[Fix](nvbug6304585): specdec README online base-model example should use Instruct model

examples/speculative_decoding/README.md

@@ -78,7 +78,7 @@ For small base models that fit in GPU memory, we can collocate them with draft m
 ```bash
 ./launch_train.sh \
     --config ../../modelopt_recipes/general/speculative_decoding/eagle3.yaml \
-    model.model_name_or_path=meta-llama/Llama-3.2-1B \
+    model.model_name_or_path=meta-llama/Llama-3.2-1B-Instruct \
     data.data_path=input_conversations/train.jsonl \
     training.output_dir=ckpts/llama-3.2-1b-online
 ```
@@ -123,7 +123,7 @@ Once we finish dumping hidden states, launch offline training pointing to the hi
 ```bash
 ./launch_train.sh \
     --config ../../modelopt_recipes/general/speculative_decoding/eagle3.yaml \
-    model.model_name_or_path=meta-llama/Llama-3.2-1B \
+    model.model_name_or_path=meta-llama/Llama-3.2-1B-Instruct \
     data.offline_data_path=$HIDDEN_STATES_DIR \
     training.output_dir=ckpts/llama-3.2-1b-offline
 ```

modelopt/torch/utils/plugins/transformers_dataset.py

@@ -160,7 +160,12 @@ def __init__(
 
         self._post_process_tokenizer()
         if self.tokenizer.chat_template is None:
-            raise ValueError("No valid chat template!")
+            raise ValueError(
+                "No valid chat template! The tokenizer for this model has no chat_template, "
+                "which is common for base / pretrained checkpoints. Use an instruction-tuned "
+                "model (e.g. a '*-Instruct' variant), or pass a custom chat_template via the "
+                "training config or LanguageDataCollator(chat_template=...)."
+            )
 
         if self.answer_only_loss:
             self._verify_generation_tags()