Merge branch 'hiyouga:main' into main

src/llamafactory/extras/constants.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
 from collections import OrderedDict, defaultdict
 from enum import Enum
 from typing import Dict, Optional
@@ -47,7 +48,7 @@
 
 IGNORE_INDEX = -100
 
-IMAGE_PLACEHOLDER = "<image>"
+IMAGE_PLACEHOLDER = os.environ.get("IMAGE_PLACEHOLDER", "<image>")
 
 LAYERNORM_NAMES = {"norm", "ln"}
 
@@ -95,7 +96,7 @@
 
 SUPPORTED_CLASS_FOR_S2ATTN = {"llama"}
 
-VIDEO_PLACEHOLDER = "<video>"
+VIDEO_PLACEHOLDER = os.environ.get("VIDEO_PLACEHOLDER", "<video>")
 
 V_HEAD_WEIGHTS_NAME = "value_head.bin"
 

src/llamafactory/model/loader.py

@@ -153,8 +153,9 @@ def load_model(
                 load_class = AutoModelForVision2Seq
             else:
                 load_class = AutoModelForCausalLM
+
             if model_args.train_from_scratch:
-                model = load_class.from_config(config)
+                model = load_class.from_config(config, trust_remote_code=True)
             else:
                 model = load_class.from_pretrained(**init_kwargs)
 

src/llamafactory/train/test_utils.py

@@ -37,9 +37,9 @@ def compare_model(model_a: "torch.nn.Module", model_b: "torch.nn.Module", diff_k
     assert set(state_dict_a.keys()) == set(state_dict_b.keys())
     for name in state_dict_a.keys():
         if any(key in name for key in diff_keys):
-            assert torch.allclose(state_dict_a[name], state_dict_b[name], rtol=1e-3, atol=1e-4) is False
+            assert torch.allclose(state_dict_a[name], state_dict_b[name], rtol=1e-2, atol=1e-3) is False
         else:
-            assert torch.allclose(state_dict_a[name], state_dict_b[name], rtol=1e-3, atol=1e-4) is True
+            assert torch.allclose(state_dict_a[name], state_dict_b[name], rtol=1e-2, atol=1e-3) is True
 
 
 def check_lora_model(model: "LoraModel") -> Tuple[Set[str], Set[str]]: