[None][chore] DO NOT REVIEW, test PR

requirements.txt

@@ -30,7 +30,7 @@ nvidia-modelopt[torch]~=0.37.0
 # torch 2.10.0+cu130 depends on nvidia-nccl-cu13==2.28.9
 nvidia-nccl-cu13>=2.28.9,<=2.29.2
 nvidia-cuda-nvrtc
-transformers==4.57.3
+transformers==5.3.0
 prometheus_client
 prometheus_fastapi_instrumentator
 pydantic>=2.9.1

tensorrt_llm/_torch/auto_deploy/export/library/transformers_sdpa_mask.py

@@ -1,6 +1,7 @@
 """Patch for transformers SDPA mask to be export-compatible."""
 
 import importlib.metadata
+from functools import partial
 
 from packaging import version
 
@@ -29,7 +30,14 @@ def _apply_patch(self):
         try:
             # imports only after version check
             from transformers import masking_utils
-            from transformers.integrations.executorch import sdpa_mask_without_vmap
+
+            # Up to ~4.53+, HF exposed this helper next to ExecuTorch export utilities.
+            # Transformers 5.x removed it; sdpa_mask now supports use_vmap=False (the default),
+            # which is export-compatible without vmap.
+            try:
+                from transformers.integrations.executorch import sdpa_mask_without_vmap
+            except ImportError:
+                sdpa_mask_without_vmap = partial(masking_utils.sdpa_mask, use_vmap=False)
 
             # recall original implementation
             self.original_values["masking_utils.sdpa_mask"] = masking_utils.sdpa_mask

tensorrt_llm/_torch/models/hf_parameter_utils.py

@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Compatibility for Hugging Face ``get_parameter_device`` / ``get_parameter_dtype``.
+
+Transformers v5 no longer exports these from ``transformers.modeling_utils``; they
+match ``ModuleUtilsMixin`` behavior for plain ``nn.Module`` stacks.
+"""
+
+from __future__ import annotations
+
+import torch.nn as nn
+
+try:
+    from transformers.modeling_utils import (get_parameter_device,
+                                             get_parameter_dtype)
+except ImportError:
+
+    def get_parameter_device(module: nn.Module) -> torch.device:
+        return next(module.parameters()).device
+
+    def get_parameter_dtype(module: nn.Module) -> torch.dtype:
+        return next(
+            param.dtype
+            for param in module.parameters()
+            if param.is_floating_point())

tensorrt_llm/_torch/models/modeling_clip.py

@@ -4,13 +4,12 @@
 import torch.nn as nn
 from transformers.activations import ACT2FN
 from transformers.modeling_outputs import BaseModelOutput
-from transformers.modeling_utils import (get_parameter_device,
-                                         get_parameter_dtype)
 from transformers.models.clip.configuration_clip import CLIPVisionConfig
 from transformers.models.clip.modeling_clip import CLIPVisionEmbeddings
 
 from tensorrt_llm._utils import prefer_pinned
 
+from .hf_parameter_utils import get_parameter_device, get_parameter_dtype
 from ..attention_backend.interface import (AttentionMetadata,
                                            PredefinedAttentionMask)
 from ..attention_backend.utils import get_attention_backend

tensorrt_llm/_torch/models/modeling_siglip.py

@@ -2,14 +2,13 @@
 
 import torch
 import torch.nn as nn
-from transformers.modeling_utils import (get_parameter_device,
-                                         get_parameter_dtype)
 from transformers.models.siglip.configuration_siglip import SiglipVisionConfig
 from transformers.models.siglip.modeling_siglip import (SiglipVisionConfig,
                                                         SiglipVisionEmbeddings)
 
 from tensorrt_llm._utils import prefer_pinned
 
+from .hf_parameter_utils import get_parameter_device, get_parameter_dtype
 from ..attention_backend.interface import AttentionMetadata
 from ..attention_backend.utils import get_attention_backend
 from ..model_config import ModelConfig

tensorrt_llm/_torch/visual_gen/models/wan/transformer_wan.py

@@ -6,8 +6,7 @@
 import torch.nn.functional as F
 from diffusers.models.embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
 from tqdm import tqdm
-from transformers.modeling_utils import get_parameter_device
-
+from tensorrt_llm._torch.models.hf_parameter_utils import get_parameter_device
 from tensorrt_llm._torch.modules.layer_norm import LayerNorm
 from tensorrt_llm._torch.modules.linear import Linear
 from tensorrt_llm._torch.modules.mlp import MLP

tensorrt_llm/models/gpt/convert.py

@@ -29,8 +29,13 @@
 import torch.nn as nn
 import yaml
 from tqdm import tqdm
-from transformers import (AutoModelForCausalLM, AutoModelForVision2Seq,
-                          AutoTokenizer)
+try:
+    from transformers import AutoModelForVision2Seq
+except ImportError:
+    # Transformers v5+: vision-to-seq auto models use AutoModelForImageTextToText
+    from transformers import AutoModelForImageTextToText as AutoModelForVision2Seq
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
 from transformers.models.gpt2.modeling_gpt2 import GPT2Block
 from transformers.pytorch_utils import Conv1D
 

tensorrt_llm/tools/multimodal_builder.py

@@ -14,10 +14,15 @@
 from tensorrt_llm._utils import torch_dtype_to_str, to_json_file
 from tensorrt_llm.builder import Builder
 from tensorrt_llm.logger import logger
+try:
+    from transformers import AutoModelForVision2Seq
+except ImportError:
+    # Transformers v5+: vision-to-seq auto models use AutoModelForImageTextToText
+    from transformers import AutoModelForImageTextToText as AutoModelForVision2Seq
+
 from transformers import (AutoConfig, AutoModel, AutoModelForCausalLM,
-                          AutoModelForVision2Seq, AutoProcessor,
-                          Blip2ForConditionalGeneration, Blip2Processor,
-                          FuyuForCausalLM, FuyuProcessor,
+                          AutoProcessor, Blip2ForConditionalGeneration,
+                          Blip2Processor, FuyuForCausalLM, FuyuProcessor,
                           LlavaForConditionalGeneration, NougatProcessor,
                           Pix2StructForConditionalGeneration,
                           VisionEncoderDecoderModel, CLIPVisionModel)