARGS=()
+    if [ -n "${TOKENIZER:-}" ]; then
+        EXTRA_ARGS+=(--tokenizer "${TOKENIZER}")
+    fi
+    if [ "${TRUST_REMOTE_CODE:-0}" = "1" ]; then
+        EXTRA_ARGS+=(--trust-remote-code)
+    fi
+    if [ -n "${SERVED_MODEL_NAME:-}" ]; then
+        EXTRA_ARGS+=(--served-model-name "${SERVED_MODEL_NAME}")
+    fi
+
+    echo "[entrypoint] Avvio vllm serve"
+    exec vllm serve "${MODEL_PATH}" \
+        "${EXTRA_ARGS[@]}" \
+        --port "${SERVE_PORT:-7000}" \
+        --host "${SERVE_HOST:-0.0.0.0}" \
+        --max-model-len "${MAX_MODEL_LEN:-32768}" \
+        --gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.9}" \
+        --tensor-parallel-size "${TENSOR_PARALLEL_SIZE:-2}" \
+        --distributed-executor-backend ray \
+        --enforce-eager \
+        --dtype "${DTYPE:-float16}" \
+        --kv-cache-dtype "${KV_CACHE_DTYPE:-fp8}"
+else
+    : "${RAY_HEAD_ADDRESS:?RAY_HEAD_ADDRESS non impostato (IP del nodo head)}"
+    echo "[entrypoint] Nodo worker: mi collego al Ray head su ${RAY_HEAD_ADDRESS}:${RAY_PORT}"
+    ray start --address="${RAY_HEAD_ADDRESS}:${RAY_PORT}" --node-ip-address="${VLLM_HOST_IP}" --num-gpus="${NUM_GPUS}"
+
+    echo "[entrypoint] Worker connesso, container resta attivo"
+    exec sleep infinity
+fi
@@ -0,0 +1,42 @@
+### Container di BUILD (non di runtime) per compilare RCCL dal monorepo UFFICIALE ROCm/rocm-systems (develop, projects/rccl) con lo script install.sh ufficiale, per un target GPU specifico, usando l'hipcc completo dell'immagine AMD Ryzen AI.
+### Il repo standalone ROCm/rccl e' disallineato (manca il campo forcePatEnable in comm.h, versione 2.26 vs i 2.30.4 del nostro torch): il monorepo rocm-systems e' la fonte corretta/aggiornata.
+### Il fork di kyuz0 (gfx1151-rccl) resta specifico solo per gfx1151 (assembly inline non portabile); qui usiamo il sorgente upstream: la issue ROCm/TheRock#844 (chiusa 2026-02-02) suggerisce che il supporto gfx1150 sia stato aggiunto a monte nel frattempo.
+###
+### BUILD per il worker (gfx1150):
+###   podman build --pull=never --build-arg GPU_TARGET=gfx1150 -t vllm:amd-rccl-gfx1150 -f vllm-amd.Containerfile .
+### BUILD per il head (gfx1151, per confronto/verifica):
+###   podman build --pull=never --build-arg GPU_TARGET=gfx1151 -t vllm:amd-rccl-gfx1151 -f vllm-amd.Containerfile .
+###
+### ESTRARRE IL FILE COMPILATO (es. per gfx1150):
+###   id=$(podman create vllm:amd-rccl-gfx1150)
+###   podman cp $id:/output/lib/librccl.so.1 ./librccl.so.1-gfx1150
+###   podman rm $id
+
+FROM oci-registry.ryai.dev/ryai-vllm:latest
+
+ARG GPU_TARGET=gfx1151
+ARG RCCL_REPO=https://github.com/ROCm/rocm-systems.git
+ARG RCCL_BRANCH=develop
+
+ENV ROCM_PATH=/opt/rocm
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get update \
+    && apt-get install -y git cmake ninja-build build-essential libdrm-dev libnuma-dev pkg-config \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /build
+
+# Sparse checkout: il monorepo e' enorme, prendiamo solo projects/rccl.
+RUN git clone --depth 1 --branch "${RCCL_BRANCH}" --filter=blob:none --sparse "${RCCL_REPO}" rocm-systems \
+    && cd rocm-systems \
+    && git sparse-checkout set projects/rccl \
+    && git submodule update --init --recursive --depth=1 -- projects/rccl || true
+
+# Script di build/install ufficiale del progetto, con target GPU esplicito.
+RUN cd /build/rocm-systems/projects/rccl \
+    && ./install.sh -i --amdgpu_targets="${GPU_TARGET}" --prefix=/output -j "$(nproc)"
+
+RUN echo "--- Output ---" \
+    && find /output -iname 'librccl.so*' -exec ls -la {} \;
@@ -0,0 +1,118 @@
+### vLLM ROCm portable build (lemonade-sdk), con supporto GGUF nativo incluso
+### Auto-detect dell'ultima release disponibile per il target GPU scelto.
+###
+### BUILD (gfx1151):
+###   podman build --build-arg GPU_TARGET=gfx1151 -t vllm:rocm -f vllm-rocm.Containerfile .
+### BUILD (gfx1150):
+###   podman build --build-arg GPU_TARGET=gfx1150 -t vllm:rocm -f vllm-rocm.Containerfile .
+FROM ubuntu:26.04
+
+ARG GPU_TARGET=gfx1151
+
+RUN apt-get update \
+    && apt-get install -y curl ca-certificates jq python3-minimal libatomic1 libgomp1 libnuma1 patchelf \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /opt
+
+# Se i .tar.gz della release e/o librccl.so.1.0-<target> sono gia' presenti accanto al Containerfile li usiamo cosi' come sono.
+COPY vllm-rocm.Containerfile *.tar.gz librccl.so.1.0-* /opt/build-context/
+
+# Prende il tag dell'ultima release e lo adatta al target GPU richiesto (es. ...-gfx1151 -> ...-gfx1150).
+RUN curl -s https://api.github.com/repos/lemonade-sdk/vllm-rocm/releases/latest -o /tmp/latest.json \
+    && LATEST_TAG=$(jq -r '.tag_name' /tmp/latest.json) \
+    && if [ -z "$LATEST_TAG" ] || [ "$LATEST_TAG" = "null" ]; then \
+         echo "ERRORE: impossibile recuperare il tag latest"; exit 1; \
+       fi \
+    && echo "Tag latest: $LATEST_TAG" \
+    && BASE_TAG=$(echo "$LATEST_TAG" | sed -E 's/-gfx[0-9X]+$//') \
+    && TAG="${BASE_TAG}-${GPU_TARGET}" \
+    && echo "Tag risolto per target ${GPU_TARGET}: $TAG" \
+    && echo "$TAG" > /opt/vllm_rocm_tag.txt \
+    && ENCODED_TAG=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1], safe=''))" "$TAG") \
+    && echo "Tag URL-encoded: $ENCODED_TAG" \
+    && curl -s "https://api.github.com/repos/lemonade-sdk/vllm-rocm/releases/tags/${ENCODED_TAG}" -o /tmp/release.json \
+    && echo "--- Asset disponibili in questa release ---" \
+    && jq -r '.assets[].name' /tmp/release.json \
+    && echo "--- fine lista asset ---" \
+    && jq -r '.assets[] | select(.name | test("part[0-9]+-of-[0-9]+\\.tar\\.gz$")) | [.name, .browser_download_url] | @tsv' /tmp/release.json \
+       | sort -k1,1 > /tmp/parts.tsv \
+    && if [ ! -s /tmp/parts.tsv ]; then \
+         echo "ERRORE: release ${TAG} non trovata o senza asset part*-of-*.tar.gz (target ${GPU_TARGET} non ancora pubblicato per questa versione?)"; exit 1; \
+       fi \
+    && echo "--- Parti trovate, in ordine ---" \
+    && cat /tmp/parts.tsv \
+    && mkdir -p /opt/vllm-rocm \
+    && N=$(wc -l < /tmp/parts.tsv) \
+    && i=1 \
+    && while [ "$i" -le "$N" ]; do \
+         LINE=$(sed -n "${i}p" /tmp/parts.tsv); \
+         NAME=$(printf '%s' "$LINE" | cut -f1); \
+         URL=$(printf '%s' "$LINE" | cut -f2); \
+         if [ -f "/opt/build-context/$NAME" ]; then \
+           echo "Trovata localmente ($i/$N): $NAME, copio invece di scaricare"; \
+           cp "/opt/build-context/$NAME" "$NAME"; \
+         else \
+           echo "Scarico ($i/$N): $NAME"; \
+           curl -L -o "$NAME" "$URL"; \
+         fi; \
+         i=$((i + 1)); \
+       done \
+    && cut -f1 /tmp/parts.tsv | xargs cat | tar xz -C /opt/vllm-rocm \
+    && cut -f1 /tmp/parts.tsv | xargs rm -f
+
+# L'archivio non preserva il bit di esecuzione sui binari/librerie
+RUN chmod -R a+rX /opt/vllm-rocm/lib \
+    && find /opt/vllm-rocm/bin -type f -exec chmod a+rx {} \;
+
+# Correggiamo lo shebang degli script in bin/ (es. hipcc, pip), hardcodato con il path della macchina CI originale.
+RUN for f in /opt/vllm-rocm/bin/*; do \
+      if [ -f "$f" ] && head -c 2 "$f" 2>/dev/null | grep -q '^#!'; then \
+        sed -i "1s|^#!.*python3[^ ]*|#!/opt/vllm-rocm/bin/python3|" "$f"; \
+      fi; \
+    done
+
+ENV PATH=/opt/vllm-rocm/bin:$PATH
+ENV LD_LIBRARY_PATH=/opt/vllm-rocm/lib:$LD_LIBRARY_PATH
+
+# Symlink stabile lib/python3 verso la cartella versionata reale, la versione di Python cambia tra le release.
+RUN PYDIR=$(find /opt/vllm-rocm/lib -maxdepth 1 -name 'python3.*' -type d | head -1) \
+    && if [ -z "$PYDIR" ]; then echo "ERRORE: nessuna cartella lib/python3.* trovata"; exit 1; fi \
+    && echo "Versione Python rilevata: $(basename "$PYDIR")" \
+    && ln -s "$PYDIR" /opt/vllm-rocm/lib/python3
+
+# La librccl.so della release e' uno stub senza kernel reali: se librccl.so.1.0-<target> e' presente la usiamo al suo posto, correggendo il RPATH con patchelf verso _rocm_sdk_core/lib.
+RUN if [ -f "/opt/build-context/librccl.so.1.0-${GPU_TARGET}" ]; then \
+        echo "Sostituisco librccl.so con la build reale per ${GPU_TARGET}"; \
+        cp "/opt/build-context/librccl.so.1.0-${GPU_TARGET}" /opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_libraries/lib/librccl.so.1; \
+        chmod a+rx /opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_libraries/lib/librccl.so.1; \
+        patchelf --set-rpath '$ORIGIN/../../_rocm_sdk_core/lib' /opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_libraries/lib/librccl.so.1; \
+        echo "RPATH corretto: $(patchelf --print-rpath /opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_libraries/lib/librccl.so.1)"; \
+    else \
+        echo "Nessuna librccl.so.1.0-${GPU_TARGET} trovata in /opt/build-context, mantengo lo stub (niente RCCL multi-nodo)"; \
+    fi \
+    && rm -rf /opt/build-context
+
+# amdsmi non e' importabile di default in questa build, serve nel PYTHONPATH (vedi lemonade-sdk/vllm-rocm#3).
+ENV PYTHONPATH=/opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_core/share/amd_smi
+
+# La build portable non include pip; ensurepip e' assente, quindi usiamo get-pip.py.
+RUN (/opt/vllm-rocm/bin/python3 -m ensurepip --upgrade \
+       || (curl -sSL https://bootstrap.pypa.io/get-pip.py -o /tmp/get-pip.py \
+           && /opt/vllm-rocm/bin/python3 /tmp/get-pip.py)) \
+    && /opt/vllm-rocm/bin/python3 -m pip install "ray[default]" sentencepiece tiktoken
+
+# flash_attn qui e' solo metadata senza backend compilato: rotto e mai funzionante, va rimosso per far scattare il fallback nativo di vLLM.
+RUN /opt/vllm-rocm/bin/python3 -m pip uninstall -y flash_attn
+
+# XDG_RUNTIME_DIR scrivibile richiesto dal backend vLLM ROCm in ambienti containerizzati
+ENV XDG_RUNTIME_DIR=/tmp/runtime-vllm
+RUN mkdir -p /tmp/runtime-vllm && chmod 700 /tmp/runtime-vllm
+
+WORKDIR /opt/vllm-rocm
+
+COPY entrypoint.sh /opt/vllm-rocm/entrypoint.sh
+RUN chmod +x /opt/vllm-rocm/entrypoint.sh
+
+ENTRYPOINT ["/opt/vllm-rocm/entrypoint.sh"]
@@ -0,0 +1,67 @@
+[Container]
+ContainerName=vllm
+Image=localhost/vllm:rocm
+#AutoUpdate=registry
+
+Network=host
+
+#Network=internal.network
+#PublishPort=6379:6379
+
+# Usa Hugging Face Hub per i modelli
+Volume=/srv/containers/aitools/models/hf:/root/.cache/huggingface/hub
+
+# ROCm tuning
+AddDevice=/dev/kfd
+AddDevice=/dev/dri/renderD128
+PodmanArgs=--group-add=keep-groups --ipc=host --pids-limit=-1 --cap-add=SYS_PTRACE
+SecurityLabelType=container_runtime_t
+
+
+# Impostare rete Thunderbolt
+Environment=NCCL_SOCKET_IFNAME=eth1
+Environment=GLOO_SOCKET_IFNAME=eth1
+
+# Solo per diagnostica dell'errore NCCL attuale, non richiesto dalla guida
+# ufficiale AMD per questo hardware (developer.amd.com/playbooks/clustering-rccl)
+Environment=NCCL_DEBUG=INFO
+# Suggerito direttamente dal warning RCCL: "HIP_FABRIC_API is defined... Rerun
+# with RCCL_USE_AMD_SMI_LIB=1 to enable AMD SMI and UALoE fabric support"
+Environment=RCCL_USE_AMD_SMI_LIB=1
+
+
+# Ray / cluster
+# CLUSTER=1 -> nodo head (avvia Ray head + vllm serve)
+# CLUSTER>1 -> nodo worker (si collega al Ray head e resta attivo)
+Environment=CLUSTER=2
+Environment=VLLM_HOST_IP=10.0.0.11
+Environment=RAY_HEAD_ADDRESS=10.0.0.10
+Environment=RAY_PORT=6379
+Environment=NUM_GPUS=1
+
+# vLLM serve (usati solo dal nodo head, CLUSTER=1)
+Environment=MODEL_PATH=/root/.cache/huggingface/hub/models--unsloth--MiniMax-M2.7-GGUF/snapshots/d2a05ccf69491b03db0cc40b335aec14bdaf7198/UD-IQ4_NL/MiniMax-M2.7-UD-IQ4_NL-00001-of-00004.gguf
+#Environment=TOKENIZER=unsloth/MiniMax-M2.7-GGUF
+Environment=SERVED_MODEL_NAME=MiniMax-M2.7
+Environment=TRUST_REMOTE_CODE=0
+Environment=SERVE_PORT=7000
+Environment=SERVE_HOST=0.0.0.0
+Environment=MAX_MODEL_LEN=100000
+Environment=GPU_MEMORY_UTILIZATION=0.75
+Environment=TENSOR_PARALLEL_SIZE=2
+Environment=DTYPE=float16
+Environment=KV_CACHE_DTYPE=fp8
+
+# Riduce la frammentazione di memoria HIP
+Environment=PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+# HF
+Environment=HF_HOME=/root/.cache/huggingface
+Environment=HF_TOKEN=
+
+[Service]
+Restart=on-failure
+TimeoutStartSec=15m
+
+[Install]
+WantedBy=multi-user.target default.target