feat(cli): add spnl bench subcommands

.github/workflows/core.yml

@@ -92,7 +92,7 @@ jobs:
             # - Excluded: local (CPU inference - pulls in CUDA deps via mistralrs)
             # - Excluded: cuda, cuda-flash-attn, cuda-flash-attn-v3 (CUDA features)
             cargo test -p spnl --features cli_support,print,lisp,run,ollama,openai,gemini,pull,yaml,metal,rag,rag-deep-debug,spnl-api,vllm,k8s,gce,ffi,pypi,run_py,tok,openssl-vendored -- --nocapture
-            cargo test -p spnl-cli --features rag,spnl-api,vllm,k8s,gce,local,metal -- --nocapture
+            cargo test -p spnl-cli --features bench,rag,spnl-api,vllm,k8s,gce,local,metal -- --nocapture
           else
             # Test default features on Linux (no GPU features)
             cargo test -p spnl -p spnl-cli -- --nocapture
@@ -108,7 +108,7 @@ jobs:
         if: runner.os == 'macOS'
         run: |
           cargo clippy -p spnl --features cli_support,print,lisp,run,ollama,openai,gemini,pull,yaml,metal,rag,rag-deep-debug,spnl-api,vllm,k8s,gce,ffi,pypi,run_py,tok,openssl-vendored --tests --no-deps -- -D warnings
-          cargo clippy -p spnl-cli --features rag,spnl-api,vllm,k8s,gce,local,metal --tests --no-deps -- -D warnings
+          cargo clippy -p spnl-cli --features bench,rag,spnl-api,vllm,k8s,gce,local,metal --tests --no-deps -- -D warnings
 
       - name: rustfmt
         run: cargo fmt --all -- --check

.github/workflows/release-cli.yml

@@ -171,7 +171,7 @@ jobs:
         run: |
           # Only use openssl-vendored for musl builds (static linking)
           # Other platforms use system OpenSSL
-          cargo build --features $LOCAL_FEATURE,rag,k8s,gce,spnl-api $BUILD_RELEASE_ARG --package spnl-cli --target $TARGET
+          cargo build --features $LOCAL_FEATURE,bench,rag,k8s,gce,spnl-api $BUILD_RELEASE_ARG --package spnl-cli --target $TARGET
 
       - name: Check for Apple signing secrets
         if: runner.os == 'macOS'
@@ -318,7 +318,7 @@ jobs:
         if: matrix.platform.arch == 'x86_64'
         shell: sh
         run: |
-          apk add --no-cache musl-dev gcc protobuf-dev curl unzip perl make
+          apk add --no-cache musl-dev gcc g++ protobuf-dev curl unzip perl make
 
       - name: Setup Rust toolchain
         shell: sh
@@ -328,7 +328,7 @@ jobs:
       - name: Build CLI
         shell: sh
         run: |
-          cargo build --features $LOCAL_FEATURE,rag,k8s,gce,spnl-api,openssl-vendored $BUILD_RELEASE_ARG --package spnl-cli --target $TARGET
+          cargo build --features $LOCAL_FEATURE,bench,rag,k8s,gce,spnl-api,openssl-vendored $BUILD_RELEASE_ARG --package spnl-cli --target $TARGET
 
       - name: Prepare binary
         shell: sh
@@ -522,7 +522,7 @@ jobs:
           echo "Building with features: $FEATURES,rag,k8s,gce,spnl-api"
           echo "CUDA_COMPUTE_CAP: $CUDA_COMPUTE_CAP"
           echo "CUDA_NVCC_FLAGS: $CUDA_NVCC_FLAGS"
-          cargo build --features $FEATURES,rag,k8s,gce,spnl-api $BUILD_RELEASE_ARG \
+          cargo build --features $FEATURES,bench,rag,k8s,gce,spnl-api $BUILD_RELEASE_ARG \
             --package spnl-cli --target $TARGET
 
       - name: Prepare binary

cli/Cargo.toml

@@ -20,6 +20,7 @@ metal = ["local","spnl/metal"]
 cuda = ["local","spnl/cuda"]
 cuda-flash-attn = ["local","spnl/cuda-flash-attn"]
 cuda-flash-attn-v3 = ["local","spnl/cuda-flash-attn-v3"]
+bench = ["dep:criterion", "dep:lipsum", "dep:petname", "dep:dirs", "dep:csv", "dep:rand", "dep:reqwest", "dep:indicatif", "dep:tokenizers"]
 
 [dependencies]
 anyhow = "1.0.99"
@@ -35,3 +36,14 @@ spnl = { version = ">=0.1.0", path = "../spnl" }
 tabled = { version = "0.20", optional = true }
 tinytemplate = "1.2.1"
 tokio = { version = "1.44.2", features = ["rt-multi-thread"] }
+
+# bench dependencies
+criterion = { version = "0.8", features = ["async_tokio"], optional = true }
+csv = { version = "1.3", optional = true }
+dirs = { version = "6.0", optional = true }
+indicatif = { version = "0.18.0", optional = true }
+lipsum = { git = "https://github.com/mgeisler/lipsum.git", optional = true }
+petname = { version = "2.0.2", optional = true }
+rand = { version = "0.9.2", optional = true }
+reqwest = { version = "0.12", features = ["blocking", "json"], optional = true }
+tokenizers = { version = "0.22.0", default-features = false, features = ["onig", "esaxx_fast", "hf-hub", "http"], optional = true }

cli/src/args.rs

@@ -17,6 +17,13 @@ pub enum Commands {
     #[cfg(feature = "local")]
     List,
 
+    /// Run benchmarks
+    #[cfg(feature = "bench")]
+    Bench {
+        #[command(subcommand)]
+        command: crate::bench::BenchCommands,
+    },
+
     /// Bring up vLLM in a Kubernetes cluster
     #[cfg(feature = "vllm")]
     Vllm {