Remove --increased-determinism CLI option but leave library change.

Author: KerfuffleV2

llama-cli/src/cli_args.rs

@@ -1,4 +1,4 @@
-use clap::{Parser, ValueEnum};
+use clap::Parser;
 use llama_rs::TokenBias;
 use once_cell::sync::Lazy;
 
@@ -103,26 +103,6 @@ pub struct Args {
     /// and then as a list of comma seperated string keys and token ID values.
     #[arg(long, default_value_t = false)]
     pub dump_prompt_tokens: bool,
-
-    /// Increase determinism at the possible cost of speed/memory. If set to auto, it
-    /// will be enabled if the --seed option is specified for more deterministic
-    /// results (otherwise changing the thread count or batch size can produce
-    /// different output).
-    #[arg(long, default_value_t = IncreasedDeterminismSetting::Auto, value_enum)]
-    pub increased_determinism: IncreasedDeterminismSetting,
-}
-
-#[derive(Copy, Clone, Debug, PartialEq, Eq, PartialOrd, Ord, ValueEnum)]
-pub enum IncreasedDeterminismSetting {
-    Auto,
-    On,
-    Off,
-}
-
-impl IncreasedDeterminismSetting {
-    pub fn is_enabled(&self, have_seed: bool) -> bool {
-        (have_seed && self != &Self::Off) || self == &Self::On
-    }
 }
 
 fn parse_bias(s: &str) -> Result<TokenBias, String> {

llama-cli/src/main.rs

@@ -112,8 +112,7 @@ fn main() {
                 TokenBias::default()
             }
         }),
-
-        increased_determinism: args.increased_determinism.is_enabled(args.seed.is_some()),
+        ..Default::default()
     };
     let inference_session_params = {
         let mem_typ = if args.float16 {

llama-rs/src/lib.rs

@@ -149,7 +149,7 @@ impl Default for InferenceParameters {
             repeat_penalty: 1.30,
             temp: 0.80,
             bias_tokens: TokenBias::default(),
-            increased_determinism: false,
+            increased_determinism: true,
         }
     }
 }