Fix intel-mkl-src dependency and use static-linking in Dockerfile (#715)

Author: alvarobartt

Cargo.toml

@@ -42,13 +42,13 @@ serde_json = "1.0"
 thiserror = "1.0"
 rand = "0.9"
 serial_test = "2.0.0"
-cudarc = { version = "0.13" , features =["cuda-12020"], default-features = false}
-intel-mkl-src = { version = "0.8"}
+cudarc = { version = "0.13", features =["cuda-12020"], default-features = false }
+intel-mkl-src = { version = "0.8", default-features = false }
 candle = { version = "0.8", package = "candle-core" }
-candle-nn = { version = "0.8"  }
+candle-nn = { version = "0.8" }
 candle-transformers = { version = "0.8" }
 candle-flash-attn = { version = "0.8" }
-candle-cublaslt= { version = "0.0.1" }
+candle-cublaslt = { version = "0.0.1" }
 candle-layer-norm = { version = "0.0.1" }
 candle-rotary = { version = "0.0.1" }
 candle-flash-attn-v1 = { version = "0.0.1" }

Dockerfile

@@ -43,7 +43,7 @@ COPY --from=planner /usr/src/recipe.json recipe.json
 
 RUN --mount=type=secret,id=actions_results_url,env=ACTIONS_RESULTS_URL \
     --mount=type=secret,id=actions_runtime_token,env=ACTIONS_RUNTIME_TOKEN \
-    cargo chef cook --release --features ort,candle,mkl --no-default-features --recipe-path recipe.json && sccache -s
+    cargo chef cook --release --features ort,candle,mkl,static-linking --no-default-features --recipe-path recipe.json && sccache -s
 
 COPY backends backends
 COPY core core
@@ -55,7 +55,7 @@ FROM builder AS http-builder
 
 RUN --mount=type=secret,id=actions_results_url,env=ACTIONS_RESULTS_URL \
     --mount=type=secret,id=actions_runtime_token,env=ACTIONS_RUNTIME_TOKEN \
-    cargo build --release --bin text-embeddings-router --features ort,candle,mkl,http --no-default-features && sccache -s
+    cargo build --release --bin text-embeddings-router --features ort,candle,mkl,static-linking,http --no-default-features && sccache -s
 
 FROM builder AS grpc-builder
 
@@ -69,7 +69,7 @@ COPY proto proto
 
 RUN --mount=type=secret,id=actions_results_url,env=ACTIONS_RESULTS_URL \
     --mount=type=secret,id=actions_runtime_token,env=ACTIONS_RUNTIME_TOKEN \
-    cargo build --release --bin text-embeddings-router --features ort,candle,mkl,grpc --no-default-features && sccache -s
+    cargo build --release --bin text-embeddings-router --features ort,candle,mkl,static-linking,grpc --no-default-features && sccache -s
 
 FROM debian:bookworm-slim AS base
 

backends/candle/Cargo.toml

@@ -8,7 +8,7 @@ homepage.workspace = true
 [dependencies]
 anyhow = { workspace = true }
 accelerate-src = { version = "0.3.2", optional = true }
-intel-mkl-src = { workspace = true, optional = true  }
+intel-mkl-src = { workspace = true, optional = true }
 candle = { workspace = true }
 candle-nn = { workspace = true }
 candle-transformers = { workspace = true }

router/Cargo.toml

@@ -83,7 +83,7 @@ default = ["candle", "http", "dynamic-linking"]
 http = ["dep:axum", "dep:axum-tracing-opentelemetry", "dep:base64", "dep:tower-http", "dep:utoipa", "dep:utoipa-swagger-ui"]
 grpc = ["metrics-exporter-prometheus/http-listener", "dep:prost", "dep:tonic", "dep:tonic-health", "dep:tonic-reflection", "dep:tonic-build", "dep:async-stream", "dep:tokio-stream"]
 metal = ["text-embeddings-backend/metal"]
-mkl = ["text-embeddings-backend/mkl"]
+mkl = ["text-embeddings-backend/mkl", "dep:intel-mkl-src"]
 accelerate = ["text-embeddings-backend/accelerate"]
 python = ["text-embeddings-backend/python"]
 ort = ["text-embeddings-backend/ort"]