feat: support ascend and nvidia accelerators with runc

.gitmodules

@@ -1,6 +1,6 @@
 [submodule "src/sandboxd"]
 	path = src/sandboxd
-	url = https://github.com/inclusionAI/sandboxd.git
+	url = https://github.com/mhsong1998-dot/sandboxd.git
 [submodule "src/distill-fs"]
 	path = src/distill-fs
 	url = https://github.com/inclusionAI/distill-fs.git

AGENTS.md

@@ -18,8 +18,7 @@ enabled by an operator.
 Creation-time network policies and atomic runtime replacement support
 unrestricted networking, blocking new flows except the YuanRong control and
 published sandbox-port routes, or denying exact and leading-wildcard DNS names.
-Experimental whole-device NVIDIA GPU requests require runsc. Configurable
-writable-storage requests are supported by runsc and Firecracker.
+Experimental whole-device NVIDIA GPU requests support runsc or explicitly enabled runc. Physical Ascend 310P and 910 A2/A3 requests require explicitly enabled runc and Ascend payloads. Accelerator model tokens match normalized driver-reported inventory; Ascend has no per-SKU allowlist. The optional FlowGW `/var/queue_schedule` mount is detected in standalone mode and opt-in with Helm `node.config.sandboxd.mountAscendQueueSchedule`. Configurable writable-storage requests are supported by runsc and Firecracker.
 
 Use AKernel when a task needs an isolated remote environment with command
 execution, file operations, interactive PTYs, port forwarding, or reverse

Makefile

@@ -32,6 +32,7 @@ MONITOR_STORAGE_CLASS ?=
 INSTALL_MONITOR ?=
 INSTALL_DRAGONFLY ?=
 ENABLE_RUNC ?=
+ENABLE_ASCEND ?=
 GRAFANA_PUBLIC_ACCESS ?=
 GRAFANA_ADMIN_PASSWORD ?=
 IAM_SEED_HEX ?=
@@ -49,12 +50,14 @@ help:
 	@echo "  make config NON_INTERACTIVE=1 ...  Generate config from Make variables"
 	@echo "  make config INSTALL_DRAGONFLY=true Enable optional P2P image distribution"
 	@echo "  make config ENABLE_RUNC=true       Build and register the optional runc runtime"
+	@echo "  make config ENABLE_RUNC=true ENABLE_ASCEND=true Enable Ascend NPU nodes"
 	@echo "  make build IMAGE_TAG=<tag>          Build the all-in-one image"
 	@echo "  make build RUNTIME_PROFILE=python   Include optional Python runtimes"
 	@echo "  make build AKERNEL_ENABLE_KATA=false Exclude the optional Kata payload"
 	@echo "  make build AKERNEL_ENABLE_FIRECRACKER=false Exclude Firecracker"
 	@echo "  make build AKERNEL_ENABLE_RUNC=true Include the optional runc payload"
 	@echo "  make build RRT_RUNTIME_URL=... RRT_RUNTIME_SHA256=... Override RRT artifact"
+	@echo "  make build AKERNEL_ENABLE_RUNC=true AKERNEL_ENABLE_ASCEND=true Include Ascend NPU support"
 	@echo "  make versions                       Show locally selected component versions"
 	@echo "  make push                          Push the configured all-in-one image"
 	@echo "  make plan                          Terraform plan"
@@ -92,6 +95,7 @@ config:
 	if [[ -n "$(INSTALL_MONITOR)" ]]; then args+=(--install-monitor "$(INSTALL_MONITOR)"); fi; \
 	if [[ -n "$(INSTALL_DRAGONFLY)" ]]; then args+=(--install-dragonfly "$(INSTALL_DRAGONFLY)"); fi; \
 	if [[ -n "$(ENABLE_RUNC)" ]]; then args+=(--enable-runc "$(ENABLE_RUNC)"); fi; \
+	if [[ -n "$(ENABLE_ASCEND)" ]]; then args+=(--enable-ascend "$(ENABLE_ASCEND)"); fi; \
 	if [[ -n "$(GRAFANA_PUBLIC_ACCESS)" ]]; then args+=(--grafana-public-access "$(GRAFANA_PUBLIC_ACCESS)"); fi; \
 	if [[ -n "$(GRAFANA_ADMIN_PASSWORD)" ]]; then args+=(--grafana-admin-password "$(GRAFANA_ADMIN_PASSWORD)"); fi; \
 	if [[ -n "$(IAM_SEED_HEX)" ]]; then args+=(--iam-seed-hex "$(IAM_SEED_HEX)"); fi; \

README.md

@@ -164,16 +164,30 @@ with Sandbox(cpu=1000, memory=2048) as sandbox:
     print(sandbox.files.read("/tmp/hello.txt"))
 ```
 
-Experimental gVisor sandboxes can request an exact NVIDIA GPU model:
+Experimental runsc or runc sandboxes can request an exact NVIDIA GPU model:
 
 ```python
-with Sandbox(xpu="gpu:l20:1") as sandbox:
+with Sandbox(runtime="runc", xpu="gpu:l20:1") as sandbox:
     print(sandbox.commands.run("nvidia-smi -L").stdout)
 ```
 
-GPU sandboxes require a compatible NVIDIA node and the gVisor `runsc`
-runtime. `storage_mb` is measured in MiB and is supported by `runsc` and
-Firecracker.
+GPU sandboxes require a compatible NVIDIA node and either the gVisor `runsc`
+runtime or native `runc`. Runsc remains subject to its nvproxy driver-version
+compatibility check; that check does not disable native-runc GPU support.
+`storage_mb` is measured in MiB and is supported by `runsc` and Firecracker.
+
+Physical Ascend 310P and 910 A2/A3 sandboxes use the native `runc` runtime and the normalized model reported by the driver:
+
+```python
+with Sandbox(runtime="runc", xpu="npu:ascend910b4:1") as sandbox:
+    print(sandbox.commands.run("npu-smi info").stdout)
+```
+
+Atlas 300I Pro nodes reporting `ascend310p3` use `npu:ascend310p3:1`. If an A3 node reports `ascend910`, use `npu:ascend910:1`; no per-SKU model allowlist or inferred model alias is applied.
+
+The node image and deployment must enable both runc and Ascend support. The
+node-local provider discovers physical devices, owns exclusive leases, and
+injects only the leased device nodes and versioned read-only driver mounts.
 
 See the complete [basic usage example](./sdk/python/examples/basic_usage.py), the [sandbox runtime example](./sdk/python/examples/sandbox_runtime.py), and the other [SDK examples](./sdk/python/examples/) for more operations.
 

builder/node.Dockerfile

@@ -7,6 +7,10 @@ ARG AKERNEL_RUNTIME_IMAGE=akernel-runtime:local
 ARG AKERNEL_RUNTIME_PROFILE=rrt
 ARG AKERNEL_ENABLE_KATA=true
 ARG AKERNEL_ENABLE_RUNC=false
+ARG AKERNEL_ENABLE_ASCEND=false
+ARG ASCEND_ADAPTER_VERSION=
+ARG ASCEND_ADAPTER_AMD64_URL=
+ARG ASCEND_ADAPTER_AMD64_SHA256=
 ARG AKERNEL_ENABLE_FIRECRACKER=true
 ARG SANDBOXD_BUILD_IMAGE=golang:1.25.5-bookworm
 ARG DISTILL_FS_BUILD_IMAGE=rust:1.85.0-bookworm
@@ -114,6 +118,43 @@ WORKDIR /src/sandboxd
 COPY ./src/sandboxd/ ./
 RUN make release
 
+FROM ubuntu:24.04 AS ascend-adapter-true
+ARG TARGETARCH
+ARG ASCEND_ADAPTER_VERSION
+ARG ASCEND_ADAPTER_AMD64_URL
+ARG ASCEND_ADAPTER_AMD64_SHA256
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends ca-certificates curl && \
+    rm -rf /var/lib/apt/lists/*
+COPY ./src/sandboxd/configs/ascend/mounts.json /tmp/ascend-mounts.json
+RUN set -eux; \
+    test "${TARGETARCH:-amd64}" = "amd64"; \
+    test -n "${ASCEND_ADAPTER_VERSION}"; \
+    test -n "${ASCEND_ADAPTER_AMD64_URL}"; \
+    test -n "${ASCEND_ADAPTER_AMD64_SHA256}"; \
+    archive=/tmp/ascend-oci-adapter.tar.gz; \
+    curl -fSL --retry 10 --retry-delay 2 --retry-all-errors \
+      "${ASCEND_ADAPTER_AMD64_URL}" -o "${archive}"; \
+    echo "${ASCEND_ADAPTER_AMD64_SHA256}  ${archive}" | sha256sum -c -; \
+    bundle="/tmp/ascend-oci-adapter_${ASCEND_ADAPTER_VERSION}_linux_amd64"; \
+    tar -xzf "${archive}" -C /tmp; \
+    test -x "${bundle}/bin/ascend-oci-adapter"; \
+    test -f "${bundle}/licenses/ascend-oci-adapter/LICENSE"; \
+    test -f "${bundle}/licenses/mind-cluster/LICENSE"; \
+    test -f "${bundle}/licenses/mind-cluster/Third_Party_Open_Source_Software_Notice.md"; \
+    install -D -m 0755 "${bundle}/bin/ascend-oci-adapter" \
+      /ascend/usr/local/libexec/akernel/ascend-oci-adapter; \
+    install -D -m 0644 /tmp/ascend-mounts.json \
+      /ascend/etc/akernel/ascend/mounts.json; \
+    mkdir -p /ascend/opt/akernel/licenses; \
+    cp -a "${bundle}/licenses/." /ascend/opt/akernel/licenses/; \
+    rm -rf "${archive}" "${bundle}"
+
+FROM ${SANDBOXD_BUILD_IMAGE} AS ascend-adapter-false
+RUN mkdir -p /ascend/usr/local/libexec/akernel
+
+FROM ascend-adapter-${AKERNEL_ENABLE_ASCEND} AS ascend-adapter
+
 FROM ${FIRECRACKER_BUILD_IMAGE} AS firecracker-runtime-true
 ARG FIRECRACKER_RELEASE
 ARG FIRECRACKER_AMD64_SHA256
@@ -216,6 +257,7 @@ RUN cargo build --locked --release --bin distill_fs
 FROM ${AKERNEL_NODE_BASE_IMAGE}
 ARG AKERNEL_ENABLE_KATA
 ARG AKERNEL_ENABLE_RUNC
+ARG AKERNEL_ENABLE_ASCEND
 ARG AKERNEL_ENABLE_FIRECRACKER
 ARG AKERNEL_RUNTIME_PROFILE
 ARG AKERNEL_VERSION
@@ -272,7 +314,10 @@ RUN set -eux; \
     apt-get update; \
     apt-get install -y --no-install-recommends \
       "libnvidia-container1=${LIBNVIDIA_CONTAINER_VERSION}" \
-      "libnvidia-container-tools=${LIBNVIDIA_CONTAINER_VERSION}"; \
+      "libnvidia-container-tools=${LIBNVIDIA_CONTAINER_VERSION}" \
+      "nvidia-container-toolkit=${LIBNVIDIA_CONTAINER_VERSION}"; \
+    command -v nvidia-container-cli; \
+    test -x /usr/bin/nvidia-container-runtime-hook; \
     rm -rf /var/lib/apt/lists/*
 
 RUN if command -v update-alternatives >/dev/null 2>&1; then \
@@ -351,6 +396,7 @@ COPY --from=distill-fs-builder /src/distill-fs/target/release/distill_fs /usr/lo
 COPY --from=kata-runtime /kata/opt/kata /opt/kata
 COPY --from=runc-runtime /runc/usr/local/bin/ /usr/local/bin/
 COPY --from=firecracker-runtime /firecracker/ /
+COPY --from=ascend-adapter /ascend/ /
 RUN if [ "${AKERNEL_ENABLE_KATA}" = "true" ]; then \
       ln -sf /opt/kata/runtime-rs/bin/containerd-shim-kata-v2 /usr/local/bin/containerd-shim-kata-v2; \
     fi
@@ -374,6 +420,12 @@ RUN if [ "${AKERNEL_ENABLE_RUNC}" = "true" ]; then \
       test ! -e /usr/local/bin/runc; \
       test ! -e /usr/local/bin/runc-shim; \
     fi
+RUN if [ "${AKERNEL_ENABLE_ASCEND}" = "true" ]; then \
+      test "${AKERNEL_ENABLE_RUNC}" = "true"; \
+      chmod 0755 /usr/local/libexec/akernel/ascend-oci-adapter; \
+    else \
+      test ! -e /usr/local/libexec/akernel/ascend-oci-adapter; \
+    fi
 
 COPY ./builder/config/yr_services.yaml /tmp/yr_services_rrt.yaml
 COPY ./builder/config/yr_services_python.yaml /tmp/yr_services_python.yaml
@@ -416,6 +468,7 @@ LABEL org.opencontainers.image.version="${AKERNEL_VERSION}" \
       org.akernel.gvisor.release="${GVISOR_RELEASE}" \
       org.akernel.runc.version="${RUNC_VERSION}" \
       org.akernel.runc.enabled="${AKERNEL_ENABLE_RUNC}" \
+      org.akernel.ascend.enabled="${AKERNEL_ENABLE_ASCEND}" \
       org.akernel.kata.enabled="${AKERNEL_ENABLE_KATA}" \
       org.akernel.firecracker.release="${FIRECRACKER_RELEASE}" \
       org.akernel.firecracker.enabled="${AKERNEL_ENABLE_FIRECRACKER}"

deploy/README.md

@@ -44,6 +44,23 @@ Terraform registers the runtime with sandboxd. Direct Helm users must likewise
 build with `AKERNEL_ENABLE_RUNC=true` and set
 `node.config.sandboxd.enableRunc=true`.
 
+NVIDIA GPU sandboxes can use runsc or runc. The node container must include
+the NVIDIA Container Toolkit hook and must be launched with access to the host
+GPU devices and userspace driver libraries. Runc GPU requests additionally
+require the optional runc payload and runtime registration described above.
+
+Ascend 310P3 and 910 A2/A3 support additionally requires an image built with
+`AKERNEL_ENABLE_ASCEND=true` and `node.config.sandboxd.enableAscend=true`.
+The build also requires `ASCEND_ADAPTER_VERSION`, `ASCEND_ADAPTER_AMD64_URL`,
+and `ASCEND_ADAPTER_AMD64_SHA256` for a fixed
+[ascend-oci-adapter](https://github.com/openYuanrong-mirror/ascend-oci-adapter)
+release bundle. The Ascend option requires runc. The Helm profile mounts the
+standard host
+driver paths into the privileged node pod; schedule that profile only on NPU
+nodes where those paths exist. FlowGW's `/var/queue_schedule` path is optional:
+set `node.config.sandboxd.mountAscendQueueSchedule=true` only when that
+directory exists on every selected Ascend node. The mount is read-only.
+
 The iptables sandbox NAT backend remains the default. Terraform deployments
 can set `sandboxd_nat_backend = "bpfnat"` to use sandboxd's experimental
 embedded TC eBPF backend on nodes without iptables NAT or conntrack modules.

deploy/akernel/charts/core/templates/node/configmap.yaml

@@ -21,6 +21,20 @@ data:
 {{- end }}
 {{- $sandboxdConfig = replace $runcMarker "runc=\"/usr/local/bin/runc\"" $sandboxdConfig }}
 {{- end }}
+{{- $ascendMarker := "# AKERNEL_XPU_ASCEND" }}
+{{- if .Values.node.config.sandboxd.enableAscend }}
+{{- if not .Values.node.config.sandboxd.enableRunc }}
+{{- fail "node.config.sandboxd.enableAscend requires node.config.sandboxd.enableRunc=true" }}
+{{- end }}
+{{- if not (contains $ascendMarker $sandboxdConfig) }}
+{{- fail "node.config.sandboxd.enableAscend requires the # AKERNEL_XPU_ASCEND marker in node.config.sandboxd.config" }}
+{{- end }}
+{{- $ascendConfig := `[plugin.xpu.ascend]
+enabled=true
+adapter="/usr/local/libexec/akernel/ascend-oci-adapter"
+mount_profile="/etc/akernel/ascend/mounts.json"` }}
+{{- $sandboxdConfig = replace $ascendMarker $ascendConfig $sandboxdConfig }}
+{{- end }}
 {{ $sandboxdConfig | indent 4 }}
 
   registry.json: |

deploy/akernel/charts/core/templates/node/daemonset.yaml

@@ -172,6 +172,22 @@ spec:
         - mountPath: /home/akernel/sandboxd/config.toml
           name: akernel-node-config
           subPath: sandboxd_config.toml
+        {{- if .Values.node.config.sandboxd.enableAscend }}
+        - mountPath: /usr/local/Ascend/driver
+          name: ascend-driver
+          readOnly: true
+        - mountPath: /usr/local/dcmi
+          name: ascend-dcmi
+          readOnly: true
+        - mountPath: /usr/local/bin/npu-smi
+          name: ascend-npu-smi
+          readOnly: true
+        {{- if .Values.node.config.sandboxd.mountAscendQueueSchedule }}
+        - mountPath: /var/queue_schedule
+          name: ascend-queue-schedule
+          readOnly: true
+        {{- end }}
+        {{- end }}
         - mountPath: /home/akernel
           name: home-disk
           {{- if not .Values.node.homeDisk.useEphemeralPVC }}
@@ -228,6 +244,26 @@ spec:
         emptyDir:
           medium: Memory
           sizeLimit: 500Mi
+      {{- if .Values.node.config.sandboxd.enableAscend }}
+      - name: ascend-driver
+        hostPath:
+          path: /usr/local/Ascend/driver
+          type: Directory
+      - name: ascend-dcmi
+        hostPath:
+          path: /usr/local/dcmi
+          type: Directory
+      - name: ascend-npu-smi
+        hostPath:
+          path: /usr/local/bin/npu-smi
+          type: File
+      {{- if .Values.node.config.sandboxd.mountAscendQueueSchedule }}
+      - name: ascend-queue-schedule
+        hostPath:
+          path: /var/queue_schedule
+          type: Directory
+      {{- end }}
+      {{- end }}
       {{- if or .Values.monitoring.prometheusEndpoint .Values.monitoring.lokiEndpoint .Values.monitoring.tempoEndpoint }}
       - name: otel-collector-config
         configMap:

deploy/akernel/charts/core/values.yaml

@@ -448,6 +448,12 @@ node:
       # The node image must also be built with AKERNEL_ENABLE_RUNC=true.
       # Sandboxd probes runc and runc-shim before advertising the runtime.
       enableRunc: false
+      # Requires an image built with AKERNEL_ENABLE_ASCEND=true and runc enabled.
+      # Deploy this profile only to Ascend nodes whose host driver paths exist.
+      enableAscend: false
+      # Mount FlowGW's optional queue-scheduling directory when it exists on
+      # every selected Ascend node.
+      mountAscendQueueSchedule: false
       config: |-
         rootDir="/home/akernel/sandboxd/root"
         storeDir="/home/akernel/sandboxd/store"
@@ -505,6 +511,8 @@ node:
         shim_binary="/usr/local/bin/runc-shim"
         kvm_device="/dev/kvm"
 
+        # AKERNEL_XPU_ASCEND
+
         # image-manager config. root holds image-manager runtime state
         # (daemons/, mount_records.db) and is wiped on pod change; input
         # configs live under /home/akernel/sandboxd/config/ so the wipe never

deploy/scripts/build-image.sh

@@ -31,6 +31,9 @@ open_yr_core_wheel_url="${OPEN_YR_CORE_WHEEL_URL:-}"
 open_yr_core_wheel_sha256="${OPEN_YR_CORE_WHEEL_SHA256:-}"
 rrt_runtime_url="${RRT_RUNTIME_URL:-}"
 rrt_runtime_sha256="${RRT_RUNTIME_SHA256:-}"
+ascend_adapter_version="${ASCEND_ADAPTER_VERSION:-}"
+ascend_adapter_amd64_url="${ASCEND_ADAPTER_AMD64_URL:-}"
+ascend_adapter_amd64_sha256="${ASCEND_ADAPTER_AMD64_SHA256:-}"
 print_component_versions=0
 
 component_revision() {
@@ -149,6 +152,21 @@ case "${AKERNEL_ENABLE_RUNC:-false}" in
   *) die "AKERNEL_ENABLE_RUNC must be true or false" ;;
 esac
 
+case "${AKERNEL_ENABLE_ASCEND:-false}" in
+  true|false) ;;
+  *) die "AKERNEL_ENABLE_ASCEND must be true or false" ;;
+esac
+if [[ "${AKERNEL_ENABLE_ASCEND:-false}" == "true" &&
+      "${AKERNEL_ENABLE_RUNC:-false}" != "true" ]]; then
+  die "AKERNEL_ENABLE_ASCEND=true requires AKERNEL_ENABLE_RUNC=true"
+fi
+if [[ "${AKERNEL_ENABLE_ASCEND:-false}" == "true" &&
+      ( -z "${ascend_adapter_version}" ||
+        -z "${ascend_adapter_amd64_url}" ||
+        -z "${ascend_adapter_amd64_sha256}" ) ]]; then
+  die "AKERNEL_ENABLE_ASCEND=true requires ASCEND_ADAPTER_VERSION, ASCEND_ADAPTER_AMD64_URL, and ASCEND_ADAPTER_AMD64_SHA256"
+fi
+
 repository="${repository:-akernel-all-in-one}"
 tag="${tag:-$(git -C "${AKERNEL_REPO_ROOT}" rev-parse --short HEAD)-$(date +%Y%m%d%H%M%S)}"
 
@@ -208,6 +226,10 @@ node_build_args=(
   --build-arg "AKERNEL_RUNTIME_PROFILE=${runtime_profile}"
   --build-arg "AKERNEL_ENABLE_KATA=${AKERNEL_ENABLE_KATA:-true}"
   --build-arg "AKERNEL_ENABLE_RUNC=${AKERNEL_ENABLE_RUNC:-false}"
+  --build-arg "AKERNEL_ENABLE_ASCEND=${AKERNEL_ENABLE_ASCEND:-false}"
+  --build-arg "ASCEND_ADAPTER_VERSION=${ascend_adapter_version}"
+  --build-arg "ASCEND_ADAPTER_AMD64_URL=${ascend_adapter_amd64_url}"
+  --build-arg "ASCEND_ADAPTER_AMD64_SHA256=${ascend_adapter_amd64_sha256}"
   --build-arg "AKERNEL_ENABLE_FIRECRACKER=${AKERNEL_ENABLE_FIRECRACKER:-true}"
   --build-arg "AKERNEL_VERSION=${akernel_version}"
   --build-arg "AKERNEL_REVISION=${akernel_revision}"