Let ModelRunner take InputMetadata as input, instead of ScheduleBatch (#1541)

test/srt/models/test_lora.py

@@ -15,7 +15,6 @@ limitations under the License.
 
 import multiprocessing as mp
 import unittest
-import uuid
 
 import torch
 
@@ -85,9 +84,9 @@ class TestLoRA(unittest.TestCase):
 
         with SRTRunner(
             base_path,
-            tp_size=tp_size,
             torch_dtype=torch_dtype,
-            is_generation=True,
+            model_type="generation",
+            tp_size=tp_size,
             lora_paths=all_lora_paths,
             max_loras_per_batch=3,
             disable_cuda_graph=True,

test/srt/run_suite.py

@@ -7,6 +7,7 @@ suites = {
     "minimal": [
         "models/test_embedding_models.py",
         "models/test_generation_models.py",
+        # "models/test_lora.py",
         "models/test_reward_models.py",
         "sampling/penaltylib",
         "test_chunked_prefill.py",