satisfy compiler

backends/client/src/v2/client.rs

@@ -143,6 +143,7 @@ impl Client {
                     seed: 0,
                     repetition_penalty: 1.2,
                     frequency_penalty: 0.1,
+                    no_repeat_ngram_size: 0,
                     watermark: true,
                     grammar: String::new(),
                     grammar_type: GrammarType::None as i32,

backends/client/src/v2/sharded_client.rs

@@ -228,6 +228,7 @@ impl Health for ShardedClient {
                 seed: 0,
                 repetition_penalty: 1.0,
                 frequency_penalty: 0.0,
+                no_repeat_ngram_size: 0,
                 watermark: false,
                 grammar: String::new(),
                 grammar_type: GrammarType::None as i32,

backends/client/src/v3/client.rs

@@ -166,6 +166,7 @@ impl Client {
                     seed: 0,
                     repetition_penalty: 1.2,
                     frequency_penalty: 0.1,
+                    no_repeat_ngram_size: 0,
                     watermark: true,
                     grammar: String::new(),
                     grammar_type: GrammarType::None as i32,

backends/client/src/v3/sharded_client.rs

@@ -231,6 +231,7 @@ impl Health for ShardedClient {
                 seed: 0,
                 repetition_penalty: 1.0,
                 frequency_penalty: 0.0,
+                no_repeat_ngram_size: 0,
                 watermark: false,
                 grammar: String::new(),
                 grammar_type: GrammarType::None as i32,

backends/v3/src/queue.rs

@@ -454,6 +454,7 @@ impl From<ValidParameters> for NextTokenChooserParameters {
             seed: value.seed,
             repetition_penalty: value.repetition_penalty,
             frequency_penalty: value.frequency_penalty,
+            no_repeat_ngram_size: value.no_repeat_ngram_size,
             watermark: value.watermark,
             grammar,
             grammar_type: grammar_type.into(),
@@ -497,6 +498,7 @@ mod tests {
                     seed: 0,
                     repetition_penalty: 0.0,
                     frequency_penalty: 0.0,
+                    no_repeat_ngram_size: 0,
                     watermark: false,
                     grammar: None,
                 },

benchmark/src/lib.rs

@@ -31,6 +31,7 @@ pub async fn run(
     typical_p: Option<f32>,
     repetition_penalty: Option<f32>,
     frequency_penalty: Option<f32>,
+    no_repeat_ngram_size: Option<u32>,
     watermark: bool,
     do_sample: bool,
     client: ShardedClient,
@@ -44,6 +45,7 @@ pub async fn run(
         seed: 0,
         repetition_penalty: repetition_penalty.unwrap_or(1.0),
         frequency_penalty: frequency_penalty.unwrap_or(0.0),
+        no_repeat_ngram_size: no_repeat_ngram_size.unwrap_or(0),
         watermark,
         grammar: String::new(),
         grammar_type: GrammarType::None as i32,
@@ -145,6 +147,7 @@ pub async fn run(
         typical_p,
         repetition_penalty,
         frequency_penalty,
+        no_repeat_ngram_size,
         watermark,
         do_sample,
     );

benchmark/src/main.rs

@@ -89,6 +89,11 @@ struct Args {
     #[clap(long, env)]
     frequency_penalty: Option<f32>,
 
+    /// Generation parameter in case you want to specifically test/debug particular
+    /// decoding strategies, for full doc refer to the `text-generation-server`
+    #[clap(long, env)]
+    no_repeat_ngram_size: Option<u32>,
+
     /// Generation parameter in case you want to specifically test/debug particular
     /// decoding strategies, for full doc refer to the `text-generation-server`
     #[clap(long, env)]
@@ -125,6 +130,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
         typical_p,
         repetition_penalty,
         frequency_penalty,
+        no_repeat_ngram_size,
         watermark,
         do_sample,
         master_shard_uds_path,
@@ -196,6 +202,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
                 typical_p,
                 repetition_penalty,
                 frequency_penalty,
+                no_repeat_ngram_size,
                 watermark,
                 do_sample,
                 sharded_client,

router/src/infer/v2/queue.rs

@@ -377,6 +377,7 @@ impl From<ValidParameters> for NextTokenChooserParameters {
             seed: value.seed,
             repetition_penalty: value.repetition_penalty,
             frequency_penalty: value.frequency_penalty,
+            no_repeat_ngram_size: value.no_repeat_ngram_size,
             watermark: value.watermark,
             grammar,
             grammar_type: grammar_type.into(),
@@ -420,6 +421,7 @@ mod tests {
                     seed: 0,
                     repetition_penalty: 0.0,
                     frequency_penalty: 0.0,
+                    no_repeat_ngram_size: 0,
                     watermark: false,
                     grammar: None,
                 },

router/src/validation.rs

@@ -195,6 +195,7 @@ impl Validation {
             temperature,
             repetition_penalty,
             frequency_penalty,
+            no_repeat_ngram_size,
             top_k,
             top_p,
             typical_p,
@@ -238,6 +239,8 @@ impl Validation {
             return Err(ValidationError::FrequencyPenalty);
         }
 
+        let no_repeat_ngram_size = no_repeat_ngram_size.unwrap_or(0);
+
         // Different because the proto default value is not a valid value
         // for the user
         let top_p = top_p
@@ -373,6 +376,7 @@ impl Validation {
             temperature,
             repetition_penalty,
             frequency_penalty,
+            no_repeat_ngram_size,
             top_k,
             top_p,
             typical_p,