fix tokenizer setting in tests

rllm-cuda/src/llm/phi.rs

@@ -216,8 +216,14 @@ impl TModelInner for MixFormerSequentialForCausalLM {
 
         // it should approximately match...
         let tok_size = self.config.tok_vocab_size as i64;
-        assert!(r.size()[1] >= tok_size);
-        assert!(r.size()[1] < tok_size + 1000);
+        if r.size()[1] < tok_size || r.size()[1] > tok_size + 1000 {
+            panic!(
+                "unexpected logits size: {:?} ({}/{})",
+                r.size(),
+                tok_size,
+                self.config.vocab_size
+            );
+        }
 
         let r = r.i((.., 0..tok_size));
         batch_info.extract_positions(&r)

rllm-cuda/src/server/mod.rs

@@ -601,13 +601,11 @@ pub async fn server_main(mut args: RllmCliArgs) -> () {
         loader_args.dtype = dtype;
     }
 
-    if args.test.len() > 0 {
-        run_tests(&args, loader_args);
-        return;
-    }
-
     match &args.tokenizer {
-        Some(v) => loader_args.tokenizer = v.clone(),
+        Some(v) => {
+            log::info!("explicit tokenizer: {}", v);
+            loader_args.tokenizer = v.clone();
+        }
         None => match guess_tokenizer(&loader_args.model_id) {
             Some(v) => {
                 log::info!("guessed tokenizer: {}", v);
@@ -621,6 +619,11 @@ pub async fn server_main(mut args: RllmCliArgs) -> () {
         },
     }
 
+    if args.test.len() > 0 {
+        run_tests(&args, loader_args);
+        return;
+    }
+
     let (tokenizer, tok_trie) =
         RllmEngine::load_tokenizer(&mut loader_args).expect("failed to load tokenizer");