Merge commit 'd803bfe2b1fe7f5e219e50ac20d6801a0a58ac75' as 'vendor/ruvector'

vendor/ruvector/crates/ruvector-attention-cli/src/commands/benchmark.rs

@@ -0,0 +1,212 @@
+use clap::Args;
+use crate::{config::Config, output::{print_benchmark_results, BenchmarkRow}};
+use ruvector_attention::{
+    attention::{ScaledDotProductAttention, MultiHeadAttention},
+    hyperbolic::HyperbolicAttention,
+    sparse::{FlashAttention, LinearAttention},
+    moe::MoEAttention,
+};
+use indicatif::{ProgressBar, ProgressStyle};
+use std::time::Instant;
+
+#[derive(Args)]
+pub struct BenchmarkArgs {
+    /// Attention types to benchmark (comma-separated)
+    #[arg(short, long, value_delimiter = ',')]
+    attention_types: Option<Vec<String>>,
+
+    /// Dimensions to test (comma-separated)
+    #[arg(short, long, value_delimiter = ',')]
+    dimensions: Option<Vec<usize>>,
+
+    /// Number of iterations per test
+    #[arg(short, long)]
+    iterations: Option<usize>,
+
+    /// Number of warmup iterations
+    #[arg(short, long)]
+    warmup: Option<usize>,
+
+    /// Sequence length
+    #[arg(short, long, default_value = "128")]
+    seq_length: usize,
+
+    /// Output results to file
+    #[arg(short, long)]
+    output: Option<std::path::PathBuf>,
+
+    /// Output format (json, csv, table)
+    #[arg(short, long, default_value = "table")]
+    format: String,
+}
+
+pub async fn run(args: BenchmarkArgs, config: &Config) -> anyhow::Result<()> {
+    let attention_types = args.attention_types.unwrap_or_else(|| {
+        vec![
+            "scaled_dot".to_string(),
+            "multi_head".to_string(),
+            "hyperbolic".to_string(),
+            "flash".to_string(),
+            "linear".to_string(),
+            "moe".to_string(),
+        ]
+    });
+
+    let dimensions = args.dimensions.unwrap_or_else(|| config.benchmark.dimensions.clone());
+    let iterations = args.iterations.unwrap_or(config.benchmark.iterations);
+    let warmup = args.warmup.unwrap_or(config.benchmark.warmup);
+
+    println!("Running benchmarks...");
+    println!("Attention types: {:?}", attention_types);
+    println!("Dimensions: {:?}", dimensions);
+    println!("Iterations: {}, Warmup: {}", iterations, warmup);
+    println!();
+
+    let total_tests = attention_types.len() * dimensions.len();
+    let pb = ProgressBar::new(total_tests as u64);
+    pb.set_style(
+        ProgressStyle::default_bar()
+            .template("[{elapsed_precise}] {bar:40.cyan/blue} {pos}/{len} {msg}")?
+            .progress_chars("##-")
+    );
+
+    let mut results = Vec::new();
+
+    for attention_type in &attention_types {
+        for &dim in &dimensions {
+            pb.set_message(format!("Testing {} (dim={})", attention_type, dim));
+
+            let timings = benchmark_attention(
+                attention_type,
+                dim,
+                args.seq_length,
+                iterations,
+                warmup
+            )?;
+
+            let mean = timings.iter().sum::<f64>() / timings.len() as f64;
+            let variance = timings.iter()
+                .map(|&x| (x - mean).powi(2))
+                .sum::<f64>() / timings.len() as f64;
+            let std_dev = variance.sqrt();
+            let throughput = 1000.0 / mean; // operations per second
+
+            results.push(BenchmarkRow {
+                attention_type: attention_type.clone(),
+                dimension: dim,
+                mean_time_ms: mean,
+                std_dev_ms: std_dev,
+                throughput,
+            });
+
+            pb.inc(1);
+        }
+    }
+
+    pb.finish_with_message("Benchmarks complete!");
+    println!();
+
+    match args.format.as_str() {
+        "json" => {
+            let json = serde_json::to_string_pretty(&results)?;
+            if let Some(path) = args.output {
+                std::fs::write(path, json)?;
+            } else {
+                println!("{}", json);
+            }
+        }
+        "csv" => {
+            let mut csv = String::from("attention_type,dimension,mean_time_ms,std_dev_ms,throughput\n");
+            for row in &results {
+                csv.push_str(&format!(
+                    "{},{},{},{},{}\n",
+                    row.attention_type, row.dimension, row.mean_time_ms, row.std_dev_ms, row.throughput
+                ));
+            }
+            if let Some(path) = args.output {
+                std::fs::write(path, csv)?;
+            } else {
+                println!("{}", csv);
+            }
+        }
+        _ => {
+            print_benchmark_results(results);
+        }
+    }
+
+    Ok(())
+}
+
+fn benchmark_attention(
+    attention_type: &str,
+    dim: usize,
+    seq_length: usize,
+    iterations: usize,
+    warmup: usize,
+) -> anyhow::Result<Vec<f64>> {
+    // Generate random test data
+    let query: Vec<Vec<f32>> = (0..seq_length)
+        .map(|_| (0..dim).map(|_| rand::random::<f32>()).collect())
+        .collect();
+    let keys: Vec<Vec<f32>> = (0..seq_length)
+        .map(|_| (0..dim).map(|_| rand::random::<f32>()).collect())
+        .collect();
+    let values: Vec<Vec<f32>> = (0..seq_length)
+        .map(|_| (0..dim).map(|_| rand::random::<f32>()).collect())
+        .collect();
+
+    let keys_refs: Vec<&[f32]> = keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = values.iter().map(|v| v.as_slice()).collect();
+
+    // Warmup
+    for _ in 0..warmup {
+        run_attention(attention_type, dim, &query, &keys_refs, &values_refs)?;
+    }
+
+    // Actual benchmark
+    let mut timings = Vec::new();
+    for _ in 0..iterations {
+        let start = Instant::now();
+        run_attention(attention_type, dim, &query, &keys_refs, &values_refs)?;
+        let elapsed = start.elapsed();
+        timings.push(elapsed.as_secs_f64() * 1000.0);
+    }
+
+    Ok(timings)
+}
+
+fn run_attention(
+    attention_type: &str,
+    dim: usize,
+    query: &[Vec<f32>],
+    keys: &[&[f32]],
+    values: &[&[f32]],
+) -> anyhow::Result<Vec<Vec<f32>>> {
+    match attention_type {
+        "scaled_dot" => {
+            let attention = ScaledDotProductAttention::new(dim, None);
+            attention.compute(query, keys, values)
+        }
+        "multi_head" => {
+            let attention = MultiHeadAttention::new(dim, 8)?;
+            attention.compute(query, keys, values)
+        }
+        "hyperbolic" => {
+            let attention = HyperbolicAttention::new(dim, 1.0)?;
+            attention.compute(query, keys, values)
+        }
+        "flash" => {
+            let attention = FlashAttention::new(dim, 64)?;
+            attention.compute(query, keys, values)
+        }
+        "linear" => {
+            let attention = LinearAttention::new(dim)?;
+            attention.compute(query, keys, values)
+        }
+        "moe" => {
+            let attention = MoEAttention::new(dim, 4, 2)?;
+            attention.compute(query, keys, values)
+        }
+        _ => Err(anyhow::anyhow!("Unknown attention type: {}", attention_type)),
+    }
+}

vendor/ruvector/crates/ruvector-attention-cli/src/commands/compute.rs

@@ -0,0 +1,183 @@
+use clap::Args;
+use crate::{config::Config, output::{Output, OutputFormat, OutputDimensions, OutputMetadata}};
+use ruvector_attention::{
+    attention::{ScaledDotProductAttention, MultiHeadAttention},
+    hyperbolic::HyperbolicAttention,
+    sparse::{FlashAttention, LinearAttention},
+    moe::MoEAttention,
+};
+use std::time::Instant;
+
+#[derive(Args)]
+pub struct ComputeArgs {
+    /// Input file (JSON/binary/msgpack)
+    #[arg(short, long)]
+    input: std::path::PathBuf,
+
+    /// Output file (optional, prints to stdout if not specified)
+    #[arg(short, long)]
+    output: Option<std::path::PathBuf>,
+
+    /// Attention type
+    #[arg(short, long, default_value = "scaled_dot")]
+    attention_type: AttentionType,
+
+    /// Number of attention heads (for multi-head attention)
+    #[arg(long, default_value = "8")]
+    num_heads: usize,
+
+    /// Number of experts (for MoE attention)
+    #[arg(long, default_value = "4")]
+    num_experts: usize,
+
+    /// Top-k experts (for MoE attention)
+    #[arg(long, default_value = "2")]
+    top_k: usize,
+
+    /// Curvature (for hyperbolic attention)
+    #[arg(long, default_value = "1.0")]
+    curvature: f32,
+
+    /// Output format
+    #[arg(short, long, default_value = "pretty")]
+    format: OutputFormat,
+
+    /// Show detailed metrics
+    #[arg(short, long)]
+    verbose: bool,
+}
+
+#[derive(Clone, clap::ValueEnum)]
+pub enum AttentionType {
+    ScaledDot,
+    MultiHead,
+    Hyperbolic,
+    Flash,
+    Linear,
+    MoE,
+}
+
+pub async fn run(args: ComputeArgs, config: &Config) -> anyhow::Result<()> {
+    tracing::info!("Loading input from {:?}", args.input);
+    let input_data = super::load_input(&args.input)?;
+
+    tracing::info!(
+        "Input dimensions: query={:?}, keys={}, values={}",
+        input_data.query.len(),
+        input_data.keys.len(),
+        input_data.values.len()
+    );
+
+    let start = Instant::now();
+    let (result, attention_type_str) = match args.attention_type {
+        AttentionType::ScaledDot => {
+            tracing::info!("Computing scaled dot-product attention");
+            let attention = ScaledDotProductAttention::new(input_data.dim, None);
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "ScaledDotProduct")
+        }
+        AttentionType::MultiHead => {
+            tracing::info!("Computing multi-head attention with {} heads", args.num_heads);
+            let attention = MultiHeadAttention::new(input_data.dim, args.num_heads)?;
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "MultiHead")
+        }
+        AttentionType::Hyperbolic => {
+            tracing::info!("Computing hyperbolic attention with curvature={}", args.curvature);
+            let attention = HyperbolicAttention::new(input_data.dim, args.curvature)?;
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "Hyperbolic")
+        }
+        AttentionType::Flash => {
+            tracing::info!("Computing flash attention");
+            let attention = FlashAttention::new(input_data.dim, 64)?;
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "Flash")
+        }
+        AttentionType::Linear => {
+            tracing::info!("Computing linear attention");
+            let attention = LinearAttention::new(input_data.dim)?;
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "Linear")
+        }
+        AttentionType::MoE => {
+            tracing::info!(
+                "Computing MoE attention with {} experts, top-{}",
+                args.num_experts,
+                args.top_k
+            );
+            let attention = MoEAttention::new(
+                input_data.dim,
+                args.num_experts,
+                args.top_k
+            )?;
+            let result = attention.compute(
+                &input_data.query,
+                &input_data.keys_refs(),
+                &input_data.values_refs()
+            )?;
+            (result, "MixtureOfExperts")
+        }
+    };
+
+    let elapsed = start.elapsed();
+
+    if args.verbose {
+        tracing::info!("Computation completed in {:.2}ms", elapsed.as_secs_f64() * 1000.0);
+    }
+
+    let dimensions = OutputDimensions {
+        batch_size: 1,
+        num_heads: args.num_heads,
+        seq_length: input_data.keys.len(),
+        embedding_dim: input_data.dim,
+    };
+
+    let metadata = OutputMetadata {
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        memory_bytes: estimate_memory_usage(&result),
+        num_parameters: calculate_parameters(&args, input_data.dim),
+    };
+
+    let output = Output::new(attention_type_str, dimensions, result, metadata);
+    output.write(args.output.as_deref(), args.format)?;
+
+    tracing::info!("Output written successfully");
+
+    Ok(())
+}
+
+fn estimate_memory_usage(result: &[Vec<f32>]) -> usize {
+    result.iter().map(|row| row.len() * std::mem::size_of::<f32>()).sum()
+}
+
+fn calculate_parameters(args: &ComputeArgs, dim: usize) -> usize {
+    match args.attention_type {
+        AttentionType::ScaledDot => dim * dim * 3, // Q, K, V projections
+        AttentionType::MultiHead => dim * dim * 3 * args.num_heads + dim * dim, // + output projection
+        AttentionType::Hyperbolic => dim * dim * 3 + dim, // + curvature params
+        AttentionType::Flash => dim * dim * 3,
+        AttentionType::Linear => dim * dim * 2, // Linear projections
+        AttentionType::MoE => dim * dim * 3 * args.num_experts + dim * args.num_experts, // + router
+    }
+}

vendor/ruvector/crates/ruvector-attention-cli/src/commands/convert.rs

@@ -0,0 +1,166 @@
+use clap::Args;
+use crate::config::Config;
+use serde::{Deserialize, Serialize};
+
+#[derive(Args)]
+pub struct ConvertArgs {
+    /// Input file
+    #[arg(short, long)]
+    input: std::path::PathBuf,
+
+    /// Output file
+    #[arg(short, long)]
+    output: std::path::PathBuf,
+
+    /// Input format (auto-detect if not specified)
+    #[arg(long)]
+    from: Option<DataFormat>,
+
+    /// Output format
+    #[arg(long)]
+    to: DataFormat,
+
+    /// Pretty print output (for text formats)
+    #[arg(short, long)]
+    pretty: bool,
+}
+
+#[derive(Clone, clap::ValueEnum)]
+pub enum DataFormat {
+    Json,
+    Binary,
+    MsgPack,
+    Csv,
+    Npy,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct Data {
+    values: Vec<Vec<f32>>,
+}
+
+pub fn run(args: ConvertArgs, _config: &Config) -> anyhow::Result<()> {
+    tracing::info!("Converting from {:?} to {:?}", args.input, args.output);
+
+    // Read input
+    let content = std::fs::read(&args.input)?;
+    let data = parse_data(&content, args.from.as_ref())?;
+
+    // Write output
+    write_data(&args.output, &data, &args.to, args.pretty)?;
+
+    tracing::info!("Conversion complete");
+
+    Ok(())
+}
+
+fn parse_data(content: &[u8], format: Option<&DataFormat>) -> anyhow::Result<Data> {
+    if let Some(fmt) = format {
+        match fmt {
+            DataFormat::Json => Ok(serde_json::from_slice(content)?),
+            DataFormat::Binary => Ok(bincode::deserialize(content)?),
+            DataFormat::MsgPack => Ok(rmp_serde::from_slice(content)?),
+            DataFormat::Csv => parse_csv(content),
+            DataFormat::Npy => parse_npy(content),
+        }
+    } else {
+        // Auto-detect
+        if let Ok(data) = serde_json::from_slice::<Data>(content) {
+            return Ok(data);
+        }
+        if let Ok(data) = rmp_serde::from_slice::<Data>(content) {
+            return Ok(data);
+        }
+        if let Ok(data) = bincode::deserialize::<Data>(content) {
+            return Ok(data);
+        }
+
+        Err(anyhow::anyhow!("Failed to auto-detect format"))
+    }
+}
+
+fn write_data(
+    path: &std::path::Path,
+    data: &Data,
+    format: &DataFormat,
+    pretty: bool,
+) -> anyhow::Result<()> {
+    match format {
+        DataFormat::Json => {
+            let content = if pretty {
+                serde_json::to_string_pretty(data)?
+            } else {
+                serde_json::to_string(data)?
+            };
+            std::fs::write(path, content)?;
+        }
+        DataFormat::Binary => {
+            let bytes = bincode::serialize(data)?;
+            std::fs::write(path, bytes)?;
+        }
+        DataFormat::MsgPack => {
+            let bytes = rmp_serde::to_vec(data)?;
+            std::fs::write(path, bytes)?;
+        }
+        DataFormat::Csv => {
+            let csv = to_csv(data)?;
+            std::fs::write(path, csv)?;
+        }
+        DataFormat::Npy => {
+            let npy = to_npy(data)?;
+            std::fs::write(path, npy)?;
+        }
+    }
+
+    Ok(())
+}
+
+fn parse_csv(content: &[u8]) -> anyhow::Result<Data> {
+    let text = std::str::from_utf8(content)?;
+    let mut values = Vec::new();
+
+    for line in text.lines().skip(1) { // Skip header
+        let row: Vec<f32> = line.split(',')
+            .filter_map(|s| s.trim().parse().ok())
+            .collect();
+        if !row.is_empty() {
+            values.push(row);
+        }
+    }
+
+    Ok(Data { values })
+}
+
+fn to_csv(data: &Data) -> anyhow::Result<String> {
+    let mut csv = String::new();
+
+    // Write header
+    if let Some(first_row) = data.values.first() {
+        csv.push_str("row");
+        for i in 0..first_row.len() {
+            csv.push_str(&format!(",col_{}", i));
+        }
+        csv.push('\n');
+    }
+
+    // Write data
+    for (i, row) in data.values.iter().enumerate() {
+        csv.push_str(&i.to_string());
+        for val in row {
+            csv.push_str(&format!(",{}", val));
+        }
+        csv.push('\n');
+    }
+
+    Ok(csv)
+}
+
+fn parse_npy(_content: &[u8]) -> anyhow::Result<Data> {
+    // Simplified NPY parsing (real implementation would use a proper NPY library)
+    Err(anyhow::anyhow!("NPY parsing not yet implemented"))
+}
+
+fn to_npy(_data: &Data) -> anyhow::Result<Vec<u8>> {
+    // Simplified NPY writing (real implementation would use a proper NPY library)
+    Err(anyhow::anyhow!("NPY writing not yet implemented"))
+}

vendor/ruvector/crates/ruvector-attention-cli/src/commands/mod.rs

@@ -0,0 +1,66 @@
+pub mod compute;
+pub mod benchmark;
+pub mod convert;
+pub mod serve;
+pub mod repl;
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct InputData {
+    pub query: Vec<Vec<f32>>,
+    pub keys: Vec<Vec<f32>>,
+    pub values: Vec<Vec<f32>>,
+    pub dim: usize,
+}
+
+impl InputData {
+    pub fn keys_refs(&self) -> Vec<&[f32]> {
+        self.keys.iter().map(|k| k.as_slice()).collect()
+    }
+
+    pub fn values_refs(&self) -> Vec<&[f32]> {
+        self.values.iter().map(|v| v.as_slice()).collect()
+    }
+}
+
+pub fn load_input(path: &std::path::Path) -> anyhow::Result<InputData> {
+    let content = std::fs::read(path)?;
+
+    // Try to parse as JSON first
+    if let Ok(data) = serde_json::from_slice::<InputData>(&content) {
+        return Ok(data);
+    }
+
+    // Try MessagePack
+    if let Ok(data) = rmp_serde::from_slice::<InputData>(&content) {
+        return Ok(data);
+    }
+
+    // Try bincode
+    if let Ok(data) = bincode::deserialize::<InputData>(&content) {
+        return Ok(data);
+    }
+
+    Err(anyhow::anyhow!("Failed to parse input file"))
+}
+
+pub fn save_output(path: &std::path::Path, data: &[Vec<f32>], format: &str) -> anyhow::Result<()> {
+    match format {
+        "json" => {
+            let json = serde_json::to_string_pretty(data)?;
+            std::fs::write(path, json)?;
+        }
+        "msgpack" => {
+            let bytes = rmp_serde::to_vec(data)?;
+            std::fs::write(path, bytes)?;
+        }
+        "binary" => {
+            let bytes = bincode::serialize(data)?;
+            std::fs::write(path, bytes)?;
+        }
+        _ => return Err(anyhow::anyhow!("Unsupported format: {}", format)),
+    }
+
+    Ok(())
+}

vendor/ruvector/crates/ruvector-attention-cli/src/commands/repl.rs

@@ -0,0 +1,220 @@
+use clap::Args;
+use crate::config::Config;
+use rustyline::{Editor, error::ReadlineError, history::DefaultHistory};
+use ruvector_attention::{
+    attention::{ScaledDotProductAttention, MultiHeadAttention},
+    hyperbolic::HyperbolicAttention,
+    sparse::{FlashAttention, LinearAttention},
+    moe::MoEAttention,
+};
+
+#[derive(Args)]
+pub struct ReplArgs {
+    /// Initial dimension
+    #[arg(short, long, default_value = "512")]
+    dim: usize,
+}
+
+enum Command {
+    Help,
+    Load(String),
+    Compute(ComputeArgs),
+    SetType(String),
+    SetDim(usize),
+    Config,
+    Quit,
+    Unknown(String),
+}
+
+struct ComputeArgs {
+    query: Vec<Vec<f32>>,
+    keys: Vec<Vec<f32>>,
+    values: Vec<Vec<f32>>,
+}
+
+struct ReplState {
+    config: Config,
+    dim: usize,
+    attention_type: String,
+    last_query: Option<Vec<Vec<f32>>>,
+    last_keys: Option<Vec<Vec<f32>>>,
+    last_values: Option<Vec<Vec<f32>>>,
+}
+
+impl ReplState {
+    fn new(config: &Config, dim: usize) -> anyhow::Result<Self> {
+        Ok(Self {
+            config: config.clone(),
+            dim,
+            attention_type: config.attention.default_type.clone(),
+            last_query: None,
+            last_keys: None,
+            last_values: None,
+        })
+    }
+
+    fn load(&mut self, path: &str) -> anyhow::Result<()> {
+        let data = super::load_input(&std::path::Path::new(path))?;
+        self.last_query = Some(data.query);
+        self.last_keys = Some(data.keys);
+        self.last_values = Some(data.values);
+        self.dim = data.dim;
+        println!("Loaded data from {}", path);
+        Ok(())
+    }
+
+    fn compute(&self, args: &ComputeArgs) -> anyhow::Result<Vec<Vec<f32>>> {
+        let keys_refs: Vec<&[f32]> = args.keys.iter().map(|k| k.as_slice()).collect();
+        let values_refs: Vec<&[f32]> = args.values.iter().map(|v| v.as_slice()).collect();
+
+        match self.attention_type.as_str() {
+            "scaled_dot" => {
+                let attention = ScaledDotProductAttention::new(self.dim, None);
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            "multi_head" => {
+                let attention = MultiHeadAttention::new(self.dim, self.config.attention.default_heads)?;
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            "hyperbolic" => {
+                let attention = HyperbolicAttention::new(self.dim, 1.0)?;
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            "flash" => {
+                let attention = FlashAttention::new(self.dim, 64)?;
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            "linear" => {
+                let attention = LinearAttention::new(self.dim)?;
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            "moe" => {
+                let attention = MoEAttention::new(self.dim, 4, 2)?;
+                attention.compute(&args.query, &keys_refs, &values_refs)
+            }
+            _ => Err(anyhow::anyhow!("Unknown attention type: {}", self.attention_type)),
+        }
+    }
+
+    fn set_attention_type(&mut self, attention_type: String) {
+        self.attention_type = attention_type;
+        println!("Attention type set to: {}", self.attention_type);
+    }
+
+    fn set_dim(&mut self, dim: usize) {
+        self.dim = dim;
+        println!("Dimension set to: {}", self.dim);
+    }
+
+    fn config(&self) -> &Config {
+        &self.config
+    }
+}
+
+pub async fn run(args: ReplArgs, config: &Config) -> anyhow::Result<()> {
+    let mut rl = Editor::<(), DefaultHistory>::new()?;
+
+    println!("RuVector Attention REPL v{}", env!("CARGO_PKG_VERSION"));
+    println!("Type 'help' for commands, 'quit' to exit\n");
+
+    let mut state = ReplState::new(config, args.dim)?;
+
+    loop {
+        match rl.readline("attention> ") {
+            Ok(line) => {
+                if line.trim().is_empty() {
+                    continue;
+                }
+
+                rl.add_history_entry(&line)?;
+
+                match parse_command(&line) {
+                    Command::Help => print_help(),
+                    Command::Load(path) => {
+                        if let Err(e) = state.load(&path) {
+                            eprintln!("Error loading file: {}", e);
+                        }
+                    }
+                    Command::Compute(args) => {
+                        match state.compute(&args) {
+                            Ok(result) => {
+                                println!("Result shape: {}x{}", result.len(), result.first().map(|r| r.len()).unwrap_or(0));
+                                println!("First row (first 5 values): {:?}",
+                                    result.first().map(|r| &r[..5.min(r.len())]));
+                            }
+                            Err(e) => eprintln!("Error computing attention: {}", e),
+                        }
+                    }
+                    Command::SetType(t) => state.set_attention_type(t),
+                    Command::SetDim(d) => state.set_dim(d),
+                    Command::Config => println!("{:#?}", state.config()),
+                    Command::Quit => break,
+                    Command::Unknown(cmd) => println!("Unknown command: '{}'. Type 'help' for available commands.", cmd),
+                }
+            }
+            Err(ReadlineError::Interrupted) | Err(ReadlineError::Eof) => break,
+            Err(err) => return Err(err.into()),
+        }
+    }
+
+    println!("Goodbye!");
+    Ok(())
+}
+
+fn parse_command(line: &str) -> Command {
+    let parts: Vec<&str> = line.trim().split_whitespace().collect();
+
+    if parts.is_empty() {
+        return Command::Unknown(String::new());
+    }
+
+    match parts[0] {
+        "help" | "h" | "?" => Command::Help,
+        "load" => {
+            if parts.len() > 1 {
+                Command::Load(parts[1].to_string())
+            } else {
+                Command::Unknown("load requires a file path".to_string())
+            }
+        }
+        "compute" => {
+            // For simplicity, use random data
+            let query = vec![vec![0.1, 0.2, 0.3]];
+            let keys = vec![vec![0.1, 0.2, 0.3], vec![0.4, 0.5, 0.6]];
+            let values = vec![vec![0.7, 0.8, 0.9], vec![1.0, 1.1, 1.2]];
+            Command::Compute(ComputeArgs { query, keys, values })
+        }
+        "type" => {
+            if parts.len() > 1 {
+                Command::SetType(parts[1].to_string())
+            } else {
+                Command::Unknown("type requires an attention type".to_string())
+            }
+        }
+        "dim" => {
+            if parts.len() > 1 {
+                if let Ok(d) = parts[1].parse() {
+                    Command::SetDim(d)
+                } else {
+                    Command::Unknown("dim requires a number".to_string())
+                }
+            } else {
+                Command::Unknown("dim requires a dimension".to_string())
+            }
+        }
+        "config" => Command::Config,
+        "quit" | "exit" | "q" => Command::Quit,
+        _ => Command::Unknown(parts[0].to_string()),
+    }
+}
+
+fn print_help() {
+    println!("Available commands:");
+    println!("  help                  - Show this help message");
+    println!("  load <file>          - Load input data from file");
+    println!("  compute              - Compute attention with loaded data");
+    println!("  type <type>          - Set attention type (scaled_dot, multi_head, hyperbolic, flash, linear, moe)");
+    println!("  dim <size>           - Set dimension size");
+    println!("  config               - Show current configuration");
+    println!("  quit                 - Exit REPL");
+}

vendor/ruvector/crates/ruvector-attention-cli/src/commands/serve.rs

@@ -0,0 +1,290 @@
+use clap::Args;
+use crate::config::Config;
+use axum::{
+    routing::{get, post},
+    Router, Json, extract::State,
+    http::StatusCode,
+    response::IntoResponse,
+};
+use serde::{Deserialize, Serialize};
+use std::sync::Arc;
+use tower_http::cors::CorsLayer;
+use ruvector_attention::{
+    attention::{ScaledDotProductAttention, MultiHeadAttention},
+    hyperbolic::HyperbolicAttention,
+    sparse::{FlashAttention, LinearAttention},
+    moe::MoEAttention,
+};
+
+#[derive(Args)]
+pub struct ServeArgs {
+    /// Host address
+    #[arg(short = 'H', long, default_value = "0.0.0.0")]
+    host: String,
+
+    /// Port number
+    #[arg(short, long, default_value = "8080")]
+    port: u16,
+
+    /// Enable CORS
+    #[arg(long)]
+    cors: bool,
+}
+
+struct ServerState {
+    config: Config,
+}
+
+#[derive(Debug, Deserialize)]
+struct AttentionRequest {
+    query: Vec<Vec<f32>>,
+    keys: Vec<Vec<f32>>,
+    values: Vec<Vec<f32>>,
+    #[serde(default)]
+    num_heads: Option<usize>,
+    #[serde(default)]
+    num_experts: Option<usize>,
+    #[serde(default)]
+    top_k: Option<usize>,
+    #[serde(default)]
+    curvature: Option<f32>,
+}
+
+#[derive(Debug, Serialize)]
+struct AttentionResponse {
+    result: Vec<Vec<f32>>,
+    compute_time_ms: f64,
+    metadata: ResponseMetadata,
+}
+
+#[derive(Debug, Serialize)]
+struct ResponseMetadata {
+    attention_type: String,
+    dimensions: (usize, usize),
+}
+
+#[derive(Debug, Serialize)]
+struct ErrorResponse {
+    error: String,
+}
+
+pub async fn run(args: ServeArgs, config: &Config) -> anyhow::Result<()> {
+    let state = Arc::new(ServerState {
+        config: config.clone(),
+    });
+
+    let mut app = Router::new()
+        .route("/health", get(health))
+        .route("/attention/scaled_dot", post(scaled_dot_attention))
+        .route("/attention/multi_head", post(multi_head_attention))
+        .route("/attention/hyperbolic", post(hyperbolic_attention))
+        .route("/attention/flash", post(flash_attention))
+        .route("/attention/linear", post(linear_attention))
+        .route("/attention/moe", post(moe_attention))
+        .route("/batch", post(batch_compute))
+        .with_state(state);
+
+    if args.cors {
+        app = app.layer(CorsLayer::permissive());
+    }
+
+    let addr = format!("{}:{}", args.host, args.port);
+    tracing::info!("Starting server at http://{}", addr);
+
+    let listener = tokio::net::TcpListener::bind(&addr).await?;
+    axum::serve(listener, app).await?;
+
+    Ok(())
+}
+
+async fn health() -> impl IntoResponse {
+    Json(serde_json::json!({
+        "status": "healthy",
+        "version": env!("CARGO_PKG_VERSION")
+    }))
+}
+
+async fn scaled_dot_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let attention = ScaledDotProductAttention::new(dim, None);
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: "ScaledDotProduct".to_string(),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn multi_head_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let num_heads = req.num_heads.unwrap_or(8);
+    let attention = MultiHeadAttention::new(dim, num_heads)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: format!("MultiHead({})", num_heads),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn hyperbolic_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let curvature = req.curvature.unwrap_or(1.0);
+    let attention = HyperbolicAttention::new(dim, curvature)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: "Hyperbolic".to_string(),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn flash_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let attention = FlashAttention::new(dim, 64)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: "Flash".to_string(),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn linear_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let attention = LinearAttention::new(dim)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: "Linear".to_string(),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn moe_attention(
+    State(_state): State<Arc<ServerState>>,
+    Json(req): Json<AttentionRequest>,
+) -> Result<Json<AttentionResponse>, (StatusCode, Json<ErrorResponse>)> {
+    let start = std::time::Instant::now();
+
+    let dim = req.query.first().map(|q| q.len()).unwrap_or(0);
+    let num_experts = req.num_experts.unwrap_or(4);
+    let top_k = req.top_k.unwrap_or(2);
+    let attention = MoEAttention::new(dim, num_experts, top_k)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let keys_refs: Vec<&[f32]> = req.keys.iter().map(|k| k.as_slice()).collect();
+    let values_refs: Vec<&[f32]> = req.values.iter().map(|v| v.as_slice()).collect();
+
+    let result = attention.compute(&req.query, &keys_refs, &values_refs)
+        .map_err(|e| error_response(e.to_string()))?;
+
+    let elapsed = start.elapsed();
+
+    Ok(Json(AttentionResponse {
+        result,
+        compute_time_ms: elapsed.as_secs_f64() * 1000.0,
+        metadata: ResponseMetadata {
+            attention_type: format!("MoE({}/{})", top_k, num_experts),
+            dimensions: (req.query.len(), dim),
+        },
+    }))
+}
+
+async fn batch_compute(
+    State(_state): State<Arc<ServerState>>,
+    Json(_req): Json<serde_json::Value>,
+) -> Result<Json<serde_json::Value>, (StatusCode, Json<ErrorResponse>)> {
+    Err(error_response("Batch compute not yet implemented".to_string()))
+}
+
+fn error_response(message: String) -> (StatusCode, Json<ErrorResponse>) {
+    (
+        StatusCode::INTERNAL_SERVER_ERROR,
+        Json(ErrorResponse { error: message }),
+    )
+}