Fix test bugs (#5484)

Author: tastelikefeet

examples/export/quantize/bert/gptq.sh

@@ -4,6 +4,13 @@ CUDA_VISIBLE_DEVICES=0 swift export \
     --output_dir output/swift_test_bert_merged \
     --merge_lora true
 
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -ne 0 ]; then
+    echo "Error: LoRA merge failed with exit code $EXIT_CODE"
+    exit $EXIT_CODE
+fi
+
 # gptq quantize
 CUDA_VISIBLE_DEVICES=0 swift export \
     --model output/swift_test_bert_merged \
@@ -13,6 +20,14 @@ CUDA_VISIBLE_DEVICES=0 swift export \
     --quant_method gptq \
     --max_length 512
 
+
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -ne 0 ]; then
+    echo "Error: GPTQ quantization failed with exit code $EXIT_CODE"
+    exit $EXIT_CODE
+fi
+
 # infer
 CUDA_VISIBLE_DEVICES=0 swift infer \
     --model output/swift_test_bert_gptq_int4

swift/llm/argument/base_args/base_args.py

@@ -104,7 +104,8 @@ def _prepare_training_args(self, training_args: Dict[str, Any]) -> None:
 
     def _init_lazy_tokenize(self):
         if self.lazy_tokenize is None:
-            if self.model_meta.is_multimodal and not self.streaming and not self.packing:
+            if (self.model_meta is not None and self.model_meta.is_multimodal and not self.streaming
+                    and not self.packing):
                 self.lazy_tokenize = True
             else:
                 self.lazy_tokenize = False
@@ -172,7 +173,7 @@ def __post_init__(self):
         QuantizeArguments.__post_init__(self)
         TemplateArguments.__post_init__(self)
         DataArguments.__post_init__(self)
-        if self.max_length is None:
+        if self.max_length is None and self.model_info is not None:
             self.max_length = self.model_info.max_model_len
         if self.packing and self.packing_length is None:
             self.packing_length = self.max_length

swift/llm/argument/base_args/template_args.py

@@ -47,7 +47,7 @@ class TemplateArguments:
     template_backend: Literal['swift', 'jinja'] = 'swift'
 
     def __post_init__(self):
-        if self.template is None and hasattr(self, 'model_meta'):
+        if self.template is None and getattr(self, 'model_meta', None):
             self.template = self.model_meta.template
         if self.use_chat_template is None:
             self.use_chat_template = True

swift/llm/argument/sampling_args.py

@@ -56,6 +56,9 @@ class SamplingArguments(BaseArguments):
     def _init_model_info(self):
         if self.sampler_engine != 'client':
             return super()._init_model_info()
+        else:
+            self.model_info = None
+            self.model_meta = None
         self.task_type = 'causal_lm'
         return
 

swift/llm/train/tuner.py

@@ -243,7 +243,7 @@ def prepare_adapter(args: TrainArguments, model, *, template=None, train_dataset
         logger.info(f'adalora_config: {adalora_config}')
     elif args.train_type == 'llamapro':
         llamapro_config = LLaMAProConfig(
-            model_type=model.model_meta.model_arch,
+            model_type=model.model_meta.model_arch.arch_name,
             num_new_blocks=args.llamapro_num_new_blocks,
             num_groups=args.llamapro_num_groups)
         model = Swift.prepare_model(model, llamapro_config)