[version] support transformers 449 (#6982)

* support transformers 449

* fix mm plugin

Former-commit-id: e9118a9df0839d24f6ddff5a0b55ef101a1d3d22

README.md

@@ -390,7 +390,7 @@ huggingface-cli login
 | ------------ | ------- | --------- |
 | python       | 3.9     | 3.10      |
 | torch        | 1.13.1  | 2.4.0     |
-| transformers | 4.41.2  | 4.45.2    |
+| transformers | 4.41.2  | 4.49.0    |
 | datasets     | 2.16.0  | 3.2.0     |
 | accelerate   | 0.34.0  | 1.2.1     |
 | peft         | 0.11.1  | 0.12.0    |
@@ -399,9 +399,9 @@ huggingface-cli login
 | Optional     | Minimum | Recommend |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
-| deepspeed    | 0.10.0  | 0.14.0    |
+| deepspeed    | 0.10.0  | 0.16.2    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.6.6     |
+| vllm         | 0.4.3   | 0.7.2     |
 | flash-attn   | 2.3.0   | 2.7.2     |
 
 ### Hardware Requirement

README_zh.md

@@ -392,7 +392,7 @@ huggingface-cli login
 | ------------ | ------- | --------- |
 | python       | 3.9     | 3.10      |
 | torch        | 1.13.1  | 2.4.0     |
-| transformers | 4.41.2  | 4.45.2    |
+| transformers | 4.41.2  | 4.49.0    |
 | datasets     | 2.16.0  | 3.2.0     |
 | accelerate   | 0.34.0  | 1.2.1     |
 | peft         | 0.11.1  | 0.12.0    |
@@ -401,9 +401,9 @@ huggingface-cli login
 | 可选项       | 至少     | 推荐      |
 | ------------ | ------- | --------- |
 | CUDA         | 11.6    | 12.2      |
-| deepspeed    | 0.10.0  | 0.14.0    |
+| deepspeed    | 0.10.0  | 0.16.2    |
 | bitsandbytes | 0.39.0  | 0.43.1    |
-| vllm         | 0.4.3   | 0.6.6     |
+| vllm         | 0.4.3   | 0.7.2     |
 | flash-attn   | 2.3.0   | 2.7.2     |
 
 ### 硬件依赖

requirements.txt

@@ -1,5 +1,5 @@
-transformers>=4.41.2,<=4.48.3,!=4.46.*,!=4.47.*,!=4.48.0,!=4.48.1,!=4.48.2;python_version<'3.10'
+transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.*;python_version<'3.10'
-transformers>=4.41.2,<=4.48.3,!=4.46.*,!=4.47.*,!=4.48.0;python_version>='3.10'
+transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.0;python_version>='3.10'
 datasets>=2.16.0,<=3.2.0
 accelerate>=0.34.0,<=1.2.1
 peft>=0.11.1,<=0.12.0

src/llamafactory/__init__.py

@@ -20,7 +20,7 @@ Level:
 
 Dependency graph:
   main:
-    transformers>=4.41.2,<=4.48.3,!=4.46.*,!=4.47.*,!=4.48.0
+    transformers>=4.41.2,<=4.49.0,!=4.46.*,!=4.47.*,!=4.48.0
     datasets>=2.16.0,<=3.2.0
     accelerate>=0.34.0,<=1.2.1
     peft>=0.11.1,<=0.12.0

src/llamafactory/data/collator.py

@@ -187,8 +187,6 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
             mm_inputs["cross_attention_mask"] = F.pad(cross_attention_mask, (0, 0, 0, 0, 0, seq_len - orig_len))
 
         features.update(mm_inputs)
-        if isinstance(features.get("pixel_values"), list):  # for pixtral inputs
-            features = features.data  # use default_collate() instead of BatchEncoding.to()
 
         if "image_bound" in features:  # for minicpmv inputs
             bsz, seq_length = features["input_ids"].shape

src/llamafactory/data/mm_plugin.py

@@ -380,10 +380,8 @@ class LlavaNextPlugin(BasePlugin):
         num_image_tokens = 0
         messages = deepcopy(messages)
         mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
-        if "image_sizes" in mm_inputs:
-            image_sizes = iter(mm_inputs["image_sizes"])
-
         if "pixel_values" in mm_inputs:
+            image_sizes = iter(mm_inputs["image_sizes"].tolist())
             height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
 
         for message in messages:
@@ -439,7 +437,7 @@ class LlavaNextVideoPlugin(BasePlugin):
         messages = deepcopy(messages)
         mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
         if "pixel_values" in mm_inputs:
-            image_sizes = iter(mm_inputs["image_sizes"])
+            image_sizes = iter(mm_inputs["image_sizes"].tolist())
             height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
             for message in messages:
                 content = message["content"]
@@ -916,16 +914,14 @@ class PixtralPlugin(BasePlugin):
         num_image_tokens = 0
         messages = deepcopy(messages)
         mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
-        image_input_sizes = mm_inputs.get("image_sizes", None)
+        if "pixel_values" in mm_inputs:
+            image_sizes = iter(mm_inputs["image_sizes"].tolist())
+
         for message in messages:
             content = message["content"]
             while IMAGE_PLACEHOLDER in content:
-                if image_input_sizes is None:
-                    raise ValueError("Cannot get image input sizes.")
-
                 if self.expand_mm_tokens:
-                    image_size = image_input_sizes[0][num_image_tokens]
+                    height, width = next(image_sizes)
-                    height, width = image_size
                     num_height_tokens = height // patch_size
                     num_width_tokens = width // patch_size
                     replace_tokens = [[image_token] * num_width_tokens + [image_break_token]] * num_height_tokens
@@ -959,9 +955,6 @@ class PixtralPlugin(BasePlugin):
     ) -> Dict[str, Union[List[int], "torch.Tensor"]]:
         self._validate_input(images, videos, audios)
         mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
-        if mm_inputs.get("pixel_values"):
-            mm_inputs["pixel_values"] = mm_inputs["pixel_values"][0]
-
         mm_inputs.pop("image_sizes", None)
         return mm_inputs
 

src/llamafactory/extras/misc.py

@@ -94,7 +94,7 @@ def check_dependencies() -> None:
     r"""
     Checks the version of the required packages.
     """
-    check_version("transformers>=4.41.2,<=4.48.3,!=4.46.0,!=4.46.1,!=4.46.2,!=4.46.3,!=4.47.0,!=4.47.1,!=4.48.0")
+    check_version("transformers>=4.41.2,<=4.49.0,!=4.46.0,!=4.46.1,!=4.46.2,!=4.46.3,!=4.47.0,!=4.47.1,!=4.48.0")
     check_version("datasets>=2.16.0,<=3.2.0")
     check_version("accelerate>=0.34.0,<=1.2.1")
     check_version("peft>=0.11.1,<=0.12.0")