diff --git a/tensorrt_llm/_torch/models/__init__.py b/tensorrt_llm/_torch/models/__init__.py
index aabd0d504831..84dcaae1b377 100644
--- a/tensorrt_llm/_torch/models/__init__.py
+++ b/tensorrt_llm/_torch/models/__init__.py
@@ -1,4 +1,5 @@
 import transformers
+from packaging import version
 
 from .modeling_auto import AutoModelForCausalLM
 from .modeling_bert import BertForSequenceClassification
@@ -89,7 +90,7 @@
     "Cohere2ForCausalLM",
 ]
 
-if transformers.__version__ >= "4.45.1":
+if version.parse(transformers.__version__) >= version.parse("4.45.1"):
     from .modeling_mllama import MllamaForConditionalGeneration  # noqa
 
     __all__.append("MllamaForConditionalGeneration")
diff --git a/tensorrt_llm/_torch/speculative/interface.py b/tensorrt_llm/_torch/speculative/interface.py
index d7129bd9b64b..8d2c9dc32d0e 100644
--- a/tensorrt_llm/_torch/speculative/interface.py
+++ b/tensorrt_llm/_torch/speculative/interface.py
@@ -7,6 +7,7 @@
 from typing import TYPE_CHECKING, List, Optional, Type
 
 import torch
+from packaging import version
 from torch import nn
 
 from tensorrt_llm.logger import logger
@@ -517,7 +518,8 @@ def __init__(self, use_separate_draft_kv_cache: bool = False):
         super().__init__()
         self.guided_decoder: Optional["CapturableGuidedDecoder"] = None
         self.force_num_accepted_tokens = get_force_num_accepted_tokens()
-        self.use_flashinfer = IS_FLASHINFER_AVAILABLE and flashinfer.__version__ >= "0.6.4"
+        self.use_flashinfer = IS_FLASHINFER_AVAILABLE and version.parse(
+            flashinfer.__version__) >= version.parse("0.6.4")
         self.seed: Optional[torch.Tensor] = None
         self.offset: Optional[torch.Tensor] = None
         self.use_separate_draft_kv_cache = use_separate_draft_kv_cache
diff --git a/tensorrt_llm/tools/multimodal_builder.py b/tensorrt_llm/tools/multimodal_builder.py
index bf948eb2506f..3918bc7a2126 100644
--- a/tensorrt_llm/tools/multimodal_builder.py
+++ b/tensorrt_llm/tools/multimodal_builder.py
@@ -6,6 +6,7 @@
 from time import time
 
 import yaml
+from packaging import version
 
 # isort: off
 import torch
@@ -1406,7 +1407,8 @@ class VisionAttentionOpt(VisionAttention):
 
         def __init__(self, config: Qwen2VLVisionConfig):
             # Fallback for compatibility with older transformers versions (for certain nvbugs/tests)
-            if transformers.__version__ >= '4.53.0':
+            if version.parse(
+                    transformers.__version__) >= version.parse('4.53.0'):
                 super().__init__(config)
                 self.head_dim = config.embed_dim // config.num_heads
             else: