diff --git a/config/platform-infra/selfmedia-voice.yaml b/config/platform-infra/selfmedia-voice.yaml
index c30744f1..0a1248ce 100644
--- a/config/platform-infra/selfmedia-voice.yaml
+++ b/config/platform-infra/selfmedia-voice.yaml
@@ -17,6 +17,7 @@ artifacts:
   containerfile: deploy/selfmedia-voice/Containerfile
   api: deploy/selfmedia-voice/api.py
   modelInit: deploy/selfmedia-voice/model-init.sh
+  sourceInit: deploy/selfmedia-voice/source-init.sh
   d518Compose: deploy/selfmedia-voice/compose.d518.yaml
   nc01Compose: deploy/selfmedia-voice/compose.nc01.yaml
   frpcTemplate: deploy/selfmedia-voice/frpc.toml.tmpl
@@ -33,6 +34,9 @@ image:
     repository: https://github.com/FunAudioLLM/CosyVoice.git
     ref: 074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc
     submodules: recursive
+    cache:
+      retryAttempts: 3
+      retryDelaySeconds: 5
   modelSource:
     provider: modelscope-git-lfs
     id: FunAudioLLM/Fun-CosyVoice3-0.5B-2512
@@ -95,6 +99,7 @@ targets:
     workDir: /root/.unidesk/platform-infra/selfmedia-voice/runtime
     composeProject: unidesk-selfmedia-voice
     cacheDir: /root/.cache/unidesk/selfmedia-voice
+    sourceDir: /root/.cache/unidesk/selfmedia-voice/source
     dataDir: /root/.unidesk/.state/selfmedia-voice
   - id: NC01
     role: frps
diff --git a/deploy/selfmedia-voice/Containerfile b/deploy/selfmedia-voice/Containerfile
index 239d75f3..e7dacb94 100644
--- a/deploy/selfmedia-voice/Containerfile
+++ b/deploy/selfmedia-voice/Containerfile
@@ -1,9 +1,7 @@
 ARG BASE_IMAGE
 FROM ${BASE_IMAGE}
 
-ARG SOURCE_REPOSITORY
 ARG SOURCE_REF
-ARG SOURCE_SUBMODULES
 ARG REFERENCE_WAV_PATH
 ARG REFERENCE_WAV_SHA256
 ARG PIP_VERSION
@@ -26,11 +24,10 @@ RUN apt-get update \
     && rm -rf /var/lib/apt/lists/* \
     && python -m pip install "pip==${PIP_VERSION}" "${SETUPTOOLS_INSTALL_SPEC}" wheel
 
-RUN git clone --no-checkout "${SOURCE_REPOSITORY}" /opt/CosyVoice \
-    && git -C /opt/CosyVoice fetch --depth 1 origin "${SOURCE_REF}" \
-    && git -C /opt/CosyVoice checkout --detach "${SOURCE_REF}" \
-    && test "$(git -C /opt/CosyVoice rev-parse HEAD)" = "${SOURCE_REF}" \
-    && if [ "${SOURCE_SUBMODULES}" = recursive ]; then git -C /opt/CosyVoice submodule update --init --recursive --depth 1; fi \
+COPY --from=cosyvoice_source . /opt/CosyVoice
+
+RUN test "$(git -C /opt/CosyVoice rev-parse HEAD)" = "${SOURCE_REF}" \
+    && test -z "$(git -C /opt/CosyVoice submodule status --recursive | sed -n '/^[+-U]/p')" \
     && printf '%s  %s\n' "${REFERENCE_WAV_SHA256}" "/opt/CosyVoice/${REFERENCE_WAV_PATH}" | sha256sum -c -
 
 RUN sed -e '/^torch==/d' \
diff --git a/deploy/selfmedia-voice/compose.d518.yaml b/deploy/selfmedia-voice/compose.d518.yaml
index fd9fac6d..ad902ad0 100644
--- a/deploy/selfmedia-voice/compose.d518.yaml
+++ b/deploy/selfmedia-voice/compose.d518.yaml
@@ -4,11 +4,11 @@ services:
     build: &voice-build
       context: .
       dockerfile: Containerfile
+      additional_contexts:
+        cosyvoice_source: ${VOICE_SOURCE_DIR}
       args:
         BASE_IMAGE: ${BASE_IMAGE}
-        SOURCE_REPOSITORY: ${SOURCE_REPOSITORY}
         SOURCE_REF: ${SOURCE_REF}
-        SOURCE_SUBMODULES: ${SOURCE_SUBMODULES}
         REFERENCE_WAV_PATH: ${REFERENCE_WAV_PATH}
         REFERENCE_WAV_SHA256: ${REFERENCE_WAV_SHA256}
         PIP_VERSION: ${PIP_VERSION}
diff --git a/deploy/selfmedia-voice/source-init.sh b/deploy/selfmedia-voice/source-init.sh
new file mode 100644
index 00000000..da3757aa
--- /dev/null
+++ b/deploy/selfmedia-voice/source-init.sh
@@ -0,0 +1,38 @@
+#!/bin/sh
+set -eu
+
+retry() {
+    attempt=1
+    while ! "$@"; do
+        if [ "$attempt" -ge "$SOURCE_FETCH_RETRY_ATTEMPTS" ]; then
+            return 1
+        fi
+        sleep "$SOURCE_FETCH_RETRY_DELAY_SECONDS"
+        attempt=$((attempt + 1))
+    done
+}
+
+mkdir -p "$(dirname "$VOICE_SOURCE_DIR")"
+if [ ! -d "$VOICE_SOURCE_DIR/.git" ]; then
+    rm -rf "$VOICE_SOURCE_DIR"
+    mkdir -p "$VOICE_SOURCE_DIR"
+    git -C "$VOICE_SOURCE_DIR" init
+    git -C "$VOICE_SOURCE_DIR" remote add origin "$SOURCE_REPOSITORY"
+fi
+
+test "$(git -C "$VOICE_SOURCE_DIR" remote get-url origin)" = "$SOURCE_REPOSITORY"
+if ! git -C "$VOICE_SOURCE_DIR" cat-file -e "$SOURCE_REF^{commit}" 2>/dev/null; then
+    retry git -C "$VOICE_SOURCE_DIR" fetch --depth 1 origin "$SOURCE_REF"
+fi
+git -C "$VOICE_SOURCE_DIR" checkout --detach "$SOURCE_REF"
+test "$(git -C "$VOICE_SOURCE_DIR" rev-parse HEAD)" = "$SOURCE_REF"
+
+if [ "$SOURCE_SUBMODULES" = recursive ]; then
+    if ! git -C "$VOICE_SOURCE_DIR" submodule status --recursive >/dev/null 2>&1 \
+        || git -C "$VOICE_SOURCE_DIR" submodule status --recursive | grep -q '^[+-U]'; then
+        retry git -C "$VOICE_SOURCE_DIR" submodule update --init --recursive --depth 1
+    fi
+    ! git -C "$VOICE_SOURCE_DIR" submodule status --recursive | grep -q '^[+-U]'
+fi
+
+printf '%s  %s\n' "$REFERENCE_WAV_SHA256" "$VOICE_SOURCE_DIR/$REFERENCE_WAV_PATH" | sha256sum -c -
diff --git a/scripts/src/platform-infra-selfmedia-voice.ts b/scripts/src/platform-infra-selfmedia-voice.ts
index 6d287b8f..f5be8ba8 100644
--- a/scripts/src/platform-infra-selfmedia-voice.ts
+++ b/scripts/src/platform-infra-selfmedia-voice.ts
@@ -18,6 +18,7 @@ interface VoiceTarget {
   composeProject: string;
   cacheDir?: string;
   dataDir?: string;
+  sourceDir?: string;
 }
 
 interface VoiceConfig {
@@ -25,11 +26,11 @@ interface VoiceConfig {
   kind: "platform-infra-selfmedia-voice";
   metadata: { id: string; owner: string; relatedIssues: number[]; mdtodo: string };
   rollout: { prerequisite: string; approved: boolean };
-  artifacts: Record<"root" | "containerfile" | "api" | "modelInit" | "d518Compose" | "nc01Compose" | "frpcTemplate" | "frpsTemplate", string>;
+  artifacts: Record<"root" | "containerfile" | "api" | "modelInit" | "sourceInit" | "d518Compose" | "nc01Compose" | "frpcTemplate" | "frpsTemplate", string>;
   image: {
     name: string;
     base: { repository: string; tag: string; digest: string; imageId: string };
-    source: { repository: string; ref: string; submodules: "recursive" };
+    source: { repository: string; ref: string; submodules: "recursive"; cache: { retryAttempts: number; retryDelaySeconds: number } };
     modelSource: { provider: "modelscope-git-lfs"; id: string; repository: string; branch: string; ref: string; requireBranchHeadMatch: boolean };
     referenceVoice: { id: string; path: string; sha256: string; promptText: string };
   };
@@ -160,6 +161,8 @@ function readConfig(): VoiceConfig {
   requireSha256(config.image.base.imageId, "image.base.imageId");
   requireCommit(config.image.source.ref, "image.source.ref");
   if (config.image.source.submodules !== "recursive") throw new Error(`${configLabel}.image.source.submodules must be recursive`);
+  requireInteger(config.image.source.cache.retryAttempts, "image.source.cache.retryAttempts");
+  requireInteger(config.image.source.cache.retryDelaySeconds, "image.source.cache.retryDelaySeconds");
   if (config.image.modelSource.provider !== "modelscope-git-lfs") throw new Error(`${configLabel}.image.modelSource.provider must be modelscope-git-lfs`);
   requireCommit(config.image.modelSource.ref, "image.modelSource.ref");
   requireString(config.image.modelSource.branch, "image.modelSource.branch");
@@ -198,6 +201,7 @@ function readConfig(): VoiceConfig {
   }
   requireAbsolute(gpu.cacheDir, `targets.${gpu.id}.cacheDir`);
   requireAbsolute(gpu.dataDir, `targets.${gpu.id}.dataDir`);
+  requireAbsolute(gpu.sourceDir, `targets.${gpu.id}.sourceDir`);
   requireInteger(config.runtime.inference.replicas, "runtime.inference.replicas");
   requireInteger(config.runtime.inference.workers, "runtime.inference.workers");
   if (config.runtime.inference.concurrency !== "serial") throw new Error(`${configLabel}.runtime.inference.concurrency must be serial`);
@@ -316,6 +320,7 @@ function renderBundle(config: VoiceConfig): { files: Record<string, string>; sum
     containerfile: artifact(config.artifacts.containerfile),
     api: artifact(config.artifacts.api),
     modelInit: artifact(config.artifacts.modelInit),
+    sourceInit: artifact(config.artifacts.sourceInit),
     d518Compose: artifact(config.artifacts.d518Compose),
     nc01Compose: artifact(config.artifacts.nc01Compose),
     frpc: replace(artifact(config.artifacts.frpcTemplate), { SERVER_ADDRESS: config.frp.server.address, CONTROL_PORT: config.frp.server.controlPort, PROXY_NAME: config.frp.tunnel.proxyName, LOCAL_ADDRESS: config.frp.tunnel.localAddress, LOCAL_PORT: config.frp.tunnel.localPort, REMOTE_PORT: config.frp.tunnel.remotePort }),
@@ -328,7 +333,7 @@ function renderBundle(config: VoiceConfig): { files: Record<string, string>; sum
 function runtimeFiles(config: VoiceConfig, bundle: ReturnType<typeof renderBundle>, secret: ReturnType<typeof readSecret>, role: VoiceTarget["role"]): Record<string, string> {
   const target = targetByRole(config, role);
   const common = { "runtime.env": runtimeEnv(config, target), "secret.env": `${config.frp.secret.sourceKey}=${secret.value}\n` };
-  return role === "frps" ? { ...common, "compose.yaml": bundle.files.nc01Compose, "frps.toml": bundle.files.frps } : { ...common, "compose.yaml": bundle.files.d518Compose, "Containerfile": bundle.files.containerfile, "api.py": bundle.files.api, "model-init.sh": bundle.files.modelInit, "frpc.toml": bundle.files.frpc };
+  return role === "frps" ? { ...common, "compose.yaml": bundle.files.nc01Compose, "frps.toml": bundle.files.frps } : { ...common, "compose.yaml": bundle.files.d518Compose, "Containerfile": bundle.files.containerfile, "api.py": bundle.files.api, "model-init.sh": bundle.files.modelInit, "source-init.sh": bundle.files.sourceInit, "frpc.toml": bundle.files.frpc };
 }
 
 function runtimeEnv(config: VoiceConfig, target: VoiceTarget): string {
@@ -342,6 +347,8 @@ function runtimeEnv(config: VoiceConfig, target: VoiceTarget): string {
     SOURCE_REPOSITORY: config.image.source.repository,
     SOURCE_REF: config.image.source.ref,
     SOURCE_SUBMODULES: config.image.source.submodules,
+    SOURCE_FETCH_RETRY_ATTEMPTS: config.image.source.cache.retryAttempts,
+    SOURCE_FETCH_RETRY_DELAY_SECONDS: config.image.source.cache.retryDelaySeconds,
     MODEL_ID: config.image.modelSource.id,
     MODEL_REPOSITORY: config.image.modelSource.repository,
     MODEL_BRANCH: config.image.modelSource.branch,
@@ -364,6 +371,7 @@ function runtimeEnv(config: VoiceConfig, target: VoiceTarget): string {
     VOICE_PORT: config.runtime.port,
     VOICE_CACHE_DIR: requireAbsolute(target.cacheDir, `targets.${target.id}.cacheDir`),
     VOICE_DATA_DIR: requireAbsolute(target.dataDir, `targets.${target.id}.dataDir`),
+    VOICE_SOURCE_DIR: requireAbsolute(target.sourceDir, `targets.${target.id}.sourceDir`),
     VOICE_DEVICE: config.runtime.inference.device,
     VOICE_FP16: String(config.runtime.inference.fp16),
     VOICE_LOAD_TRT: String(config.runtime.inference.loadTrt),
@@ -382,7 +390,8 @@ function runtimeEnv(config: VoiceConfig, target: VoiceTarget): string {
 
 async function convergeTarget(configRoot: UniDeskConfig, config: VoiceConfig, target: VoiceTarget, files: Record<string, string>) {
   const encoded = Object.entries(files).map(([name, value]) => `printf %s ${shQuote(Buffer.from(value).toString("base64"))} | base64 -d > ${shQuote(name)}`).join("\n");
-  const script = `set -eu\numask 077\nmkdir -p ${shQuote(target.workDir)}\ncd ${shQuote(target.workDir)}\n${encoded}\nchmod 600 secret.env\ndocker compose --project-name ${shQuote(target.composeProject)} --env-file runtime.env -f compose.yaml config >/dev/null\ndocker compose --project-name ${shQuote(target.composeProject)} --env-file runtime.env -f compose.yaml up -d --build\n`;
+  const prepareSource = target.role === "gpu-service" ? "chmod 700 source-init.sh\nset -a\n. ./runtime.env\nset +a\n./source-init.sh\n" : "";
+  const script = `set -eu\numask 077\nmkdir -p ${shQuote(target.workDir)}\ncd ${shQuote(target.workDir)}\n${encoded}\nchmod 600 secret.env\n${prepareSource}docker compose --project-name ${shQuote(target.composeProject)} --env-file runtime.env -f compose.yaml config >/dev/null\ndocker compose --project-name ${shQuote(target.composeProject)} --env-file runtime.env -f compose.yaml up -d --build\n`;
   return await capture(configRoot, target.route, ["sh"], script, { runtimeTimeoutMs: config.runtime.operations.applyTimeoutSeconds * 1000 });
 }
 
@@ -439,7 +448,7 @@ function targetByRole(config: VoiceConfig, role: VoiceTarget["role"]): VoiceTarg
 }
 
 function targetSummary(target: VoiceTarget): Record<string, unknown> {
-  return { id: target.id, role: target.role, route: target.route, workDir: target.workDir, composeProject: target.composeProject, cacheDir: target.cacheDir ?? null, dataDir: target.dataDir ?? null };
+  return { id: target.id, role: target.role, route: target.route, workDir: target.workDir, composeProject: target.composeProject, cacheDir: target.cacheDir ?? null, dataDir: target.dataDir ?? null, sourceDir: target.sourceDir ?? null };
 }
 
 function requireString(value: unknown, label: string): string {