fix mean of half tensor if sum is greater than hlaf.max (#4327)

sum of half does acc in float32 already, add an arg to not downcast to half and use that in mean

test/test_dtype.py

@@ -621,6 +621,11 @@ class TestAutoCastType(unittest.TestCase):
     t.reshape(2, 1).expand(2, 10001).max().backward()
     np.testing.assert_allclose(t.grad.numpy(), [1, 0])
 
+  @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
+  def test_mean_half_precision(self):
+    t = Tensor([60000, 60000, 60000], dtype=dtypes.half)
+    np.testing.assert_allclose(t.mean().numpy(), 60000)
+
 class TestImplicitFunctionTypeChange(unittest.TestCase):
   def test_functions(self):
     result = []

tinygrad/features/multi.py

@@ -113,15 +113,15 @@ class MultiLazyBuffer:
   def _shape_to_single_shard(self, shape:Tuple[sint, ...], lb:LazyBuffer) -> Tuple[sint, ...]:
     return tuple(lb.shape[self.axis] if a == self.axis else s for a,s in enumerate(shape))
 
-  def r(self, op:ReduceOps, axis:Tuple[int, ...], acc_dt:Optional[DType]=None) -> MultiLazyBuffer:
+  def r(self, op:ReduceOps, axis:Tuple[int, ...], acc_dt:Optional[DType]=None, downcast_half:bool=True) -> MultiLazyBuffer:
     if self.axis is not None and self.axis in axis:
       # all-reduce on sharded axes
       new_shape = tuple(1 if i in axis else s for i,s in enumerate(self.shape))
-      reduced_parts = [x.r(op, axis, acc_dt) if r else x.const(0, shape=new_shape) for x,r in zip(self.lbs, self.real)]
+      reduced_parts = [x.r(op, axis, acc_dt, downcast_half) if r else x.const(0, shape=new_shape) for x,r in zip(self.lbs, self.real)]
       if all(self.real): return MultiLazyBuffer(all_reduce(op, reduced_parts), None)
       return MultiLazyBuffer(reduced_parts, None, self.real)
     # reduce on non sharded axes, piecewise is fine. if axis is None this is also correct
-    return MultiLazyBuffer([x.r(op, axis, acc_dt) for x in self.lbs], self.axis, self.real)
+    return MultiLazyBuffer([x.r(op, axis, acc_dt, downcast_half) for x in self.lbs], self.axis, self.real)
 
   # *** movement ops ***
 

tinygrad/function.py

@@ -146,14 +146,14 @@ class Where(Function):
 # ************* reduce ops *************
 
 class Sum(Function):
-  def forward(self, x:LazyBuffer, axis:Tuple[int, ...], acc_dtype:Optional[DType]=None) -> LazyBuffer:
+  def forward(self, x:LazyBuffer, axis:Tuple[int, ...], acc_dtype:Optional[DType]=None, downcast_half:bool=True) -> LazyBuffer:
     self.input_shape = x.shape
-    return x.r(ReduceOps.SUM, axis, acc_dtype)
+    return x.r(ReduceOps.SUM, axis, acc_dtype, downcast_half)
 
   def backward(self, grad_output:LazyBuffer) -> LazyBuffer: return grad_output.expand(self.input_shape)
 
 class Max(Function):
-  def forward(self, x:LazyBuffer, axis:Tuple[int, ...], acc_dtype:Optional[DType]=None) -> LazyBuffer:
+  def forward(self, x:LazyBuffer, axis:Tuple[int, ...], acc_dtype:Optional[DType]=None, downcast_half:bool=True) -> LazyBuffer:
     self.x, self.ret, self.axis = x, x.r(ReduceOps.MAX, axis), axis
     return self.ret
 

tinygrad/lazy.py

@@ -160,7 +160,7 @@ class LazyBuffer:
     new_shape = tuple(1 if i in axis else s for i,s in enumerate(self.shape))
     return create_lazybuffer(self.device, ShapeTracker.from_shape(new_shape), self.dtype, op, (axis, acc_dt), (self,))
 
-  def r(self, op:ReduceOps, axis:Tuple[int, ...], acc_dt:Optional[DType]=None) -> LazyBuffer:
+  def r(self, op:ReduceOps, axis:Tuple[int, ...], acc_dt:Optional[DType]=None, downcast_half:bool=True) -> LazyBuffer:
     new_shape = tuple(1 if i in axis else s for i,s in enumerate(self.shape))
     # TODO: this logic should move to the scheduler
     if self.size == 0 and 0 not in new_shape: return self.const({ReduceOps.SUM: 0.0, ReduceOps.MAX: -math.inf}[op], new_shape)
@@ -175,7 +175,7 @@ class LazyBuffer:
                least_upper_dtype(self.dtype, dtypes.float)
     if acc_dt is not None and acc_dt != self.dtype:
       # cast back to float16 or bfloat16 to match torch / jax behavior
-      return self.cast(acc_dt).r(op, axis, acc_dt).cast(self.dtype if self.dtype in [dtypes.float16, dtypes.bfloat16] else acc_dt)
+      return self.cast(acc_dt).r(op, axis, acc_dt).cast(self.dtype if downcast_half and self.dtype in [dtypes.float16, dtypes.bfloat16] else acc_dt)
 
     # TODO: can we split symbolic shape if the reduce axis is not symbolic?
     if not getenv("SPLIT_REDUCEOP", 1) or not all_int(self.shape) or (0 in self.shape) or \

tinygrad/tensor.py

@@ -907,21 +907,23 @@ class Tensor:
 
   # ***** reduce ops *****
 
-  def _reduce(self, fxn:Type[Function], axis:Optional[Union[int, Tuple[int, ...]]]=None, keepdim=False, acc_dtype:Optional[DType]=None) -> Tensor:
+  def _reduce(self, fxn:Type[Function], axis:Optional[Union[int, Tuple[int, ...]]]=None,
+              keepdim=False, acc_dtype:Optional[DType]=None, downcast_half:bool=True) -> Tensor:
     axis_: Tuple[int, ...] = tuple(range(len(self.shape))) if axis is None else ((axis,) if isinstance(axis, int) else tuple(axis))
     axis_ = tuple(x if x >= 0 else x+len(self.shape) for x in axis_)
     shape = tuple(s for i,s in enumerate(self.shape) if i not in axis_)
-    ret = fxn.apply(self, axis=axis_, acc_dtype=acc_dtype)
+    ret = fxn.apply(self, axis=axis_, acc_dtype=acc_dtype, downcast_half=downcast_half)
     return ret if keepdim else ret.reshape(shape)
 
-  def sum(self, axis=None, keepdim=False, acc_dtype:Optional[DType]=None): return self._reduce(F.Sum, axis, keepdim, acc_dtype)
+  def sum(self, axis=None, keepdim=False, acc_dtype:Optional[DType]=None, downcast_half:bool=True):
+    return self._reduce(F.Sum, axis, keepdim, acc_dtype, downcast_half)
   def max(self, axis=None, keepdim=False): return self._reduce(F.Max, axis, keepdim)
   def min(self, axis=None, keepdim=False): return -((-self).max(axis=axis, keepdim=keepdim))
 
   def mean(self, axis=None, keepdim=False):
     assert all_int(self.shape), "does not support symbolic shape"
-    out = self.sum(axis=axis, keepdim=keepdim)
+    out = self.sum(axis=axis, keepdim=keepdim, downcast_half=False)
-    return out.div(prod(self.shape) / prod(out.shape)) if 0 not in out.shape else out
+    return out.div(prod(self.shape) / prod(out.shape)).cast(self.dtype) if 0 not in out.shape else out.cast(self.dtype)
   def var(self, axis=None, keepdim=False, correction=1):
     assert all_int(self.shape), "does not support symbolic shape"
     square_sum = ((self - self.mean(axis=axis, keepdim=True)).square()).sum(axis=axis, keepdim=keepdim)