Support both tokenizor.json and vocab.json+merges.txt formats

1 files changed, 28 insertions, 4 deletions

diff --git a/makima/src/tts/qwen3/mod.rs b/makima/src/tts/qwen3/mod.rs
index 9bac794..1520be6 100644
--- a/makima/src/tts/qwen3/mod.rs
+++ b/makima/src/tts/qwen3/mod.rs
@@ -99,10 +99,34 @@ impl Qwen3Tts {
             Qwen3TtsConfig::default()
         };
 
-        // Load text tokenizer
-        let tokenizer_path = model_dir.join("tokenizer.json");
-        let tokenizer = Tokenizer::from_file(&tokenizer_path)
-            .map_err(|e| TtsError::Tokenizer(format!("failed to load tokenizer: {e}")))?;
+        // Load text tokenizer (supports both tokenizer.json and vocab.json+merges.txt formats)
+        let tokenizer_json_path = model_dir.join("tokenizer.json");
+        let tokenizer = if tokenizer_json_path.exists() {
+            Tokenizer::from_file(&tokenizer_json_path)
+                .map_err(|e| TtsError::Tokenizer(format!("failed to load tokenizer.json: {e}")))?
+        } else {
+            // Fall back to vocab.json + merges.txt (HuggingFace Qwen3-TTS format)
+            let vocab_path = model_dir.join("vocab.json");
+            let merges_path = model_dir.join("merges.txt");
+
+            if !vocab_path.exists() || !merges_path.exists() {
+                return Err(TtsError::Tokenizer(format!(
+                    "tokenizer files not found: need either tokenizer.json or vocab.json+merges.txt in {}",
+                    model_dir.display()
+                )));
+            }
+
+            tokenizers::Tokenizer::from_file(&vocab_path)
+                .or_else(|_| {
+                    // Build BPE tokenizer from vocab and merges
+                    use tokenizers::models::bpe::BPE;
+                    let bpe = BPE::from_file(&vocab_path.to_string_lossy(), &merges_path.to_string_lossy())
+                        .build()
+                        .map_err(|e| TtsError::Tokenizer(format!("failed to build BPE tokenizer: {e}")))?;
+                    Ok(Tokenizer::new(bpe))
+                })
+                .map_err(|e: TtsError| TtsError::Tokenizer(format!("failed to load tokenizer: {e}")))?
+        };
 
         // Load LM weights from safetensors
         let lm_weights_path = model_dir.join("model.safetensors");