Upgrade transformers from 5.5.4 to 5.6.0 (#23525)

python/pyproject.toml

@@ -72,7 +72,7 @@ dependencies = [
   "torchvision",
   "tqdm",
   "mistral_common>=1.11.0",
-  "transformers==5.5.4",
+  "transformers==5.6.0",
   "uvicorn",
   "uvloop",
   "watchfiles",

python/pyproject_cpu.toml

@@ -62,7 +62,7 @@ dependencies = [
   "torchvision==0.24.0",
   "tqdm",
   "mistral_common>=1.11.0",
-  "transformers==5.5.4",
+  "transformers==5.6.0",
   "triton==3.5.0",
   "uvicorn",
   "uvloop",

python/pyproject_npu.toml

@@ -59,7 +59,7 @@ dependencies = [
   "torchao==0.9.0",
   "tqdm",
   "mistral_common>=1.11.0",
-  "transformers==5.5.4",
+  "transformers==5.6.0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.32",

python/pyproject_other.toml

@@ -59,7 +59,7 @@ runtime_common = [
   "torchao==0.9.0",
   "tqdm",
   "mistral_common>=1.11.0",
-  "transformers==5.5.4",
+  "transformers==5.6.0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.32",

python/pyproject_xpu.toml

@@ -63,7 +63,7 @@ dependencies = [
   "torchao==0.9.0+xpu",
   "tqdm",
   "mistral_common>=1.11.0",
-  "transformers==5.5.4",
+  "transformers==5.6.0",
   "uvicorn",
   "uvloop",
   # "xgrammar==0.1.24", , xgrammar depends on CUDA PyTorch and Triton only

python/sglang/multimodal_gen/runtime/models/encoders/gemma_3.py

@@ -931,6 +931,16 @@ class Gemma3TextModel(nn.Module):
 
 
 class Gemma3ForConditionalGeneration(nn.Module):
+    # transformers 5.6.0 flattened SiglipVisionModel, dropping the
+    # `vision_model` intermediate wrapper. Our reimpl keeps it, so remap
+    # HF source keys back into our nested namespace when transferring weights.
+    param_names_mapping = {
+        r"^(vision_tower\.)(embeddings|encoder|post_layernorm|head)\.": r"\1vision_model.\2.",
+    }
+    reverse_param_names_mapping = {
+        r"^(vision_tower\.)vision_model\.(embeddings|encoder|post_layernorm|head)\.": r"\1\2.",
+    }
+
     def __init__(
         self,
         config: Gemma3Config,

python/sglang/srt/models/llava.py

@@ -495,6 +495,9 @@ class LlavaBaseForCausalLM(nn.Module):
             "model.mm_projector.0": "multi_modal_projector.linear_1",
             "model.mm_projector.2": "multi_modal_projector.linear_2",
             "model.vision_tower.vision_tower": "vision_tower",
+            # transformers 5.6.0 flattened CLIPVisionModel/SiglipVisionModel,
+            # dropping the `vision_model` intermediate wrapper.
+            "vision_tower.vision_model.": "vision_tower.",
             # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
             "model.image_newline": "language_model.model.image_newline",
         }

python/sglang/srt/models/llavavid.py

@@ -255,6 +255,9 @@ class LlavaVidForCausalLM(nn.Module):
             "model.vision_resampler.mm_projector.0": "multi_modal_projector.linear_1",
             "model.vision_resampler.mm_projector.2": "multi_modal_projector.linear_2",
             "model.vision_tower.vision_tower": "vision_tower",
+            # transformers 5.6.0 flattened CLIPVisionModel/SiglipVisionModel,
+            # dropping the `vision_model` intermediate wrapper.
+            "vision_tower.vision_model.": "vision_tower.",
             # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
             "model.image_newline": "language_model.model.image_newline",
         }

python/sglang/srt/models/yivl.py

@@ -73,6 +73,9 @@ class YiVLForCausalLM(LlavaLlamaForCausalLM):
             "model.mm_projector.3": "multi_modal_projector.linear_2",
             "model.mm_projector.4": "multi_modal_projector.ln_2",
             "model.vision_tower.vision_tower": "vision_tower",  # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
+            # transformers 5.6.0 flattened CLIPVisionModel/SiglipVisionModel,
+            # dropping the `vision_model` intermediate wrapper.
+            "vision_tower.vision_model.": "vision_tower.",
         }
         params_dict = dict(self.named_parameters())
         weights = list(weights)