diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
index 7f163a5076..5c8b36bf6b 100644
--- a/.github/workflows/benchmark.yml
+++ b/.github/workflows/benchmark.yml
@@ -53,7 +53,7 @@ jobs:
     - name: Run GPT2 w HALF
       run: JIT=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
     - name: Run GPT2 w HALF/BEAM
-      run: JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
     - name: Train MNIST
       run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=97.3 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - name: Run 10 CIFAR training steps
@@ -129,7 +129,7 @@ jobs:
     - name: Run GPT2 w HALF
       run: CUDA=1 JIT=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
     - name: Run GPT2 w HALF/BEAM
-      run: CUDA=1 JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 JIT_BATCH_SIZE=4 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: CUDA=1 JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 JIT_BATCH_SIZE=4 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
     - name: Train MNIST
       run: time PYTHONPATH=. CUDA=1 TARGET_EVAL_ACC_PCT=97.3 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - name: Run 10 CIFAR training steps
diff --git a/test/external/verify_kernel.py b/test/external/verify_kernel.py
index 756f701bda..de01478ca9 100644
--- a/test/external/verify_kernel.py
+++ b/test/external/verify_kernel.py
@@ -9,7 +9,7 @@ from tinygrad.features.search import time_linearizer
 # Use this with the LOGKERN options to verify that all executed kernels are valid and evaluate to the same ground truth results
 
 # Example for GPT2:
-# 1) Run the model to log all kernels: `PYTHONPATH=. LOGKERN=/tmp/gpt2_kerns.txt JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing`   # noqa: E501
+# 1) Run the model to log all kernels: `PYTHONPATH=. LOGKERN=/tmp/gpt2_kerns.txt JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing`   # noqa: E501
 # 2) Validate the kernel correctness: `PYTHONPATH=. python3 ./test/external/verify_kernel.py --file /tmp/gpt2_kerns.txt`
 
 if __name__ == "__main__":
diff --git a/test/test_const_folding.py b/test/test_const_folding.py
index 1dcdc5ada9..611ef54728 100644
--- a/test/test_const_folding.py
+++ b/test/test_const_folding.py
@@ -106,14 +106,12 @@ class TestMovedConstFolding(unittest.TestCase):
     _check_ast_count(1, Tensor([1.0, 2, 3, 4]) * Tensor.ones(2).pad(((1, 1),)))
 
   def test_cast_padded(self):
-    # NOTE: this is folded due to CAST_BEFORE_VIEW
-    _check_ast_count(0, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
+    _check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
     np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16).numpy(), [0, 1, 1, 1, 1, 0])
-    _check_ast_count(0, Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16))
-    np.testing.assert_equal(Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16).numpy(), [0, 65535, 65535, 65535, 65535, 0])
-    # not folded
     _check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64))
     np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64).numpy(), [0, 1, 1, 1, 1, 0])
+    _check_ast_count(1, Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16))
+    np.testing.assert_equal(Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16).numpy(), [0, 65535, 65535, 65535, 65535, 0])
 
 class TestReduceOpsConstFolding(unittest.TestCase):
   def test_const_sum(self):
diff --git a/tinygrad/lazy.py b/tinygrad/lazy.py
index 4466203348..265751a337 100644
--- a/tinygrad/lazy.py
+++ b/tinygrad/lazy.py
@@ -84,9 +84,6 @@ class LazyBuffer:
     if self.device.startswith("DISK") and not bitcast: raise RuntimeError("attempted to cast disk buffer (bitcast only)")
     if self.is_unrealized_unmasked_const() and not bitcast:
       return create_lazybuffer(self.device, self.st, dtype, LoadOps.CONST, dtypes.as_const(self.base.arg, dtype))
-    # TODO: applying this makes gpt2 slower
-    if getenv("CAST_BEFORE_VIEW", 1) and dtype.itemsize <= self.dtype.itemsize and self != self.base:
-      return self.base.cast(dtype, bitcast)._view(self.st)
     new_shape = self.shape
     if bitcast and self.dtype.itemsize != dtype.itemsize:
       if not self.device.startswith("DISK"): raise RuntimeError("shape changing bitcast only supported on DISK right now")