From f35bea1fb475ce291c2c5db851b9ed232ea85677 Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 13:45:54 -0700
Subject: [PATCH 1/9] Bump diffusers to >=0.37 and drop older-diffusers code
 paths

0.37 is the first release with QwenEmbedRope's max_txt_seq_len, which the Qwen-Image
FastGen plugin passes. Also raise huggingface_hub to >=1.5 and accelerate to >=1.1,
which transformers>=5.5 already requires, and remove the unused
skip_if_onnx_version_above_1_18 helper.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 CHANGELOG.rst                                 |  1 +
 .../diffusers/quantization/models_utils.py    | 14 +----
 examples/diffusers/quantization/utils.py      |  2 +-
 modelopt/torch/export/unified_export_hf.py    |  2 +-
 .../plugins/diffusion/diffusers.py            | 55 ++++++----------
 pyproject.toml                                |  6 +-
 tests/_test_utils/import_helper.py            | 17 -----
 tests/_test_utils/torch/diffusers_models.py   | 62 +++----------------
 8 files changed, 36 insertions(+), 123 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 90aa9a89b4c..8295ddb6788 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -47,6 +47,7 @@ Changelog
 - ``export_mcore_gpt_to_hf`` given a Hugging Face Hub model ID no longer writes the tokenizer, processor, generation config or remote-code files. Pass a local directory instead, e.g. from ``modelopt.torch.export.ensure_local_checkpoint``.
 - Drop support for Python 3.10, which has reached end-of-life. ModelOpt now requires Python 3.11 or later.
 - Bump minimum required PyTorch version to 2.10.
+- Bump minimum required ``diffusers`` version to 0.37 for the ``[hf]`` extra.
 - Diffusion HF exports now omit attention quantizer buffers by default, except for FLUX.1, which retains enabled Q/K/V ``_amax`` keys for runtime compatibility; other consumers of those keys can opt in with ``export_attention_quantizers=True`` (CLI: ``--extra-param export_attention_quantizers=true``). Disabled attention quantizers remain excluded, and diffusion components, including LTX-2, no longer emit KV-cache quantization metadata.
 - ``hf_quant_config.json`` of an INT4 or W4A8 AWQ checkpoint now sets ``pre_quant_scale`` per layer, only for layers that have the tensor. A model whose fused-MoE experts have none (AWQ never applies to them) is now exported as ``MIXED_PRECISION`` with per-layer entries instead of a uniform algorithm.
 - The Windows FVD example rejects I3D checkpoints requiring unrestricted pickle loading. Custom TorchScript archives now require ``--weights-sha256`` with a trusted checksum; tensor state dictionaries continue to load directly.
diff --git a/examples/diffusers/quantization/models_utils.py b/examples/diffusers/quantization/models_utils.py
index 7a0750bc9a3..2c5c2d12795 100644
--- a/examples/diffusers/quantization/models_utils.py
+++ b/examples/diffusers/quantization/models_utils.py
@@ -21,23 +21,13 @@
 import torch
 from diffusers import (
     DiffusionPipeline,
+    Flux2Pipeline,
     FluxPipeline,
     LTXConditionPipeline,
+    QwenImagePipeline,
     StableDiffusion3Pipeline,
     WanPipeline,
 )
-
-try:
-    from diffusers import Flux2Pipeline
-except ImportError:
-    Flux2Pipeline = None
-
-# Qwen-Image classes were added in a recent diffusers release; import lazily so
-# this example still imports on older diffusers versions.
-try:
-    from diffusers import QwenImagePipeline
-except ImportError:
-    QwenImagePipeline = None
 from utils import (
     filter_func_default,
     filter_func_flux_dev,
diff --git a/examples/diffusers/quantization/utils.py b/examples/diffusers/quantization/utils.py
index b7a79e49e70..24dcf4f355a 100644
--- a/examples/diffusers/quantization/utils.py
+++ b/examples/diffusers/quantization/utils.py
@@ -20,12 +20,12 @@
 import torch
 import torch.nn.functional as F
 from datasets import load_dataset
+from diffusers.models.attention import AttentionModuleMixin
 from diffusers.models.attention_processor import Attention, AttnProcessor
 from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
 from diffusers.utils import load_image
 
 import modelopt.torch.quantization as mtq
-from modelopt.torch.quantization.plugins.diffusion.diffusers import AttentionModuleMixin
 
 USE_PEFT = True
 try:
diff --git a/modelopt/torch/export/unified_export_hf.py b/modelopt/torch/export/unified_export_hf.py
index 14f3c82fd36..ffbd62ddfba 100644
--- a/modelopt/torch/export/unified_export_hf.py
+++ b/modelopt/torch/export/unified_export_hf.py
@@ -1355,7 +1355,7 @@ def _export_diffusers_checkpoint(
             # Image-conditioned Wan also requires a successful dummy forward for fusion.
             is_wan_i2v_component = (
                 HAS_DIFFUSERS
-                and isinstance(component, getattr(diffusers, "WanTransformer3DModel", ()))
+                and isinstance(component, diffusers.WanTransformer3DModel)
                 and getattr(getattr(component, "config", None), "image_dim", None) is not None
             )
             _fuse_qkv_linears_diffusion(component, strict=is_qwen_component or is_wan_i2v_component)
diff --git a/modelopt/torch/quantization/plugins/diffusion/diffusers.py b/modelopt/torch/quantization/plugins/diffusion/diffusers.py
index d096ddda7a3..1f3952dfadf 100644
--- a/modelopt/torch/quantization/plugins/diffusion/diffusers.py
+++ b/modelopt/torch/quantization/plugins/diffusion/diffusers.py
@@ -20,30 +20,18 @@
 from types import ModuleType
 from typing import TYPE_CHECKING
 
-import diffusers
 import onnx
 import torch
+from diffusers.models.attention_dispatch import AttentionBackendName, attention_backend
 from diffusers.models.attention_processor import Attention
 from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
-from packaging.version import parse as parse_version
-
-if parse_version(diffusers.__version__) >= parse_version("0.35.0"):
-    from diffusers.models.attention import AttentionModuleMixin
-    from diffusers.models.attention_dispatch import AttentionBackendName, attention_backend
-    from diffusers.models.transformers.transformer_flux import FluxAttention
-    from diffusers.models.transformers.transformer_ltx import LTXAttention
-    from diffusers.models.transformers.transformer_wan import WanAttention
-
-    try:
-        from diffusers.models.transformers.transformer_flux2 import (
-            Flux2Attention,
-            Flux2ParallelSelfAttention,
-        )
-    except ImportError:
-        Flux2Attention = None
-        Flux2ParallelSelfAttention = None
-else:
-    AttentionModuleMixin = type("_dummy_type_no_instance", (), {})  # pylint: disable=invalid-name
+from diffusers.models.transformers.transformer_flux import FluxAttention
+from diffusers.models.transformers.transformer_flux2 import (
+    Flux2Attention,
+    Flux2ParallelSelfAttention,
+)
+from diffusers.models.transformers.transformer_ltx import LTXAttention
+from diffusers.models.transformers.transformer_wan import WanAttention
 from torch.autograd import Function
 from torch.nn import functional as F
 from torch.onnx import symbolic_helper
@@ -193,24 +181,21 @@ def _setup(self):
 QuantModuleRegistry.register({Attention: "Attention"})(_QuantAttention)
 
 
-if AttentionModuleMixin.__module__.startswith(diffusers.__name__):
+class _QuantAttentionModuleMixin(_QuantAttention):
+    """Quantized AttentionModuleMixin for performing attention-related computations."""
 
-    class _QuantAttentionModuleMixin(_QuantAttention):
-        """Quantized AttentionModuleMixin for performing attention-related computations."""
+    def forward(self, *args, **kwargs):
+        with attention_backend(AttentionBackendName.NATIVE):
+            return super().forward(*args, **kwargs)
 
-        def forward(self, *args, **kwargs):
-            with attention_backend(AttentionBackendName.NATIVE):
-                return super().forward(*args, **kwargs)
 
-    QuantModuleRegistry.register({FluxAttention: "FluxAttention"})(_QuantAttentionModuleMixin)
-    QuantModuleRegistry.register({WanAttention: "WanAttention"})(_QuantAttentionModuleMixin)
-    QuantModuleRegistry.register({LTXAttention: "LTXAttention"})(_QuantAttentionModuleMixin)
-    if Flux2Attention is not None:
-        QuantModuleRegistry.register({Flux2Attention: "Flux2Attention"})(_QuantAttentionModuleMixin)
-    if Flux2ParallelSelfAttention is not None:
-        QuantModuleRegistry.register({Flux2ParallelSelfAttention: "Flux2ParallelSelfAttention"})(
-            _QuantAttentionModuleMixin
-        )
+QuantModuleRegistry.register({FluxAttention: "FluxAttention"})(_QuantAttentionModuleMixin)
+QuantModuleRegistry.register({WanAttention: "WanAttention"})(_QuantAttentionModuleMixin)
+QuantModuleRegistry.register({LTXAttention: "LTXAttention"})(_QuantAttentionModuleMixin)
+QuantModuleRegistry.register({Flux2Attention: "Flux2Attention"})(_QuantAttentionModuleMixin)
+QuantModuleRegistry.register({Flux2ParallelSelfAttention: "Flux2ParallelSelfAttention"})(
+    _QuantAttentionModuleMixin
+)
 
 
 original_scaled_dot_product_attention = F.scaled_dot_product_attention
diff --git a/pyproject.toml b/pyproject.toml
index 3c07d934ea1..d911a6e2260 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -76,12 +76,12 @@ onnx = [
     "onnxruntime-ep-nv-tensorrt-rtx-cu13==0.4.0; platform_system == 'Windows' and platform_machine == 'ARM64'",
 ]
 hf = [
-    "accelerate>=1.0.0",
+    "accelerate>=1.1.0",
     "datasets>=3.0.0",
     "deepspeed>=0.9.6; platform_system != 'Darwin' and platform_system != 'Windows'",
-    "diffusers>=0.32.2",
+    "diffusers>=0.37.0",
     "httpx>=0.23.0,<1",
-    "huggingface_hub>=0.24.0",
+    "huggingface_hub>=1.5.0",
     "nltk",
     "peft>=0.17.0",
     "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
diff --git a/tests/_test_utils/import_helper.py b/tests/_test_utils/import_helper.py
index a2b46a5b1c2..4480e050d1a 100644
--- a/tests/_test_utils/import_helper.py
+++ b/tests/_test_utils/import_helper.py
@@ -13,12 +13,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import ctypes
-import importlib.metadata
 import os
 import shutil
 
 import pytest
-from packaging import version
 
 
 def skip_if_no_tensorrt():
@@ -66,18 +64,3 @@ def skip_if_no_mamba():
         import mamba_ssm  # noqa: F401
     except ImportError:
         pytest.skip("Mamba required for Megatron test", allow_module_level=True)
-
-
-def skip_if_onnx_version_above_1_18():
-    package_name = "onnx"
-    required_version = "1.18.0"
-
-    try:
-        installed_version = importlib.metadata.version(package_name)
-    except importlib.metadata.PackageNotFoundError:
-        pytest.skip(f"{package_name} is not installed")
-
-    if version.parse(installed_version) > version.parse(required_version):
-        pytest.skip(
-            f"{package_name} version {installed_version} is greater than required {required_version}"
-        )
diff --git a/tests/_test_utils/torch/diffusers_models.py b/tests/_test_utils/torch/diffusers_models.py
index 849d1ae6f2e..b2954200ebb 100644
--- a/tests/_test_utils/torch/diffusers_models.py
+++ b/tests/_test_utils/torch/diffusers_models.py
@@ -21,37 +21,14 @@
 
 pytest.importorskip("diffusers")
 from diffusers import UNet2DConditionModel
-
-try:
-    from diffusers.models.transformers import DiTTransformer2DModel, FluxTransformer2DModel
-except Exception:  # pragma: no cover - optional diffusers models
-    DiTTransformer2DModel = None
-    FluxTransformer2DModel = None
-
-try:
-    from diffusers.models.transformers import Flux2Transformer2DModel
-except Exception:  # pragma: no cover - optional diffusers models
-    Flux2Transformer2DModel = None
-
-try:
-    from diffusers.models.transformers import WanTransformer3DModel
-except Exception:  # pragma: no cover - optional diffusers models
-    WanTransformer3DModel = None
-
-try:
-    from diffusers.models.autoencoders import AutoencoderKLWan
-except Exception:  # pragma: no cover - optional diffusers models
-    AutoencoderKLWan = None
-
-try:
-    from diffusers.models.transformers import QwenImageTransformer2DModel
-except Exception:  # pragma: no cover - optional diffusers models
-    QwenImageTransformer2DModel = None
-
-try:
-    from diffusers.models.autoencoders import AutoencoderKLQwenImage
-except Exception:  # pragma: no cover - optional diffusers models
-    AutoencoderKLQwenImage = None
+from diffusers.models.autoencoders import AutoencoderKLQwenImage, AutoencoderKLWan
+from diffusers.models.transformers import (
+    DiTTransformer2DModel,
+    Flux2Transformer2DModel,
+    FluxTransformer2DModel,
+    QwenImageTransformer2DModel,
+    WanTransformer3DModel,
+)
 
 import modelopt.torch.opt as mto
 
@@ -79,9 +56,6 @@ def get_tiny_unet(**config_kwargs) -> UNet2DConditionModel:
 
 def get_tiny_dit(**config_kwargs):
     """Create a tiny DiTTransformer2DModel for testing."""
-    if DiTTransformer2DModel is None:
-        pytest.skip("DiTTransformer2DModel is not available in this diffusers version.")
-
     kwargs = {
         "num_attention_heads": 2,
         "attention_head_dim": 8,
@@ -99,9 +73,6 @@ def get_tiny_dit(**config_kwargs):
 
 def get_tiny_flux(**config_kwargs):
     """Create a tiny FluxTransformer2DModel for testing."""
-    if FluxTransformer2DModel is None:
-        pytest.skip("FluxTransformer2DModel is not available in this diffusers version.")
-
     kwargs = {
         "patch_size": 1,
         "in_channels": 4,
@@ -121,9 +92,6 @@ def get_tiny_flux(**config_kwargs):
 
 def get_tiny_flux2(**config_kwargs):
     """Create a tiny Flux2Transformer2DModel for testing."""
-    if Flux2Transformer2DModel is None:
-        pytest.skip("Flux2Transformer2DModel is not available in this diffusers version.")
-
     kwargs = {
         "patch_size": 1,
         "in_channels": 16,
@@ -182,9 +150,6 @@ def df_modelopt_state_and_output_tester(model_ref, model_test):
 
 def get_tiny_wan22_transformer(**config_kwargs):
     """Create a tiny WanTransformer3DModel for testing."""
-    if WanTransformer3DModel is None:
-        pytest.skip("WanTransformer3DModel is not available in this diffusers version.")
-
     kwargs = {
         "patch_size": (1, 2, 2),
         "num_attention_heads": 2,
@@ -205,9 +170,6 @@ def get_tiny_wan22_transformer(**config_kwargs):
 
 def get_tiny_wan22_vae(**config_kwargs):
     """Create a tiny AutoencoderKLWan for testing."""
-    if AutoencoderKLWan is None:
-        pytest.skip("AutoencoderKLWan is not available in this diffusers version.")
-
     kwargs = {
         "base_dim": 3,
         "z_dim": 16,
@@ -324,9 +286,6 @@ def get_tiny_qwen_image_transformer(**config_kwargs):
       (the bundled text encoder is bypassed), so pair this with
       ``--data.dataloader.text_embed_dim=<joint_attention_dim>``.
     """
-    if QwenImageTransformer2DModel is None:
-        pytest.skip("QwenImageTransformer2DModel is not available in this diffusers version.")
-
     kwargs = {
         "patch_size": 2,
         "in_channels": 64,  # vae z_dim (16) * 2x2 patch
@@ -351,9 +310,6 @@ def get_tiny_qwen_image_transformer(**config_kwargs):
 
 def get_tiny_qwen_image_vae(**config_kwargs):
     """Create a tiny AutoencoderKLQwenImage for testing (z_dim=16 to match the transformer)."""
-    if AutoencoderKLQwenImage is None:
-        pytest.skip("AutoencoderKLQwenImage is not available in this diffusers version.")
-
     kwargs = {
         "base_dim": 8,
         "z_dim": 16,  # = transformer out_channels
@@ -438,8 +394,6 @@ def create_tiny_qwen_image_pipeline_dir(tmp_path: Path) -> Path:
     (16) so the pipeline runs end-to-end during quantization calibration. The saved
     dir loads with ``QwenImagePipeline.from_pretrained(path)``.
     """
-    if QwenImageTransformer2DModel is None or AutoencoderKLQwenImage is None:
-        pytest.skip("QwenImage diffusers classes not available in this diffusers version.")
     from diffusers import FlowMatchEulerDiscreteScheduler, QwenImagePipeline
 
     transformers = pytest.importorskip("transformers")

From 2e3c90b8ac82e058ec00059ba211882e46e9ff7d Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 14:16:34 -0700
Subject: [PATCH 2/9] Require diffusers>=0.39 for Qwen-Image PDD

diffusers 0.39 moved the joint text/image mask from the root forward into the
attention processor. The PDD forward replaces the root forward, so on older
diffusers padded text tokens leak into attention; fail loudly instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 modelopt/torch/fastgen/plugins/qwen_image_pdd.py       | 6 ++++++
 tests/unit/torch/fastgen/test_qwen_image_pdd_plugin.py | 9 ++++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/modelopt/torch/fastgen/plugins/qwen_image_pdd.py b/modelopt/torch/fastgen/plugins/qwen_image_pdd.py
index b8b512dde9a..ac88a4a66d0 100644
--- a/modelopt/torch/fastgen/plugins/qwen_image_pdd.py
+++ b/modelopt/torch/fastgen/plugins/qwen_image_pdd.py
@@ -40,6 +40,7 @@
 from typing import Any
 
 import torch
+from packaging.version import Version
 from torch import nn
 
 from ..config import PDDConfig
@@ -224,8 +225,13 @@ def enable_qwen_image_pdd_forward(transformer: nn.Module) -> nn.Module:
         return transformer
 
     # Diffusers is an optional dependency used only by the Qwen example.
+    import diffusers
     from diffusers import QwenImageTransformer2DModel
 
+    # Older diffusers builds the joint mask in the root forward, which this forward replaces.
+    if Version(diffusers.__version__) < Version("0.39"):
+        raise RuntimeError(f"Qwen-Image PDD requires diffusers>=0.39, got {diffusers.__version__}.")
+
     if not isinstance(transformer, QwenImageTransformer2DModel):
         raise TypeError(
             "Qwen PDD forward binding requires QwenImageTransformer2DModel, "
diff --git a/tests/unit/torch/fastgen/test_qwen_image_pdd_plugin.py b/tests/unit/torch/fastgen/test_qwen_image_pdd_plugin.py
index dc6a2bc89dd..4b5511e90e3 100644
--- a/tests/unit/torch/fastgen/test_qwen_image_pdd_plugin.py
+++ b/tests/unit/torch/fastgen/test_qwen_image_pdd_plugin.py
@@ -79,7 +79,7 @@ def _mr210_rollout_oracle(
 
 
 def _tiny_diffusers_qwen():
-    diffusers = pytest.importorskip("diffusers")
+    diffusers = pytest.importorskip("diffusers", minversion="0.39")
     return diffusers.QwenImageTransformer2DModel(
         patch_size=2,
         in_channels=8,
@@ -306,6 +306,13 @@ def test_qwen_pdd_forward_binding_preserves_root_state_and_deepcopy() -> None:
         enable_qwen_image_pdd_forward(conflicting)
 
 
+def test_qwen_pdd_forward_binding_rejects_old_diffusers(monkeypatch) -> None:
+    model = _tiny_diffusers_qwen()
+    monkeypatch.setattr("diffusers.__version__", "0.38.0")
+    with pytest.raises(RuntimeError, match=r"diffusers>=0\.39"):
+        enable_qwen_image_pdd_forward(model)
+
+
 def test_mr210_qwen_conversion_preserves_every_initialized_head() -> None:
     base = _tiny_diffusers_qwen().eval().to(torch.bfloat16)
     student = copy.deepcopy(base)

From 40981f876515d53ff11fe686b1e774a62f48be33 Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 14:16:51 -0700
Subject: [PATCH 3/9] Bump transformers to >=5.8 and nltk to >=3.9

Drop the transformers<5.6 expert-parallel test skips and the pre-punkt_tab nltk path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 CHANGELOG.rst                                   |  2 +-
 examples/llm_sparsity/weight_sparsity/eval.py   |  6 +-----
 modelopt/torch/__init__.py                      |  2 +-
 .../calibration/ruler_dataset.py                |  3 +--
 modelopt/torch/utils/moe.py                     |  5 ++---
 noxfile.py                                      |  2 +-
 pyproject.toml                                  |  4 ++--
 tests/gpu/torch/utils/test_moe_a2a.py           |  5 +----
 tests/unit/torch/utils/test_moe_a2a.py          | 17 ++++-------------
 9 files changed, 14 insertions(+), 32 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 8295ddb6788..8f135a88763 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -47,7 +47,7 @@ Changelog
 - ``export_mcore_gpt_to_hf`` given a Hugging Face Hub model ID no longer writes the tokenizer, processor, generation config or remote-code files. Pass a local directory instead, e.g. from ``modelopt.torch.export.ensure_local_checkpoint``.
 - Drop support for Python 3.10, which has reached end-of-life. ModelOpt now requires Python 3.11 or later.
 - Bump minimum required PyTorch version to 2.10.
-- Bump minimum required ``diffusers`` version to 0.37 for the ``[hf]`` extra.
+- Bump minimum required ``transformers`` version to 5.8, ``diffusers`` to 0.37 and ``nltk`` to 3.9 for the ``[hf]`` extra.
 - Diffusion HF exports now omit attention quantizer buffers by default, except for FLUX.1, which retains enabled Q/K/V ``_amax`` keys for runtime compatibility; other consumers of those keys can opt in with ``export_attention_quantizers=True`` (CLI: ``--extra-param export_attention_quantizers=true``). Disabled attention quantizers remain excluded, and diffusion components, including LTX-2, no longer emit KV-cache quantization metadata.
 - ``hf_quant_config.json`` of an INT4 or W4A8 AWQ checkpoint now sets ``pre_quant_scale`` per layer, only for layers that have the tensor. A model whose fused-MoE experts have none (AWQ never applies to them) is now exported as ``MIXED_PRECISION`` with per-layer entries instead of a uniform algorithm.
 - The Windows FVD example rejects I3D checkpoints requiring unrestricted pickle loading. Custom TorchScript archives now require ``--weights-sha256`` with a trusted checksum; tensor state dictionaries continue to load directly.
diff --git a/examples/llm_sparsity/weight_sparsity/eval.py b/examples/llm_sparsity/weight_sparsity/eval.py
index a5f2fb91b2d..3aea2104e2c 100644
--- a/examples/llm_sparsity/weight_sparsity/eval.py
+++ b/examples/llm_sparsity/weight_sparsity/eval.py
@@ -25,7 +25,6 @@
 import torch
 import transformers
 from accelerate import Accelerator
-from packaging.version import Version
 from torch.utils.data import DataLoader, Dataset
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -244,10 +243,7 @@ def main():
 
     accelerator = Accelerator()
     with accelerator.main_process_first():
-        if Version(nltk.__version__) < Version("3.8.2"):
-            nltk.download("punkt")
-        else:
-            nltk.download("punkt_tab")
+        nltk.download("punkt_tab")
 
     tokenizer = prepare_tokenizer(accelerator, args.model_dir, args.model_max_length)
     dataset = get_dataset(accelerator, args.data_path, tokenizer)
diff --git a/modelopt/torch/__init__.py b/modelopt/torch/__init__.py
index 1182304f541..1caebbfbe3c 100644
--- a/modelopt/torch/__init__.py
+++ b/modelopt/torch/__init__.py
@@ -48,7 +48,7 @@
 try:
     from transformers import __version__ as _transformers_version
 
-    if _Version(_transformers_version) < _Version("5.5") or _Version(
+    if _Version(_transformers_version) < _Version("5.8") or _Version(
         _transformers_version
     ) >= _Version("5.19"):
         _warnings.warn(
diff --git a/modelopt/torch/sparsity/attention_sparsity/calibration/ruler_dataset.py b/modelopt/torch/sparsity/attention_sparsity/calibration/ruler_dataset.py
index abbbc399d6b..e3a0cb526b6 100644
--- a/modelopt/torch/sparsity/attention_sparsity/calibration/ruler_dataset.py
+++ b/modelopt/torch/sparsity/attention_sparsity/calibration/ruler_dataset.py
@@ -268,9 +268,8 @@ def generate_niah_sample(
     from nltk.tokenize import sent_tokenize
 
     try:
-        nltk.data.find("tokenizers/punkt")
+        nltk.data.find("tokenizers/punkt_tab")
     except LookupError:
-        nltk.download("punkt", quiet=True)
         nltk.download("punkt_tab", quiet=True)
 
     if random_seed is not None:
diff --git a/modelopt/torch/utils/moe.py b/modelopt/torch/utils/moe.py
index 80924b72f26..3351b68f7a1 100644
--- a/modelopt/torch/utils/moe.py
+++ b/modelopt/torch/utils/moe.py
@@ -291,9 +291,8 @@ def register_torch_a2a():
 
     These two styles are functionally transformers' ``megamoe_router`` / ``megamoe_experts`` (added
     in 5.10.4 for DeepGEMM Mega-MoE), reimplemented here so the transport works below 5.10.4 too.
-    The base classes (``RouterParallel``, ``TensorParallelLayer``) are stock wherever the EP infra
-    exists (transformers >= 5.6, the framework floor enforced in ``load_and_prepare_ep``), so this
-    works on any such version; ``grouped_gemm`` (expert weight sharding) is stock too.
+    The base classes (``RouterParallel``, ``TensorParallelLayer``) are stock in every transformers
+    version ModelOpt supports; ``grouped_gemm`` (expert weight sharding) is stock too.
     """
     global _TORCH_A2A_REGISTERED
     if _TORCH_A2A_REGISTERED:
diff --git a/noxfile.py b/noxfile.py
index 8528f7136e0..78627e773a6 100644
--- a/noxfile.py
+++ b/noxfile.py
@@ -48,7 +48,7 @@
 # ``.[all,dev-test]`` install to constrain that env).
 TRANSFORMERS_VERSIONS = {
     "tf_latest": ("transformers~=5.18.0",),
-    "tf_min": ("transformers~=5.5.0",),
+    "tf_min": ("transformers~=5.8.0",),
 }
 
 
diff --git a/pyproject.toml b/pyproject.toml
index d911a6e2260..ae913c62772 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -82,11 +82,11 @@ hf = [
     "diffusers>=0.37.0",
     "httpx>=0.23.0,<1",
     "huggingface_hub>=1.5.0",
-    "nltk",
+    "nltk>=3.9",
     "peft>=0.17.0",
     "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
     "tiktoken",
-    "transformers>=5.5,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py
+    "transformers>=5.8,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py
     "wonderwords",
 ]
 
diff --git a/tests/gpu/torch/utils/test_moe_a2a.py b/tests/gpu/torch/utils/test_moe_a2a.py
index 9b8a3be53a4..ef85c0b9b44 100644
--- a/tests/gpu/torch/utils/test_moe_a2a.py
+++ b/tests/gpu/torch/utils/test_moe_a2a.py
@@ -20,7 +20,6 @@
 import pytest
 import torch
 import torch.distributed as dist
-from packaging.version import Version
 
 import modelopt.torch.utils.moe as moe
 
@@ -67,7 +66,5 @@ def test_torch_a2a_gives_grouped_mms_result_bit_for_bit(dist_workers):
     """fp32 routing weights, as Mixtral's and DeepSeek-V3's routers return, make each token's top-k
     sum an fp32 one, whose bf16 result depends on the order of its adds. At EP=1 the torch_a2a
     forward adds them in grouped_mm's fixed order, not in the order of CUDA atomics."""
-    transformers = pytest.importorskip("transformers")
-    if Version(transformers.__version__) < Version("5.6"):
-        pytest.skip("expert parallelism needs transformers >= 5.6")
+    pytest.importorskip("transformers")
     dist_workers.run(_grouped_mm_bit_for_bit_job)
diff --git a/tests/unit/torch/utils/test_moe_a2a.py b/tests/unit/torch/utils/test_moe_a2a.py
index a53c334ec51..2d79715a673 100644
--- a/tests/unit/torch/utils/test_moe_a2a.py
+++ b/tests/unit/torch/utils/test_moe_a2a.py
@@ -32,7 +32,6 @@
 import pytest
 import torch
 import torch.distributed as dist
-from packaging.version import Version
 
 import modelopt.torch.utils.moe as moe
 from modelopt.torch.utils.moe import a2a_moe
@@ -217,9 +216,7 @@ def test_the_router_style_registers_under_transformers_5_16_0s_name(monkeypatch)
 def test_experts_whose_biases_the_plan_left_whole_are_refused():
     """A plan that shards GPT-OSS's expert weights but not their biases leaves every rank all the
     experts' biases, which the forward indexes by local expert id; it refuses them instead."""
-    transformers = pytest.importorskip("transformers")
-    if Version(transformers.__version__) < Version("5.6"):
-        pytest.skip("expert parallelism needs transformers >= 5.6")
+    pytest.importorskip("transformers")
     from transformers import GptOssConfig
     from transformers.models.gpt_oss.modeling_gpt_oss import GptOssExperts
 
@@ -254,9 +251,7 @@ def _no_tokens_for_rank1_job(rank, size):
 def test_experts_that_receive_no_tokens_run_on_an_empty_batch(cpu_dist_workers):
     """Every token routed to rank 0's experts: rank 1 runs transformers' fused experts on an empty
     batch, and the forward and gradients still match the whole module's on every rank."""
-    transformers = pytest.importorskip("transformers")
-    if Version(transformers.__version__) < Version("5.6"):
-        pytest.skip("expert parallelism needs transformers >= 5.6")
+    pytest.importorskip("transformers")
     cpu_dist_workers.run(_no_tokens_for_rank1_job)
 
 
@@ -372,9 +367,7 @@ def test_fused_experts_backward_through_torch_a2a_matches_the_whole_module(cpu_d
     """transformers' fused experts, split over two ranks, train through the torch_a2a forward exactly
     as the whole module does on one process: same output, and same gradients for the hidden states,
     the routing weights and each rank's experts."""
-    transformers = pytest.importorskip("transformers")
-    if Version(transformers.__version__) < Version("5.6"):
-        pytest.skip("expert parallelism needs transformers >= 5.6")
+    pytest.importorskip("transformers")
     cpu_dist_workers.run(_real_experts_backward_job)
 
 
@@ -397,7 +390,5 @@ def _fp32_routing_weights_job(rank, size):
 def test_fp32_routing_weights_give_grouped_mms_result_at_ep1(cpu_dist_workers):
     """A router may return fp32 weights for bf16 experts, as Mixtral's does. grouped_mm weights the
     experts' results in fp32 then; at EP=1 the torch_a2a forward must give its result bit for bit."""
-    transformers = pytest.importorskip("transformers")
-    if Version(transformers.__version__) < Version("5.6"):
-        pytest.skip("expert parallelism needs transformers >= 5.6")
+    pytest.importorskip("transformers")
     cpu_dist_workers.run(_fp32_routing_weights_job)

From daec6dee4f17afc48df681b691e24e775b3fba1f Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 14:48:04 -0700
Subject: [PATCH 4/9] Drop transformers 4.x compatibility leftovers

Use AutoModelForImageTextToText directly, import native NemotronH, Qwen3.5 and
conversion_mapping at module top, and remove test skips for APIs and models every
transformers>=5.8 has.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 modelopt/torch/models/nemotron_h/mtp.py       | 15 +++-----
 modelopt/torch/speculative/utils.py           |  8 +----
 .../_test_utils/torch/transformers_models.py  | 34 +++++--------------
 .../export/test_quant_aware_conversion_gpu.py |  9 +----
 .../export/test_qwen_vision_recipe_export.py  |  3 --
 tests/unit/recipe/test_qwen_vision_recipe.py  |  3 --
 .../torch/export/test_hf_spec_rope_export.py  |  4 +--
 .../export/test_quant_aware_conversion.py     |  7 ----
 .../torch/export/test_unified_export_hf.py    |  4 +--
 .../torch/utils/test_hf_checkpoint_utils.py   |  3 +-
 10 files changed, 19 insertions(+), 71 deletions(-)

diff --git a/modelopt/torch/models/nemotron_h/mtp.py b/modelopt/torch/models/nemotron_h/mtp.py
index 672aab1c324..6018d91aa2d 100644
--- a/modelopt/torch/models/nemotron_h/mtp.py
+++ b/modelopt/torch/models/nemotron_h/mtp.py
@@ -30,6 +30,11 @@
 from transformers import AutoConfig
 from transformers.dynamic_module_utils import get_class_from_dynamic_module
 from transformers.masking_utils import create_causal_mask
+from transformers.models.nemotron_h.modeling_nemotron_h import (
+    MIXER_TYPES,
+    NemotronHBlock,
+    NemotronHRMSNorm,
+)
 
 from modelopt.torch.quantization.model_calib import (
     _needs_activation_forward_for_max_calib,
@@ -84,9 +89,6 @@ def prefers_bundled_code(hf_config, checkpoint_path: str, config_kwargs: dict) -
         # Config-parsing incompatibilities only; e.g. an OSError from the Hub must propagate.
         return True
     # Parsing can succeed while model construction fails on a block type the built-in lacks.
-    # Native Nemotron-H is optional: older Transformers can still load unrelated models.
-    from transformers.models.nemotron_h.modeling_nemotron_h import MIXER_TYPES
-
     return any(t not in MIXER_TYPES for t in getattr(builtin_config, "layers_block_type", ()))
 
 
@@ -141,13 +143,6 @@ class _NemotronHMTP(torch.nn.Module):
 
     def __init__(self, config):
         super().__init__()
-        # Native Nemotron-H is optional: older Transformers can still load unrelated models.
-        from transformers.models.nemotron_h.modeling_nemotron_h import (
-            MIXER_TYPES,
-            NemotronHBlock,
-            NemotronHRMSNorm,
-        )
-
         block_config = copy.deepcopy(config)
         # Some remote-code revisions drop this field while materializing ``llm_config``.
         # The detected tensor layout is the canonical attention + MoE MTP tail.
diff --git a/modelopt/torch/speculative/utils.py b/modelopt/torch/speculative/utils.py
index 910536fdb10..23c75203338 100644
--- a/modelopt/torch/speculative/utils.py
+++ b/modelopt/torch/speculative/utils.py
@@ -778,13 +778,7 @@ def _warn_overrides_on_fake_base():
         return FakeBaseModel.from_source(model_name_or_path, trust_remote_code=trust_remote_code)
 
     if _is_vlm:
-        # Transformers 5 renamed AutoModelForVision2Seq to
-        # AutoModelForImageTextToText.  Prefer the pre-5 name so this loader
-        # continues to support the Transformers 4 environments used by older
-        # speculative-decoding jobs.
-        model_cls = getattr(transformers, "AutoModelForVision2Seq", None)
-        if model_cls is None:
-            model_cls = transformers.AutoModelForImageTextToText
+        model_cls = transformers.AutoModelForImageTextToText
     else:
         model_cls = transformers.AutoModelForCausalLM
 
diff --git a/tests/_test_utils/torch/transformers_models.py b/tests/_test_utils/torch/transformers_models.py
index bdf910981d2..46543bab271 100644
--- a/tests/_test_utils/torch/transformers_models.py
+++ b/tests/_test_utils/torch/transformers_models.py
@@ -41,16 +41,22 @@
     LlamaForSequenceClassification,
     MixtralConfig,
     NemotronConfig,
+    NemotronHConfig,
     PreTrainedModel,
+    Qwen3_5Config,
+    Qwen3_5MoeConfig,
     Qwen3Config,
     Qwen3MoeConfig,
     Qwen3VLConfig,
-    Qwen3VLTextConfig,
     T5Config,
     T5ForConditionalGeneration,
     ViTConfig,
     ViTImageProcessor,
 )
+from transformers.conversion_mapping import (
+    get_checkpoint_conversion_mapping,
+    register_checkpoint_conversion_mapping,
+)
 
 import modelopt.torch.opt as mto
 
@@ -227,14 +233,7 @@ def get_tiny_qwen3vl(**config_kwargs) -> PreTrainedModel:
         "vocab_size": 32,
         "rope_scaling": {"rope_type": "default", "mrope_section": [1, 1, 2]},
     }
-    # Transformers 4.x names this field rope_scaling; 5.x renamed it to rope_parameters.
-    # Supplying it avoids the 4.57 Qwen3-VL constructor dereferencing a None rope config.
-    rope_field = (
-        "rope_parameters"
-        if "rope_parameters" in getattr(Qwen3VLTextConfig, "__annotations__", {})
-        else "rope_scaling"
-    )
-    text_kwargs[rope_field] = {"rope_type": "default", "mrope_section": [1, 1, 2]}
+    text_kwargs["rope_parameters"] = {"rope_type": "default", "mrope_section": [1, 1, 2]}
     text_kwargs.update(config_kwargs)
     # Pass as dicts — transformers 5.3.0 Qwen3VLConfig.__init__ only handles
     # vision_config/text_config when they are dicts or None, not instances.
@@ -351,9 +350,6 @@ def __len__(self):
 
 
 def _get_tiny_qwen3_5_vl(moe: bool = False, *, tokenizer=None, **config_kwargs) -> PreTrainedModel:
-    # Lazy imports — Qwen3.5-VL requires a recent transformers version.
-    from transformers import Qwen3_5Config, Qwen3_5MoeConfig
-
     set_seed(SEED)
 
     # Vocab + vision token ids normally derive from the reference tokenizer. Tests that do not
@@ -529,16 +525,6 @@ def _match_released_nemotron_h_embedding_name() -> None:
     with every real checkpoint on one tensor. Dropped upstream in 5.16, where this is a no-op
     since there is no such rule left to remove.
     """
-    # Optional dependency: conversion_mapping is transformers 5.x only, while NemotronH also
-    # exists on the 4.57 floor of the support matrix.
-    try:
-        from transformers.conversion_mapping import (
-            get_checkpoint_conversion_mapping,
-            register_checkpoint_conversion_mapping,
-        )
-    except ImportError:
-        return
-
     mapping = get_checkpoint_conversion_mapping("nemotron_h")
     if mapping is None:
         return
@@ -550,10 +536,6 @@ def _match_released_nemotron_h_embedding_name() -> None:
 def get_tiny_nemotron_h(**config_kwargs) -> PreTrainedModel:
     set_seed(SEED)
 
-    # Lazy import — NemotronHConfig only exists in newer transformers builds, and this
-    # module is imported broadly (including by the min-transformers CI job).
-    from transformers import NemotronHConfig
-
     _match_released_nemotron_h_embedding_name()
 
     # Tiny NemotronH hybrid. hybrid_override_pattern letters: M=Mamba, E=MoE/FFN, *=Attention.
diff --git a/tests/gpu/torch/export/test_quant_aware_conversion_gpu.py b/tests/gpu/torch/export/test_quant_aware_conversion_gpu.py
index 32ffa5d2f1d..4d7c3d212c2 100644
--- a/tests/gpu/torch/export/test_quant_aware_conversion_gpu.py
+++ b/tests/gpu/torch/export/test_quant_aware_conversion_gpu.py
@@ -31,6 +31,7 @@
 import torch
 from _test_utils.torch.transformers_models import get_tiny_mixtral
 from safetensors import safe_open
+from transformers.core_model_loading import revert_weight_conversion
 
 import modelopt.torch.quantization as mtq
 from modelopt.torch.export import export_hf_checkpoint
@@ -55,14 +56,6 @@ def _tiny_mixtral():
 
 
 def test_export_tensor_names_match_hub_after_conversion_reverse():
-    try:
-        from transformers.conversion_mapping import get_checkpoint_conversion_mapping
-        from transformers.core_model_loading import revert_weight_conversion
-    except ImportError:
-        pytest.skip("transformers build has no conversion_mapping API")
-    if not get_checkpoint_conversion_mapping("mixtral"):
-        pytest.skip("transformers build has no mixtral conversion_mapping")
-
     # Canonical hub names: transformers' own reverse on the unquantized reference.
     ref = _tiny_mixtral()
     hub_names = set(revert_weight_conversion(ref, ref.state_dict()).keys())
diff --git a/tests/gpu/torch/export/test_qwen_vision_recipe_export.py b/tests/gpu/torch/export/test_qwen_vision_recipe_export.py
index 389375428b1..53872377948 100644
--- a/tests/gpu/torch/export/test_qwen_vision_recipe_export.py
+++ b/tests/gpu/torch/export/test_qwen_vision_recipe_export.py
@@ -17,7 +17,6 @@
 
 import pytest
 import torch
-import transformers
 from _test_utils.torch import transformers_models
 from safetensors import safe_open
 
@@ -32,8 +31,6 @@
 def _get_tiny_qwen_vlm(model_type):
     if model_type == "qwen3_vl":
         return transformers_models.get_tiny_qwen3vl()
-    if not all(hasattr(transformers, name) for name in ("Qwen3_5Config", "Qwen3_5MoeConfig")):
-        pytest.skip("Qwen3.5-VL requires a newer Transformers release")
     return transformers_models.get_tiny_qwen3_5_vl_offline()
 
 
diff --git a/tests/unit/recipe/test_qwen_vision_recipe.py b/tests/unit/recipe/test_qwen_vision_recipe.py
index 264e3b73239..6de8403d840 100644
--- a/tests/unit/recipe/test_qwen_vision_recipe.py
+++ b/tests/unit/recipe/test_qwen_vision_recipe.py
@@ -14,7 +14,6 @@
 # limitations under the License.
 
 import pytest
-import transformers
 from _test_utils.torch import transformers_models
 
 import modelopt.torch.quantization as mtq
@@ -27,8 +26,6 @@
 def _get_tiny_qwen_vlm(model_type):
     if model_type == "qwen3_vl":
         return transformers_models.get_tiny_qwen3vl()
-    if not all(hasattr(transformers, name) for name in ("Qwen3_5Config", "Qwen3_5MoeConfig")):
-        pytest.skip("Qwen3.5-VL requires a newer Transformers release")
     return transformers_models.get_tiny_qwen3_5_vl_offline()
 
 
diff --git a/tests/unit/torch/export/test_hf_spec_rope_export.py b/tests/unit/torch/export/test_hf_spec_rope_export.py
index cceddce9c9c..737aaa46d82 100644
--- a/tests/unit/torch/export/test_hf_spec_rope_export.py
+++ b/tests/unit/torch/export/test_hf_spec_rope_export.py
@@ -216,8 +216,8 @@ def test_missing_everywhere_returns_the_default(self):
     def test_reads_a_real_config_whichever_layout_it_uses(self):
         """A real config resolves on every supported transformers version.
 
-        Asserts the outcome, not the layout: 5.12 keeps the value only in the dict while
-        the minimum supported version (4.57) has only the flat field and no dict at all.
+        Asserts the outcome, not the layout, which varies across supported releases (5.12
+        keeps the value only in the dict).
         The layouts themselves are pinned above, built explicitly, so they stay covered
         where transformers is absent -- this is the only test here that needs it.
         """
diff --git a/tests/unit/torch/export/test_quant_aware_conversion.py b/tests/unit/torch/export/test_quant_aware_conversion.py
index 1a798b2c1fd..ba86a58f736 100644
--- a/tests/unit/torch/export/test_quant_aware_conversion.py
+++ b/tests/unit/torch/export/test_quant_aware_conversion.py
@@ -638,13 +638,6 @@ def test_build_reverse_rules_from_mixtral_conversion_mapping_cpu():
     pytest.importorskip("transformers")
     from _test_utils.torch.transformers_models import get_tiny_mixtral
 
-    try:
-        from transformers.conversion_mapping import get_checkpoint_conversion_mapping
-    except ImportError:
-        pytest.skip("transformers build has no conversion_mapping API")
-    if not get_checkpoint_conversion_mapping("mixtral"):
-        pytest.skip("transformers build has no mixtral conversion_mapping")
-
     model = get_tiny_mixtral(**_MIXTRAL_KWARGS)
 
     p = "model.layers.0"
diff --git a/tests/unit/torch/export/test_unified_export_hf.py b/tests/unit/torch/export/test_unified_export_hf.py
index 57923dbadaf..1b8d55373fa 100644
--- a/tests/unit/torch/export/test_unified_export_hf.py
+++ b/tests/unit/torch/export/test_unified_export_hf.py
@@ -136,9 +136,7 @@ def test_hf_all_tied_weights_keys_contract():
     (``tie_word_embeddings=True`` -> lm_head aliases the embedding). If transformers renames it or
     flips the direction, this breaks instead of silently skipping tied-weight dedup.
     """
-    pytest.importorskip(
-        "transformers", minversion="5.0"
-    )  # attribute only exists on transformers>=5.0
+    pytest.importorskip("transformers")
     from transformers import AutoModelForCausalLM, LlamaConfig
 
     cfg = LlamaConfig(
diff --git a/tests/unit/torch/utils/test_hf_checkpoint_utils.py b/tests/unit/torch/utils/test_hf_checkpoint_utils.py
index ef992a9f832..b14f26b4acb 100644
--- a/tests/unit/torch/utils/test_hf_checkpoint_utils.py
+++ b/tests/unit/torch/utils/test_hf_checkpoint_utils.py
@@ -228,8 +228,7 @@ def test_sanitize_hf_config_for_deployment_trims_nextn_layer_types():
 @pytest.mark.parametrize("native", [False, True])
 @pytest.mark.parametrize("keep_tokenizer", [False, True])
 def test_export_config_reloads_nemotron_h_with_source_auto_map(tmp_path, native, keep_tokenizer):
-    # Native NemotronH is optional on the minimum-Transformers test lane.
-    transformers = pytest.importorskip("transformers", minversion="5.5")
+    transformers = pytest.importorskip("transformers")
     from modelopt.torch.export.unified_export_hf import (
         _write_hf_export_config,
         save_non_weight_artifacts,

From ecf455954bab2f28fe9124698b6e52c614c207fe Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 23:41:36 -0700
Subject: [PATCH 5/9] Read NemotronH's mixer registry at call time

A module-level MIXER_TYPES import kept a stale mapping once a test patched it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 modelopt/torch/models/nemotron_h/mtp.py | 30 ++++++++++++++++---------
 1 file changed, 19 insertions(+), 11 deletions(-)

diff --git a/modelopt/torch/models/nemotron_h/mtp.py b/modelopt/torch/models/nemotron_h/mtp.py
index 6018d91aa2d..2ac7d0509c3 100644
--- a/modelopt/torch/models/nemotron_h/mtp.py
+++ b/modelopt/torch/models/nemotron_h/mtp.py
@@ -30,11 +30,7 @@
 from transformers import AutoConfig
 from transformers.dynamic_module_utils import get_class_from_dynamic_module
 from transformers.masking_utils import create_causal_mask
-from transformers.models.nemotron_h.modeling_nemotron_h import (
-    MIXER_TYPES,
-    NemotronHBlock,
-    NemotronHRMSNorm,
-)
+from transformers.models.nemotron_h import modeling_nemotron_h
 
 from modelopt.torch.quantization.model_calib import (
     _needs_activation_forward_for_max_calib,
@@ -89,7 +85,10 @@ def prefers_bundled_code(hf_config, checkpoint_path: str, config_kwargs: dict) -
         # Config-parsing incompatibilities only; e.g. an OSError from the Hub must propagate.
         return True
     # Parsing can succeed while model construction fails on a block type the built-in lacks.
-    return any(t not in MIXER_TYPES for t in getattr(builtin_config, "layers_block_type", ()))
+    return any(
+        t not in modeling_nemotron_h.MIXER_TYPES
+        for t in getattr(builtin_config, "layers_block_type", ())
+    )
 
 
 def _normalize_mtp_block_types(block_types, mixer_types):
@@ -147,17 +146,26 @@ def __init__(self, config):
         # Some remote-code revisions drop this field while materializing ``llm_config``.
         # The detected tensor layout is the canonical attention + MoE MTP tail.
         block_types = list(getattr(config, "mtp_layers_block_type", None) or ("attention", "moe"))
-        block_config.layers_block_type = _normalize_mtp_block_types(block_types, MIXER_TYPES)
+        block_config.layers_block_type = _normalize_mtp_block_types(
+            block_types, modeling_nemotron_h.MIXER_TYPES
+        )
         self.layers = torch.nn.ModuleList(
-            [NemotronHBlock(block_config, layer_idx) for layer_idx in range(len(block_types))]
+            [
+                modeling_nemotron_h.NemotronHBlock(block_config, layer_idx)
+                for layer_idx in range(len(block_types))
+            ]
         )
         first_layer, last_layer = self.layers[0], self.layers[-1]
-        first_layer.enorm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
-        first_layer.hnorm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        first_layer.enorm = modeling_nemotron_h.NemotronHRMSNorm(
+            config.hidden_size, eps=config.layer_norm_epsilon
+        )
+        first_layer.hnorm = modeling_nemotron_h.NemotronHRMSNorm(
+            config.hidden_size, eps=config.layer_norm_epsilon
+        )
         first_layer.eh_proj = torch.nn.Linear(
             config.hidden_size * 2, config.hidden_size, bias=False
         )
-        last_layer.final_layernorm = NemotronHRMSNorm(
+        last_layer.final_layernorm = modeling_nemotron_h.NemotronHRMSNorm(
             config.hidden_size, eps=config.layer_norm_epsilon
         )
 

From 1c5f591e9bf005bc7a96cbd443c3f1f54167cd7e Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Wed, 28 Oct 2026 23:50:26 -0700
Subject: [PATCH 6/9] Bump transformers to >=5.9

transformers 5.8.0 and 5.8.1 skip their own weight init for models built from a
config, leaving every nn.Linear at PyTorch's default init.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 CHANGELOG.rst              | 2 +-
 modelopt/torch/__init__.py | 2 +-
 noxfile.py                 | 2 +-
 pyproject.toml             | 2 +-
 4 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 8f135a88763..5bcdce2da0f 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -47,7 +47,7 @@ Changelog
 - ``export_mcore_gpt_to_hf`` given a Hugging Face Hub model ID no longer writes the tokenizer, processor, generation config or remote-code files. Pass a local directory instead, e.g. from ``modelopt.torch.export.ensure_local_checkpoint``.
 - Drop support for Python 3.10, which has reached end-of-life. ModelOpt now requires Python 3.11 or later.
 - Bump minimum required PyTorch version to 2.10.
-- Bump minimum required ``transformers`` version to 5.8, ``diffusers`` to 0.37 and ``nltk`` to 3.9 for the ``[hf]`` extra.
+- Bump minimum required ``transformers`` version to 5.9, ``diffusers`` to 0.37 and ``nltk`` to 3.9 for the ``[hf]`` extra.
 - Diffusion HF exports now omit attention quantizer buffers by default, except for FLUX.1, which retains enabled Q/K/V ``_amax`` keys for runtime compatibility; other consumers of those keys can opt in with ``export_attention_quantizers=True`` (CLI: ``--extra-param export_attention_quantizers=true``). Disabled attention quantizers remain excluded, and diffusion components, including LTX-2, no longer emit KV-cache quantization metadata.
 - ``hf_quant_config.json`` of an INT4 or W4A8 AWQ checkpoint now sets ``pre_quant_scale`` per layer, only for layers that have the tensor. A model whose fused-MoE experts have none (AWQ never applies to them) is now exported as ``MIXED_PRECISION`` with per-layer entries instead of a uniform algorithm.
 - The Windows FVD example rejects I3D checkpoints requiring unrestricted pickle loading. Custom TorchScript archives now require ``--weights-sha256`` with a trusted checksum; tensor state dictionaries continue to load directly.
diff --git a/modelopt/torch/__init__.py b/modelopt/torch/__init__.py
index 1caebbfbe3c..55975bb4ec0 100644
--- a/modelopt/torch/__init__.py
+++ b/modelopt/torch/__init__.py
@@ -48,7 +48,7 @@
 try:
     from transformers import __version__ as _transformers_version
 
-    if _Version(_transformers_version) < _Version("5.8") or _Version(
+    if _Version(_transformers_version) < _Version("5.9") or _Version(
         _transformers_version
     ) >= _Version("5.19"):
         _warnings.warn(
diff --git a/noxfile.py b/noxfile.py
index 78627e773a6..ebda78425aa 100644
--- a/noxfile.py
+++ b/noxfile.py
@@ -48,7 +48,7 @@
 # ``.[all,dev-test]`` install to constrain that env).
 TRANSFORMERS_VERSIONS = {
     "tf_latest": ("transformers~=5.18.0",),
-    "tf_min": ("transformers~=5.8.0",),
+    "tf_min": ("transformers~=5.9.0",),
 }
 
 
diff --git a/pyproject.toml b/pyproject.toml
index ae913c62772..8f38814a1fe 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -86,7 +86,7 @@ hf = [
     "peft>=0.17.0",
     "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
     "tiktoken",
-    "transformers>=5.8,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py
+    "transformers>=5.9,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py
     "wonderwords",
 ]
 

From 3fca1bf920e1debc7432191e0769183e49ccff20 Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Thu, 29 Oct 2026 00:05:34 -0700
Subject: [PATCH 7/9] Bump peft to >=0.19 and accelerate to >=1.6; trim
 redundant example pins

transformers>=5.6 refuses peft<0.18.2 in add_adapter/enable_adapters, which the QAT
trainer calls, and ModelOpt's HF trainers read Accelerator.is_fsdp2 (accelerate 1.6+).
Drop example pins the new floors already cover.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 CHANGELOG.rst                                  | 2 +-
 examples/gpt-oss/requirements.txt              | 2 --
 examples/speculative_decoding/requirements.txt | 3 +--
 examples/torch_trt/requirements.txt            | 3 +--
 pyproject.toml                                 | 4 ++--
 5 files changed, 5 insertions(+), 9 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 5bcdce2da0f..8a6715572d4 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -47,7 +47,7 @@ Changelog
 - ``export_mcore_gpt_to_hf`` given a Hugging Face Hub model ID no longer writes the tokenizer, processor, generation config or remote-code files. Pass a local directory instead, e.g. from ``modelopt.torch.export.ensure_local_checkpoint``.
 - Drop support for Python 3.10, which has reached end-of-life. ModelOpt now requires Python 3.11 or later.
 - Bump minimum required PyTorch version to 2.10.
-- Bump minimum required ``transformers`` version to 5.9, ``diffusers`` to 0.37 and ``nltk`` to 3.9 for the ``[hf]`` extra.
+- Bump minimum required ``transformers`` version to 5.9, ``diffusers`` to 0.37, ``peft`` to 0.19, ``accelerate`` to 1.6 and ``nltk`` to 3.9 for the ``[hf]`` extra.
 - Diffusion HF exports now omit attention quantizer buffers by default, except for FLUX.1, which retains enabled Q/K/V ``_amax`` keys for runtime compatibility; other consumers of those keys can opt in with ``export_attention_quantizers=True`` (CLI: ``--extra-param export_attention_quantizers=true``). Disabled attention quantizers remain excluded, and diffusion components, including LTX-2, no longer emit KV-cache quantization metadata.
 - ``hf_quant_config.json`` of an INT4 or W4A8 AWQ checkpoint now sets ``pre_quant_scale`` per layer, only for layers that have the tensor. A model whose fused-MoE experts have none (AWQ never applies to them) is now exported as ``MIXED_PRECISION`` with per-layer entries instead of a uniform algorithm.
 - The Windows FVD example rejects I3D checkpoints requiring unrestricted pickle loading. Custom TorchScript archives now require ``--weights-sha256`` with a trusted checksum; tensor state dictionaries continue to load directly.
diff --git a/examples/gpt-oss/requirements.txt b/examples/gpt-oss/requirements.txt
index 260f256e6e4..0ee4601c161 100644
--- a/examples/gpt-oss/requirements.txt
+++ b/examples/gpt-oss/requirements.txt
@@ -1,5 +1,3 @@
 kernels>=0.9.0,<0.13
 trackio>=0.41
-# transformers>=5.3 avoids CVE-2026-4372 (RCE via the `kernels` Hub download path, which this example installs)
-transformers>=5.3
 trl>=1.0,<1.15
diff --git a/examples/speculative_decoding/requirements.txt b/examples/speculative_decoding/requirements.txt
index d333178fad9..3df5ac52fdf 100644
--- a/examples/speculative_decoding/requirements.txt
+++ b/examples/speculative_decoding/requirements.txt
@@ -1,4 +1,3 @@
 accelerate>=1.12.0
-datasets
 peft==0.18.1
-transformers>=5.0,<5.13
+transformers<5.13
diff --git a/examples/torch_trt/requirements.txt b/examples/torch_trt/requirements.txt
index 436366630e3..e2fc4c8095d 100644
--- a/examples/torch_trt/requirements.txt
+++ b/examples/torch_trt/requirements.txt
@@ -3,6 +3,5 @@ datasets>=2.14.4
 # torch / torchvision / torch-tensorrt trio together: torchvision pins an exact
 # torch, so if the base install pulls a newer torch (e.g. 2.13) the torch-tensorrt
 # downgrade would otherwise leave torchvision mismatched and break `import`.
-torch-tensorrt>=2.4.0,<2.13
+torch-tensorrt>=2.10,<2.13
 torchvision<0.28
-transformers>=4.56
diff --git a/pyproject.toml b/pyproject.toml
index 8f38814a1fe..bdeda676e27 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -76,14 +76,14 @@ onnx = [
     "onnxruntime-ep-nv-tensorrt-rtx-cu13==0.4.0; platform_system == 'Windows' and platform_machine == 'ARM64'",
 ]
 hf = [
-    "accelerate>=1.1.0",
+    "accelerate>=1.6.0",
     "datasets>=3.0.0",
     "deepspeed>=0.9.6; platform_system != 'Darwin' and platform_system != 'Windows'",
     "diffusers>=0.37.0",
     "httpx>=0.23.0,<1",
     "huggingface_hub>=1.5.0",
     "nltk>=3.9",
-    "peft>=0.17.0",
+    "peft>=0.19.0",
     "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
     "tiktoken",
     "transformers>=5.9,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py

From f8ce66e19605f86bd6bb9a78aa9733ce656b10a5 Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Thu, 29 Oct 2026 00:24:23 -0700
Subject: [PATCH 8/9] Drop the speculative-decoding example's peft==0.18.1 pin

It downgraded peft below the [hf] extra's new >=0.19 floor; the EAGLE LoRA tests pass on
peft 0.18.1 through 0.21.2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 examples/speculative_decoding/requirements.txt | 1 -
 1 file changed, 1 deletion(-)

diff --git a/examples/speculative_decoding/requirements.txt b/examples/speculative_decoding/requirements.txt
index 3df5ac52fdf..b9068fd6f12 100644
--- a/examples/speculative_decoding/requirements.txt
+++ b/examples/speculative_decoding/requirements.txt
@@ -1,3 +1,2 @@
 accelerate>=1.12.0
-peft==0.18.1
 transformers<5.13

From c6b635662efb5dc37b76bb7c6c08560045db04f3 Mon Sep 17 00:00:00 2001
From: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Date: Thu, 29 Oct 2026 01:28:27 -0700
Subject: [PATCH 9/9] Keep the transformers floor at 5.5 and peft at 0.18.1 for
 TensorRT-LLM

TensorRT-LLM 1.3.0 pins transformers==5.5.4 and peft<0.19, so higher floors upgrade them
in its container and break TensorRT-LLM's own Qwen3-VL loading. Restore the
transformers<5.6 expert-parallel test skips.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
---
 CHANGELOG.rst                          |  2 +-
 modelopt/torch/__init__.py             |  2 +-
 modelopt/torch/utils/moe.py            |  4 ++--
 noxfile.py                             |  2 +-
 pyproject.toml                         |  5 +++--
 tests/gpu/torch/utils/test_moe_a2a.py  |  5 ++++-
 tests/unit/torch/utils/test_moe_a2a.py | 17 +++++++++++++----
 7 files changed, 25 insertions(+), 12 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 8a6715572d4..5b8b221cca7 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -47,7 +47,7 @@ Changelog
 - ``export_mcore_gpt_to_hf`` given a Hugging Face Hub model ID no longer writes the tokenizer, processor, generation config or remote-code files. Pass a local directory instead, e.g. from ``modelopt.torch.export.ensure_local_checkpoint``.
 - Drop support for Python 3.10, which has reached end-of-life. ModelOpt now requires Python 3.11 or later.
 - Bump minimum required PyTorch version to 2.10.
-- Bump minimum required ``transformers`` version to 5.9, ``diffusers`` to 0.37, ``peft`` to 0.19, ``accelerate`` to 1.6 and ``nltk`` to 3.9 for the ``[hf]`` extra.
+- Bump minimum required ``diffusers`` version to 0.37, ``peft`` to 0.18.1, ``accelerate`` to 1.6 and ``nltk`` to 3.9 for the ``[hf]`` extra.
 - Diffusion HF exports now omit attention quantizer buffers by default, except for FLUX.1, which retains enabled Q/K/V ``_amax`` keys for runtime compatibility; other consumers of those keys can opt in with ``export_attention_quantizers=True`` (CLI: ``--extra-param export_attention_quantizers=true``). Disabled attention quantizers remain excluded, and diffusion components, including LTX-2, no longer emit KV-cache quantization metadata.
 - ``hf_quant_config.json`` of an INT4 or W4A8 AWQ checkpoint now sets ``pre_quant_scale`` per layer, only for layers that have the tensor. A model whose fused-MoE experts have none (AWQ never applies to them) is now exported as ``MIXED_PRECISION`` with per-layer entries instead of a uniform algorithm.
 - The Windows FVD example rejects I3D checkpoints requiring unrestricted pickle loading. Custom TorchScript archives now require ``--weights-sha256`` with a trusted checksum; tensor state dictionaries continue to load directly.
diff --git a/modelopt/torch/__init__.py b/modelopt/torch/__init__.py
index 55975bb4ec0..1182304f541 100644
--- a/modelopt/torch/__init__.py
+++ b/modelopt/torch/__init__.py
@@ -48,7 +48,7 @@
 try:
     from transformers import __version__ as _transformers_version
 
-    if _Version(_transformers_version) < _Version("5.9") or _Version(
+    if _Version(_transformers_version) < _Version("5.5") or _Version(
         _transformers_version
     ) >= _Version("5.19"):
         _warnings.warn(
diff --git a/modelopt/torch/utils/moe.py b/modelopt/torch/utils/moe.py
index 3351b68f7a1..fc62fe9f71c 100644
--- a/modelopt/torch/utils/moe.py
+++ b/modelopt/torch/utils/moe.py
@@ -291,8 +291,8 @@ def register_torch_a2a():
 
     These two styles are functionally transformers' ``megamoe_router`` / ``megamoe_experts`` (added
     in 5.10.4 for DeepGEMM Mega-MoE), reimplemented here so the transport works below 5.10.4 too.
-    The base classes (``RouterParallel``, ``TensorParallelLayer``) are stock in every transformers
-    version ModelOpt supports; ``grouped_gemm`` (expert weight sharding) is stock too.
+    The base classes (``RouterParallel``, ``TensorParallelLayer``) are stock wherever the EP infra
+    exists (transformers >= 5.6); ``grouped_gemm`` (expert weight sharding) is stock too.
     """
     global _TORCH_A2A_REGISTERED
     if _TORCH_A2A_REGISTERED:
diff --git a/noxfile.py b/noxfile.py
index ebda78425aa..8528f7136e0 100644
--- a/noxfile.py
+++ b/noxfile.py
@@ -48,7 +48,7 @@
 # ``.[all,dev-test]`` install to constrain that env).
 TRANSFORMERS_VERSIONS = {
     "tf_latest": ("transformers~=5.18.0",),
-    "tf_min": ("transformers~=5.9.0",),
+    "tf_min": ("transformers~=5.5.0",),
 }
 
 
diff --git a/pyproject.toml b/pyproject.toml
index bdeda676e27..ebbe3a79270 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -83,10 +83,11 @@ hf = [
     "httpx>=0.23.0,<1",
     "huggingface_hub>=1.5.0",
     "nltk>=3.9",
-    "peft>=0.19.0",
+    "peft>=0.18.1",
     "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
     "tiktoken",
-    "transformers>=5.9,<5.19",                                                        # Should match modelopt/torch/__init__.py and noxfile.py
+    # TensorRT-LLM 1.3.0 ships transformers 5.5; keep in sync with modelopt/torch/__init__.py and noxfile.py
+    "transformers>=5.5,<5.19",
     "wonderwords",
 ]
 
diff --git a/tests/gpu/torch/utils/test_moe_a2a.py b/tests/gpu/torch/utils/test_moe_a2a.py
index ef85c0b9b44..9b8a3be53a4 100644
--- a/tests/gpu/torch/utils/test_moe_a2a.py
+++ b/tests/gpu/torch/utils/test_moe_a2a.py
@@ -20,6 +20,7 @@
 import pytest
 import torch
 import torch.distributed as dist
+from packaging.version import Version
 
 import modelopt.torch.utils.moe as moe
 
@@ -66,5 +67,7 @@ def test_torch_a2a_gives_grouped_mms_result_bit_for_bit(dist_workers):
     """fp32 routing weights, as Mixtral's and DeepSeek-V3's routers return, make each token's top-k
     sum an fp32 one, whose bf16 result depends on the order of its adds. At EP=1 the torch_a2a
     forward adds them in grouped_mm's fixed order, not in the order of CUDA atomics."""
-    pytest.importorskip("transformers")
+    transformers = pytest.importorskip("transformers")
+    if Version(transformers.__version__) < Version("5.6"):
+        pytest.skip("expert parallelism needs transformers >= 5.6")
     dist_workers.run(_grouped_mm_bit_for_bit_job)
diff --git a/tests/unit/torch/utils/test_moe_a2a.py b/tests/unit/torch/utils/test_moe_a2a.py
index 2d79715a673..a53c334ec51 100644
--- a/tests/unit/torch/utils/test_moe_a2a.py
+++ b/tests/unit/torch/utils/test_moe_a2a.py
@@ -32,6 +32,7 @@
 import pytest
 import torch
 import torch.distributed as dist
+from packaging.version import Version
 
 import modelopt.torch.utils.moe as moe
 from modelopt.torch.utils.moe import a2a_moe
@@ -216,7 +217,9 @@ def test_the_router_style_registers_under_transformers_5_16_0s_name(monkeypatch)
 def test_experts_whose_biases_the_plan_left_whole_are_refused():
     """A plan that shards GPT-OSS's expert weights but not their biases leaves every rank all the
     experts' biases, which the forward indexes by local expert id; it refuses them instead."""
-    pytest.importorskip("transformers")
+    transformers = pytest.importorskip("transformers")
+    if Version(transformers.__version__) < Version("5.6"):
+        pytest.skip("expert parallelism needs transformers >= 5.6")
     from transformers import GptOssConfig
     from transformers.models.gpt_oss.modeling_gpt_oss import GptOssExperts
 
@@ -251,7 +254,9 @@ def _no_tokens_for_rank1_job(rank, size):
 def test_experts_that_receive_no_tokens_run_on_an_empty_batch(cpu_dist_workers):
     """Every token routed to rank 0's experts: rank 1 runs transformers' fused experts on an empty
     batch, and the forward and gradients still match the whole module's on every rank."""
-    pytest.importorskip("transformers")
+    transformers = pytest.importorskip("transformers")
+    if Version(transformers.__version__) < Version("5.6"):
+        pytest.skip("expert parallelism needs transformers >= 5.6")
     cpu_dist_workers.run(_no_tokens_for_rank1_job)
 
 
@@ -367,7 +372,9 @@ def test_fused_experts_backward_through_torch_a2a_matches_the_whole_module(cpu_d
     """transformers' fused experts, split over two ranks, train through the torch_a2a forward exactly
     as the whole module does on one process: same output, and same gradients for the hidden states,
     the routing weights and each rank's experts."""
-    pytest.importorskip("transformers")
+    transformers = pytest.importorskip("transformers")
+    if Version(transformers.__version__) < Version("5.6"):
+        pytest.skip("expert parallelism needs transformers >= 5.6")
     cpu_dist_workers.run(_real_experts_backward_job)
 
 
@@ -390,5 +397,7 @@ def _fp32_routing_weights_job(rank, size):
 def test_fp32_routing_weights_give_grouped_mms_result_at_ep1(cpu_dist_workers):
     """A router may return fp32 weights for bf16 experts, as Mixtral's does. grouped_mm weights the
     experts' results in fp32 then; at EP=1 the torch_a2a forward must give its result bit for bit."""
-    pytest.importorskip("transformers")
+    transformers = pytest.importorskip("transformers")
+    if Version(transformers.__version__) < Version("5.6"):
+        pytest.skip("expert parallelism needs transformers >= 5.6")
     cpu_dist_workers.run(_fp32_routing_weights_job)