From 2003da34f88f1cf5e0d1b3311873700785bef65c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 12 Sep 2025 03:13:15 -0700
Subject: [PATCH] importlib_version

---
 unsloth/__init__.py      | 7 +++++--
 unsloth/models/_utils.py | 2 ++
 unsloth/models/llama.py  | 7 +------
 unsloth/models/vision.py | 7 +------
 4 files changed, 9 insertions(+), 14 deletions(-)

diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 5529e4eeef..8255e505a8 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -14,7 +14,7 @@
 
 import warnings, importlib, sys
 from packaging.version import Version
-import os, re, subprocess, inspect
+import os, re, subprocess, inspect, functools
 import numpy as np
 
 # Fix some issues before importing other packages
@@ -69,9 +69,12 @@ except Exception as exception:
     raise exception
 pass
 
+@functools.cache
 def is_hip():
     return bool(getattr(getattr(torch, "version", None), "hip", None))
+pass
 
+@functools.cache
 def get_device_type():
     if hasattr(torch, "cuda") and torch.cuda.is_available():
         if is_hip():
@@ -83,6 +86,7 @@ def get_device_type():
 pass
 DEVICE_TYPE : str = get_device_type()
 
+@functools.cache
 def get_device_count():
     if DEVICE_TYPE in ("cuda", "hip"):
         return torch.cuda.device_count()
@@ -137,7 +141,6 @@ pass
 import importlib.util
 from pathlib import Path
 from importlib.metadata import version as importlib_version
-from packaging.version import Version
 from .import_fixes import fix_xformers_performance_issue
 fix_xformers_performance_issue(); del fix_xformers_performance_issue;
 from .import_fixes import fix_vllm_aimv2_issue
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index e3b70d93a5..e3ac56ac83 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -24,6 +24,7 @@ __all__ = [
     "xformers_attention",
     "xformers_version",
     "__version__",
+    "importlib_version",
     "HAS_FLASH_ATTENTION",
     "HAS_FLASH_ATTENTION_SOFTCAPPING",
     "USE_MODELSCOPE",
@@ -81,6 +82,7 @@ import re
 import functools
 import warnings, subprocess, re, inspect, psutil, os, math
 from unsloth_zoo.utils import Version
+from importlib.metadata import version as importlib_version
 from unsloth import DEVICE_TYPE, DEVICE_COUNT
 
 from unsloth_zoo.tokenizer_utils import (
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index af3ab8d3da..f7a53d05fd 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -19,7 +19,7 @@ import functools
 from typing import Any, Dict, Optional, Tuple, List, Union
 from ._utils import *
 from ._utils import patch_unsloth_smart_gradient_checkpointing
-from ._utils import __version__
+from ._utils import __version__, importlib_version
 from ._utils import move_to_device
 from ._utils import _prepare_model_for_qat
 from torch.nn.functional import scaled_dot_product_attention
@@ -1867,23 +1867,18 @@ class FastLlamaModel:
             gpu_stats = torch.cuda.get_device_properties(0)
             gpu_version = torch.version.cuda
             gpu_stats_snippet = f"CUDA: {gpu_stats.major}.{gpu_stats.minor}. CUDA Toolkit: {gpu_version}."
-
-            from importlib.metadata import version as importlib_version
             try:    vllm_version = f" vLLM: {importlib_version('vllm')}."
             except: vllm_version = ""
         elif DEVICE_TYPE == "hip":
             gpu_stats = torch.cuda.get_device_properties(0)
             gpu_version = torch.version.hip
             gpu_stats_snippet = f"ROCm Toolkit: {gpu_version}."
-
-            from importlib.metadata import version as importlib_version
             try:    vllm_version = f" vLLM: {importlib_version('vllm')}."
             except: vllm_version = ""
         elif DEVICE_TYPE == "xpu":
             gpu_stats = torch.xpu.get_device_properties(0)
             gpu_version = torch.version.xpu
             gpu_stats_snippet = f"Intel Toolkit: {gpu_version}."
-
             try:    vllm_version = f" vLLM: {importlib_version('vllm')}."
             except: vllm_version = ""
         else:
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 6c986a3bd3..7a07a8d0fd 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -28,7 +28,7 @@ pass
 from ..kernels import (
     post_patch_loss_function,
 )
-from ._utils import __version__
+from ._utils import __version__, importlib_version
 from ._utils import *
 from ..save import patch_saving_functions
 from peft import LoraConfig, TaskType, get_peft_model as _get_peft_model
@@ -277,23 +277,18 @@ class FastBaseModel:
             gpu_stats = torch.cuda.get_device_properties(0)
             gpu_version = torch.version.cuda
             gpu_stats_snippet = f"CUDA: {gpu_stats.major}.{gpu_stats.minor}. CUDA Toolkit: {gpu_version}."
-
-            from importlib.metadata import version as importlib_version
             try:    vllm_version = f" vLLM: {importlib_version('vllm')}."
             except: vllm_version = ""
         elif DEVICE_TYPE == "hip":
             gpu_stats = torch.cuda.get_device_properties(0)
             gpu_version = torch.version.hip
             gpu_stats_snippet = f"ROCm Toolkit: {gpu_version}."
-
-            from importlib.metadata import version as importlib_version
             try:    vllm_version = f" vLLM: {importlib_version('vllm')}."
             except: vllm_version = ""
         elif DEVICE_TYPE == "xpu":
             gpu_stats = torch.xpu.get_device_properties(0)
             gpu_version = torch.version.xpu
             gpu_stats_snippet = f"Intel Toolkit: {gpu_version}."
-
             # TODO: After adding vLLM support for XPU, changed this
             vllm_version = ""
         else: