Merge origin/main into litellm_rust_qdrant_semantic_cache

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm-rust/Cargo.lock

@@ -2654,6 +2654,21 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "litellm-cache-gcs"
+version = "0.1.0"
+dependencies = [
+ "futures-util",
+ "litellm-auth-gcp",
+ "litellm-auth-types",
+ "litellm-cache",
+ "percent-encoding",
+ "reqwest 0.12.28",
+ "serde_json",
+ "tokio",
+ "wiremock",
+]
+
 [[package]]
 name = "litellm-cache-memory"
 version = "0.1.0"
@@ -2878,6 +2893,7 @@ dependencies = [
  "litellm-cache",
  "litellm-cache-azure-blob",
  "litellm-cache-disk",
+ "litellm-cache-gcs",
  "litellm-cache-memory",
  "litellm-cache-qdrant-semantic",
  "litellm-cache-redis",

litellm-rust/Cargo.toml

@@ -32,6 +32,7 @@ litellm-cache = { path = "crates/cache" }
 litellm-cache-azure-blob = { path = "crates/cache-azure-blob" }
 litellm-cache-memory = { path = "crates/cache-memory" }
 litellm-cache-redis = { path = "crates/cache-redis" }
+litellm-cache-gcs = { path = "crates/cache-gcs" }
 litellm-cache-disk = { path = "crates/cache-disk" }
 litellm-cache-response = { path = "crates/cache-response" }
 litellm-cache-qdrant-semantic = { path = "crates/cache-qdrant-semantic" }
@@ -73,6 +74,7 @@ base64 = "0.22"
 moka = { version = "0.12.16", features = ["future"] }
 strum = { version = "0.28.0", features = ["derive"] }
 url = "2.5.8"
+percent-encoding = "2.3"
 webpki-roots = "1"
 time = { version = "0.3.53", features = ["parsing"] }
 criterion = "0.8.2"

litellm-rust/crates/auth-gcp/src/lib.rs

@@ -128,6 +128,14 @@ impl VertexAuth {
         }
     }
 
+    pub async fn access_token(
+        &self,
+        config: &VertexConfig,
+        env_lookup: &(dyn Fn(&str) -> Option<String> + Sync),
+    ) -> Result<String, Error> {
+        self.load_provider(config, env_lookup).await?.token().await
+    }
+
     pub async fn validate_environment(
         &self,
         headers: Vec<(String, String)>,

litellm-rust/crates/cache-gcs/Cargo.toml

@@ -0,0 +1,20 @@
+[package]
+name = "litellm-cache-gcs"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[dependencies]
+futures-util.workspace = true
+litellm-auth-gcp.workspace = true
+litellm-auth-types.workspace = true
+litellm-cache.workspace = true
+percent-encoding.workspace = true
+reqwest.workspace = true
+tokio.workspace = true
+
+[dev-dependencies]
+serde_json.workspace = true
+tokio.workspace = true
+wiremock = "0.6.5"

litellm-rust/crates/cache-gcs/src/cache.rs

@@ -0,0 +1,260 @@
+use std::{future::Future, sync::Arc, time::Duration};
+
+use futures_util::future::try_join_all;
+use litellm_cache::{
+    BaseCache, BatchCache, BatchEntry, CacheCodec, CacheConnectionResult, Error, ExactCacheContext,
+    FlushCache,
+};
+use percent_encoding::{AsciiSet, NON_ALPHANUMERIC, percent_encode};
+use reqwest::Client;
+
+use crate::{GcpTokenSource, TokenSource};
+
+pub const DEFAULT_ENDPOINT: &str = "https://storage.googleapis.com";
+
+const OBJECT_NAME_ENCODE_SET: &AsciiSet = &NON_ALPHANUMERIC
+    .remove(b'-')
+    .remove(b'_')
+    .remove(b'.')
+    .remove(b'~');
+
+pub fn key_prefix(gcs_path: Option<&str>) -> String {
+    match gcs_path {
+        Some(path) if !path.is_empty() => format!("{}/", path.trim_end_matches('/')),
+        _ => String::new(),
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct GcsConfig {
+    pub bucket_name: String,
+    pub gcs_path: Option<String>,
+    pub path_service_account: Option<String>,
+    pub endpoint: String,
+}
+
+impl GcsConfig {
+    pub fn new(bucket_name: impl Into<String>) -> Self {
+        Self {
+            bucket_name: bucket_name.into(),
+            gcs_path: None,
+            path_service_account: None,
+            endpoint: DEFAULT_ENDPOINT.to_string(),
+        }
+    }
+}
+
+pub struct GcsCache<S: CacheCodec> {
+    config: GcsConfig,
+    key_prefix: String,
+    client: Client,
+    token: Arc<dyn TokenSource>,
+    codec: S,
+}
+
+impl<S: CacheCodec> GcsCache<S> {
+    pub fn new(config: GcsConfig, codec: S) -> Result<Self, Error> {
+        let token = Arc::new(GcpTokenSource::new(config.path_service_account.clone()));
+        Self::with_token_source(config, codec, token)
+    }
+
+    pub fn with_token_source(
+        config: GcsConfig,
+        codec: S,
+        token: Arc<dyn TokenSource>,
+    ) -> Result<Self, Error> {
+        let client = Client::builder().build().map_err(|_| Error::Unavailable)?;
+        let key_prefix = key_prefix(config.gcs_path.as_deref());
+        Ok(Self {
+            config,
+            key_prefix,
+            client,
+            token,
+            codec,
+        })
+    }
+
+    pub fn bucket_name(&self) -> &str {
+        &self.config.bucket_name
+    }
+
+    pub fn key_prefix(&self) -> &str {
+        &self.key_prefix
+    }
+
+    pub fn path_service_account(&self) -> Option<&str> {
+        self.config.path_service_account.as_deref()
+    }
+
+    pub fn object_name(&self, key: &str) -> String {
+        format!("{}{}", self.key_prefix, key)
+    }
+
+    fn encoded_object_name(&self, key: &str) -> String {
+        percent_encode(self.object_name(key).as_bytes(), OBJECT_NAME_ENCODE_SET).to_string()
+    }
+
+    fn endpoint(&self, path: &str) -> String {
+        format!("{}{}", self.config.endpoint.trim_end_matches('/'), path)
+    }
+
+    async fn async_set(&self, key: &str, value: S::Value) -> Result<(), Error> {
+        let token = self.token.bearer_token().await?;
+        let payload = self.codec.encode(&value)?;
+        let url = self.endpoint(&format!(
+            "/upload/storage/v1/b/{}/o?uploadType=media&name={}",
+            self.config.bucket_name,
+            self.encoded_object_name(key)
+        ));
+        let response = self
+            .client
+            .post(url)
+            .bearer_auth(token)
+            .header(reqwest::header::CONTENT_TYPE, "application/json")
+            .body(payload)
+            .send()
+            .await
+            .map_err(|_| Error::Unavailable)?;
+        if !response.status().is_success() {
+            return Err(Error::Unavailable);
+        }
+        Ok(())
+    }
+
+    async fn async_get(&self, key: &str) -> Result<Option<S::Value>, Error> {
+        let token = self.token.bearer_token().await?;
+        let url = self.endpoint(&format!(
+            "/storage/v1/b/{}/o/{}?alt=media",
+            self.config.bucket_name,
+            self.encoded_object_name(key)
+        ));
+        let response = self
+            .client
+            .get(url)
+            .bearer_auth(token)
+            .header(reqwest::header::CONTENT_TYPE, "application/json")
+            .send()
+            .await
+            .map_err(|_| Error::Unavailable)?;
+        if response.status() == reqwest::StatusCode::NOT_FOUND {
+            return Ok(None);
+        }
+        if !response.status().is_success() {
+            return Err(Error::Unavailable);
+        }
+        let body = response.bytes().await.map_err(|_| Error::Unavailable)?;
+        self.codec
+            .decode(&body)
+            .map(Some)
+            .map_err(|_| Error::InvalidEntry)
+    }
+
+    fn run_sync<T, F>(future: F) -> Result<T, Error>
+    where
+        F: Future<Output = Result<T, Error>> + Send,
+        T: Send,
+    {
+        let run = || {
+            tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .map_err(|_| Error::Unavailable)
+                .and_then(|runtime| runtime.block_on(future))
+        };
+        if let Ok(handle) = tokio::runtime::Handle::try_current() {
+            if handle.runtime_flavor() == tokio::runtime::RuntimeFlavor::MultiThread {
+                return tokio::task::block_in_place(run);
+            }
+            return std::thread::scope(|scope| {
+                scope
+                    .spawn(run)
+                    .join()
+                    .map_err(|_| Error::Unavailable)
+                    .and_then(|result| result)
+            });
+        }
+        run()
+    }
+}
+
+impl<S: CacheCodec> BaseCache for GcsCache<S> {
+    type Value = S::Value;
+    type Context = ExactCacheContext;
+
+    fn get_ttl(&self, _: &Self::Context) -> Option<Duration> {
+        None
+    }
+
+    fn set_cache(&self, key: &str, value: Self::Value, _: &Self::Context) -> Result<(), Error> {
+        Self::run_sync(self.async_set(key, value))
+    }
+
+    fn get_cache(&self, key: &str, _: &Self::Context) -> Result<Option<Self::Value>, Error> {
+        Self::run_sync(self.async_get(key))
+    }
+
+    async fn async_set_cache(
+        &self,
+        key: &str,
+        value: Self::Value,
+        _: Self::Context,
+    ) -> Result<(), Error> {
+        self.async_set(key, value).await
+    }
+
+    async fn async_get_cache(
+        &self,
+        key: &str,
+        _: &Self::Context,
+    ) -> Result<Option<Self::Value>, Error> {
+        self.async_get(key).await
+    }
+
+    async fn async_set_cache_pipeline(
+        &self,
+        entries: Vec<(String, Self::Value)>,
+        context: Self::Context,
+    ) -> Result<(), Error> {
+        try_join_all(entries.into_iter().map(|(key, value)| {
+            let context = context.clone();
+            async move { self.async_set_cache(&key, value, context).await }
+        }))
+        .await
+        .map(|_| ())
+    }
+
+    async fn disconnect(&self) -> Result<(), Error> {
+        Ok(())
+    }
+
+    async fn test_connection(&self) -> Result<CacheConnectionResult, Error> {
+        Err(Error::UnsupportedOperation)
+    }
+}
+
+impl<S: CacheCodec> BatchCache for GcsCache<S> {
+    async fn async_batch_get_cache(
+        &self,
+        keys: Vec<String>,
+        context: Self::Context,
+    ) -> Result<Vec<BatchEntry<Self::Value>>, Error> {
+        try_join_all(keys.into_iter().map(|key| {
+            let context = context.clone();
+            async move {
+                match self.async_get_cache(&key, &context).await {
+                    Ok(Some(value)) => Ok(BatchEntry::Hit(value)),
+                    Ok(None) => Ok(BatchEntry::Miss),
+                    Err(Error::InvalidEntry) => Ok(BatchEntry::Invalid),
+                    Err(error) => Err(error),
+                }
+            }
+        }))
+        .await
+    }
+}
+
+impl<S: CacheCodec> FlushCache for GcsCache<S> {
+    fn flush_cache(&self) -> Result<(), Error> {
+        Ok(())
+    }
+}

litellm-rust/crates/cache-gcs/src/lib.rs

@@ -0,0 +1,5 @@
+mod cache;
+mod token;
+
+pub use cache::{DEFAULT_ENDPOINT, GcsCache, GcsConfig, key_prefix};
+pub use token::{GcpTokenSource, StaticTokenSource, TokenSource};

litellm-rust/crates/cache-gcs/src/token.rs

@@ -0,0 +1,44 @@
+use std::{future::Future, pin::Pin};
+
+use litellm_auth_gcp::{VertexAuth, VertexConfig};
+use litellm_auth_types::{InputSource, SecretValue, Sourced};
+use litellm_cache::Error;
+
+pub trait TokenSource: Send + Sync + 'static {
+    fn bearer_token(&self) -> Pin<Box<dyn Future<Output = Result<String, Error>> + Send + '_>>;
+}
+
+pub struct GcpTokenSource {
+    auth: VertexAuth,
+    config: VertexConfig,
+}
+
+impl GcpTokenSource {
+    pub fn new(path_service_account: Option<String>) -> Self {
+        let credentials = path_service_account
+            .map(|path| Sourced::new(SecretValue::new(path), InputSource::Deployment));
+        Self {
+            auth: VertexAuth::default(),
+            config: VertexConfig::new(credentials, None, None),
+        }
+    }
+}
+
+impl TokenSource for GcpTokenSource {
+    fn bearer_token(&self) -> Pin<Box<dyn Future<Output = Result<String, Error>> + Send + '_>> {
+        Box::pin(async move {
+            self.auth
+                .access_token(&self.config, &|name| std::env::var(name).ok())
+                .await
+                .map_err(|_| Error::Unavailable)
+        })
+    }
+}
+
+pub struct StaticTokenSource(pub String);
+
+impl TokenSource for StaticTokenSource {
+    fn bearer_token(&self) -> Pin<Box<dyn Future<Output = Result<String, Error>> + Send + '_>> {
+        Box::pin(async move { Ok(self.0.clone()) })
+    }
+}

litellm-rust/crates/cache-gcs/tests/cache.rs

@@ -0,0 +1,324 @@
+use std::{sync::Arc, time::Duration};
+
+use litellm_cache::{
+    BaseCache, BatchCache, BatchEntry, CacheContext, Error, ExactCacheContext, FlushCache,
+    JsonCodec,
+};
+use litellm_cache_gcs::{GcsCache, GcsConfig, StaticTokenSource, TokenSource, key_prefix};
+use serde_json::json;
+use wiremock::{
+    Mock, MockServer, ResponseTemplate,
+    matchers::{body_bytes, header, method, path, query_param},
+};
+
+fn config(server: &MockServer, gcs_path: Option<&str>) -> GcsConfig {
+    GcsConfig {
+        bucket_name: "bucket".into(),
+        gcs_path: gcs_path.map(str::to_string),
+        path_service_account: None,
+        endpoint: server.uri(),
+    }
+}
+
+fn cache(server: &MockServer, gcs_path: Option<&str>) -> GcsCache<JsonCodec<serde_json::Value>> {
+    GcsCache::with_token_source(
+        config(server, gcs_path),
+        JsonCodec::new(),
+        Arc::new(StaticTokenSource("tok".into())),
+    )
+    .unwrap()
+}
+
+#[tokio::test]
+async fn set_writes_encoded_object_and_headers() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/upload/storage/v1/b/bucket/o"))
+        .and(query_param("uploadType", "media"))
+        .and(header("authorization", "Bearer tok"))
+        .and(header("content-type", "application/json"))
+        .and(body_bytes(br#"{"value":"entry"}"#))
+        .respond_with(ResponseTemplate::new(200))
+        .expect(1)
+        .mount(&server)
+        .await;
+    cache(&server, Some("cache/"))
+        .set_cache(
+            "team:a b/c",
+            json!({"value": "entry"}),
+            &ExactCacheContext::default(),
+        )
+        .unwrap();
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(requests.len(), 1);
+    assert_eq!(
+        requests[0].url.query(),
+        Some("uploadType=media&name=cache%2Fteam%3Aa%20b%2Fc")
+    );
+}
+
+#[tokio::test]
+async fn get_maps_statuses_and_decode_failures() {
+    let server = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/hit"))
+        .and(query_param("alt", "media"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(json!({"value": "entry"})))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/missing"))
+        .respond_with(ResponseTemplate::new(404))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/server-error"))
+        .respond_with(ResponseTemplate::new(500))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/invalid"))
+        .respond_with(ResponseTemplate::new(200).set_body_string("not json"))
+        .mount(&server)
+        .await;
+
+    let cache = cache(&server, None);
+    assert_eq!(
+        cache
+            .get_cache("hit", &ExactCacheContext::default())
+            .unwrap(),
+        Some(json!({"value": "entry"}))
+    );
+    assert_eq!(
+        cache
+            .get_cache("missing", &ExactCacheContext::default())
+            .unwrap(),
+        None
+    );
+    assert_eq!(
+        cache
+            .get_cache("server-error", &ExactCacheContext::default())
+            .unwrap_err(),
+        Error::Unavailable
+    );
+    assert_eq!(
+        cache
+            .get_cache("invalid", &ExactCacheContext::default())
+            .unwrap_err(),
+        Error::InvalidEntry
+    );
+}
+
+#[test]
+fn key_prefix_normalizes_paths() {
+    assert_eq!(key_prefix(None), "");
+    assert_eq!(key_prefix(Some("a/b/")), "a/b/");
+    assert_eq!(key_prefix(Some("a/b")), "a/b/");
+    assert_eq!(key_prefix(Some("")), "");
+}
+
+#[tokio::test]
+async fn object_names_use_python_quote_encoding() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/upload/storage/v1/b/bucket/o"))
+        .and(query_param("uploadType", "media"))
+        .respond_with(ResponseTemplate::new(200))
+        .expect(2)
+        .mount(&server)
+        .await;
+    let cache = cache(&server, Some("p/"));
+    cache
+        .set_cache(
+            "a~b-c_d.e/f g%h",
+            json!({"value": "punctuation"}),
+            &ExactCacheContext::default(),
+        )
+        .unwrap();
+    cache
+        .set_cache(
+            "ключ",
+            json!({"value": "utf8"}),
+            &ExactCacheContext::default(),
+        )
+        .unwrap();
+    let requests = server.received_requests().await.unwrap();
+    let queries: Vec<_> = requests
+        .iter()
+        .filter_map(|request| request.url.query())
+        .collect();
+    assert!(queries.contains(&"uploadType=media&name=p%2Fa~b-c_d.e%2Ff%20g%25h"));
+    assert!(queries.contains(&"uploadType=media&name=p%2F%D0%BA%D0%BB%D1%8E%D1%87"));
+}
+
+#[tokio::test]
+async fn ignores_ttl_and_writes_pipeline_concurrently() {
+    let server = MockServer::start().await;
+    for key in ["one", "two", "three"] {
+        Mock::given(method("POST"))
+            .and(path("/upload/storage/v1/b/bucket/o"))
+            .and(query_param("uploadType", "media"))
+            .and(query_param("name", key))
+            .respond_with(ResponseTemplate::new(200))
+            .expect(1)
+            .mount(&server)
+            .await;
+    }
+    let cache = cache(&server, None);
+    assert_eq!(cache.get_ttl(&ExactCacheContext::default()), None);
+    assert_eq!(
+        cache.get_ttl(&ExactCacheContext::default().with_ttl(Some(Duration::from_secs(5)))),
+        None
+    );
+    cache
+        .async_set_cache_pipeline(
+            vec![
+                ("one".into(), json!({"key": "one"})),
+                ("two".into(), json!({"key": "two"})),
+                ("three".into(), json!({"key": "three"})),
+            ],
+            ExactCacheContext::default().with_ttl(Some(Duration::from_secs(5))),
+        )
+        .await
+        .unwrap();
+}
+
+#[tokio::test]
+async fn async_batch_get_preserves_hits_misses_and_invalid_entries() {
+    let server = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/hit"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(json!({"value": "entry"})))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/missing"))
+        .respond_with(ResponseTemplate::new(404))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/invalid"))
+        .respond_with(ResponseTemplate::new(200).set_body_string("not json"))
+        .mount(&server)
+        .await;
+
+    assert_eq!(
+        cache(&server, None)
+            .async_batch_get_cache(
+                vec!["hit".into(), "missing".into(), "invalid".into()],
+                ExactCacheContext::default(),
+            )
+            .await
+            .unwrap(),
+        vec![
+            BatchEntry::Hit(json!({"value": "entry"})),
+            BatchEntry::Miss,
+            BatchEntry::Invalid,
+        ]
+    );
+}
+
+#[tokio::test]
+async fn lifecycle_operations_are_noops_and_connection_test_is_unsupported() {
+    let server = MockServer::start().await;
+    let cache = cache(&server, None);
+    assert_eq!(cache.flush_cache(), Ok(()));
+    assert_eq!(cache.disconnect().await, Ok(()));
+    assert_eq!(
+        cache.test_connection().await,
+        Err(Error::UnsupportedOperation)
+    );
+}
+
+#[test]
+fn sync_operations_work_without_an_active_runtime() {
+    let runtime = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()
+        .unwrap();
+    let server = runtime.block_on(MockServer::start());
+    runtime.block_on(
+        Mock::given(method("POST"))
+            .and(path("/upload/storage/v1/b/bucket/o"))
+            .respond_with(ResponseTemplate::new(200))
+            .mount(&server),
+    );
+    runtime.block_on(
+        Mock::given(method("GET"))
+            .and(path("/storage/v1/b/bucket/o/key"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(json!({"value": "entry"})))
+            .mount(&server),
+    );
+    let cache = cache(&server, None);
+    cache
+        .set_cache(
+            "key",
+            json!({"value": "entry"}),
+            &ExactCacheContext::default(),
+        )
+        .unwrap();
+    assert_eq!(
+        cache
+            .get_cache("key", &ExactCacheContext::default())
+            .unwrap(),
+        Some(json!({"value": "entry"}))
+    );
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn sync_operations_work_inside_a_multi_thread_runtime() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/upload/storage/v1/b/bucket/o"))
+        .respond_with(ResponseTemplate::new(200))
+        .mount(&server)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/storage/v1/b/bucket/o/key"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(json!({"value": "entry"})))
+        .mount(&server)
+        .await;
+    let cache = cache(&server, None);
+    cache
+        .set_cache(
+            "key",
+            json!({"value": "entry"}),
+            &ExactCacheContext::default(),
+        )
+        .unwrap();
+    assert_eq!(
+        cache
+            .get_cache("key", &ExactCacheContext::default())
+            .unwrap(),
+        Some(json!({"value": "entry"}))
+    );
+}
+
+struct FailingTokenSource;
+
+impl TokenSource for FailingTokenSource {
+    fn bearer_token(
+        &self,
+    ) -> std::pin::Pin<Box<dyn std::future::Future<Output = Result<String, Error>> + Send + '_>>
+    {
+        Box::pin(async { Err(Error::Unavailable) })
+    }
+}
+
+#[tokio::test]
+async fn token_source_failure_skips_http() {
+    let server = MockServer::start().await;
+    let cache = GcsCache::with_token_source(
+        config(&server, None),
+        JsonCodec::<serde_json::Value>::new(),
+        Arc::new(FailingTokenSource),
+    )
+    .unwrap();
+    assert_eq!(
+        cache
+            .get_cache("key", &ExactCacheContext::default())
+            .unwrap_err(),
+        Error::Unavailable
+    );
+    assert_eq!(server.received_requests().await.unwrap().len(), 0);
+}

litellm-rust/crates/cache/src/error.rs

@@ -6,7 +6,7 @@ pub enum Error {
     InvalidEntry,
     #[error("flushing Redis requires an explicit namespace")]
     UnscopedFlush,
-    #[error("cache backend does not support this operation")]
+    #[error("operation is not supported by this cache")]
     UnsupportedOperation,
     #[error("semantic cache requires request messages")]
     MissingPrompt,

litellm-rust/crates/python-bridge/Cargo.toml

@@ -24,6 +24,7 @@ litellm-cache.workspace = true
 litellm-cache-azure-blob.workspace = true
 litellm-cache-memory.workspace = true
 litellm-cache-redis.workspace = true
+litellm-cache-gcs.workspace = true
 litellm-cache-disk.workspace = true
 litellm-cache-response.workspace = true
 litellm-cache-qdrant-semantic.workspace = true

litellm-rust/crates/python-bridge/src/cache/config.rs

@@ -80,6 +80,13 @@ pub(super) struct RedisCacheConfig {
     pub(super) connection: RedisConnectionConfig,
 }
 
+#[derive(Debug, PartialEq)]
+pub(super) struct GcsCacheConfig {
+    pub(super) bucket_name: String,
+    pub(super) key_prefix: String,
+    pub(super) path_service_account: Option<String>,
+}
+
 pub(super) struct AzureBlobCacheConfig {
     pub(super) account_url: String,
     pub(super) container: String,
@@ -119,6 +126,7 @@ impl QdrantSemanticCacheConfig {
 pub(super) enum CacheBackendConfig {
     Memory(MemoryCacheConfig),
     Redis(Box<RedisCacheConfig>),
+    Gcs(GcsCacheConfig),
     Disk(DiskCacheConfig),
     QdrantSemantic(Box<QdrantSemanticCacheConfig>),
     AzureBlob(AzureBlobCacheConfig),
@@ -136,6 +144,7 @@ pub(super) enum UnsupportedCacheConfig {
     RedisCredentials,
     RedisConnection,
     RedisOption,
+    GcsBucket,
     DiskStore,
     QdrantEndpoint,
     SemanticEmbedding,
@@ -149,6 +158,7 @@ impl UnsupportedCacheConfig {
             Self::RedisCredentials => "native Redis credentials require Python",
             Self::RedisConnection => "native Redis connection type is not implemented",
             Self::RedisOption => "native Redis configuration requires Python",
+            Self::GcsBucket => "native GCS cache requires a configured bucket name",
             Self::DiskStore => "native disk cache requires the built-in diskcache store",
             Self::QdrantEndpoint => {
                 "native Qdrant requires the default REST port so the gRPC port can be derived"
@@ -196,6 +206,13 @@ impl NativeCacheConfig {
                 }))),
                 Err(reason) => Ok(CacheConfigProjection::Unsupported(reason)),
             },
+            Some(CacheType::Gcs) => match project_gcs(&backend)? {
+                Ok(backend) => Ok(CacheConfigProjection::Native(Box::new(Self {
+                    policy,
+                    backend: CacheBackendConfig::Gcs(backend),
+                }))),
+                Err(reason) => Ok(CacheConfigProjection::Unsupported(reason)),
+            },
             Some(CacheType::Disk) => match project_disk(&backend)? {
                 Ok(backend) => Ok(CacheConfigProjection::Native(Box::new(Self {
                     policy,
@@ -216,20 +233,41 @@ impl NativeCacheConfig {
                     backend: CacheBackendConfig::AzureBlob(backend),
                 }))
             }),
-            Some(
-                CacheType::RedisSemantic
-                | CacheType::ValkeySemantic
-                | CacheType::S3
-                | CacheType::Gcs,
-            )
-            | None => Ok(CacheConfigProjection::Unsupported(
-                UnsupportedCacheConfig::Backend,
-            )),
+            Some(CacheType::RedisSemantic | CacheType::ValkeySemantic | CacheType::S3) | None => {
+                Ok(CacheConfigProjection::Unsupported(
+                    UnsupportedCacheConfig::Backend,
+                ))
+            }
         }
     }
 
     pub(super) fn service_mismatch(&self, service: &NativeResponseCache) -> Option<&'static str> {
         match &self.backend {
+            CacheBackendConfig::Gcs(_) if service.kind() != "gcs" => {
+                Some("facade and native backend types must match")
+            }
+            CacheBackendConfig::Gcs(config)
+                if service
+                    .gcs_backend()
+                    .is_none_or(|backend| backend.bucket_name() != config.bucket_name) =>
+            {
+                Some("facade and native backend buckets must match")
+            }
+            CacheBackendConfig::Gcs(config)
+                if service
+                    .gcs_backend()
+                    .is_none_or(|backend| backend.key_prefix() != config.key_prefix) =>
+            {
+                Some("facade and native backend key prefixes must match")
+            }
+            CacheBackendConfig::Gcs(config)
+                if service.gcs_backend().is_none_or(|backend| {
+                    backend.path_service_account() != config.path_service_account.as_deref()
+                }) =>
+            {
+                Some("facade and native backend credentials must match")
+            }
+            CacheBackendConfig::Gcs(_) => None,
             CacheBackendConfig::Memory(_) if service.kind() != "memory" => {
                 Some("facade and native backend types must match")
             }
@@ -421,6 +459,23 @@ fn project_azure_blob(backend: &Bound<'_, PyAny>) -> PyResult<AzureBlobCacheConf
     })
 }
 
+#[inline(never)]
+fn project_gcs(
+    backend: &Bound<'_, PyAny>,
+) -> PyResult<Result<GcsCacheConfig, UnsupportedCacheConfig>> {
+    let bucket_name = match backend.getattr("bucket_name")?.extract::<Option<String>>() {
+        Ok(Some(bucket_name)) if !bucket_name.is_empty() => bucket_name,
+        _ => return Ok(Err(UnsupportedCacheConfig::GcsBucket)),
+    };
+    Ok(Ok(GcsCacheConfig {
+        bucket_name,
+        key_prefix: backend.getattr("key_prefix")?.extract::<String>()?,
+        path_service_account: backend
+            .getattr("path_service_account")?
+            .extract::<Option<String>>()?,
+    }))
+}
+
 #[inline(never)]
 fn project_memory(backend: &Bound<'_, PyAny>) -> PyResult<MemoryCacheConfig> {
     let max_size_kib = backend.getattr("max_size_per_item")?.extract::<usize>()?;
@@ -863,7 +918,7 @@ mod tests {
 
     use super::{
         CacheBackendConfig, CacheConfigProjection, CachePolicy, CertificateRequirement,
-        DiskCacheConfig, NativeCacheConfig, RedisProtocol, UnsupportedCacheConfig,
+        DiskCacheConfig, GcsCacheConfig, NativeCacheConfig, RedisProtocol, UnsupportedCacheConfig,
     };
     use crate::cache::native::NativeResponseCache;
 
@@ -1067,6 +1122,56 @@ sys.modules['litellm.caching._embedding_router'] = embedding_router
         });
     }
 
+    #[test]
+    fn projects_gcs_configuration() {
+        Python::initialize();
+        Python::attach(|py| {
+            let facade = facade(
+                py,
+                "backend = SimpleNamespace(bucket_name='bucket', key_prefix='cache/', path_service_account='credentials.json')\n\
+                 facade = SimpleNamespace(type='gcs', mode='default-on', ttl=None, namespace=None, supported_call_types=None, redis_flush_size=None, semantic_cache_scope='key', cache=backend)",
+            );
+            let CacheConfigProjection::Native(config) =
+                NativeCacheConfig::project(&facade).unwrap()
+            else {
+                panic!("GCS cache should be supported");
+            };
+            let CacheBackendConfig::Gcs(gcs) = config.backend else {
+                panic!("expected GCS configuration");
+            };
+            assert_eq!(
+                gcs,
+                GcsCacheConfig {
+                    bucket_name: "bucket".into(),
+                    key_prefix: "cache/".into(),
+                    path_service_account: Some("credentials.json".into()),
+                }
+            );
+        });
+    }
+
+    #[test]
+    fn rejects_gcs_without_a_bucket_name() {
+        Python::initialize();
+        Python::attach(|py| {
+            let facade = facade(
+                py,
+                "backend = SimpleNamespace(bucket_name=None, key_prefix='', path_service_account=None)\n\
+                 facade = SimpleNamespace(type='gcs', mode='default-on', ttl=None, namespace=None, supported_call_types=None, redis_flush_size=None, semantic_cache_scope='key', cache=backend)",
+            );
+            let CacheConfigProjection::Unsupported(reason) =
+                NativeCacheConfig::project(&facade).unwrap()
+            else {
+                panic!("GCS cache without a bucket should be unsupported");
+            };
+            assert!(matches!(&reason, UnsupportedCacheConfig::GcsBucket));
+            assert_eq!(
+                reason.message(),
+                "native GCS cache requires a configured bucket name"
+            );
+        });
+    }
+
     #[test]
     fn projects_builtin_disk_configuration_and_rejects_custom_stores() {
         Python::initialize();

litellm-rust/crates/python-bridge/src/cache/facade.rs

@@ -355,6 +355,7 @@ impl FacadeGuard {
                 "QdrantSemanticCache",
                 "qdrant-semantic",
             ),
+            ("gcs", _) => ("litellm.caching.gcs_cache", "GCSCache", "gcs"),
             ("disk", _) => ("litellm.caching.disk_cache", "DiskCache", "disk"),
             ("azure-blob", _) => (
                 "litellm.caching.azure_blob_cache",
@@ -381,13 +382,16 @@ impl FacadeGuard {
             return Err(PyTypeError::new_err(message));
         }
         let backend_config_names = match kind {
-            "memory" | "redis" | "azure-blob" | "disk" => &[
+            "memory" | "redis" | "azure-blob" | "disk" | "gcs" => &[
                 "namespace",
                 "default_ttl",
                 "max_size_in_memory",
                 "max_size_per_item",
                 "redis_kwargs",
                 "redis_flush_size",
+                "bucket_name",
+                "key_prefix",
+                "path_service_account",
             ][..],
             "qdrant_semantic" => &[
                 "qdrant_api_base",

litellm-rust/crates/python-bridge/src/cache/handle.rs

@@ -1,5 +1,6 @@
 use std::env;
 
+use litellm_cache_gcs::{DEFAULT_ENDPOINT, GcsConfig};
 use litellm_cache_redis::{RedisNode, RedisTopology};
 use litellm_http::ClientVariant;
 
@@ -74,6 +75,31 @@ impl CacheTestHandle {
         })
     }
 
+    #[staticmethod]
+    #[pyo3(signature = (bucket_name, *, gcs_path=None, path_service_account=None, endpoint=None, token=None))]
+    fn gcs(
+        py: Python<'_>,
+        bucket_name: String,
+        gcs_path: Option<String>,
+        path_service_account: Option<String>,
+        endpoint: Option<String>,
+        token: Option<String>,
+    ) -> PyResult<Self> {
+        let config = GcsConfig {
+            bucket_name,
+            gcs_path,
+            path_service_account,
+            endpoint: endpoint.unwrap_or_else(|| DEFAULT_ENDPOINT.to_string()),
+        };
+        let service = release_gil(py, move || NativeResponseCache::gcs(config, token))
+            .map_err(cache_error)?;
+        Ok(Self {
+            service,
+            guard: None,
+            pid: std::process::id(),
+        })
+    }
+
     #[staticmethod]
     #[pyo3(signature = (directory))]
     fn disk(py: Python<'_>, directory: String) -> PyResult<Self> {

litellm-rust/crates/python-bridge/src/cache/mod.rs

@@ -10,7 +10,7 @@ mod resolver;
 
 use litellm_cache::Error;
 use pyo3::{
-    exceptions::{PyRuntimeError, PyValueError},
+    exceptions::{PyNotImplementedError, PyRuntimeError, PyValueError},
     prelude::*,
 };
 
@@ -21,6 +21,7 @@ pub(crate) use self::{
 fn cache_error(error: Error) -> PyErr {
     match error {
         Error::InvalidEntry => PyValueError::new_err(error.to_string()),
+        Error::UnsupportedOperation => PyNotImplementedError::new_err(error.to_string()),
         _ => PyRuntimeError::new_err(error.to_string()),
     }
 }

litellm-rust/crates/python-bridge/src/cache/native.rs

@@ -3,6 +3,7 @@ use std::{path::Path, sync::Arc, time::Duration};
 use litellm_cache::{CacheCodec, CacheConnectionResult, Error, SemanticCacheContext};
 use litellm_cache_azure_blob::AzureBlobCache;
 use litellm_cache_disk::DiskCache;
+use litellm_cache_gcs::{GcsCache, GcsConfig, StaticTokenSource};
 use litellm_cache_memory::InMemoryCache;
 use litellm_cache_qdrant_semantic::{Embedder, OpenAiEmbedder, QdrantSemanticCache};
 use litellm_cache_redis::{RedisCache, RedisTopology};
@@ -20,6 +21,7 @@ pub(super) enum NativeResponseCache {
         cache: Arc<ResponseCache<RedisCache<ResponseCacheCodec>>>,
         buffer: Option<Arc<WriteBuffer>>,
     },
+    Gcs(Arc<ResponseCache<GcsCache<ResponseCacheCodec>>>),
     Disk(Arc<ResponseCache<DiskCache<ResponseCacheCodec>>>),
     QdrantSemantic(Arc<ResponseCache<QdrantSemanticCache<OpenAiEmbedder, ResponseCacheCodec>>>),
     AzureBlob(Arc<ResponseCache<AzureBlobCache<ResponseCacheCodec>>>),
@@ -59,6 +61,18 @@ impl NativeResponseCache {
         Ok(Self::Disk(Arc::new(ResponseCache::new(Arc::new(cache)))))
     }
 
+    pub fn gcs(config: GcsConfig, token: Option<String>) -> Result<Self, Error> {
+        let backend = match token {
+            Some(token) => GcsCache::with_token_source(
+                config,
+                ResponseCacheCodec,
+                Arc::new(StaticTokenSource(token)),
+            )?,
+            None => GcsCache::new(config, ResponseCacheCodec)?,
+        };
+        Ok(Self::Gcs(Arc::new(ResponseCache::new(Arc::new(backend)))))
+    }
+
     pub async fn qdrant_semantic(
         config: QdrantSemanticCacheConfig,
         client: reqwest::Client,
@@ -103,7 +117,11 @@ impl NativeResponseCache {
                 cache.backend().account_url(),
                 cache.backend().container_name(),
             )),
-            Self::Memory(_) | Self::Redis { .. } | Self::Disk(_) | Self::QdrantSemantic(_) => None,
+            Self::Memory(_)
+            | Self::Redis { .. }
+            | Self::Gcs(_)
+            | Self::Disk(_)
+            | Self::QdrantSemantic(_) => None,
         }
     }
 }
@@ -113,6 +131,7 @@ impl NativeResponseCache {
         match self {
             Self::Memory(_) => "memory",
             Self::Redis { .. } => "redis",
+            Self::Gcs(_) => "gcs",
             Self::Disk(_) => "disk",
             Self::QdrantSemantic(_) => "qdrant_semantic",
             Self::AzureBlob(_) => "azure-blob",
@@ -123,6 +142,7 @@ impl NativeResponseCache {
         match self {
             Self::Memory(cache) => cache.default_ttl(),
             Self::Redis { cache, .. } => cache.default_ttl(),
+            Self::Gcs(cache) => cache.default_ttl(),
             Self::Disk(cache) => cache.default_ttl(),
             Self::QdrantSemantic(_) => None,
             Self::AzureBlob(cache) => cache.default_ttl(),
@@ -131,14 +151,22 @@ impl NativeResponseCache {
 
     pub fn namespace(&self) -> Option<&str> {
         match self {
-            Self::Memory(_) | Self::Disk(_) | Self::QdrantSemantic(_) | Self::AzureBlob(_) => None,
+            Self::Memory(_)
+            | Self::Disk(_)
+            | Self::QdrantSemantic(_)
+            | Self::AzureBlob(_)
+            | Self::Gcs(_) => None,
             Self::Redis { cache, .. } => cache.backend().namespace(),
         }
     }
 
     pub fn topology(&self) -> Option<&RedisTopology> {
         match self {
-            Self::Memory(_) | Self::Disk(_) | Self::QdrantSemantic(_) | Self::AzureBlob(_) => None,
+            Self::Memory(_)
+            | Self::Disk(_)
+            | Self::QdrantSemantic(_)
+            | Self::AzureBlob(_)
+            | Self::Gcs(_) => None,
             Self::Redis { cache, .. } => Some(cache.backend().topology()),
         }
     }
@@ -146,7 +174,7 @@ impl NativeResponseCache {
     pub fn capacity(&self) -> Option<usize> {
         match self {
             Self::Memory(cache) => Some(cache.backend().max_size_in_memory()),
-            Self::Redis { .. } | Self::Disk(_) => None,
+            Self::Redis { .. } | Self::Gcs(_) | Self::Disk(_) => None,
             Self::QdrantSemantic(_) => None,
             Self::AzureBlob(_) => None,
         }
@@ -155,7 +183,7 @@ impl NativeResponseCache {
     pub fn max_entry_bytes(&self) -> Option<usize> {
         match self {
             Self::Memory(cache) => cache.backend().max_entry_bytes(),
-            Self::Redis { .. } | Self::Disk(_) => None,
+            Self::Redis { .. } | Self::Gcs(_) | Self::Disk(_) => None,
             Self::QdrantSemantic(_) => None,
             Self::AzureBlob(_) => None,
         }
@@ -174,9 +202,11 @@ impl NativeResponseCache {
     pub fn directory(&self) -> Option<&Path> {
         match self {
             Self::Disk(cache) => Some(cache.backend().directory()),
-            Self::Memory(_) | Self::Redis { .. } | Self::QdrantSemantic(_) | Self::AzureBlob(_) => {
-                None
-            }
+            Self::Memory(_)
+            | Self::Redis { .. }
+            | Self::Gcs(_)
+            | Self::QdrantSemantic(_)
+            | Self::AzureBlob(_) => None,
         }
     }
 
@@ -216,6 +246,7 @@ impl NativeResponseCache {
         match self {
             Self::Memory(cache) => cache.lookup(&exact(request), now),
             Self::Redis { cache, .. } => cache.lookup(&exact(request), now),
+            Self::Gcs(cache) => cache.lookup(&exact(request), now),
             Self::Disk(cache) => cache.lookup(&exact(request), now),
             Self::QdrantSemantic(cache) => cache.lookup(request, now),
             Self::AzureBlob(cache) => cache.lookup(&exact(request), now),
@@ -231,6 +262,7 @@ impl NativeResponseCache {
         match self {
             Self::Memory(cache) => cache.store(&exact(request), response, now),
             Self::Redis { cache, .. } => cache.store(&exact(request), response, now),
+            Self::Gcs(cache) => cache.store(&exact(request), response, now),
             Self::Disk(cache) => cache.store(&exact(request), response, now),
             Self::QdrantSemantic(cache) => cache.store(request, response, now),
             Self::AzureBlob(cache) => cache.store(&exact(request), response, now),
@@ -249,6 +281,9 @@ impl NativeResponseCache {
             Self::Redis { cache, .. } => {
                 cache.lookup_batch(&requests.iter().map(exact).collect::<Vec<_>>(), now)
             }
+            Self::Gcs(cache) => {
+                cache.lookup_batch(&requests.iter().map(exact).collect::<Vec<_>>(), now)
+            }
             Self::Disk(cache) => {
                 cache.lookup_batch(&requests.iter().map(exact).collect::<Vec<_>>(), now)
             }
@@ -267,6 +302,7 @@ impl NativeResponseCache {
         match self {
             Self::Memory(cache) => cache.async_lookup(&exact(request), now).await,
             Self::Redis { cache, .. } => cache.async_lookup(&exact(request), now).await,
+            Self::Gcs(cache) => cache.async_lookup(&exact(request), now).await,
             Self::Disk(cache) => cache.async_lookup(&exact(request), now).await,
             Self::QdrantSemantic(cache) => cache.async_lookup(request, now).await,
             Self::AzureBlob(cache) => cache.async_lookup(&exact(request), now).await,
@@ -292,6 +328,7 @@ impl NativeResponseCache {
                 let request = exact(request);
                 buffer.async_store(cache, &request, response, now).await
             }
+            Self::Gcs(cache) => cache.async_store(&exact(request), response, now).await,
             Self::Disk(cache) => cache.async_store(&exact(request), response, now).await,
             Self::QdrantSemantic(cache) => cache.async_store(request, response, now).await,
             Self::AzureBlob(cache) => cache.async_store(&exact(request), response, now).await,
@@ -312,6 +349,10 @@ impl NativeResponseCache {
                 let requests = requests.iter().map(exact).collect::<Vec<_>>();
                 cache.async_lookup_batch(&requests, now).await
             }
+            Self::Gcs(cache) => {
+                let requests = requests.iter().map(exact).collect::<Vec<_>>();
+                cache.async_lookup_batch(&requests, now).await
+            }
             Self::Disk(cache) => {
                 let requests = requests.iter().map(exact).collect::<Vec<_>>();
                 cache.async_lookup_batch(&requests, now).await
@@ -352,6 +393,17 @@ impl NativeResponseCache {
                     )
                     .await
             }
+            Self::Gcs(cache) => {
+                cache
+                    .async_store_batch(
+                        entries
+                            .into_iter()
+                            .map(|(request, value)| (exact(&request), value))
+                            .collect(),
+                        now,
+                    )
+                    .await
+            }
             Self::Disk(cache) => {
                 cache
                     .async_store_batch(
@@ -387,6 +439,7 @@ impl NativeResponseCache {
                 }
                 cache.async_flush().await
             }
+            Self::Gcs(cache) => cache.async_flush().await,
             Self::Disk(cache) => cache.async_flush().await,
             Self::QdrantSemantic(_) => Err(Error::UnsupportedOperation),
             Self::AzureBlob(cache) => cache.async_flush().await,
@@ -397,9 +450,17 @@ impl NativeResponseCache {
         match self {
             Self::Memory(cache) => cache.test_connection().await,
             Self::Redis { cache, .. } => cache.test_connection().await,
+            Self::Gcs(cache) => cache.test_connection().await,
             Self::Disk(cache) => cache.test_connection().await,
             Self::QdrantSemantic(_) => Err(Error::UnsupportedOperation),
             Self::AzureBlob(cache) => cache.test_connection().await,
         }
     }
+
+    pub fn gcs_backend(&self) -> Option<&GcsCache<ResponseCacheCodec>> {
+        match self {
+            Self::Gcs(cache) => Some(cache.backend()),
+            _ => None,
+        }
+    }
 }

litellm/llms/xai/chat/transformation.py

@@ -118,6 +118,7 @@ class XAIChatConfig(OpenAIGPTConfig):
         base_openai_params: Final = [
             "logit_bias",
             "logprobs",
+            "max_completion_tokens",
             "max_tokens",
             "n",
             "parallel_tool_calls",

tests/test_litellm/llms/xai/test_xai_chat_transformation.py

@@ -90,6 +90,16 @@ class TestXAIReasoningTokenFolding:
         assert response.usage.total_tokens == 999
 
 
+def test_max_completion_tokens_is_accepted_and_mapped_to_max_tokens() -> None:
+    optional_params = litellm.get_optional_params(
+        model="grok-4.20",
+        custom_llm_provider="xai",
+        max_completion_tokens=64,
+    )
+    assert optional_params["max_tokens"] == 64, optional_params
+    assert "max_completion_tokens" not in optional_params, optional_params
+
+
 class TestXAIParallelToolCalls:
     """Test suite for XAI parallel tool calls functionality."""
 

tests/test_litellm_rust/support/fake_gcs.py

@@ -0,0 +1,152 @@
+from __future__ import annotations
+
+import json
+import threading
+from collections.abc import Mapping
+from dataclasses import dataclass
+from functools import partial
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from socket import socket
+from types import MappingProxyType
+from typing import Final, cast
+from urllib.parse import unquote, urlsplit
+
+
+@dataclass(frozen=True, slots=True)
+class RecordedRequest:
+    method: str
+    path: str
+    query: str
+    headers: Mapping[str, str]
+    body: bytes
+
+
+class _FakeGcsHandler(BaseHTTPRequestHandler):
+    def __init__(
+        self,
+        request: socket | tuple[bytes, socket],
+        client_address: tuple[str, int],
+        server: ThreadingHTTPServer,
+        *,
+        fake: FakeGcs,
+    ) -> None:
+        self._fake: Final = fake
+        super().__init__(request, client_address, server)
+
+    def _handle(self) -> None:
+        parsed: Final = urlsplit(self.path)
+        content_length: Final = int(self.headers.get("Content-Length", "0"))
+        body: Final = self.rfile.read(content_length) if content_length else b""
+        headers: Final = MappingProxyType(
+            {name.title(): value for name, value in self.headers.items()}
+        )
+        self._fake.record(
+            RecordedRequest(
+                method=self.command,
+                path=parsed.path,
+                query=parsed.query,
+                headers=headers,
+                body=body,
+            )
+        )
+        if self.headers.get("Authorization") != f"Bearer {self._fake.token}":
+            self._send_json(401, {"error": "unauthorized"})
+            return
+
+        upload_prefix: Final = "/upload/storage/v1/b/"
+        download_prefix: Final = "/storage/v1/b/"
+        if parsed.path.startswith(upload_prefix) and parsed.path.endswith("/o"):
+            self._upload(parsed.path[len(upload_prefix) : -2], parsed.query, body)
+            return
+        if parsed.path.startswith(download_prefix):
+            self._download(parsed.path[len(download_prefix) :], parsed.query)
+            return
+        self._send_json(404, {"error": "not found"})
+
+    def _upload(self, path: str, query: str, body: bytes) -> None:
+        values: Final = {
+            unquote(pair.partition("=")[0]): unquote(pair.partition("=")[2])
+            for pair in query.split("&")
+            if pair
+        }
+        if not path or values.get("uploadType") != "media" or "name" not in values:
+            self._send_json(404, {"error": "not found"})
+            return
+        self._fake.put_object(path, values["name"], body)
+        self._send_json(200, {"name": values["name"], "bucket": path})
+
+    def _download(self, path: str, query: str) -> None:
+        bucket, separator, encoded_name = path.partition("/o/")
+        if not separator or query != "alt=media":
+            self._send_json(404, {"error": "not found"})
+            return
+        name: Final = unquote(encoded_name)
+        if name.endswith("/server-error") or name == "server-error":
+            self._send_json(500, {"error": "server error"})
+            return
+        body: Final = self._fake.get_object(bucket, name)
+        if body is None:
+            self._send_json(404, {"error": "not found"})
+            return
+        self._send(200, body, "application/octet-stream")
+
+    def _send_json(self, status: int, value: object) -> None:
+        payload: Final = json.dumps(value).encode()
+        self._send(status, payload, "application/json")
+
+    def _send(self, status: int, body: bytes, content_type: str) -> None:
+        self.send_response(status)
+        self.send_header("Content-Type", content_type)
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, format: str, *args: object) -> None:
+        pass
+
+    do_GET = _handle
+    do_POST = _handle
+
+
+class FakeGcs:
+    def __init__(self) -> None:
+        self._objects: dict[tuple[str, str], bytes] = {}  # mutable-ok: fake object store
+        self._requests: list[RecordedRequest] = []  # mutable-ok: recorded request history
+        self._server = ThreadingHTTPServer(
+            ("127.0.0.1", 0),
+            partial(_FakeGcsHandler, fake=self),
+        )
+        self._worker = threading.Thread(target=self._server.serve_forever, daemon=True)
+        self._worker.start()
+        self.token: Final = "test-token"
+
+    @property
+    def url(self) -> str:
+        address: Final = cast(tuple[str, int], self._server.server_address)
+        host, port = address
+        return f"http://{host}:{port}"
+
+    @property
+    def objects(self) -> Mapping[tuple[str, str], bytes]:
+        return MappingProxyType(self._objects)
+
+    @property
+    def requests(self) -> tuple[RecordedRequest, ...]:
+        return tuple(self._requests)
+
+    def put(self, bucket: str, name: str, body: bytes) -> None:
+        self.put_object(bucket, name, body)
+
+    def close(self) -> None:
+        self._server.shutdown()
+        self._server.server_close()
+        self._worker.join(timeout=5)
+
+    def record(self, request: RecordedRequest) -> None:
+        self._requests.append(request)
+
+    def put_object(self, bucket: str, name: str, body: bytes) -> None:
+        self._objects[(bucket, name)] = body
+
+    def get_object(self, bucket: str, name: str) -> bytes | None:
+        return self._objects.get((bucket, name))

tests/test_litellm_rust/test_cache.py

@@ -26,11 +26,13 @@ from azure.storage.blob import ContainerClient
 import litellm
 from litellm.caching.azure_blob_cache import AzureBlobCache
 from litellm.caching.caching import Cache, disable_cache, enable_cache, update_cache
+from litellm.caching.gcs_cache import GCSCache
 from litellm.caching.disk_cache import DiskCache
 from litellm.caching.in_memory_cache import InMemoryCache
 from litellm.caching.redis_cluster_cache import RedisClusterCache
 from litellm.rust_bridge import _native
 from litellm.types.caching import LiteLLMCacheType
+from tests.test_litellm_rust.support.fake_gcs import FakeGcs
 from tests.test_litellm_rust.support.isolation import rebound
 
 pytestmark: Final = pytest.mark.requires_rust_extension
@@ -38,6 +40,7 @@ pytestmark: Final = pytest.mark.requires_rust_extension
 
 class CacheLookup(Protocol):
     def get_cache(self, **kwargs: object) -> object: ...
+    def flush_cache(self) -> object: ...
 
 
 def request(key: str = "key") -> dict[str, object]:
@@ -122,6 +125,15 @@ def redis_url() -> Generator[str]:
         worker.join(timeout=5)
 
 
+@pytest.fixture
+def fake_gcs() -> Generator[FakeGcs]:
+    server: Final = FakeGcs()
+    try:
+        yield server
+    finally:
+        server.close()
+
+
 @pytest.fixture
 def azure_blob_facade() -> Generator[Cache]:
     account_url: Final = os.environ.get("AZURE_BLOB_CACHE_ACCOUNT_URL")
@@ -699,6 +711,222 @@ async def test_disk_native_batch_lookup_and_store_report_partial_hits(tmp_path:
     }
 
 
+async def test_gcs_reads_python_entries_and_writes_python_compatible_objects(
+    fake_gcs: FakeGcs, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.delenv("GCS_PATH_SERVICE_ACCOUNT", raising=False)
+    monkeypatch.delenv("GCS_BUCKET_NAME", raising=False)
+    response: Final = {"choices": [{"text": "cached"}], "usage": {"total_tokens": 3}, "flag": True, "empty": None}
+    fake_gcs.put(
+        "bucket",
+        "cache/sync",
+        json.dumps({"timestamp": time.time(), "response": json.dumps(response)}).encode(),
+    )
+    fake_gcs.put("bucket", "cache/async", json.dumps({"timestamp": time.time(), "response": response}).encode())
+    fake_gcs.put("bucket", "cache/raw", json.dumps(response).encode())
+    fake_gcs.put("bucket", "cache/invalid", b"not a cache entry")
+    binding: Final = _native._CacheTestResolver(
+        SimpleNamespace(
+            cache=_native._CacheTestHandle.gcs(
+                "bucket",
+                gcs_path="cache",
+                endpoint=fake_gcs.url,
+                token=fake_gcs.token,
+            )
+        )
+    ).resolve()
+
+    assert binding.lookup(request("sync")) == response
+    assert await binding.async_lookup(request("async")) == response
+    assert binding.lookup(request("raw")) == response
+    assert await binding.async_lookup(request("invalid")) is None
+    assert binding.lookup(request("missing")) is None
+
+    await binding.async_store({**request("native"), "ttl_seconds": 12.0}, response)
+    stored: Final = fake_gcs.objects[("bucket", "cache/native")]
+    stored_value: Final = cast(dict[str, object], json.loads(stored))
+    assert stored_value["response"] == response
+    assert isinstance(stored_value["timestamp"], float)
+    upload: Final = next(item for item in fake_gcs.requests if item.method == "POST")
+    assert upload.path == "/upload/storage/v1/b/bucket/o"
+    assert upload.query == "uploadType=media&name=cache%2Fnative"
+    assert upload.headers["Authorization"] == f"Bearer {fake_gcs.token}"
+    assert upload.headers["Content-Type"] == "application/json"
+    upload_text: Final = f"{upload.path}?{upload.query}{upload.headers}"
+    assert "ttl" not in upload_text.lower()
+    assert "expiry" not in upload_text.lower()
+    download: Final = next(item for item in fake_gcs.requests if item.path.endswith("/cache%2Fsync"))
+    assert download.path == "/storage/v1/b/bucket/o/cache%2Fsync"
+    assert download.query == "alt=media"
+
+    binding.store(request("sync2"), response)
+    assert binding.lookup(request("sync2")) == response
+    assert GCSCache(bucket_name="bucket", gcs_path="cache").key_prefix == "cache/"
+    assert GCSCache(bucket_name="bucket", gcs_path="cache/").key_prefix == "cache/"
+    assert GCSCache(bucket_name="bucket").key_prefix == ""
+
+
+async def test_gcs_batch_lookup_preserves_order_and_treats_malformed_entries_as_misses(fake_gcs: FakeGcs) -> None:
+    fake_gcs.put("bucket", "cache/hit", json.dumps({"timestamp": time.time(), "response": {"value": 1}}).encode())
+    fake_gcs.put("bucket", "cache/invalid", b"not a cache entry")
+    binding: Final = _native._CacheTestResolver(
+        SimpleNamespace(
+            cache=_native._CacheTestHandle.gcs(
+                "bucket",
+                gcs_path="cache",
+                endpoint=fake_gcs.url,
+                token=fake_gcs.token,
+            )
+        )
+    ).resolve()
+    requests: Final = [request("hit"), request("missing"), request("invalid")]
+    expected: Final = {"values": [{"value": 1}, None, None], "missing_indices": [1, 2]}
+
+    assert await binding.async_lookup_batch(requests) == expected
+    assert binding.lookup_batch(requests) == expected
+    await binding.async_store_batch([request("first"), request("second")], [{"value": 1}, {"value": 2}])
+    assert ("bucket", "cache/first") in fake_gcs.objects
+    assert ("bucket", "cache/second") in fake_gcs.objects
+
+
+async def test_gcs_facade_binds_only_exact_matching_configuration(
+    fake_gcs: FakeGcs, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.delenv("GCS_PATH_SERVICE_ACCOUNT", raising=False)
+    monkeypatch.delenv("GCS_BUCKET_NAME", raising=False)
+    monkeypatch.setenv("GOOGLE_APPLICATION_CREDENTIALS", "/nonexistent")
+    facade: Final = Cache(type=LiteLLMCacheType.GCS, gcs_bucket_name="bucket", gcs_path="cache/")
+    assert type(facade.cache) is GCSCache
+
+    mismatched_bucket: Final = _native._CacheTestHandle.gcs(
+        "other",
+        gcs_path="cache",
+        endpoint=fake_gcs.url,
+        token=fake_gcs.token,
+    )
+    with pytest.raises(TypeError, match="buckets must match"):
+        mismatched_bucket._bind_facade(facade)
+    mismatched_prefix: Final = _native._CacheTestHandle.gcs(
+        "bucket",
+        gcs_path="x",
+        endpoint=fake_gcs.url,
+        token=fake_gcs.token,
+    )
+    with pytest.raises(TypeError, match="key prefixes must match"):
+        mismatched_prefix._bind_facade(facade)
+    mismatched_credentials: Final = _native._CacheTestHandle.gcs(
+        "bucket",
+        gcs_path="cache",
+        path_service_account="sa.json",
+        endpoint=fake_gcs.url,
+        token=fake_gcs.token,
+    )
+    with pytest.raises(TypeError, match="credentials must match"):
+        mismatched_credentials._bind_facade(facade)
+    with pytest.raises(TypeError, match="types must match"):
+        _native._CacheTestHandle.memory()._bind_facade(facade)
+
+    matching: Final = _native._CacheTestHandle.gcs(
+        "bucket",
+        gcs_path="cache",
+        endpoint=fake_gcs.url,
+        token=fake_gcs.token,
+    )
+    matching._bind_facade(facade)
+    resolver: Final = _native._CacheTestResolver(SimpleNamespace(cache=facade))
+    binding: Final = resolver.resolve()
+    assert binding.kind == "native"
+    await binding.async_store(request("native"), {"value": "native"})
+    assert await binding.async_lookup(request("native")) == {"value": "native"}
+    assert cast(CacheLookup, facade).get_cache(cache_key="native") is None
+
+    with rebound(facade.cache, "bucket_name", "other"):
+        assert resolver.resolve().kind == "python_callback"
+    with rebound(facade.cache, "key_prefix", "x/"):
+        assert resolver.resolve().kind == "python_callback"
+    with rebound(facade.cache, "path_service_account", "sa.json"):
+        assert resolver.resolve().kind == "python_callback"
+    def no_get_cache(*args: object, **kwargs: object) -> None:
+        return None
+
+    with rebound(facade.cache, "get_cache", no_get_cache):
+        assert resolver.resolve().kind == "python_callback"
+    with rebound(facade, "ttl", 12):
+        assert resolver.resolve().kind == "python_callback"
+
+    class CustomGcs(GCSCache):
+        pass
+
+    with rebound(facade, "cache", CustomGcs(bucket_name="bucket", gcs_path="cache/")):
+        assert resolver.resolve().kind == "python_callback"
+    custom_facade: Final = Cache(type=LiteLLMCacheType.GCS, gcs_bucket_name="bucket", gcs_path="cache/")
+    with rebound(custom_facade, "cache", CustomGcs(bucket_name="bucket", gcs_path="cache/")):
+        with pytest.raises(TypeError, match="types must match"):
+            matching._bind_facade(custom_facade)
+
+    missing_bucket: Final = Cache(type=LiteLLMCacheType.GCS)
+    with pytest.raises(TypeError, match="requires a configured bucket name"):
+        matching._bind_facade(missing_bucket)
+
+
+async def test_gcs_flush_is_a_no_op_and_ping_is_not_implemented(
+    fake_gcs: FakeGcs, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.delenv("GCS_PATH_SERVICE_ACCOUNT", raising=False)
+    monkeypatch.delenv("GCS_BUCKET_NAME", raising=False)
+    binding: Final = _native._CacheTestResolver(
+        SimpleNamespace(
+            cache=_native._CacheTestHandle.gcs(
+                "bucket",
+                gcs_path="cache",
+                endpoint=fake_gcs.url,
+                token=fake_gcs.token,
+            )
+        )
+    ).resolve()
+    await binding.async_store(request("key"), {"value": "stored"})
+    await binding.async_flush()
+    assert ("bucket", "cache/key") in fake_gcs.objects
+    assert await binding.async_lookup(request("key")) == {"value": "stored"}
+    with pytest.raises(NotImplementedError):
+        await binding.ping()
+
+    facade: Final = Cache(type=LiteLLMCacheType.GCS, gcs_bucket_name="bucket", gcs_path="cache/")
+    with pytest.raises(AttributeError):
+        await facade.ping()
+    assert cast(CacheLookup, facade.cache).flush_cache() is None
+
+
+async def test_gcs_unauthorized_and_server_errors_surface_as_runtime_errors(fake_gcs: FakeGcs) -> None:
+    wrong_token: Final = _native._CacheTestResolver(
+        SimpleNamespace(
+            cache=_native._CacheTestHandle.gcs(
+                "bucket",
+                gcs_path="cache",
+                endpoint=fake_gcs.url,
+                token="wrong-token",
+            )
+        )
+    ).resolve()
+    with pytest.raises(RuntimeError):
+        wrong_token.lookup(request("missing"))
+    assert not fake_gcs.objects
+
+    binding: Final = _native._CacheTestResolver(
+        SimpleNamespace(
+            cache=_native._CacheTestHandle.gcs(
+                "bucket",
+                gcs_path="cache",
+                endpoint=fake_gcs.url,
+                token=fake_gcs.token,
+            )
+        )
+    ).resolve()
+    with pytest.raises(RuntimeError):
+        binding.lookup(request("server-error"))
+    assert binding.lookup(request("missing")) is None
+
+
 async def test_redis_cluster_facade_serves_multi_slot_batches_and_scoped_flush_natively(
     cluster_nodes: tuple[tuple[str, int], ...],
 ) -> None: