Revert "perf: stop generation on default turn boundaries"

This reverts commit caf4654c0a2a856c8f71dd465d4a4b1d67cde5bc.

crates/inference-core/src/session.rs

@@ -27,8 +27,6 @@ pub struct GenerationSession {
     vocab: *const llama_vocab,
     sampler: *mut llama_sys::llama_sampler,
     stop_sequences: Vec<String>,
-    stop_suffix_len: usize,
-    generated_tail: String,
 
     // generation state
     n_decoded: i32, // number of tokens decoded so far (prefill + gen)
@@ -165,21 +163,12 @@ impl GenerationSession {
         );
 
         // create session state
-        let stop_suffix_len = sampler_cfg
-            .stop_sequences
-            .iter()
-            .map(|stop| stop.len())
-            .max()
-            .unwrap_or(0);
-
         Ok(Self {
             lib,
             ctx,
             vocab,
             sampler: sampler_chain, // use sampler_chain here
             stop_sequences: sampler_cfg.stop_sequences.clone(),
-            stop_suffix_len,
-            generated_tail: String::new(),
             n_decoded: n_prompt_tokens,
             prompt_tokens: n_prompt_tokens as u32,
             max_tokens: sampler_cfg.max_tokens,
@@ -238,10 +227,14 @@ impl GenerationSession {
             let token_text = self.detokenize_token(new_token);
 
             // Check stop sequences
-            if self.matches_stop_sequence(&token_text) {
-                self.finished = true;
-                self.finalize_timing();
-                return None;
+            // (simple check: if generated text so far ends with a stop seq)
+            // For streaming we check per-token — good enough for Phase 2.
+            for stop in &self.stop_sequences {
+                if token_text.trim_end() == stop.as_str() {
+                    self.finished = true;
+                    self.finalize_timing();
+                    return None;
+                }
             }
 
             Some(token_text)
@@ -298,28 +291,6 @@ impl GenerationSession {
                 .into_owned()
         }
     }
-
-    fn matches_stop_sequence(&mut self, token_text: &str) -> bool {
-        if self.stop_sequences.is_empty() {
-            return false;
-        }
-
-        self.generated_tail.push_str(token_text);
-        tail_ends_with_stop(
-            &mut self.generated_tail,
-            self.stop_suffix_len,
-            &self.stop_sequences,
-        )
-    }
-}
-
-fn tail_ends_with_stop(tail: &mut String, max_len: usize, stop_sequences: &[String]) -> bool {
-    if tail.len() > max_len {
-        let trim_from = tail.len() - max_len;
-        tail.drain(..trim_from);
-    }
-
-    stop_sequences.iter().any(|stop| tail.ends_with(stop))
 }
 
 fn decode_token_piece_with<F>(mut call: F) -> String
@@ -347,7 +318,7 @@ where
 
 #[cfg(test)]
 mod tests {
-    use super::{decode_token_piece_with, tail_ends_with_stop};
+    use super::decode_token_piece_with;
 
     #[test]
     fn retries_when_piece_buffer_is_too_small() {
@@ -393,22 +364,6 @@ mod tests {
         assert_eq!(decoded, " Quantum computer");
         assert_eq!(calls, 2);
     }
-
-    #[test]
-    fn matches_stop_sequence_across_multiple_tokens() {
-        let stops = vec!["\nAssistant:".to_owned()];
-        let mut tail = String::new();
-
-        tail.push_str("\nAssis");
-        assert!(!tail_ends_with_stop(
-            &mut tail,
-            "\nAssistant:".len(),
-            &stops
-        ));
-
-        tail.push_str("tant:");
-        assert!(tail_ends_with_stop(&mut tail, "\nAssistant:".len(), &stops));
-    }
 }
 
 impl Drop for GenerationSession {