Bump Transformers version to 5.15.0 (#51668)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

requirements/test/cpu.txt

@@ -1169,7 +1169,7 @@ tqdm==4.67.3
     #   segmentation-models-pytorch
     #   sentence-transformers
     #   transformers
-transformers==5.14.1
+transformers==5.15.0
     # via
     #   -r requirements/test/../common.txt
     #   -r requirements/test/cuda.in

requirements/test/cuda.in

@@ -39,7 +39,7 @@ open_clip_torch==2.32.0 # Required for nemotron_vl test, Nemotron Parse in test_
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
-transformers==5.14.1
+transformers==5.15.0
 tokenizers==0.22.2
 schemathesis>=4.0.0 # Required for openai schema test.
 # quantization

requirements/test/cuda.txt

@@ -1272,7 +1272,7 @@ tqdm==4.67.3
     #   segmentation-models-pytorch
     #   sentence-transformers
     #   transformers
-transformers==5.14.1
+transformers==5.15.0
     # via
     #   -c requirements/common.txt
     #   -r requirements/test/../common.txt

requirements/test/nightly-torch.txt

@@ -29,7 +29,7 @@ opencv-python-headless >= 4.13.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
-transformers==5.14.1
+transformers==5.15.0
 tokenizers==0.22.2
 schemathesis>=4.0.0 # Required for openai schema test.
 # quantization

requirements/test/rocm.in

@@ -35,7 +35,7 @@ open_clip_torch==2.32.0 # Required for nemotron_vl test, Nemotron Parse in test_
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
-transformers==5.14.1
+transformers==5.15.0
 tokenizers==0.22.2
 schemathesis>=4.0.0 # Required for openai schema test
 # quantization

requirements/test/rocm.txt

@@ -1222,7 +1222,7 @@ tqdm==4.67.3
     #   sentence-transformers
     #   tilelang
     #   transformers
-transformers==5.14.1
+transformers==5.15.0
     # via
     #   -c requirements/common.txt
     #   -r requirements/test/../common.txt

requirements/test/xpu.in

@@ -17,7 +17,7 @@ accelerate
 arctic-inference
 lm_eval[api]>=0.4.12
 modelscope<1.38
-transformers==5.14.1
+transformers==5.15.0
 
 # --- Audio Processing ---
 librosa

requirements/test/xpu.txt

@@ -940,7 +940,7 @@ tqdm==4.67.3
     #   pqdm
     #   sentence-transformers
     #   transformers
-transformers==5.14.1
+transformers==5.15.0
     # via
     #   -c requirements/common.txt
     #   -r requirements/test/../common.txt

tests/config/test_model_arch_config.py

@@ -371,19 +371,17 @@ def test_gemma4_uniform_head_dims_are_homogeneous():
 
 
 class _HeterogeneousConfig(PretrainedConfig):
-    """A stand-in for the Transformers >= 5.15.0 heterogeneous config API.
+    """A heterogeneous config with no convertor of its own.
 
-    No released Transformers has it, so the default per-layer seam has no other
-    way to be exercised. Mirrors the parts vLLM uses: per-layer configs are
-    shallow copies with the varying attributes applied and heterogeneity
-    stripped, so they do not recurse.
-    """
+    Mirrors the parts vLLM uses: per-layer configs are shallow copies with the
+    varying attributes applied and heterogeneity stripped, so they do not recurse."""
 
     is_heterogeneous = True
 
     def __init__(self, per_layer: dict[str, list], **kwargs):
-        super().__init__(**kwargs)
+        # Set before `super().__init__` because it validates `per_layer_config`
         self._per_layer = per_layer
+        super().__init__(**kwargs)
 
     @property
     def per_layer_config(self) -> list[PretrainedConfig]:

tests/models/transformers/test_backend.py

@@ -151,7 +151,7 @@ def test_mla(vllm_runner: type[VllmRunner], example_prompts: list[str]) -> None:
             f"transformers>={required}, but got {installed}"
         )
 
-    model = get_model("DeepseekV2ForCausalLM")  # DeepSeek-V2-Lite, MLA + MoE
+    model = "hmellor/tiny-random-DeepseekV2ForCausalLM"
     args = (example_prompts, 32, 5)
     kwargs: dict[str, Any] = {"max_model_len": 2048, "enforce_eager": True}
 

vllm/model_executor/models/transformers/multimodal.py

@@ -390,6 +390,14 @@ class MultiModalProcessor(BaseMultiModalProcessor[MultiModalProcessingInfo]):
         mm_placeholders: dict[str, list[PlaceholderRange]] = {}
         split_sizes = mm_tokens_per_modality["num_image_tokens"]
         if split_sizes:
+            image_token_ids = getattr(hf_processor, "image_token_ids", None)
+            if image_token_ids is None:
+                # Transformers <5.10.0
+                image_token_ids = [hf_processor.image_token_id]
+            image_token_ids = torch.tensor(
+                [i for i in image_token_ids if i is not None]
+            )
+
             chunked_mm_positions = torch.split(mm_positions, split_sizes)
             mm_tokens = torch.tensor(prompt_ids)[mm_token_type_ids[0].bool()]
             chunked_mm_tokens = torch.split(mm_tokens, split_sizes)
@@ -397,7 +405,7 @@ class MultiModalProcessor(BaseMultiModalProcessor[MultiModalProcessingInfo]):
                 PlaceholderRange(
                     offset=positions[0].item(),
                     length=positions.shape[0],
-                    is_embed=(mm_tokens == hf_processor.image_token_id).bool(),
+                    is_embed=torch.isin(mm_tokens, image_token_ids),
                 )
                 for positions, mm_tokens in zip(chunked_mm_positions, chunked_mm_tokens)
             ]