tinygrad 0.10.0__py3-none-any.whl → 0.10.2__py3-none-any.whl

tinygrad/nn/__init__.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 import math
-from typing import Optional, Union, Tuple, List
-from tinygrad.tensor import Tensor, dtypes
+from tinygrad.tensor import Tensor
+from tinygrad.dtype import dtypes
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import prod, make_tuple, flatten
 from tinygrad.nn import optim, state, datasets  # noqa: F401
@@ -34,14 +34,14 @@ class BatchNorm:
   def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1):
     self.eps, self.track_running_stats, self.momentum = eps, track_running_stats, momentum
 
-    self.weight: Optional[Tensor] = Tensor.ones(sz) if affine else None
-    self.bias: Optional[Tensor] = Tensor.zeros(sz) if affine else None
+    self.weight: Tensor|None = Tensor.ones(sz) if affine else None
+    self.bias: Tensor|None = Tensor.zeros(sz) if affine else None
 
     self.num_batches_tracked = Tensor.zeros(1, dtype='long' if is_dtype_supported(dtypes.long) else 'int', requires_grad=False)
     if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, requires_grad=False), Tensor.ones(sz, requires_grad=False)
 
-  def calc_stats(self, x:Tensor) -> Tuple[Tensor, Tensor]:
-    shape_mask: List[int] = [1, -1, *([1]*(x.ndim-2))]
+  def calc_stats(self, x:Tensor) -> tuple[Tensor, Tensor]:
+    shape_mask: list[int] = [1, -1, *([1]*(x.ndim-2))]
     if self.track_running_stats and not Tensor.training: return self.running_mean, self.running_var.reshape(shape=shape_mask).expand(x.shape)
     # This requires two full memory accesses to x
     # https://github.com/pytorch/pytorch/blob/c618dc13d2aa23625cb0d7ada694137532a4fa33/aten/src/ATen/native/cuda/Normalization.cuh
@@ -56,12 +56,12 @@ class BatchNorm:
     # NOTE: wow, this is done all throughout training in most PyTorch models
     if self.track_running_stats and Tensor.training:
       self.running_mean.assign((1-self.momentum) * self.running_mean + self.momentum * batch_mean.detach())
-      self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * prod(x.shape)/(prod(x.shape)-x.shape[1]) * batch_var.detach())
+      self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * x.numel()/(x.numel()-x.shape[1]) * batch_var.detach())
       self.num_batches_tracked += 1
     return x.batchnorm(self.weight, self.bias, batch_mean, batch_var.add(self.eps).rsqrt())
 BatchNorm2d = BatchNorm3d = BatchNorm
 
-def Conv1d(in_channels:int, out_channels:int, kernel_size:int, stride=1, padding:Union[int, str]=0, dilation=1, groups=1, bias=True) -> Conv2d:
+def Conv1d(in_channels:int, out_channels:int, kernel_size:int, stride=1, padding:int|str=0, dilation=1, groups=1, bias=True) -> Conv2d:
   """
   Applies a 1D convolution over an input signal composed of several input planes.
 
@@ -95,7 +95,7 @@ class Conv2d:
   print(t.numpy())
   ```
   """
-  def __init__(self, in_channels:int, out_channels:int, kernel_size:Union[int, Tuple[int, ...]], stride=1, padding:Union[int, Tuple[int, ...], str]=0,
+  def __init__(self, in_channels:int, out_channels:int, kernel_size:int|tuple[int, ...], stride=1, padding:int|tuple[int, ...]|str=0,
                dilation=1, groups=1, bias=True):
     self.kernel_size = make_tuple(kernel_size, 2)
     if isinstance(padding, str):
@@ -106,10 +106,9 @@ class Conv2d:
     self.stride, self.dilation, self.groups, self.padding = stride, dilation, groups, padding
     scale = 1 / math.sqrt(in_channels * prod(self.kernel_size))
     self.weight = Tensor.uniform(out_channels, in_channels//groups, *self.kernel_size, low=-scale, high=scale)
-    self.bias: Optional[Tensor] = Tensor.uniform(out_channels, low=-scale, high=scale) if bias else None
+    self.bias: Tensor|None = Tensor.uniform(out_channels, low=-scale, high=scale) if bias else None
 
-  def __call__(self, x:Tensor) -> Tensor:
-    return x.conv2d(self.weight, self.bias, padding=self.padding, stride=self.stride, dilation=self.dilation, groups=self.groups)
+  def __call__(self, x:Tensor) -> Tensor: return x.conv2d(self.weight, self.bias, self.groups, self.stride, self.dilation, self.padding)
 
 def ConvTranspose1d(in_channels:int, out_channels:int, kernel_size:int, stride=1, padding=0, output_padding=0, dilation=1,
                       groups=1, bias=True) -> ConvTranspose2d:
@@ -146,7 +145,7 @@ class ConvTranspose2d(Conv2d):
   print(t.numpy())
   ```
   """
-  def __init__(self, in_channels:int, out_channels:int, kernel_size:Union[int, Tuple[int, ...]], stride=1, padding=0, output_padding=0,
+  def __init__(self, in_channels:int, out_channels:int, kernel_size:int|tuple[int, ...], stride=1, padding=0, output_padding=0,
                 dilation=1, groups=1, bias=True):
     super().__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias)
     scale = 1 / math.sqrt(in_channels * prod(self.kernel_size))
@@ -154,8 +153,7 @@ class ConvTranspose2d(Conv2d):
     self.output_padding = output_padding
 
   def __call__(self, x:Tensor) -> Tensor:
-    return x.conv_transpose2d(self.weight, self.bias, padding=self.padding, output_padding=self.output_padding, stride=self.stride,
-                              dilation=self.dilation, groups=self.groups)
+    return x.conv_transpose2d(self.weight, self.bias, self.groups, self.stride, self.dilation, self.padding, self.output_padding)
 
 class Linear:
   """
@@ -178,8 +176,7 @@ class Linear:
     self.weight = Tensor.uniform(out_features, in_features, low=-bound, high=bound)
     self.bias = Tensor.uniform(out_features, low=-bound, high=bound) if bias else None
 
-  def __call__(self, x:Tensor) -> Tensor:
-    return x.linear(self.weight.transpose(), self.bias)
+  def __call__(self, x:Tensor) -> Tensor: return x.linear(self.weight.transpose(), self.bias)
 
 class GroupNorm:
   """
@@ -200,8 +197,8 @@ class GroupNorm:
   """
   def __init__(self, num_groups:int, num_channels:int, eps=1e-5, affine=True):
     self.num_groups, self.num_channels, self.eps = num_groups, num_channels, eps
-    self.weight: Optional[Tensor] = Tensor.ones(num_channels) if affine else None
-    self.bias: Optional[Tensor] = Tensor.zeros(num_channels) if affine else None
+    self.weight: Tensor|None = Tensor.ones(num_channels) if affine else None
+    self.bias: Tensor|None = Tensor.zeros(num_channels) if affine else None
 
   def __call__(self, x:Tensor) -> Tensor:
     # reshape for layernorm to work as group norm
@@ -210,7 +207,7 @@ class GroupNorm:
 
     if self.weight is None or self.bias is None: return x
     # elementwise_affine on channels
-    return x * self.weight.reshape(1, -1, *[1] * (len(x.shape)-2)) + self.bias.reshape(1, -1, *[1] * (len(x.shape)-2))
+    return x * self.weight.reshape(1, -1, *[1] * (x.ndim-2)) + self.bias.reshape(1, -1, *[1] * (x.ndim-2))
 
 class InstanceNorm:
   """
@@ -231,13 +228,13 @@ class InstanceNorm:
   """
   def __init__(self, num_features:int, eps=1e-5, affine=True):
     self.num_features, self.eps = num_features, eps
-    self.weight: Optional[Tensor] = Tensor.ones(num_features) if affine else None
-    self.bias: Optional[Tensor] = Tensor.zeros(num_features) if affine else None
+    self.weight: Tensor|None = Tensor.ones(num_features) if affine else None
+    self.bias: Tensor|None = Tensor.zeros(num_features) if affine else None
 
   def __call__(self, x:Tensor) -> Tensor:
     x = x.reshape(x.shape[0], self.num_features, -1).layernorm(eps=self.eps).reshape(x.shape)
     if self.weight is None or self.bias is None: return x
-    return x * self.weight.reshape(1, -1, *[1] * (len(x.shape)-2)) + self.bias.reshape(1, -1, *[1] * (len(x.shape)-2))
+    return x * self.weight.reshape(1, -1, *[1] * (x.ndim-2)) + self.bias.reshape(1, -1, *[1] * (x.ndim-2))
 
 class LayerNorm:
   """
@@ -256,10 +253,11 @@ class LayerNorm:
   print(t.mean().item(), t.std().item())
   ```
   """
-  def __init__(self, normalized_shape:Union[int, Tuple[int, ...]], eps=1e-5, elementwise_affine=True):
-    self.normalized_shape: Tuple[int, ...] = (normalized_shape,) if isinstance(normalized_shape, int) else tuple(normalized_shape)
+  def __init__(self, normalized_shape:int|tuple[int, ...], eps=1e-5, elementwise_affine=True):
+    self.normalized_shape: tuple[int, ...] = make_tuple(normalized_shape, 1)
     self.axis, self.eps, self.elementwise_affine = tuple(-1-i for i in range(len(self.normalized_shape))), eps, elementwise_affine
-    self.weight, self.bias = (Tensor.ones(*self.normalized_shape), Tensor.zeros(*self.normalized_shape)) if elementwise_affine else (None, None)
+    self.weight: Tensor|None = Tensor.ones(*self.normalized_shape) if elementwise_affine else None
+    self.bias: Tensor|None = Tensor.zeros(*self.normalized_shape) if elementwise_affine else None
 
   def __call__(self, x:Tensor) -> Tensor:
     assert self.normalized_shape == x.shape[-len(self.normalized_shape):], f"last dimensions of {x.shape} must match {self.normalized_shape}"
@@ -322,10 +320,9 @@ class Embedding:
     self.vocab_sz, self.embed_sz, self.weight = vocab_size, embed_size, Tensor.glorot_uniform(vocab_size, embed_size)
 
   def __call__(self, idx:Tensor) -> Tensor:
-    if idx.numel() == 0: return Tensor.empty(idx.shape+(self.embed_sz,), device=self.weight.device)
-    arange_shp, weight_shp, big_shp = (self.vocab_sz, 1), (self.vocab_sz, self.embed_sz), idx.shape+(self.vocab_sz, self.embed_sz,)
-    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).reshape(arange_shp)
-    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1,)).expand(big_shp), self.weight.reshape(weight_shp).expand(big_shp)
+    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).unsqueeze(-1)
+    big_shp = idx.shape+(self.vocab_sz, self.embed_sz)
+    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1)).expand(big_shp), self.weight.expand(big_shp)
     return (arange == idx).mul(vals).sum(-2, acc_dtype=vals.dtype)
 
 class LSTMCell:
@@ -341,9 +338,10 @@ class LSTMCell:
     stdv = 1.0 / math.sqrt(hidden_size)
     self.weight_ih = Tensor.uniform(hidden_size*4, input_size, low=-stdv, high=stdv)
     self.weight_hh = Tensor.uniform(hidden_size*4, hidden_size, low=-stdv, high=stdv)
-    self.bias_ih, self.bias_hh = (Tensor.zeros(hidden_size*4), Tensor.zeros(hidden_size*4)) if bias else (None, None)
+    self.bias_ih: Tensor|None = Tensor.zeros(hidden_size*4) if bias else None
+    self.bias_hh: Tensor|None = Tensor.zeros(hidden_size*4) if bias else None
 
-  def __call__(self, x:Tensor, hc:Optional[Tuple[Tensor, Tensor]]=None) -> Tuple[Tensor, Tensor]:
+  def __call__(self, x:Tensor, hc:tuple[Tensor, Tensor]|None=None) -> tuple[Tensor, Tensor]:
     if hc is None: hc = (Tensor.zeros(x.size(0), self.weight_hh.size(1), dtype=x.dtype, device=x.device),)*2
     gates = x.linear(self.weight_ih.T, self.bias_ih) + hc[0].linear(self.weight_hh.T, self.bias_hh)
     i, f, g, o = gates.chunk(4, dim=1)

tinygrad/nn/datasets.py

@@ -1,5 +1,4 @@
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import fetch
 from tinygrad.nn.state import tar_extract
 
 def mnist(device=None, fashion=False):
@@ -9,7 +8,7 @@ def mnist(device=None, fashion=False):
          _mnist("t10k-images-idx3-ubyte.gz")[0x10:].reshape(-1,1,28,28).to(device), _mnist("t10k-labels-idx1-ubyte.gz")[8:].to(device)
 
 def cifar(device=None):
-  tt = tar_extract(fetch('https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz', gunzip=True))
+  tt = tar_extract(Tensor.from_url('https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz', gunzip=True))
   train = Tensor.cat(*[tt[f"cifar-10-batches-bin/data_batch_{i}.bin"].reshape(-1, 3073).to(device) for i in range(1,6)])
   test = tt["cifar-10-batches-bin/test_batch.bin"].reshape(-1, 3073).to(device)
   return train[:, 1:].reshape(-1,3,32,32), train[:, 0], test[:, 1:].reshape(-1,3,32,32), test[:, 0]

tinygrad/nn/optim.py

@@ -1,6 +1,5 @@
 # sorted in order of increasing complexity
-from typing import List
-from tinygrad.helpers import dedup, flatten, getenv
+from tinygrad.helpers import dedup, flatten, getenv, unwrap
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes, least_upper_dtype
 
@@ -8,15 +7,15 @@ class Optimizer:
   """
   Base class for all optimizers.
   """
-  def __init__(self, params: List[Tensor], lr: float):
+  def __init__(self, params: list[Tensor], lr: float):
     # if it's None, but being put into an optimizer, set it to True
     for x in params:
       if x.requires_grad is None: x.requires_grad = True
 
-    self.params: List[Tensor] = dedup([x for x in params if x.requires_grad])
+    self.params: list[Tensor] = dedup([x for x in params if x.requires_grad])
     assert len(self.params) != 0, "optimizer must have at least one param"
     self.device = self.params[0].device
-    self.buffers: List[Tensor] = dedup([x for x in params if not x.requires_grad])   # buffers are still realized
+    self.buffers: list[Tensor] = dedup([x for x in params if not x.requires_grad])   # buffers are still realized
     # store lr in at least float32 precision
     self.lr = Tensor(lr if getenv("CONST_LR") else [lr], requires_grad=False, device=self.device,
                      dtype=least_upper_dtype(dtypes.default_float, dtypes.float32))
@@ -32,15 +31,17 @@ class Optimizer:
     Performs a single optimization step.
     """
     Tensor.realize(*self.schedule_step())
-  def schedule_step(self) -> List[Tensor]:
+
+  def schedule_step(self) -> list[Tensor]:
     """
     Returns the tensors that need to be realized to perform a single optimization step.
     """
-    assert Tensor.training, (
+    if not Tensor.training: raise RuntimeError(
             f"""Tensor.training={Tensor.training}, Tensor.training must be enabled to use the optimizer.
                 - help: Consider setting Tensor.training=True before calling Optimizer.step().""")
-    return self._step()+self.params+self.buffers
-  def _step(self) -> List[Tensor]: raise NotImplementedError
+    return self.schedule_step_with_grads([unwrap(t.grad) for t in self.params])+self.params+self.buffers
+
+  def schedule_step_with_grads(self, grads:list[Tensor]) -> list[Tensor]: raise NotImplementedError
 
 class OptimizerGroup(Optimizer):
   """
@@ -51,10 +52,10 @@ class OptimizerGroup(Optimizer):
     self.params, self.buffers = flatten([o.params for o in self.optimizers]), flatten([o.buffers for o in self.optimizers])
   def __getitem__(self, i): return self.optimizers[i]
   def zero_grad(self): [o.zero_grad() for o in self.optimizers]
-  def _step(self) -> List[Tensor]: return [x for o in self.optimizers for x in o._step()]
+  def schedule_step(self) -> list[Tensor]: return [x for o in self.optimizers for x in o.schedule_step()]
 
 # LARS is essentially just trust ratio to SGD so if we just set the trust coeff 0.0 its just standard SGD.
-def SGD(params: List[Tensor], lr=0.001, momentum=0.0, weight_decay=0.0, nesterov=False, classic=False):
+def SGD(params: list[Tensor], lr=0.001, momentum=0.0, weight_decay=0.0, nesterov=False, classic=False):
   """
   Stochastic Gradient Descent (SGD) optimizer with optional momentum and weight decay.
 
@@ -71,17 +72,13 @@ class LARS(Optimizer):
   - Described: https://paperswithcode.com/method/lars
   - Paper: https://arxiv.org/abs/1708.03888v3
   """
-  def __init__(self, params:List[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, nesterov=False, classic=True, tcoef=0.001):
+  def __init__(self, params:list[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, nesterov=False, classic=True, tcoef=0.001):
     super().__init__(params, lr)
     self.momentum, self.wd, self.nesterov, self.classic, self.tcoef = momentum, weight_decay, nesterov, classic, tcoef
     self.b = [Tensor.zeros(*t.shape, dtype=t.dtype, device=t.device, requires_grad=False) for t in self.params] if self.momentum else []
 
-  def _step(self) -> List[Tensor]:
-    for i, t in enumerate(self.params):
-      assert t.grad is not None
-      # contiguous is needed since the grads can allegedly form a "diamond"
-      # TODO: fix this in lazy.py
-      g = t.grad.contiguous()
+  def schedule_step_with_grads(self, grads:list[Tensor]) -> list[Tensor]:
+    for i, (t, g) in enumerate(zip(self.params, grads)):
       if self.tcoef != 0:
         r1 = t.detach().square().sum().sqrt()
         r2 = g.square().sum().sqrt()
@@ -99,7 +96,7 @@ class LARS(Optimizer):
     return self.b
 
 # LAMB is essentially just the trust ratio part of LARS applied to Adam/W so if we just set the trust ratio to 1.0 its just Adam/W.
-def AdamW(params: List[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8, weight_decay=0.01):
+def AdamW(params: list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8, weight_decay=0.01):
   """
   AdamW optimizer with optional weight decay.
 
@@ -107,7 +104,7 @@ def AdamW(params: List[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8, weight_dec
   - Paper: https://arxiv.org/abs/1711.05101v3
   """
   return LAMB(params, lr, b1, b2, eps, weight_decay, adam=True)
-def Adam(params: List[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8):
+def Adam(params: list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8):
   """
   Adam optimizer.
 
@@ -123,20 +120,19 @@ class LAMB(Optimizer):
   - Described: https://paperswithcode.com/method/lamb
   - Paper: https://arxiv.org/abs/1904.00962
   """
-  def __init__(self, params: List[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, adam=False):
+  def __init__(self, params: list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, adam=False):
     super().__init__(params, lr)
     self.b1, self.b2, self.eps, self.wd, self.adam = b1, b2, eps, weight_decay, adam
     self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False).contiguous() for _ in [b1, b2])
     self.m = [Tensor.zeros(*t.shape, dtype=dtypes.float32, device=t.device, requires_grad=False).contiguous() for t in self.params]
     self.v = [Tensor.zeros(*t.shape, dtype=dtypes.float32, device=t.device, requires_grad=False).contiguous() for t in self.params]
 
-  def _step(self) -> List[Tensor]:
+  def schedule_step_with_grads(self, grads:list[Tensor]) -> list[Tensor]:
     self.b1_t *= self.b1
     self.b2_t *= self.b2
-    for i, t in enumerate(self.params):
-      assert t.grad is not None
-      self.m[i].assign(self.b1 * self.m[i] + (1.0 - self.b1) * t.grad)
-      self.v[i].assign(self.b2 * self.v[i] + (1.0 - self.b2) * (t.grad * t.grad))
+    for i, (t, g) in enumerate(zip(self.params, grads)):
+      self.m[i].assign(self.b1 * self.m[i] + (1.0 - self.b1) * g)
+      self.v[i].assign(self.b2 * self.v[i] + (1.0 - self.b2) * (g * g))
       m_hat = self.m[i] / (1.0 - self.b1_t)
       v_hat = self.v[i] / (1.0 - self.b2_t)
       up = (m_hat / (v_hat.sqrt() + self.eps)) + self.wd * t.detach()

tinygrad/nn/state.py

@@ -1,24 +1,54 @@
-import os, json, pathlib, zipfile, pickle, tarfile, struct, functools
-from typing import Dict, Union, List, Optional, Any, Tuple, Callable
+import json, pathlib, zipfile, pickle, tarfile, struct, functools, io
+from collections import OrderedDict
+from typing import Union, Optional, Any, Callable, BinaryIO, Iterable
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.helpers import prod, argsort, DEBUG, Timing, CI, unwrap, GlobalCounters, tqdm
+from tinygrad.helpers import prod, argsort, DEBUG, Timing, CI, unwrap, GlobalCounters, tqdm, round_up, T
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.multi import MultiLazyBuffer
+
+class TensorIO(io.RawIOBase, BinaryIO):
+  def __init__(self, t: Tensor):
+    if t.ndim != 1 or t.dtype != dtypes.uint8: raise ValueError("Tensor must be 1d and of dtype uint8!")
+    self._position, self._tensor = 0, t
+
+  def readable(self) -> bool: return True
+  def read(self, size: int = -1) -> bytes:
+    if (buf:=super().read(size)) is None: raise ValueError("io.RawIOBase.read returned None") # only happens if readinto returns None (never)
+    return buf
+  def readinto(self, buffer: Any) -> int:
+    data = self._tensor[self._position:self._position+len(buffer)].data()
+    buffer[:len(data)] = data
+    self._position += len(data)
+    return len(data)
+
+  def seekable(self) -> bool: return True
+  def seek(self, offset: int, whence: int = 0) -> int:
+    self._position = min(len(self._tensor), max(0, [offset, self._position+offset, len(self._tensor)+offset][whence]))
+    return self._position
+
+  # required to correctly implement BinaryIO
+  def __enter__(self): return self
+  def write(self, s: Any): raise io.UnsupportedOperation("TensorIO.write not supported")
+  def writelines(self, lines: Iterable[Any]): raise io.UnsupportedOperation("TensorIO.writelines not supported")
 
 safe_dtypes = {"BOOL":dtypes.bool, "I8":dtypes.int8, "U8":dtypes.uint8, "I16":dtypes.int16, "U16":dtypes.uint16, "I32":dtypes.int, "U32":dtypes.uint,
                "I64":dtypes.int64, "U64":dtypes.uint64, "F16":dtypes.float16, "BF16":dtypes.bfloat16, "F32":dtypes.float32, "F64":dtypes.float64}
 inverse_safe_dtypes = {v:k for k,v in safe_dtypes.items()}
 
-def safe_load_metadata(fn:Union[Tensor,str]) -> Tuple[Tensor, int, Any]:
+def accept_filename(func: Callable[[Tensor], T]) -> Callable[[Union[Tensor, str, pathlib.Path]], T]:
+  @functools.wraps(func)
+  def wrapper(fn: Union[Tensor, str, pathlib.Path]) -> T: return func(Tensor(pathlib.Path(fn)) if not isinstance(fn, Tensor) else fn)
+  return wrapper
+
+@accept_filename
+def safe_load_metadata(t:Tensor) -> tuple[Tensor, int, dict[str, Any]]:
   """
   Loads a .safetensor file from disk, returning the data, metadata length, and metadata.
   """
-  t = fn if isinstance(fn, Tensor) else Tensor.empty(os.stat(fn).st_size, dtype=dtypes.uint8, device=f"disk:{fn}")
-  json_len = t[0:8].bitcast(dtypes.int64).item()
-  return t, json_len, json.loads(t[8:8+json_len].data().tobytes())
+  data_start = int.from_bytes(t[0:8].data(), "little") + 8
+  return t, data_start, json.loads(t[8:data_start].data().tobytes())
 
-def safe_load(fn:Union[Tensor,str]) -> Dict[str, Tensor]:
+def safe_load(fn:Union[Tensor, str, pathlib.Path]) -> dict[str, Tensor]:
   """
   Loads a .safetensor file from disk, returning the state_dict.
 
@@ -26,16 +56,12 @@ def safe_load(fn:Union[Tensor,str]) -> Dict[str, Tensor]:
   state_dict = nn.state.safe_load("test.safetensor")
   ```
   """
-  t, json_len, metadata = safe_load_metadata(fn)
-  ret = {}
-  for k,v in metadata.items():
-    if k == "__metadata__": continue
-    dtype = safe_dtypes[v['dtype']]
-    sz = (v['data_offsets'][1]-v['data_offsets'][0])
-    ret[k] = t[8+json_len+v['data_offsets'][0]:8+json_len+v['data_offsets'][0]+sz].bitcast(dtype).reshape(v['shape'])
-  return ret
-
-def safe_save(tensors:Dict[str, Tensor], fn:str, metadata:Optional[Dict[str, Any]]=None):
+  t, data_start, metadata = safe_load_metadata(fn)
+  data = t[data_start:]
+  return { k: data[v['data_offsets'][0]:v['data_offsets'][1]].bitcast(safe_dtypes[v['dtype']]).reshape(v['shape'])
+          for k, v in metadata.items() if k != "__metadata__" }
+
+def safe_save(tensors:dict[str, Tensor], fn:str, metadata:Optional[dict[str, Any]]=None):
   """
   Saves a state_dict to disk in a .safetensor file with optional metadata.
 
@@ -50,7 +76,7 @@ def safe_save(tensors:Dict[str, Tensor], fn:str, metadata:Optional[Dict[str, Any
     headers[k] = {'dtype': inverse_safe_dtypes[v.dtype], 'shape': list(v.shape), 'data_offsets':[offset, offset+v.nbytes()]}
     offset += v.nbytes()
   j = json.dumps(headers, separators=(',', ':'))
-  j += "\x20"*((8-len(j)%8)%8)
+  j += "\x20"*(round_up(len(j),8)-len(j))
   pathlib.Path(fn).unlink(missing_ok=True)
   t = Tensor.empty(8+len(j)+offset, dtype=dtypes.uint8, device=f"disk:{fn}")
   t[0:8].bitcast(dtypes.int64).assign([len(j)])
@@ -59,8 +85,7 @@ def safe_save(tensors:Dict[str, Tensor], fn:str, metadata:Optional[Dict[str, Any
 
 # state dict
 
-from collections import OrderedDict
-def get_state_dict(obj, prefix:str='', tensor_type=Tensor) -> Dict[str, Tensor]:
+def get_state_dict(obj, prefix:str='', tensor_type=Tensor) -> dict[str, Tensor]:
   """
   Returns a state_dict of the object, with optional prefix.
 
@@ -84,7 +109,8 @@ def get_state_dict(obj, prefix:str='', tensor_type=Tensor) -> Dict[str, Tensor]:
   elif isinstance(obj, dict):
     for k,v in obj.items(): state_dict.update(get_state_dict(v, f"{prefix}{str(k)}.", tensor_type))
   return state_dict
-def get_parameters(obj) -> List[Tensor]:
+
+def get_parameters(obj) -> list[Tensor]:
   """
   ```python exec="true" source="above" session="tensor" result="python"
   class Net:
@@ -98,7 +124,7 @@ def get_parameters(obj) -> List[Tensor]:
   """
   return list(get_state_dict(obj).values())
 
-def load_state_dict(model, state_dict:Dict[str, Tensor], strict=True, verbose=True, consume=False) -> None:
+def load_state_dict(model, state_dict:dict[str, Tensor], strict=True, verbose=True, consume=False) -> None:
   """
   Loads a state_dict into a model.
 
@@ -114,7 +140,7 @@ def load_state_dict(model, state_dict:Dict[str, Tensor], strict=True, verbose=Tr
   ```
   """
   start_mem_used = GlobalCounters.mem_used
-  with Timing("loaded weights in ", lambda et_ns: f", {(GlobalCounters.mem_used-start_mem_used)/1e9:.2f} GB loaded at {(GlobalCounters.mem_used-start_mem_used)/et_ns:.2f} GB/s"):  # noqa: E501
+  with Timing("loaded weights in ", lambda et_ns: f", {(B:=(GlobalCounters.mem_used-start_mem_used))/1e9:.2f} GB loaded at {B/et_ns:.2f} GB/s"):
     model_state_dict = get_state_dict(model)
     if DEBUG >= 1 and len(state_dict) > len(model_state_dict):
       print("WARNING: unused weights in state_dict", sorted(list(state_dict.keys() - model_state_dict.keys())))
@@ -123,27 +149,30 @@ def load_state_dict(model, state_dict:Dict[str, Tensor], strict=True, verbose=Tr
       if k not in state_dict and not strict:
         if DEBUG >= 1: print(f"WARNING: not loading {k}")
         continue
-      if isinstance((mlb:=v.lazydata), MultiLazyBuffer):
-        if isinstance(state_dict[k].lazydata, MultiLazyBuffer): v.replace(state_dict[k]).realize()
-        else: v.replace(state_dict[k].shard(mlb.device, mlb.axis)).realize()
+      if v.shape != state_dict[k].shape:
+        raise ValueError(f'Shape mismatch in layer `{k}`: Expected shape {v.shape}, but found {state_dict[k].shape} in state dict.')
+      if isinstance(v.device, tuple):
+        if isinstance(state_dict[k].device, tuple): v.replace(state_dict[k]).realize()
+        else: v.replace(state_dict[k].shard(v.device, v.lazydata.axis)).realize()
       else: v.replace(state_dict[k].to(v.device)).realize()
       if consume: del state_dict[k]
 
-def tar_extract(fn:os.PathLike) -> Dict[str, Tensor]:
+@accept_filename
+def tar_extract(t: Tensor) -> dict[str, Tensor]:
   """
   Extracts files from a tar archive and returns them as dictionary of names (keys) and tensors (values).
 
   ```python
-  tensors = nn.state.tar_extract("archive.tar")
+  tensors = nn.state.tar_extract(Tensor(pathlib.Path("archive.tar")))
   ```
   """
-  t = Tensor(pathlib.Path(fn))
-  with tarfile.open(fn, "r") as tar:
+  with tarfile.open(fileobj=TensorIO(t), mode="r") as tar:
     return {member.name:t[member.offset_data:member.offset_data+member.size] for member in tar if member.type == tarfile.REGTYPE}
 
 # torch support!
 
-def torch_load(fn:str) -> Dict[str, Tensor]:
+@accept_filename
+def torch_load(t:Tensor) -> dict[str, Tensor]:
   """
   Loads a torch .pth file from disk.
 
@@ -151,10 +180,8 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
   state_dict = nn.state.torch_load("test.pth")
   ```
   """
-  t = Tensor.empty(os.stat(fn).st_size, dtype=dtypes.uint8, device=f"disk:{fn}")
-
-  offsets: Dict[Union[str, int], int] = {}
-  lens: Dict[Union[str, int], int] = {}
+  offsets: dict[Union[str, int], int] = {}
+  lens: dict[Union[str, int], int] = {}
   def _rebuild_tensor_v2(storage, storage_offset, size, stride, requires_grad=None, backward_hooks=None, metadata=None):
     #print(storage, storage_offset, size, stride, requires_grad, backward_hooks, metadata)
     lens[storage[2]] = storage[4] * storage[1].itemsize
@@ -168,8 +195,8 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
     if tuple(permute_indexes) != tuple(range(len(permute_indexes))):
       intermediate_shape = tuple([shape_strides[x][0] for x in argsort(permute_indexes)])
       assert tuple([shape_strides[i][1] for i in argsort(permute_indexes)]) == strides_for_shape(intermediate_shape), "nonpermutable strides"
-      if DEBUG >= 3: print(f"WARNING: this torch load is slow. CLANG to permute {intermediate_shape} with {permute_indexes}")
-      assert storage[1] != dtypes.bfloat16, "can't CLANG permute BF16"
+      if DEBUG >= 3: print(f"WARNING: this torch load is slow. to permute {intermediate_shape} with {permute_indexes}")
+      assert storage[1] != dtypes.bfloat16, "can't permute BF16"
       # TODO: find a nice way to support all shapetracker on disktensors
       ret = ret.to(None).reshape(intermediate_shape).permute(permute_indexes)
 
@@ -178,7 +205,7 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
   class Parameter:
     def __setstate__(self, state): self.tensor = state[0]
 
-  deserialized_objects: Dict[str, Any] = {}
+  deserialized_objects: dict[str, Any] = {}
   intercept = {"HalfStorage": dtypes.float16, "FloatStorage": dtypes.float32, "BFloat16Storage": dtypes.bfloat16,
                "IntStorage": dtypes.int32, "BoolStorage": dtypes.bool,
                "LongStorage": dtypes.int64, "_rebuild_tensor_v2": _rebuild_tensor_v2, "FloatTensor": None, "Parameter": Parameter}
@@ -193,8 +220,11 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
       return intercept[name] if module_root == "torch" else super().find_class(module, name)
     def persistent_load(self, pid): return deserialized_objects.get(pid, pid)
 
-  if zipfile.is_zipfile(fn):
-    myzip = zipfile.ZipFile(fn, 'r')
+  fobj = io.BufferedReader(TensorIO(t))
+  def passthrough_reset(v: bool): return fobj.seek(0, 0) or v
+
+  if passthrough_reset(zipfile.is_zipfile(fobj)): # NOTE: passthrough_reset required to support python < 3.14
+    myzip = zipfile.ZipFile(fobj, 'r')
     base_name = myzip.namelist()[0].split('/', 1)[0]
     for n in myzip.namelist():
       if n.startswith(f'{base_name}/data/'):
@@ -202,8 +232,8 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
           offsets[n.split("/")[-1]] = myfile._orig_compress_start # type: ignore
     with myzip.open(f'{base_name}/data.pkl') as myfile:
       return TorchPickle(myfile).load()
-  elif tarfile.is_tarfile(fn):
-    with tarfile.open(fn, "r") as tar:
+  elif passthrough_reset(tarfile.is_tarfile(fobj)): # NOTE: passthrough_reset required to support python < 3.11
+    with tarfile.open(fileobj=fobj, mode="r") as tar:
       storages_offset = tar.getmember('storages').offset_data
       f = unwrap(tar.extractfile('storages'))
       for i in range(TorchPickle(f).load()):  # num_storages
@@ -218,14 +248,13 @@ def torch_load(fn:str) -> Dict[str, Tensor]:
         deserialized_objects[str(key)] = _rebuild_tensor_v2((None, storage_type, storage_id, None, -1), storage_offset, size, stride)
       return {k:v.tensor if isinstance(v, Parameter) else v for k,v in TorchPickle(unwrap(tar.extractfile('pickle'))).load().items()}
   else:
-    with open(fn, "rb") as f:
-      pkl = TorchPickle(f)
-      _, _, _, rwd, _, ids, base_offset = pkl.load(), pkl.load(), pkl.load(), f.tell(), pkl.load(), pkl.load(), f.tell()
-      for i in ids:
-        offsets[i] = base_offset + 8
-        base_offset += 8 + lens[i]
-      f.seek(rwd)
-      return TorchPickle(f).load()
+    pkl = TorchPickle(fobj)
+    _, _, _, rwd, _, ids, base_offset = pkl.load(), pkl.load(), pkl.load(), fobj.tell(), pkl.load(), pkl.load(), fobj.tell()
+    for i in ids:
+      offsets[i] = base_offset + 8
+      base_offset += 8 + lens[i]
+    fobj.seek(rwd)
+    return TorchPickle(fobj).load()
 
 def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
   """
@@ -260,7 +289,8 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
       return d * (xl.bitwise_or(xh).bitcast(dtypes.int8) - 32).flatten(-2) * scales
   raise ValueError(f"GGML type '{ggml_type}' is not supported!")
 
-def gguf_load(tensor: Tensor) -> Tuple[Dict, Dict[str, Tensor]]:
+@accept_filename
+def gguf_load(tensor: Tensor) -> tuple[dict, dict[str, Tensor]]:
   """
   Loads a gguf file from a tensor.
 
@@ -270,31 +300,26 @@ def gguf_load(tensor: Tensor) -> Tuple[Dict, Dict[str, Tensor]]:
   kv_data, state_dict = gguf_load(gguf_tensor)
   ```
   """
-  if tensor.dtype != dtypes.uint8 or len(tensor.shape) != 1: raise ValueError("GGUF tensor must be 1d and of dtype uint8!")
-  pos, read_buffer, rb_start, kv_data, state_dict = 0, memoryview(bytes()), 0, {}, {}
-  def read_bytes(n: int):
-    nonlocal pos, read_buffer, rb_start
-    if rb_start + len(read_buffer) < pos + n: rb_start, read_buffer = pos, tensor[pos:(pos+max(n, 1000_000))].data()
-    return read_buffer[pos-rb_start:(pos:=pos+n)-rb_start]
-  def read_unpack(fmt: str, n: int): return struct.unpack(fmt, read_bytes(n))[0]
-  def read_str(): return str(read_bytes(read_uint64()), "utf-8")
+  reader, kv_data, state_dict = io.BufferedReader(TensorIO(tensor), 1_000_000), {}, {}
+  def read_unpack(fmt: str, n: int): return struct.unpack(fmt, reader.read(n))[0]
+  def read_str(): return str(reader.read(read_uint64()), "utf-8")
   def read_arr():
     reader, n = readers[read_int32()], read_uint64()
     return [ reader() for _ in range(n) ]
 
-  readers: Dict[int, Callable[[], Any]] = { 8: read_str, 9: read_arr, **{ t: functools.partial(read_unpack, "<"+f, nb) for t, f, nb in [ (0,"c",1),
-    (1,"b",1), (2,"H",2), (3,"h",2), (4,"I",4), (5,"i",4), (6,"f",4), (7,"?",1), (10,"Q",8), (11,"q",8), (12,"d",8) ] } }
+  readers: dict[int, Callable[[], Any]] = { 8: read_str, 9: read_arr, **{ t: functools.partial(read_unpack, "<"+f, nb) for t,f,nb in \
+    [ (0,"c",1), (1,"b",1), (2,"H",2), (3,"h",2), (4,"I",4), (5,"i",4), (6,"f",4), (7,"?",1), (10,"Q",8), (11,"q",8), (12,"d",8) ] } }
   read_uint32, read_int32, read_uint64, read_int64 = readers[4], readers[5], readers[10], readers[11]
 
-  magic, version, n_tensors, n_kv = read_bytes(4), read_int32(), read_int64(), read_int64()
+  magic, version, n_tensors, n_kv = reader.read(4), read_int32(), read_int64(), read_int64()
   if magic != b"GGUF" or version not in [2, 3]: raise ValueError("Invalid GGUF format!")
   for _ in range(n_kv):
     k, typ = read_str(), read_int32()
     kv_data[k] = readers[typ]()
 
   t_infos = [ (read_str(), tuple(read_uint64() for _ in range(read_uint32())), read_int32(), read_uint64()) for _ in range(n_tensors) ]
-  alignment = kv_data.get("general.alignment", 32)
-  data_start = pos = pos + (alignment - pos % alignment if pos % alignment != 0 else 0)
+  alignment, pos = kv_data.get("general.alignment", 32), reader.tell()
+  data_start = round_up(pos, alignment)
 
   for name, dims, typ, off in t_infos: state_dict[name] = ggml_data_to_tensor(tensor[data_start + off:], prod(dims), typ).reshape(*reversed(dims))