diff --git a/disassemblers/adreno/__init__.py b/disassemblers/adreno/__init__.py
index 4ef71a6b66..ac16b3811f 100644
--- a/disassemblers/adreno/__init__.py
+++ b/disassemblers/adreno/__init__.py
@@ -8,7 +8,7 @@ def disasm(buf):
   global fxn
   if fxn is None:
     shared = pathlib.Path(__file__).parent / "disasm.so"
-    if not os.path.isfile(shared):
+    if not shared.is_file():
       os.system(f'cd {pathlib.Path(__file__).parent} && gcc -shared disasm-a3xx.c -o disasm.so')
     fxn = ctypes.CDLL(shared.as_posix())['disasm']
   #hexdump(buf)
diff --git a/examples/compile_efficientnet.py b/examples/compile_efficientnet.py
index 5c2cb57f98..b75e206c40 100644
--- a/examples/compile_efficientnet.py
+++ b/examples/compile_efficientnet.py
@@ -1,20 +1,22 @@
+from pathlib import Path
 from models.efficientnet import EfficientNet
 from tinygrad.tensor import Tensor
 from tinygrad.nn.state import safe_save
 from extra.utils import fetch
 from extra.export_model import export_model
 from tinygrad.helpers import getenv
-import ast, os
+import ast
 
 if __name__ == "__main__":
   model = EfficientNet(0)
   model.load_from_pretrained()
   mode = "clang" if getenv("CLANG", "") != "" else "webgpu" if getenv("WEBGPU", "") != "" else ""
   prg, inp_size, out_size, state = export_model(model, Tensor.randn(1,3,224,224), mode)
+  dirname = Path(__file__).parent
   if getenv("CLANG", "") == "":
-    safe_save(state, os.path.join(os.path.dirname(__file__), "net.safetensors"))
+    safe_save(state, (dirname / "net.safetensors").as_posix())
     ext = "js" if getenv("WEBGPU", "") != "" else "json"
-    with open(os.path.join(os.path.dirname(__file__), f"net.{ext}"), "w") as text_file:
+    with open(dirname / f"net.{ext}", "w") as text_file:
       text_file.write(prg)
   else:
     cprog = [prg]
diff --git a/examples/llama.py b/examples/llama.py
index 51c7c9328d..e42e7fd227 100755
--- a/examples/llama.py
+++ b/examples/llama.py
@@ -227,7 +227,7 @@ def concat_weights(models):
 def load(fn:str):
   if fn.endswith('.index.json'):
     with open(fn) as fp: weight_map = json.load(fp)['weight_map']
-    parts = {n: load(f'{os.path.dirname(fn)}/{os.path.basename(n)}') for n in set(weight_map.values())}
+    parts = {n: load(Path(fn).parent / Path(n).name) for n in set(weight_map.values())}
     return {k: parts[n][k] for k, n in weight_map.items()}
   elif fn.endswith('.safetensors'):
     return safe_load(fn)
@@ -428,7 +428,7 @@ After you are done speaking, output [EOS]. You are not Chad.
 
 
   LLAMA_SUFFIX = {1: "", 2: "-2"}[args.gen]
-  MODEL_PATH = args.model or Path(__file__).parent.parent / f"weights/LLaMA{LLAMA_SUFFIX}/{args.size}"
+  MODEL_PATH = args.model or Path(__file__).parents[1] / f"weights/LLaMA{LLAMA_SUFFIX}/{args.size}"
   TOKENIZER_PATH = (MODEL_PATH if MODEL_PATH.is_dir() else MODEL_PATH.parent) / "tokenizer.model"
   print(f"using LLaMA{LLAMA_SUFFIX}-{args.size} model")
   llama = LLaMa.build(MODEL_PATH, TOKENIZER_PATH, model_gen=args.gen, model_size=args.size, quantize=args.quantize)
diff --git a/examples/mlperf/model_eval.py b/examples/mlperf/model_eval.py
index c46601ec62..5a65f05ba0 100644
--- a/examples/mlperf/model_eval.py
+++ b/examples/mlperf/model_eval.py
@@ -173,7 +173,7 @@ def eval_bert():
   from examples.mlperf.metrics import f1_score
   from transformers import BertTokenizer
 
-  tokenizer = BertTokenizer(str(Path(__file__).parent.parent.parent / "weights/bert_vocab.txt"))
+  tokenizer = BertTokenizer(str(Path(__file__).parents[2] / "weights/bert_vocab.txt"))
 
   c = 0
   f1 = 0.0
diff --git a/examples/so_vits_svc.py b/examples/so_vits_svc.py
index 37a81be6fb..6763efdfdf 100644
--- a/examples/so_vits_svc.py
+++ b/examples/so_vits_svc.py
@@ -1,6 +1,6 @@
 # original implementation: https://github.com/svc-develop-team/so-vits-svc
 from __future__ import annotations
-import sys, os, logging, time, io, math, argparse, operator, numpy as np
+import sys, logging, time, io, math, argparse, operator, numpy as np
 from functools import partial, reduce
 from pathlib import Path
 from typing import Tuple, Optional, Type
@@ -468,14 +468,14 @@ def repeat_expand_2d_left(content, target_len): # content : [h, t]
   return Tensor.stack(cols).transpose(0, 1)
 
 def load_fairseq_cfg(checkpoint_path):
-  assert os.path.isfile(checkpoint_path)
+  assert Path(checkpoint_path).is_file()
   state = torch_load(checkpoint_path)
   cfg = state["cfg"] if ("cfg" in state and state["cfg"] is not None) else None
   if cfg is None: raise RuntimeError(f"No cfg exist in state keys = {state.keys()}")
   return HParams(**cfg)
 
 def load_checkpoint_enc(checkpoint_path, model: ContentVec, optimizer=None, skip_list=[]):
-  assert os.path.isfile(checkpoint_path)
+  assert Path(checkpoint_path).is_file()
   start_time = time.time()
   checkpoint_dict = torch_load(checkpoint_path)
   saved_state_dict = checkpoint_dict['model']
@@ -550,7 +550,7 @@ def get_encoder(ssl_dim) -> Type[SpeechEncoder]:
 # DEMO USAGE (uses audio sample from LJ-Speech):
 # python3 examples/so_vits_svc.py --model saul_goodman
 #########################################################################################
-SO_VITS_SVC_PATH = Path(__file__).parent.parent / "weights/So-VITS-SVC"
+SO_VITS_SVC_PATH = Path(__file__).parents[1] / "weights/So-VITS-SVC"
 VITS_MODELS = { # config_path, weights_path, config_url, weights_url
   "saul_goodman" : (SO_VITS_SVC_PATH / "config_saul_gman.json", SO_VITS_SVC_PATH / "pretrained_saul_gman.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/G_80000.pth"),
   "drake" : (SO_VITS_SVC_PATH / "config_drake.json", SO_VITS_SVC_PATH / "pretrained_drake.pth", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/config_aubrey.json", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/pretrained_aubrey.pth"),
@@ -563,13 +563,13 @@ ENCODER_MODELS = { # weights_path, weights_url
   "contentvec": (SO_VITS_SVC_PATH / "contentvec_checkpoint.pt", "https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt")
 }
 ENCODER_MODEL = "contentvec"
-DEMO_PATH, DEMO_URL = Path(__file__).parent.parent / "temp/LJ037-0171.wav", "https://keithito.com/LJ-Speech-Dataset/LJ037-0171.wav"
+DEMO_PATH, DEMO_URL = Path(__file__).parents[1] / "temp/LJ037-0171.wav", "https://keithito.com/LJ-Speech-Dataset/LJ037-0171.wav"
 if __name__=="__main__":
   logging.basicConfig(stream=sys.stdout, level=(logging.INFO if DEBUG < 1 else logging.DEBUG))
   parser = argparse.ArgumentParser()
   parser.add_argument("-m", "--model", default=None, help=f"Specify the model to use. All supported models: {VITS_MODELS.keys()}", required=True)
   parser.add_argument("-f", "--file", default=DEMO_PATH, help=f"Specify the path of the input file")
-  parser.add_argument("--out_dir", default=str(Path(__file__).parent.parent / "temp"), help="Specify the output path.")
+  parser.add_argument("--out_dir", default=str(Path(__file__).parents[1] / "temp"), help="Specify the output path.")
   parser.add_argument("--out_path", default=None, help="Specify the full output path. Overrides the --out_dir and --name parameter.")
   parser.add_argument("--base_name", default="test", help="Specify the base of the output file name. Default is 'test'.")
   parser.add_argument("--speaker", default=None, help="If not specified, the first available speaker is chosen. Usually there is only one speaker per model.")
@@ -600,7 +600,7 @@ if __name__=="__main__":
 
   ### Loading audio and slicing ###
   if audio_path == DEMO_PATH: download_if_not_present(DEMO_PATH, DEMO_URL)
-  assert os.path.isfile(audio_path) and Path(audio_path).suffix == ".wav"
+  assert Path(audio_path).is_file() and Path(audio_path).suffix == ".wav"
   chunks = preprocess.cut(audio_path, db_thresh=slice_db)
   audio_data, audio_sr = preprocess.chunks2audio(audio_path, chunks)
 
diff --git a/examples/stable_diffusion.py b/examples/stable_diffusion.py
index df9e98549d..761c167cb5 100644
--- a/examples/stable_diffusion.py
+++ b/examples/stable_diffusion.py
@@ -1,6 +1,5 @@
 # https://arxiv.org/pdf/2112.10752.pdf
 # https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md
-import os
 import tempfile
 from pathlib import Path
 import gzip, argparse, math, re
@@ -424,7 +423,7 @@ class CLIPTextTransformer:
 # Clip tokenizer, taken from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py (MIT license)
 @lru_cache()
 def default_bpe():
-  fn = Path(__file__).parent.parent / "weights/bpe_simple_vocab_16e6.txt.gz"
+  fn = Path(__file__).parents[1] / "weights/bpe_simple_vocab_16e6.txt.gz"
   download_file("https://github.com/openai/CLIP/raw/main/clip/bpe_simple_vocab_16e6.txt.gz", fn)
   return fn
 
@@ -558,13 +557,13 @@ class StableDiffusion:
 # cond_stage_model.transformer.text_model
 
 # this is sd-v1-4.ckpt
-FILENAME = Path(__file__).parent.parent / "weights/sd-v1-4.ckpt"
+FILENAME = Path(__file__).parents[1] / "weights/sd-v1-4.ckpt"
 
 if __name__ == "__main__":
   parser = argparse.ArgumentParser(description='Run Stable Diffusion', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
   parser.add_argument('--steps', type=int, default=5, help="Number of steps in diffusion")
   parser.add_argument('--prompt', type=str, default="a horse sized cat eating a bagel", help="Phrase to render")
-  parser.add_argument('--out', type=str, default=os.path.join(tempfile.gettempdir(), "rendered.png"), help="Output filename")
+  parser.add_argument('--out', type=str, default=Path(tempfile.gettempdir()) / "rendered.png", help="Output filename")
   parser.add_argument('--noshow', action='store_true', help="Don't show the image")
   parser.add_argument('--fp16', action='store_true', help="Cast the weights to float16")
   parser.add_argument('--timing', action='store_true', help="Print timing per step")
diff --git a/examples/vgg7.py b/examples/vgg7.py
index 706048c172..604f01937c 100644
--- a/examples/vgg7.py
+++ b/examples/vgg7.py
@@ -1,8 +1,8 @@
 import sys
-import os
 import random
 import json
 import numpy
+from pathlib import Path
 from PIL import Image
 from tinygrad.tensor import Tensor
 from tinygrad.nn.optim import SGD
@@ -80,8 +80,7 @@ if cmd == "import":
 
   vgg7.load_waifu2x_json(json.load(open(src, "rb")))
 
-  if not os.path.isdir(model):
-    os.mkdir(model)
+  Path(model).mkdir(exist_ok=True)
   load_and_save(model, True)
 elif cmd == "execute":
   model = sys.argv[2]
@@ -102,8 +101,7 @@ elif cmd == "execute_full":
 elif cmd == "new":
   model = sys.argv[2]
 
-  if not os.path.isdir(model):
-    os.mkdir(model)
+  Path(model).mkdir(exist_ok=True)
   load_and_save(model, True)
 elif cmd == "train":
   model = sys.argv[2]
diff --git a/examples/vgg7_helpers/kinne.py b/examples/vgg7_helpers/kinne.py
index f2a2495d9a..af05e42437 100644
--- a/examples/vgg7_helpers/kinne.py
+++ b/examples/vgg7_helpers/kinne.py
@@ -1,6 +1,6 @@
 from tinygrad.tensor import Tensor
 import numpy
-import os
+from pathlib import Path
 
 # Format Details:
 #  A KINNE parameter set is stored as a set of files named "snoop_bin_*.bin",
@@ -35,8 +35,8 @@ class KinneDir:
     It is important that if you wish to save in the current directory,
      you use ".", not the empty string.
     """
-    if save and not os.path.isdir(base):
-      os.mkdir(base)
+    if save:
+      Path(base).mkdir(exist_ok=True)
     self.base = base + "/snoop_bin_"
     self.next_part_index = 0
     self.save = save
diff --git a/examples/vits.py b/examples/vits.py
index 7b88490c9d..5583d9764f 100644
--- a/examples/vits.py
+++ b/examples/vits.py
@@ -1,4 +1,4 @@
-import json, logging, math, os, re, sys, time, wave, argparse, numpy as np
+import json, logging, math, re, sys, time, wave, argparse, numpy as np
 from functools import reduce
 from pathlib import Path
 from typing import List
@@ -522,7 +522,7 @@ def load_model(symbols, hps, model) -> Synthesizer:
   _ = load_checkpoint(weights_path, net_g, None)
   return net_g
 def load_checkpoint(checkpoint_path, model: Synthesizer, optimizer=None, skip_list=[]):
-  assert os.path.isfile(checkpoint_path)
+  assert Path(checkpoint_path).is_file()
   start_time = time.time()
   checkpoint_dict = torch_load(checkpoint_path)
   iteration, learning_rate = checkpoint_dict['iteration'], checkpoint_dict['learning_rate']
@@ -556,8 +556,8 @@ def load_checkpoint(checkpoint_path, model: Synthesizer, optimizer=None, skip_li
   return model, optimizer, learning_rate, iteration
 
 def download_if_not_present(file_path: Path, url: str):
-  if not os.path.isfile(file_path):
-    logging.info(f"Did not find {file_path}, downloading...")
+  if not file_path.is_file():
+    logging.info(f"Did not find {file_path.as_posix()}, downloading...")
     download_file(url, file_path)
   return file_path
 
@@ -649,7 +649,7 @@ class TextMapper: # Based on https://github.com/keithito/tacotron
 # anime lady 1 | --model_to_use uma_trilingual --speaker_id 36
 # anime lady 2 | --model_to_use uma_trilingual --speaker_id 121
 #########################################################################################
-VITS_PATH = Path(__file__).parent.parent / "weights/VITS/"
+VITS_PATH = Path(__file__).parents[1] / "weights/VITS/"
 MODELS = { # config_path, weights_path, config_url, weights_url
   "ljs": (VITS_PATH / "config_ljs.json", VITS_PATH / "pretrained_ljs.pth", "https://raw.githubusercontent.com/jaywalnut310/vits/main/configs/ljs_base.json", "https://drive.google.com/uc?export=download&id=1q86w74Ygw2hNzYP9cWkeClGT5X25PvBT&confirm=t"),
   "vctk": (VITS_PATH / "config_vctk.json", VITS_PATH / "pretrained_vctk.pth", "https://raw.githubusercontent.com/jaywalnut310/vits/main/configs/vctk_base.json", "https://drive.google.com/uc?export=download&id=11aHOlhnxzjpdWDpsz1vFDCzbeEfoIxru&confirm=t"),
@@ -665,7 +665,7 @@ if __name__ == '__main__':
   parser.add_argument("--model_to_use", default="vctk", help="Specify the model to use. Default is 'vctk'.")
   parser.add_argument("--speaker_id", type=int, default=6, help="Specify the speaker ID. Default is 6.")
   parser.add_argument("--out_path", default=None, help="Specify the full output path. Overrides the --out_dir and --name parameter.")
-  parser.add_argument("--out_dir", default=str(Path(__file__).parent.parent / "temp"), help="Specify the output path.")
+  parser.add_argument("--out_dir", default=str(Path(__file__).parents[1] / "temp"), help="Specify the output path.")
   parser.add_argument("--base_name", default="test", help="Specify the base of the output file name. Default is 'test'.")
   parser.add_argument("--text_to_synthesize", default="""Hello person. If the code you are contributing isn't some of the highest quality code you've written in your life, either put in the effort to make it great, or don't bother.""", help="Specify the text to synthesize. Default is a greeting message.")
   parser.add_argument("--noise_scale", type=float, default=0.667, help="Specify the noise scale. Default is 0.667.")
diff --git a/examples/whisper.py b/examples/whisper.py
index e36824c5ed..b1e50e8c79 100644
--- a/examples/whisper.py
+++ b/examples/whisper.py
@@ -136,7 +136,7 @@ LANGUAGES = {
   "as": "assamese", "tt": "tatar", "haw": "hawaiian", "ln": "lingala", "ha": "hausa", "ba": "bashkir", "jw": "javanese", "su": "sundanese",
 }
 
-BASE = pathlib.Path(__file__).parent.parent / "weights"
+BASE = pathlib.Path(__file__).parents[1] / "weights"
 def get_encoding(n_vocab_in):
   download_file("https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/gpt2.tiktoken", BASE / "gpt2.tiktoken")
   ranks = {base64.b64decode(token): int(rank) for token, rank in (line.split() for line in open(BASE / "gpt2.tiktoken") if line)}
diff --git a/examples/yolov8-onnx.py b/examples/yolov8-onnx.py
index 3bab3c2956..f75b5cb333 100644
--- a/examples/yolov8-onnx.py
+++ b/examples/yolov8-onnx.py
@@ -2,11 +2,12 @@
 import os
 from ultralytics import YOLO
 import onnx
+from pathlib import Path
 from extra.onnx import get_run_onnx
 from tinygrad.tensor import Tensor
 
 os.chdir("/tmp")
-if not os.path.isfile("yolov8n-seg.onnx"):
+if not Path("yolov8n-seg.onnx").is_file():
   model = YOLO("yolov8n-seg.pt")
   model.export(format="onnx", imgsz=[480,640])
 onnx_model = onnx.load(open("yolov8n-seg.onnx", "rb"))
diff --git a/examples/yolov8.py b/examples/yolov8.py
index b8e232765e..24a5a1bf85 100644
--- a/examples/yolov8.py
+++ b/examples/yolov8.py
@@ -6,7 +6,6 @@ from extra.utils import get_child, fetch, download_file
 from pathlib import Path
 import cv2
 from collections import defaultdict
-import os
 import time, io, sys
 from tinygrad.nn.state import safe_load, load_state_dict
 
@@ -398,13 +397,12 @@ if __name__ == '__main__':
   yolo_variant = sys.argv[2] if len(sys.argv) >= 3 else (print("No variant given, so choosing 'n' as the default. Yolov8 has different variants, you can choose from ['n', 's', 'm', 'l', 'x']") or 'n')
   print(f'running inference for YOLO version {yolo_variant}')
 
-  output_folder_path = './outputs_yolov8'
-  if not os.path.exists(output_folder_path):
-    os.makedirs(output_folder_path)
+  output_folder_path = Path('./outputs_yolov8')
+  output_folder_path.mkdir(parents=True, exist_ok=True)
   #absolute image path or URL
   image_location = [np.frombuffer(io.BytesIO(fetch(img_path)).read(), np.uint8)]
   image = [cv2.imdecode(image_location[0], 1)]
-  out_paths = [os.path.join(output_folder_path, img_path.split("/")[-1].split('.')[0] + "_output" + '.' + img_path.split("/")[-1].split('.')[1])]
+  out_paths = [(output_folder_path / f"{Path(img_path).stem}_output{Path(img_path).suffix}").as_posix()]
   if not isinstance(image[0], np.ndarray):
     print('Error in image loading. Check your image file.')
     sys.exit(1)
@@ -414,7 +412,7 @@ if __name__ == '__main__':
   depth, width, ratio = get_variant_multiples(yolo_variant)
   yolo_infer = YOLOv8(w=width, r=ratio, d=depth, num_classes=80)
 
-  weights_location = Path(__file__).parent.parent / "weights" / f'yolov8{yolo_variant}.safetensors'
+  weights_location = Path(__file__).parents[1] / "weights" / f'yolov8{yolo_variant}.safetensors'
   download_file(f'https://gitlab.com/r3sist/yolov8_weights/-/raw/master/yolov8{yolo_variant}.safetensors', weights_location)
 
   state_dict = safe_load(weights_location)
diff --git a/extra/accel/ane/lib/ane.py b/extra/accel/ane/lib/ane.py
index f7f710a69e..7b7710486d 100755
--- a/extra/accel/ane/lib/ane.py
+++ b/extra/accel/ane/lib/ane.py
@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-import os
+from pathlib import Path
 from ctypes import *
 import json
 import collections
@@ -8,13 +8,13 @@ import faulthandler
 import struct
 faulthandler.enable()
 
-basedir = os.path.dirname(os.path.abspath(os.path.realpath(__file__)))
+basedir = Path(__file__).resolve().parent
 
 libane = None
 aneregs = None
 def init_libane():
   global libane, aneregs
-  libane = cdll.LoadLibrary(os.path.join(basedir, "libane.dylib"))
+  libane = cdll.LoadLibrary((basedir / "libane.dylib").as_posix())
 
   libane.ANE_Compile.argtypes = [c_char_p, c_int]
   libane.ANE_Compile.restype = c_void_p
@@ -29,7 +29,7 @@ def init_libane():
 
   #libane.ANE_RegDebug.restype = c_char_p
 
-  with open(os.path.join(basedir, "aneregs.json")) as f:
+  with open(basedir / "aneregs.json") as f:
     aneregs = json.load(f)
 
 ANE_Struct = [
diff --git a/extra/augment.py b/extra/augment.py
index c520edae1c..06e7906c7d 100644
--- a/extra/augment.py
+++ b/extra/augment.py
@@ -1,9 +1,10 @@
 import numpy as np
 from PIL import Image
-import os
+from pathlib import Path
 import sys
-sys.path.append(os.getcwd())
-sys.path.append(os.path.join(os.getcwd(), 'test'))
+cwd = Path.cwd()
+sys.path.append(cwd.as_posix())
+sys.path.append((cwd / 'test').as_posix())
 from extra.datasets import fetch_mnist
 from tqdm import trange
 
diff --git a/extra/datasets/__init__.py b/extra/datasets/__init__.py
index cfe92948a6..4d4a1a4f34 100644
--- a/extra/datasets/__init__.py
+++ b/extra/datasets/__init__.py
@@ -1,15 +1,17 @@
-import os, random, gzip, tarfile, pickle
+import random, gzip, tarfile, pickle
 import numpy as np
+from pathlib import Path
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import dtypes
 from extra.utils import download_file
 
 def fetch_mnist():
   parse = lambda file: np.frombuffer(gzip.open(file).read(), dtype=np.uint8).copy()
-  X_train = parse(os.path.dirname(__file__)+"/mnist/train-images-idx3-ubyte.gz")[0x10:].reshape((-1, 28*28)).astype(np.float32)
-  Y_train = parse(os.path.dirname(__file__)+"/mnist/train-labels-idx1-ubyte.gz")[8:]
-  X_test = parse(os.path.dirname(__file__)+"/mnist/t10k-images-idx3-ubyte.gz")[0x10:].reshape((-1, 28*28)).astype(np.float32)
-  Y_test = parse(os.path.dirname(__file__)+"/mnist/t10k-labels-idx1-ubyte.gz")[8:]
+  dirname = Path(__file__).parent.resolve()
+  X_train = parse(dirname / "mnist/train-images-idx3-ubyte.gz")[0x10:].reshape((-1, 28*28)).astype(np.float32)
+  Y_train = parse(dirname / "mnist/train-labels-idx1-ubyte.gz")[8:]
+  X_test = parse(dirname / "mnist/t10k-images-idx3-ubyte.gz")[0x10:].reshape((-1, 28*28)).astype(np.float32)
+  Y_test = parse(dirname / "mnist/t10k-labels-idx1-ubyte.gz")[8:]
   return X_train, Y_train, X_test, Y_test
 
 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
@@ -31,7 +33,7 @@ def fetch_cifar(shuffle=False):
       Y[idx:idx+bs].assign(y[order])
       idx += bs
     return X, Y  
-  fn = os.path.dirname(__file__)+"/cifar-10-python.tar.gz"
+  fn = Path(__file__).parent.resolve() / "cifar-10-python.tar.gz"
   download_file('https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz', fn)
   tt = tarfile.open(fn, mode='r:gz')
   db = [pickle.load(tt.extractfile(f'cifar-10-batches-py/data_batch_{i}'), encoding="bytes") for i in range(1,6)]
diff --git a/extra/datasets/squad.py b/extra/datasets/squad.py
index f9c1c7c915..1b95385ddf 100644
--- a/extra/datasets/squad.py
+++ b/extra/datasets/squad.py
@@ -141,7 +141,7 @@ def iterate(tokenizer, start=0):
     yield features, example
 
 if __name__ == "__main__":
-  tokenizer = BertTokenizer(str(Path(__file__).parent.parent.parent / "weights" / "bert_vocab.txt"))
+  tokenizer = BertTokenizer(str(Path(__file__).parents[2] / "weights" / "bert_vocab.txt"))
 
   X, Y = next(iterate(tokenizer))
   print(" ".join(X[0]["tokens"]))
diff --git a/extra/utils.py b/extra/utils.py
index 2d0382d6d5..af0757e47e 100644
--- a/extra/utils.py
+++ b/extra/utils.py
@@ -1,7 +1,8 @@
 import pickle
 import numpy as np
 from tqdm import tqdm
-import tempfile, platform, os
+import tempfile, platform
+from pathlib import Path
 from collections import defaultdict
 from tinygrad.helpers import prod, getenv, DEBUG, dtypes
 from tinygrad.ops import GlobalCounters
@@ -11,7 +12,7 @@ from tinygrad.shape.shapetracker import strides_for_shape
 OSX = platform.system() == "Darwin"
 WINDOWS = platform.system() == "Windows"
 
-def temp(x:str) -> str: return os.path.join(tempfile.gettempdir(), x)
+def temp(x:str) -> str: return (Path(tempfile.gettempdir()) / x).as_posix()
 
 def fetch(url):
   if url.startswith("/") or url.startswith("."):
@@ -33,19 +34,18 @@ def fetch_as_file(url):
   return fp
 
 def download_file(url, fp, skip_if_exists=True):
-  import requests, pathlib
-  if skip_if_exists and os.path.isfile(fp) and os.stat(fp).st_size > 0:
+  import requests
+  if skip_if_exists and Path(fp).is_file() and Path(fp).stat().st_size > 0:
     return
   r = requests.get(url, stream=True)
   assert r.status_code == 200
   progress_bar = tqdm(total=int(r.headers.get('content-length', 0)), unit='B', unit_scale=True, desc=url)
-  (path := pathlib.Path(fp).parent).mkdir(parents=True, exist_ok=True)
+  (path := Path(fp).parent).mkdir(parents=True, exist_ok=True)
   with tempfile.NamedTemporaryFile(dir=path, delete=False) as f:
     for chunk in r.iter_content(chunk_size=16384):
       progress_bar.update(f.write(chunk))
     f.close()
-    os.rename(f.name, fp)
-
+    Path(f.name).rename(fp)
 
 def my_unpickle(fb0):
   key_prelookup = defaultdict(list)
diff --git a/models/bert.py b/models/bert.py
index 7a7bda5481..e0d32dc167 100644
--- a/models/bert.py
+++ b/models/bert.py
@@ -10,9 +10,9 @@ class BertForQuestionAnswering:
     self.qa_outputs = Linear(hidden_size, 2)
 
   def load_from_pretrained(self):
-    fn = Path(__file__).parent.parent / "weights/bert_for_qa.pt"
+    fn = Path(__file__).parents[1] / "weights/bert_for_qa.pt"
     download_file("https://zenodo.org/record/3733896/files/model.pytorch?download=1", fn)
-    fn_vocab = Path(__file__).parent.parent / "weights/bert_vocab.txt"
+    fn_vocab = Path(__file__).parents[1] / "weights/bert_vocab.txt"
     download_file("https://zenodo.org/record/3733896/files/vocab.txt?download=1", fn_vocab)
 
     import torch
diff --git a/models/rnnt.py b/models/rnnt.py
index 1da9f487e3..88523dcce9 100644
--- a/models/rnnt.py
+++ b/models/rnnt.py
@@ -60,7 +60,7 @@ class RNNT:
     return out.realize()
 
   def load_from_pretrained(self):
-    fn = Path(__file__).parent.parent / "weights/rnnt.pt"
+    fn = Path(__file__).parents[1] / "weights/rnnt.pt"
     download_file("https://zenodo.org/record/3662521/files/DistributedDataParallel_1576581068.9962234-epoch-100.pt?download=1", fn)
 
     import torch
diff --git a/models/unet3d.py b/models/unet3d.py
index 555c622214..b67225f9fe 100644
--- a/models/unet3d.py
+++ b/models/unet3d.py
@@ -46,7 +46,7 @@ class UNet3D:
     return x
 
   def load_from_pretrained(self):
-    fn = Path(__file__).parent.parent / "weights" / "unet-3d.ckpt"
+    fn = Path(__file__).parents[1] / "weights" / "unet-3d.ckpt"
     download_file("https://zenodo.org/record/5597155/files/3dunet_kits19_pytorch.ptc?download=1", fn)
     state_dict = torch.jit.load(fn, map_location=torch.device("cpu")).state_dict()
     for k, v in state_dict.items():
diff --git a/openpilot/compile.py b/openpilot/compile.py
index e459e6900a..17bada5da5 100644
--- a/openpilot/compile.py
+++ b/openpilot/compile.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 import os, time, io, pathlib, sys, traceback
-sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+sys.path.insert(0, str(pathlib.Path(__file__).parents[1]))
 
 if os.getenv("OPT", None) is None:
   os.environ['OPT'] = '99'
diff --git a/setup.py b/setup.py
index 54b9eb1196..e1e5db6864 100644
--- a/setup.py
+++ b/setup.py
@@ -1,10 +1,10 @@
 #!/usr/bin/env python3
 
-import os
+from pathlib import Path
 from setuptools import setup
 
-directory = os.path.abspath(os.path.dirname(__file__))
-with open(os.path.join(directory, 'README.md'), encoding='utf-8') as f:
+directory = Path(__file__).resolve().parent
+with open(directory / 'README.md', encoding='utf-8') as f:
   long_description = f.read()
 
 setup(name='tinygrad',
diff --git a/sz.py b/sz.py
index 272d6e9ec7..3514e9d32e 100755
--- a/sz.py
+++ b/sz.py
@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 import os
+from pathlib import Path
 import token
 import tokenize
 import itertools
@@ -13,11 +14,11 @@ if __name__ == "__main__":
   for path, subdirs, files in os.walk("tinygrad"):
     for name in files:
       if not name.endswith(".py"): continue
-      filepath = os.path.join(path, name)
+      filepath = Path(path) / name
       with tokenize.open(filepath) as file_:
         tokens = [t for t in tokenize.generate_tokens(file_.readline) if t.type in TOKEN_WHITELIST]
         token_count, line_count = len(tokens), len(set([t.start[0] for t in tokens]))
-        table.append([filepath, line_count, token_count/line_count])
+        table.append([filepath.as_posix(), line_count, token_count/line_count])
 
   print(tabulate([headers] + sorted(table, key=lambda x: -x[1]), headers="firstrow", floatfmt=".1f")+"\n")
 
diff --git a/test/external/external_model_benchmark.py b/test/external/external_model_benchmark.py
index becfe2376a..1342e88551 100644
--- a/test/external/external_model_benchmark.py
+++ b/test/external/external_model_benchmark.py
@@ -45,7 +45,7 @@ def benchmark(mnm, nm, fxn):
   CSV[nm] = min(tms)*1e-6
   return min(tms), ret
 
-#BASE = pathlib.Path(__file__).parent.parent.parent / "weights" / "onnx"
+#BASE = pathlib.Path(__file__).parents[2] / "weights" / "onnx"
 BASE = pathlib.Path("/tmp/onnx")
 def benchmark_model(m, validate_outs=False):
   global open_csv, CSV
diff --git a/test/external/external_test_yolov8.py b/test/external/external_test_yolov8.py
index 98bf555cf6..df5ae04e15 100644
--- a/test/external/external_test_yolov8.py
+++ b/test/external/external_test_yolov8.py
@@ -11,7 +11,7 @@ from tinygrad.nn.state import safe_load, load_state_dict
 class TestYOLOv8(unittest.TestCase):
   def test_all_load_weights(self):
     for variant in ['n', 's', 'm', 'l', 'x']:
-      weights_location = Path(__file__).parent.parent.parent / "weights" / f'yolov8{variant}.safetensors'
+      weights_location = Path(__file__).parents[2] / "weights" / f'yolov8{variant}.safetensors'
       download_file(f'https://gitlab.com/r3sist/yolov8_weights/-/raw/master/yolov8{variant}.safetensors', weights_location)
 
       depth, width, ratio = get_variant_multiples(variant)
@@ -23,7 +23,7 @@ class TestYOLOv8(unittest.TestCase):
   def test_predictions(self):
     test_image_urls = ['https://raw.githubusercontent.com/ultralytics/yolov5/master/data/images/bus.jpg', 'https://www.aljazeera.com/wp-content/uploads/2022/10/2022-04-28T192650Z_1186456067_UP1EI4S1I0P14_RTRMADP_3_SOCCER-ENGLAND-MUN-CHE-REPORT.jpg']
     variant = 'n'
-    weights_location = Path(__file__).parent.parent.parent / "weights" / f'yolov8{variant}.safetensors'
+    weights_location = Path(__file__).parents[2] / "weights" / f'yolov8{variant}.safetensors'
     depth, width, ratio = get_variant_multiples(variant)
     TinyYolov8 = YOLOv8(w=width, r=ratio, d=depth, num_classes=80)
     state_dict = safe_load(weights_location)
@@ -40,13 +40,13 @@ class TestYOLOv8(unittest.TestCase):
 
   def test_forward_pass_torch_onnx(self):
     variant = 'n'
-    weights_location_onnx = Path(__file__).parent.parent.parent / "weights" / f'yolov8{variant}.onnx'
-    weights_location_pt = Path(__file__).parent.parent.parent / "weights" / f'yolov8{variant}.pt'
-    weights_location = Path(__file__).parent.parent.parent / "weights" / f'yolov8{variant}.safetensors'
+    weights_location_onnx = Path(__file__).parents[2] / "weights" / f'yolov8{variant}.onnx'
+    weights_location_pt = Path(__file__).parents[2] / "weights" / f'yolov8{variant}.pt'
+    weights_location = Path(__file__).parents[2] / "weights" / f'yolov8{variant}.safetensors'
 
     download_file(f'https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8{variant}.pt', weights_location_pt)
     # the ultralytics export prints a lot of unneccesary things
-    if not os.path.isfile(weights_location_onnx):
+    if not weights_location_onnx.is_file():
       model = ultralytics.YOLO(model=weights_location_pt, task='Detect')
       model.export(format="onnx",imgsz=[640, 480])
 
diff --git a/test/extra/test_utils.py b/test/extra/test_utils.py
index c73a8ca455..b18aa0c814 100644
--- a/test/extra/test_utils.py
+++ b/test/extra/test_utils.py
@@ -64,10 +64,10 @@ class TestDownloadFile(unittest.TestCase):
     mock_response.status_code = 200
     mock_response.headers = {'content-length': '8'}
     mock_requests.return_value = mock_response
-    self.assertFalse(os.path.exists(self.test_file.parent))
+    self.assertFalse(self.test_file.parent.exists())
     download_file("https://www.mock.com/fake.txt", self.test_file, skip_if_exists=False)
-    self.assertTrue(os.path.exists(self.test_file.parent))
-    self.assertTrue(os.path.isfile(self.test_file))
+    self.assertTrue(self.test_file.parent.exists())
+    self.assertTrue(self.test_file.is_file())
     self.assertEqual('12345678', self.test_file.read_text())
 
 class TestUtils(unittest.TestCase):
diff --git a/test/unit/test_disk_tensor.py b/test/unit/test_disk_tensor.py
index afb360eff2..5e5ad88529 100644
--- a/test/unit/test_disk_tensor.py
+++ b/test/unit/test_disk_tensor.py
@@ -28,7 +28,7 @@ class TestTorchLoad(unittest.TestCase):
   # TODO: support pytorch tar format with minimal lines
   #def test_load_resnet(self): compare_weights_both('https://download.pytorch.org/models/resnet50-19c8e357.pth')
 
-test_fn = pathlib.Path(__file__).parent.parent.parent / "weights/LLaMA/7B/consolidated.00.pth"
+test_fn = pathlib.Path(__file__).parents[2] / "weights/LLaMA/7B/consolidated.00.pth"
 #test_size = test_fn.stat().st_size
 test_size = 1024*1024*1024*2
 
diff --git a/tinygrad/runtime/ops_cuda.py b/tinygrad/runtime/ops_cuda.py
index fe5eafacc8..cfb4fe8add 100644
--- a/tinygrad/runtime/ops_cuda.py
+++ b/tinygrad/runtime/ops_cuda.py
@@ -1,4 +1,5 @@
-import subprocess, time, re, hashlib, tempfile, os, functools
+import subprocess, time, re, hashlib, tempfile, functools
+from pathlib import Path
 from typing import Optional
 import numpy as np
 from pycuda.compiler import compile as cuda_compile # type: ignore
@@ -66,7 +67,7 @@ class CUDAProgram:
     if DEBUG >= 5: print(pretty_ptx(prg))
     if DEBUG >= 6:
       try:
-        fn = os.path.join(tempfile.gettempdir(), f"tinycuda_{hashlib.md5(prg.encode('utf-8')).hexdigest()}")
+        fn = (Path(tempfile.gettempdir()) / f"tinycuda_{hashlib.md5(prg.encode('utf-8')).hexdigest()}").as_posix()
         with open(fn + ".ptx", "wb") as f: f.write(prg.encode('utf-8'))
         subprocess.run(["ptxas", f"-arch={arch()}", "-o", fn, fn+".ptx"], check=True)
         print(subprocess.check_output(['nvdisasm', fn]).decode('utf-8'))
diff --git a/tinygrad/runtime/ops_gpu.py b/tinygrad/runtime/ops_gpu.py
index ceb674c0fe..75f658cb29 100644
--- a/tinygrad/runtime/ops_gpu.py
+++ b/tinygrad/runtime/ops_gpu.py
@@ -13,7 +13,7 @@ OSX_TIMING_RATIO = (125/3) if OSX else 1.0   # see test/external_osx_profiling.p
 
 # TODO: if you fork and exit the child process after creating anything with cl on AMD, it hangs on e.wait()
 ROCM_LLVM_PATH = pathlib.Path("/opt/rocm/llvm/bin")
-#ROCM_LLVM_PATH = pathlib.Path(__file__).parent.parent.parent.parent / "extra/rocm/build/llvm-project/bin"
+#ROCM_LLVM_PATH = pathlib.Path(__file__).parents[3] / "extra/rocm/build/llvm-project/bin"
 if DEBUG >= 5:
   early_exec = fromimport("extra.helpers", "enable_early_exec")()
 
diff --git a/tinygrad/runtime/ops_metal.py b/tinygrad/runtime/ops_metal.py
index 3a1ba19a36..f093322734 100644
--- a/tinygrad/runtime/ops_metal.py
+++ b/tinygrad/runtime/ops_metal.py
@@ -60,7 +60,7 @@ class MetalProgram:
       unwrap(arc.addComputePipelineFunctionsWithDescriptor_error_(desc, None))
       unwrap(arc.serializeToURL_error_(Cocoa.NSURL.URLWithString_("file:///tmp/shader.bin"), None))
       # clone https://github.com/dougallj/applegpu.git in tinygrad/disassemblers
-      os.system(f"cd {pathlib.Path(__file__).parent.parent.parent}/disassemblers/applegpu && python3 compiler_explorer.py /tmp/shader.bin")
+      os.system(f"cd {pathlib.Path(__file__).parents[2]}/disassemblers/applegpu && python3 compiler_explorer.py /tmp/shader.bin")
     self.pipeline_state = unwrap(METAL.device.newComputePipelineStateWithFunction_error_(self.fxn, None))
 
   def __call__(self, global_size, local_size, *bufs, wait=False):