diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index bbea0ab112..20d18b139a 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -310,8 +310,6 @@ jobs:
       run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
     - name: Test symbolic with Python emulator
       run: PYTHONPATH=. PYTHON=1 python3 test/test_symbolic_ops.py
-    - name: test_linearizer_failures with Python emulator
-      run: PYTHONPATH=. PYTHON=1 python3 -m pytest -rA test/test_linearizer_failures.py::TestLinearizerFailures::test_failure_1
     - name: test_renderer_failures with Python emulator
       run: PYTHONPATH=. PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
 
@@ -523,8 +521,8 @@ jobs:
           REMOTEDEV=GPU IMAGE=2 REMOTE=1 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
       - name: Test Optimization Helpers
         run: PYTHONPATH="." DEBUG=1 python3 extra/optimization/test_helpers.py
-      - name: Test Action Space
-        run: PYTHONPATH="." DEBUG=1 GPU=1 python3 extra/optimization/get_action_space.py
+      #- name: Test Action Space
+      #  run: PYTHONPATH="." DEBUG=1 GPU=1 python3 extra/optimization/get_action_space.py
       - name: Test Beam Search
         run: PYTHONPATH="." GPU=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
       - name: Test MLPerf stuff
@@ -623,7 +621,7 @@ jobs:
       run: |
           WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit \
           --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
-          --ignore=test/test_fuzz_shape_ops.py --ignore=test/test_linearizer_failures.py --durations=20
+          --ignore=test/test_fuzz_shape_ops.py --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -774,8 +772,8 @@ jobs:
       run: PYTHONPATH="." METAL=1 python test/external/external_test_speed_llama.py
     - name: Test Beam Search
       run: PYTHONPATH="." METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Fuzz Test linearizer
-      run: PYTHONPATH="." METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
+    #- name: Fuzz Test linearizer
+    #  run: PYTHONPATH="." METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
     - name: Run TRANSCENDENTAL math
       run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
     - name: Run pytest (amd)
diff --git a/extra/optimization/generate_dataset.sh b/extra/optimization/generate_dataset.sh
index 66473154f7..e4a7fe03d1 100755
--- a/extra/optimization/generate_dataset.sh
+++ b/extra/optimization/generate_dataset.sh
@@ -18,7 +18,7 @@ python3 examples/beautiful_cartpole.py
 python3 examples/mlperf/model_spec.py
 python3 examples/yolov8.py ./test/models/efficientnet/Chicken.jpg
 examples/openpilot/go.sh
-JIT=2 BIG=1 MPS=1 pytest -n=auto test/ --ignore=test/test_fusion_op.py --ignore=test/test_linearizer_failures.py --ignore=test/test_gc.py --ignore=test/test_speed_v_torch.py --ignore=test/test_jit.py
+JIT=2 BIG=1 MPS=1 pytest -n=auto test/ --ignore=test/test_fusion_op.py --ignore=test/test_gc.py --ignore=test/test_speed_v_torch.py --ignore=test/test_jit.py
 JIT=2 BIG=1 MPS=1 python -m pytest test/test_gc.py
 JIT=2 BIG=1 MPS=1 python -m pytest test/test_jit.py
 JIT=2 BIG=1 MPS=1 python -m pytest test/test_speed_v_torch.py
@@ -26,4 +26,4 @@ JIT=2 BIG=1 MPS=1 python -m pytest test/test_speed_v_torch.py
 # extract, sort and uniq
 extra/optimization/extract_dataset.py
 sort -u /tmp/ops > /tmp/sops
-ls -lh /tmp/ops /tmp/sops	
+ls -lh /tmp/ops /tmp/sops
diff --git a/test/test_linearizer.py b/test/test_linearizer.py
index ba50f9dbfe..3eb98fd6e1 100644
--- a/test/test_linearizer.py
+++ b/test/test_linearizer.py
@@ -3,7 +3,6 @@ import numpy as np
 import unittest
 from dataclasses import replace
 
-from test.helpers import ast_const
 from tinygrad.opt.kernel import Opt, OptOps, KernelOptError, Kernel
 from tinygrad.codegen.lowerer import get_grouped_dims
 from tinygrad.uop.ops import UOp, Ops, GroupOp, KernelInfo
@@ -137,420 +136,6 @@ class TestLinearizer(unittest.TestCase):
     x = Tensor.randn(4,).realize()
     helper_linearizer_ast(store.sink(), [x], wanna_output=[x.numpy()+1], opts=[])
 
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(32, dtype=dtypes.float).realize()
-    st_x = x.uop.st
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1.view(st_x.reshape((1, 32)).expand((32, 32))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (1,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1.view(st_x.reshape((32, 1))),))
-    diff = second_x + first_reduce*ast_const(dtypes.float, -1, (32, 1))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (0,)))
-    store = UOp(Ops.STORE, dtypes.void, (g0.view(ShapeTracker.from_shape((1, 1))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)], # grouping
-      [Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 8)],
-      [Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 16)],
-      [Opt(OptOps.GROUPTOP, 0, 32), Opt(OptOps.GROUPTOP, 0, 32)],
-      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2)], # unroll reduce
-      [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)],
-      [Opt(OptOps.UNROLL, 0, 8), Opt(OptOps.UNROLL, 1, 8)] if Device.DEFAULT not in {"NV", "METAL"} else [], # can't do float8,
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)], # grouping + unrolling
-      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      [Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UNROLL, 2, 8), Opt(OptOps.UNROLL, 2, 8)],
-      [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 0, 8)],
-    ]
-    wanna_output = (x.numpy()-x.numpy().sum(-1, keepdims=True)).sum(-1).reshape(1,1)
-    lins = helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-    self._test_no_nested_ranges(lins, [0])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_mid_dim_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 32, 5, dtype=dtypes.float).realize()
-    st_x = x.uop.st
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1.view(st_x.reshape((27, 1, 32, 5)).expand((27, 32, 32, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1.view(st_x.reshape((27, 32, 1, 5))),))
-    diff = second_x + first_reduce*ast_const(dtypes.float, -1, (27, 32, 1, 5))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [
-      # locals
-      [Opt(OptOps.LOCAL, 0, 3)],
-      [Opt(OptOps.LOCAL, 0, 9)],
-      [Opt(OptOps.LOCAL, 0, 27)],
-      # grouping
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      [Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 8)],
-      [Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 16)],
-      [Opt(OptOps.GROUPTOP, 0, 32), Opt(OptOps.GROUPTOP, 0, 32)],
-      # # unroll
-      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2)],
-      [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)],
-      [Opt(OptOps.UNROLL, 0, 8), Opt(OptOps.UNROLL, 1, 8)] if Device.DEFAULT not in {"NV", "METAL"} else [],
-      # # upcasting
-      [Opt(OptOps.UPCAST, 0, 3)],
-      [Opt(OptOps.UPCAST, 0, 9)],
-      # locals with grouping
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      # locals with unroll
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2)],
-      # locals with upcasting
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.UPCAST, 0, 9)],
-      # grouping with unrolling
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-      [Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UNROLL, 2, 8), Opt(OptOps.UNROLL, 2, 8)],
-      # grouping with upcasting
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UPCAST, 0, 3)],
-      # locals with grouping with unroll
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UNROLL, 2, 8), Opt(OptOps.UNROLL, 2, 8)],
-      # locals with grouping with upcasting
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.UPCAST, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      [Opt(OptOps.LOCAL, 0, 9), Opt(OptOps.UPCAST, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      # grouping with unrolling and upcasting
-      [Opt(OptOps.UPCAST, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-      [Opt(OptOps.UPCAST, 0, 3), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UNROLL, 2, 8), Opt(OptOps.UNROLL, 2, 8)],
-      # locals + grouping + unrolling + upcasting
-      [Opt(OptOps.LOCAL, 0, 3), Opt(OptOps.UPCAST, 0, 3), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2),
-        Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-    ]
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    lins = helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-    self._test_no_nested_ranges(lins, [0])
-
-  def test_triple_multireduce(self):
-    Tensor.manual_seed(0)
-    x0 = Tensor.randn(27, 32, 5, dtype=dtypes.float).realize()
-    x1 = Tensor.randn(27, 32, 5, dtype=dtypes.float).realize()
-    x2 = Tensor.randn(27, 32, 5, dtype=dtypes.float).realize()
-    g0, g1, g2, g3 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(4)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1.view(x0.uop.st.reshape((27, 1, 1, 32, 5)).expand((27, 32, 32, 32, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (3,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g2.view(x1.uop.st.reshape((27, 1, 32, 1, 5)).expand((27, 32, 32, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 32, 32, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (2,)))
-    third_x = UOp(Ops.LOAD, dtypes.float, (g3.view(x2.uop.st.reshape((27, 32, 1, 1, 5))),))
-    mul = (third_x*second_reduce)
-    third_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (mul,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 1, 5))), third_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    wanna_output = (x2.numpy()*(x1.numpy()-x0.numpy().sum(axis=1, keepdims=True)).sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,1,5)
-    lins = helper_linearizer_ast(sink, [x0,x1,x2], wanna_output=[wanna_output])
-    self._test_no_nested_ranges(lins, [0])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  @unittest.skip("this is not supported, it worked by luck")
-  def test_double_reduce_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(8, 32, 8, 16, dtype=dtypes.float).realize()
-    st = x.uop.st
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1, st.reshape((8, 1, 32, 8, 1, 16)).expand((8, 32, 32, 8, 16, 16)).to_uop()))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2, 5)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1, st.reshape((8, 32, 1, 8, 16, 1)).to_uop()))
-    neg_first_reduce = first_reduce * ast_const(dtypes.float, -1, (8, 32, 1, 8, 16, 1))
-    squares = (second_x+neg_first_reduce)
-    squares_sum = UOp(Ops.REDUCE_AXIS, dtypes.float, (squares,), (Ops.ADD, (1, 4)))
-    store = UOp(Ops.STORE, src=(g0, ShapeTracker.from_shape((8, 1, 1, 8, 1, 1)).to_uop(), squares_sum,))
-    sink = UOp(Ops.SINK, src=(store,))
-    wanna_output = (x.numpy()-x.numpy().sum(axis=(1,3), keepdims=True)).sum(axis=(1,3)).reshape((8,1,1,8,1,1))
-    opts = [
-      # openCL / GPU=1 is 256 max threads
-      # grouping
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)], # first dim of both reduces
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 3, 2)], # both dims of the second reduce
-      [Opt(OptOps.GROUPTOP, 2, 2), Opt(OptOps.GROUPTOP, 3, 2)], # second dim of both reduces
-      [Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.GROUPTOP, 3, 2)], # both dims of the first reduce
-      # group all reduce dims
-      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.GROUPTOP, 2, 2), Opt(OptOps.GROUPTOP, 3, 2)],
-      # checking how it works with 2 grouped reduces + unrolling
-      [Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.GROUPTOP, 2, 4), Opt(OptOps.GROUPTOP, 3, 4),
-        Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-      # Checking how it works with 2 grouped reduces + locals.
-      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 0, 4),
-       Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.GROUPTOP, 2, 2), Opt(OptOps.GROUPTOP, 3, 2)],
-      # Checking how it works with 2 grouped reduces + locals + unroll.
-      [Opt(OptOps.LOCAL, 0, 2),
-       Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.GROUPTOP, 2, 4), Opt(OptOps.GROUPTOP, 3, 4),
-       Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-      # Checking how it works with 2 grouped reduces + locals + upcast.
-      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.UPCAST, 0, 2),
-       Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.GROUPTOP, 2, 2), Opt(OptOps.GROUPTOP, 3, 2)],
-      # Checking how it works with 2 grouped reduces + locals + upcast + unroll.
-      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.UPCAST, 0, 2),
-       Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.GROUPTOP, 2, 4), Opt(OptOps.GROUPTOP, 3, 4),
-       Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)],
-    ]
-    lins = helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-    self._test_no_nested_ranges(lins, [0, 1])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_partial_opt_multireduce(self):
-    # check how it works with one reduce optimized and one unoptimized
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 15, 5, dtype=dtypes.float).softmax(1).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1.view(x.uop.st.reshape((27, 1, 15, 5)).expand((27, 15, 15, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1.view(x.uop.st.reshape((27, 15, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 15, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [
-      [Opt(OptOps.GROUPTOP, 0, 3)], # grouping
-      [Opt(OptOps.GROUPTOP, 1, 3)],
-      [Opt(OptOps.GROUPTOP, 0, 15)],
-      [Opt(OptOps.GROUPTOP, 1, 15)],
-      [Opt(OptOps.UNROLL, 0, 3)],
-      [Opt(OptOps.UNROLL, 1, 3)],
-    ]
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    lins = helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-    self._test_no_nested_ranges(lins, [0])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_multireduce_with_parallel(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(4, 32, dtype=dtypes.float).realize()
-    x_p = Tensor.randn(4, 32, dtype=dtypes.float).realize()
-    g0, g1, g2 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(3)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1.view(x.uop.st.reshape((4, 1, 32)).expand((4, 32, 32))),))
-    first_x_p = UOp(Ops.LOAD, dtypes.float, (g2.view(x_p.uop.st.reshape((4, 1, 32)).expand((4, 32, 32))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    first_reduce_p = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x_p.alu(Ops.EXP2),), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1.view(x.uop.st.reshape((4, 32, 1))),))
-    diff = (second_x+(first_reduce + first_reduce_p)*ast_const(dtypes.float, -1, (4, 32, 1)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((4, 1, 1))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [
-      # [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)], # grouping
-      # [Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 8)],
-      # [Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 16)],
-      # [Opt(OptOps.GROUPTOP, 0, 32), Opt(OptOps.GROUPTOP, 0, 32)],
-      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2)], # unroll reduce
-      [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)],
-      [Opt(OptOps.UNROLL, 0, 8), Opt(OptOps.UNROLL, 1, 8)] if Device.DEFAULT not in {"NV", "METAL"} else [], # can't do float8,
-      # [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 2, 2), Opt(OptOps.UNROLL, 3, 2)], # grouping + unrolling
-      # [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
-      # [Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UNROLL, 2, 8), Opt(OptOps.UNROLL, 2, 8)],
-      # [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 0, 8)],
-    ]
-    wanna_output = (x.numpy()-(x.numpy().sum(-1, keepdims=True)+np.exp2(x_p.numpy()).sum(-1, keepdims=True))).sum(-1).reshape(4, 1,1)
-    lins = helper_linearizer_ast(sink, [x,x_p], wanna_output=[wanna_output], opts=opts)
-    self._test_no_nested_ranges(lins, [0])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_multiout_multireduce(self):
-    # check how multireduce works with multioutput
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 15, 5, dtype=dtypes.float).realize()
-    g0, g1, g2 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(3)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g2.view(x.uop.st.reshape((27, 1, 15, 5)).expand((27, 15, 15, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g2.view(x.uop.st.reshape((27, 15, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 15, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store0 = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    second_out = second_reduce * ast_const(dtypes.float, 1/15, (27, 1, 1, 5))
-    store1 = UOp(Ops.STORE, src=(g1.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_out))
-    sink = UOp(Ops.SINK, src=(store0, store1))
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output, wanna_output/15])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_multiout_intermediate_multireduce(self):
-    # check how it outputing at different stages of the multireduce works
-    # TODO: Fails because the stores shapes do not match: store1.shape = (27,15,1,5) != store0.shape = (27,1,1,5)
-    #       so the output shapes are different (FAIL!),
-    #       if we change the shape of store1 to be contiguous, it will match store0 but not the value it's storing (FAIL!)
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 15, 5, dtype=dtypes.float).realize()
-    g0, g1, g2 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(3)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g2.view(x.uop.st.reshape((27, 1, 15, 5)).expand((27, 15, 15, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g2.view(x.uop.st.reshape((27, 15, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 15, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store0 = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    store1 = UOp(Ops.STORE, src=(g1.view(ShapeTracker(views=(View(shape=(27,15,1,5), strides=(5,0,1,1), offset=0, mask=None, contiguous=False),))), first_reduce)) # noqa: E501
-    wanna_output0 = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    wanna_output1 = x.numpy().sum(axis=1).reshape(27,1,1,5)
-
-    sink = UOp(Ops.SINK, src=(store0, store1))
-    with self.assertRaises(RuntimeError): # AST is invalid
-      helper_linearizer_ast(sink, [x], wanna_output=[wanna_output0, wanna_output1])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_complete_unroll_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 3, 5, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((27, 1, 3, 5)).expand((27, 3, 3, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((27, 3, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 3, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [[Opt(OptOps.UNROLL, 0, 3), Opt(OptOps.UNROLL, 0, 3)]]
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_upcast_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 3, 5, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((27, 1, 3, 5)).expand((27, 3, 3, 5))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((27, 3, 1, 5))),))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 3, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((27, 1, 1, 5))), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [[Opt(OptOps.UPCAST, 0, 3)]]
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skip("can't group with multiple reduces yet")
-  def test_early_endif(self):
-    # make sure the if block of a grouped reduce can be closed early and the result loaded back in
-    Tensor.manual_seed(0)
-    x = Tensor.randn(27, 12, 5, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g1, x.uop.st.reshape((27, 1, 12, 5)).expand((27, 12, 12, 5)).to_uop()))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g1, x.uop.st.reshape((27, 12, 1, 5)).to_uop()))
-    diff = (second_x+first_reduce*ast_const(dtypes.float, -1, (27, 12, 1, 5)))
-    second_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (diff,), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0, ShapeTracker.from_shape((27, 1, 1, 5)).to_uop(), second_reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    opts = [[Opt(OptOps.GROUPTOP, 0, 3), Opt(OptOps.GROUPTOP, 1, 3)]]
-    wanna_output = (x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(27,1,1,5)
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output], opts=opts)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_mean_std_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(15, 25, 35, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((15, 25, 1, 35)).expand((15, 25, 35, 35))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (3,)))
-    neg_mean = first_reduce * ast_const(dtypes.float, -1/35, (15, 25, 35, 1))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((15, 25, 35, 1))),))
-    squares = (second_x+neg_mean)*(second_x+neg_mean)
-    squares_sum = UOp(Ops.REDUCE_AXIS, dtypes.float, (squares,), (Ops.ADD, (2,)))
-    variance = squares_sum * ast_const(dtypes.float, 1/35, (15, 25, 1, 1))
-    std = variance.alu(Ops.SQRT)
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((15, 25, 1, 1))), std))
-    sink = UOp(Ops.SINK, src=(store,))
-    wanna_output = x.numpy().std(axis=2, ddof=0).reshape((15,25,1,1))
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_mean_std_multireduce_mid_dim(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(15, 25, 35, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((15, 1, 25, 35)).expand((15, 25, 25, 35))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (2,)))
-    neg_mean = first_reduce * ast_const(dtypes.float, -0.04, (15, 25, 1, 35))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((15, 25, 1, 35))),))
-    squares = (second_x+neg_mean)*(second_x+neg_mean)
-    squares_sum = UOp(Ops.REDUCE_AXIS, dtypes.float, (squares,), (Ops.ADD, (1,)))
-    variance = squares_sum * ast_const(dtypes.float, 0.04, (15, 1, 1, 35))
-    std = variance.alu(Ops.SQRT)
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((15, 1, 1, 35))), std))
-    sink = UOp(Ops.SINK, src=(store,))
-    wanna_output = x.numpy().std(axis=1, ddof=0).reshape((15,1,1,35))
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  @unittest.expectedFailure
-  def test_mean_std_multireduce_multiout(self):
-    # TODO: Similar error to test_multiout_intermediate_multireduce (implicit expand vs shape mismatch)
-    Tensor.manual_seed(0)
-    x = Tensor.randn(15, 25, 35, dtype=dtypes.float).realize()
-    g0, g1, g2 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(3)]
-    first_x = UOp(Ops.LOAD, dtypes.float, (g2, x.uop.st.reshape((15, 25, 1, 35)).expand((15, 25, 35, 35)).to_uop()))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (3,)))
-    neg_mean = first_reduce * ast_const(dtypes.float, -1/35, (15, 25, 35, 1))
-    second_x = UOp(Ops.LOAD, dtypes.float, (g2, x.uop.st.reshape((15, 25, 35, 1)).to_uop()))
-    squares = (second_x+neg_mean)*(second_x+neg_mean)
-    squares_sum = UOp(Ops.REDUCE_AXIS, dtypes.float, (squares,), (Ops.ADD, (2,)))
-    variance = squares_sum * ast_const(dtypes.float, 1/35, (15, 25, 1, 1))
-    std = variance.alu(Ops.SQRT)
-    store_mean = UOp(Ops.STORE, src=(g1, ShapeTracker.from_shape((15, 25, 1, 1)).to_uop(), neg_mean))
-    store_std = UOp(Ops.STORE, src=(g0, ShapeTracker.from_shape((15, 25, 1, 1)).to_uop(), std))
-    sink = UOp(Ops.SINK, src=(store_std, store_mean))
-    wanna_output = [x.numpy().std(axis=2, ddof=0).reshape(15,25,1,1), x.numpy().mean(axis=2).reshape(15,25,1,1)]
-
-    lins = helper_linearizer_ast(sink, [x], wanna_output=wanna_output)
-    for k in lins:
-      assert len([u for u in k.uops if u.op is Ops.DEFINE_ACC]) == 2, "got more than two accs (implies the kernel didn't reuse the mean reduce)"
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"PTX", "AMD", "NV"}, "ocelot/remu doesn't have multiple wave syncs yet")
-  def test_var_multireduce(self):
-    Tensor.manual_seed(0)
-    x = Tensor.randn(3, 27, 32, dtype=dtypes.float).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    # push reduce (3, 27, 32) -> (3, 27, 1) -> (3, 27, 32) expand to LOAD
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((3, 27, 1, 32)).expand((3, 27, 32, 32))),))
-    first_reduce = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.ADD, (3,)))
-    neg_mean = first_reduce * ast_const(dtypes.float, -0.03125, (3, 27, 32, 1))
-    # store = UOp(Ops.STORE, src=(g0, ShapeTracker.from_shape((3, 27, 32, 1)).to_uop(), mean))
-    # verify_lazyop(store)
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((3, 27, 32, 1))),))
-    squares = (second_x+neg_mean)*(second_x+neg_mean)
-    squares_sum = UOp(Ops.REDUCE_AXIS, dtypes.float, (squares,), (Ops.ADD, (2,)))
-    variance = squares_sum * ast_const(dtypes.float, 0.03125, (3, 27, 1, 1))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((3, 27, 1, 1))), variance))
-    sink = UOp(Ops.SINK, src=(store,))
-    wanna_output = x.numpy().var(axis=2, ddof=0).reshape((3,27,1,1))
-    helper_linearizer_ast(sink, [x], wanna_output=[wanna_output])
-    # tinygrad ref
-    y_tiny = x.var(axis=2, correction=0).reshape(3,27,1,1)
-    np.testing.assert_allclose(y_tiny.numpy(), wanna_output, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_softmax_multireduce(self):
-    x = Tensor.rand(4, 32).realize()
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    first_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((4, 1, 32,)).expand((4, 32, 32))),))
-    max_x = UOp(Ops.REDUCE_AXIS, dtypes.float, (first_x,), (Ops.MAX, (2,)))
-    second_x = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((4, 32, 1,))),))
-    centered_x = second_x+max_x*ast_const(dtypes.float, -1, (4, 32, 1))
-    exp_x = centered_x.alu(Ops.EXP2)
-    sum_exp_x = UOp(Ops.REDUCE_AXIS, dtypes.float, (exp_x,), (Ops.ADD, (1,)))
-    # y = exp_x * sum_exp_x.alu(Ops.RECIP) # kernels cannot do a return to full shape
-    recip_sum_exp_x = sum_exp_x.alu(Ops.RECIP)
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((4,1,1))), recip_sum_exp_x))
-    sink = UOp(Ops.SINK, src=(store,))
-    expected = 1/np.exp2(x.numpy() - x.numpy().max(axis=-1, keepdims=True)).sum(axis=-1, keepdims=True).reshape(4,1,1)
-    helper_linearizer_ast(sink, [x], wanna_output=[expected])
-
   @unittest.skipIf(CI and Device.DEFAULT in {"PTX", "AMD", "NV"}, "very slow")
   def test_indexing_multireduce(self):
     dataset = Tensor.rand(16384, 256).realize()
@@ -560,271 +145,6 @@ class TestLinearizer(unittest.TestCase):
     real_index = dataset.numpy()[idxs.numpy()].reshape(4, 256, 1, 1)
     helper_linearizer_ast(sink, [dataset, idxs], wanna_output=[real_index])
 
-  # AssertionError: repeated stores in uops
-  def test_argmax_multireduce_axis0(self):
-    t = Tensor.randn(10, 20).realize()
-    t_max = t.max((0,)).realize()
-    real_argmax = np.argmax(t.numpy(), axis=0, keepdims=False).reshape(1, 20, 1)
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(20), arg=ShapeTracker(views=(View(shape=(1, 20, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(-1), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.ADD, dtypes.int, arg=None, src=(
-            UOp(Ops.CONST, dtypes.int, arg=10, src=(
-              x6:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 20, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)), # noqa: E501
-            UOp(Ops.MUL, dtypes.int, arg=None, src=(
-              x8:=UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                 x6,)),
-              UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.MAX, (0,)), src=(
-                UOp(Ops.MUL, dtypes.int, arg=None, src=(
-                  UOp(Ops.CAST, dtypes.int, arg=None, src=(
-                    UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                      UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(200), arg=ShapeTracker(views=(View(shape=(10, 20, 1), strides=(20, 1, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(-1), arg=1, src=()),)),)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(20), arg=ShapeTracker(views=(View(shape=(10, 20, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(-1), arg=2, src=()),)),)),)),
-                      UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                        x21:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(10, 20, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)), # noqa: E501
-                  UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                    UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (2,)), src=(
-                      UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                        UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(11, 19), strides=(0, 0), offset=0, mask=((0, 11), (9, 19)), contiguous=False), View(shape=(10, 20, 10), strides=(1, 0, 20), offset=0, mask=None, contiguous=False))), src=()),)), # noqa: E501
-                        UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                          x28:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(10, 20, 10), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)), # noqa: E501
-                        UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                           x28,)),)),)),
-                    UOp(Ops.CONST, dtypes.int, arg=10, src=(
-                       x21,)),)),)),)),)),)),
-           x8,)),)),))
-    helper_linearizer_ast(ast, [t, t_max], wanna_output=[real_argmax])
-
-  def test_argmax_multireduce_flat(self):
-    t = Tensor.randn(10, 20).realize()
-    t_max = t.max().realize()
-    real_argmax = np.argmax(t.numpy())
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(1), arg=ShapeTracker(views=(View(shape=(1, 1), strides=(0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(-1), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.ADD, dtypes.int, arg=None, src=(
-            UOp(Ops.CONST, dtypes.int, arg=200, src=(
-              x6:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 1), strides=(0, 0), offset=0, mask=None, contiguous=True),)), src=()),)), # noqa: E501
-            UOp(Ops.MUL, dtypes.int, arg=None, src=(
-              x8:=UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                 x6,)),
-              UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.MAX, (0,)), src=(
-                UOp(Ops.MUL, dtypes.int, arg=None, src=(
-                  UOp(Ops.CAST, dtypes.int, arg=None, src=(
-                    UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                      UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(200), arg=ShapeTracker(views=(View(shape=(200, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(-1), arg=1, src=()),)),)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(200, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(-1), arg=2, src=()),)),)),)),
-                      UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                        x21:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(200, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)), # noqa: E501
-                  UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                    UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-                      UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                        UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(201, 399), strides=(0, 0), offset=0, mask=((0, 201), (199, 399)), contiguous=False), View(shape=(200, 200), strides=(1, 400), offset=0, mask=None, contiguous=False))), src=()),)), # noqa: E501
-                        UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                          x28:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(200, 200), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)), # noqa: E501
-                        UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                           x28,)),)),)),
-                    UOp(Ops.CONST, dtypes.int, arg=200, src=(
-                       x21,)),)),)),)),)),)),
-           x8,)),)),))
-    helper_linearizer_ast(ast, [t, t_max], wanna_output=[real_argmax])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_padto_sum_multireduce(self):
-    Tensor.manual_seed(0)
-    N = 17
-    x = Tensor.rand(N, N).realize()
-    opts = [
-      [Opt(OptOps.PADTO, 0, 32)],
-      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
-      # TODO: multireduce pads
-      # causes an issue because the acc won't be masked in the second reduce
-      # [Opt(OptOps.PADTO, 1, 32), Opt(OptOps.PADTO, 2, 32)]
-    ]
-
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    x_ld0 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((1, N, N)).expand((N,N,N))),))
-    x_ld1 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, 1, N))),))
-    r0 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld0,), (Ops.ADD, (1,)))
-    r1 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld1+r0*ast_const(dtypes.float, -1, (N, 1, N)),),(Ops.ADD, (0,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((1,1,N))), r1))
-    sink = UOp(Ops.SINK, src=(store,))
-    helper_linearizer_ast(sink, [x], wanna_output=[(x.numpy()-x.numpy().sum(axis=0, keepdims=True)).sum(axis=0).reshape(1,1,N)], opts=opts)
-
-    x_ld0 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, 1, N)).expand((N,N,N))),))
-    x_ld1 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, N, 1))),))
-    r0 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld0,), (Ops.ADD, (2,)))
-    r1 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld1+r0*ast_const(dtypes.float, -1, (N, N, 1)),), (Ops.ADD, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((N,1,1))), r1))
-    sink = UOp(Ops.SINK, src=(store,))
-    helper_linearizer_ast(sink, [x], wanna_output=[(x.numpy()-x.numpy().sum(axis=1, keepdims=True)).sum(axis=1).reshape(N,1,1)], opts=opts)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_padto_max_multireduce(self):
-    Tensor.manual_seed(0)
-    N = 17
-    x = Tensor.rand(N, N).realize()
-    opts = [
-      [Opt(OptOps.PADTO, 0, 32)],
-      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),]
-    ]
-
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(2)]
-    x_ld0 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((1, N, N)).expand((N,N,N))),))
-    x_ld1 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, 1, N))),))
-    r0 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld0,), (Ops.MAX, (1,)))
-    r1 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld1+r0*ast_const(dtypes.float, -1, (N, 1, N)),), (Ops.MAX, (0,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((1,1,N))), r1))
-    sink = UOp(Ops.SINK, src=(store,))
-    helper_linearizer_ast(sink, [x], wanna_output=[(x.numpy()-x.numpy().max(axis=0, keepdims=True)).max(axis=0).reshape(1,1,N)], opts=opts)
-
-    x_ld0 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, 1, N)).expand((N,N,N))),))
-    x_ld1 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(x.uop.st.reshape((N, N, 1))),))
-    r0 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld0,), (Ops.MAX, (2,)))
-    r1 = UOp(Ops.REDUCE_AXIS, dtypes.float, (x_ld1+r0*ast_const(dtypes.float, -1, (N, N, 1)),), (Ops.MAX, (1,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((N,1,1))), r1))
-    sink = UOp(Ops.SINK, src=(store,))
-    helper_linearizer_ast(sink, [x], wanna_output=[(x.numpy()-x.numpy().max(axis=1, keepdims=True)).max(axis=1).reshape(N,1,1)], opts=opts)
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI doesn't support multiple sync threads yet")
-  def test_padto_where_multireduce(self):
-    # ternary operators try to use both ridxs
-
-    # we need to make sure the ternary operators nest properly
-    N = 17
-    x = Tensor.rand(N, N).realize()
-    a = Tensor.rand(1, 1).realize()
-    b = Tensor.rand(1, 1).realize()
-    opts = [[Opt(OptOps.PADTO, 0, 32)],[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],]
-
-    wanna_output = np.where(0.5*17 < (x.numpy()+np.where(0.75*17 < x.numpy().sum(axis=1,keepdims=True), a.numpy(), b.numpy())).sum(axis=1),0.0,1.0).reshape((N,1,1)) # noqa: E501
-    ld0 = x.uop.st.reshape((N, 1, N)).expand((N,N,N))
-    ld1 = x.uop.st.reshape((N, N, 1))
-    ast = UOp(Ops.SINK, src=(
-      UOp(Ops.STORE, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, 1, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0),)),
-        UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-          UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-            ast_const(dtypes.float, 0.5*N, (N, 1, 1)),
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(), arg=ld1, src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1),)),)),
-                UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                  UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                    ast_const(dtypes.float, 0.75*N, (N, N, 1)),
-                    UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-                      UOp(Ops.LOAD, dtypes.float, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ld0, src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1),)),)),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=2),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=3),)),)),)),)),)),)),
-          ast_const(dtypes.float, 0.0, (N, 1, 1)),
-          ast_const(dtypes.float, 1.0, (N, 1, 1)),)),)),))
-    helper_linearizer_ast(ast, [x,a,b], opts=opts, wanna_output=[wanna_output])
-
-    ld0 = x.uop.st.reshape((1, N, N)).expand((N,N,N))
-    ld1 = x.uop.st.reshape((N, 1, N))
-    wanna_output = np.where(0.5*17 < (x.numpy()+np.where(0.75*17 < x.numpy().sum(axis=0,keepdims=True), a.numpy(), b.numpy())).sum(axis=0),0.0,1.0).reshape(1,1,N) # noqa: E501
-    ast = UOp(Ops.SINK, src=(
-      UOp(Ops.STORE, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(1, 1, N), strides=(0, 0, 1), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0, src=()),)),
-        UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-          UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-            ast_const(dtypes.float, 0.5*N, (1, 1, N)),
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0,)), src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(), arg=ld1, src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1, src=()),)),)),
-                UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                  UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                    ast_const(dtypes.float, 0.75*N, (N, 1, N)),
-                    UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-                      UOp(Ops.LOAD, dtypes.float, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ld0, src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1, src=()),)),)),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, 1, N), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=2, src=()),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, 1, N), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=3, src=()),)),)),)),)),)),)),
-          ast_const(dtypes.float, 0.0, (1, 1, N)),
-          ast_const(dtypes.float, 1.0, (1, 1, N)),)),)),))
-    helper_linearizer_ast(ast, [x,a,b], opts=opts, wanna_output=[wanna_output])
-    # pad reduce axis
-    helper_linearizer_ast(ast, [x,a,b], opts=[[Opt(OptOps.PADTO, 1, 32)],], wanna_output=[wanna_output])
-
-    ld0 = x.uop.st.reshape((1,1,N,N)).expand((N,N,N,N))
-    ld1 = x.uop.st.reshape((N,N,1,1))
-    wanna_output = np.where(0.5*17 < (x.numpy()+np.where(0.75*17 < x.numpy().sum(keepdims=True), a.numpy(), b.numpy())).sum(keepdims=True),0.0,1.0).reshape((1,1,1,1))# noqa: E501
-    ast = UOp(Ops.SINK, src=(
-      UOp(Ops.STORE, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(1, 1, 1, 1), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0, src=()),)),
-        UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-          UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-            ast_const(dtypes.float, 0.5*N, (1, 1, 1, 1)),
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 1)), src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, 1, 1), strides=(N, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1),)),)),
-                UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                  UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                  ast_const(dtypes.float, 0.75*N, (N, N, 1, 1)),
-                    UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2, 3)), src=(
-                      UOp(Ops.LOAD, dtypes.float, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, N, N), strides=(0, 0, N, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=1),)),)),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, 1, 1), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=2),)),)),
-                  UOp(Ops.LOAD, dtypes.float, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(), arg=ShapeTracker(views=(View(shape=(N, N, 1, 1), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=3),)),)),)),)),)),)),
-          ast_const(dtypes.float, 0.0, (1, 1, 1, 1)),
-          ast_const(dtypes.float, 1.0, (1, 1, 1, 1)),)),)),))
-    helper_linearizer_ast(ast, [x,a,b], opts=[[Opt(OptOps.PADTO, 0, 32)],], wanna_output=[wanna_output])
-
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_end_local(self):
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), arg=i) for i in range(2)]
-    load = UOp(Ops.LOAD, dtypes.int, (g1.view(ShapeTracker.from_shape((32,))),))
-    reduce = UOp(Ops.REDUCE_AXIS, dtypes.int, (load,), (Ops.ADD, (0,)))
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker.from_shape((1,))), reduce))
-    sink = UOp(Ops.SINK, src=(store,))
-    load_t = Tensor.full(load.st_arg.shape, 1).contiguous().realize()
-    k = helper_linearizer_ast(sink, [load_t], wanna_output=[load_t.numpy().sum()])[1]
-    self.assertEqual(k.uops[-2].op, Ops.ENDIF)
-    self.assertEqual(k.uops[-1].op, Ops.SINK)
-    self.assertLess(k.uops.index([x for x in k.uops if x.op is Ops.STORE][-1]), k.uops.index(k.uops[-1]))
-
   def test_two_nested_range(self):
     a = Tensor.randn(2, ).realize()
     out = a.reshape(2, 1).expand(2, 3).sum()
@@ -914,24 +234,6 @@ class TestLinearizer(unittest.TestCase):
     assert num_loads <= 4, "more load uops than needed"
     assert num_loads >= 4, "unexpected number of uops, maybe this test needs updating?"
 
-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
-  def test_load_cache_const_bufs(self):
-    # make sure const buffers are differentiated from local and mem buffers
-    ST, DT = ShapeTracker(views=(View(shape=((1,)), strides=(0, 0), offset=0, mask=None, contiguous=False),)).to_uop(), dtypes.int
-    VAL = ast_const(DT, 2, ST.arg.shape)
-    g0, g1 = [UOp(Ops.DEFINE_GLOBAL, DT.ptr(), arg=i) for i in range(2)]
-
-    # data1[0] + VAL
-    a = UOp(Ops.LOAD, DT, (g1.view(ST.arg),)) + VAL
-    # (literal const 1) + VAL
-    b = ast_const(DT, 1, ST.arg.shape) + VAL
-
-    store = UOp(Ops.STORE, src=(g0.view(ST.arg), (a+b)))
-    sink = UOp(Ops.SINK, src=(store,))
-    sink = sink.replace(arg=KernelInfo(opts_to_apply=tuple()))
-    program = get_program(sink, Device[Device.DEFAULT].renderer)
-    assert len(program.uops) <= 10, "too many uops"
-
   def test_upcast_cse(self):
     # when upcasting, within a subtree, there may be common expressions.
 
@@ -1986,29 +1288,6 @@ class TestKernelOpts(unittest.TestCase):
       with self.assertRaises(AssertionError):
         assert k.apply_tensor_cores(use_tensor_cores=1, extra_opts=x), "no valid tensor core" # for METAL in runners
 
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
-  def test_buf_index_not_found_tensor_core(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(256), arg=ShapeTracker(views=(View(shape=(1, 256), strides=(0, 1), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(256), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.int.ptr(256), arg=ShapeTracker(views=(View(shape=(1243, 256), strides=(0, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(256), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.int.ptr(1243), arg=ShapeTracker(views=(View(shape=(1243, 256), strides=(1, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1243), arg=2, src=()),)),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(1243), arg=ShapeTracker(views=(View(shape=(1243, 256), strides=(1, 0), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1243), arg=3, src=()),)),)),)),)),)),))
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    with self.assertRaises(KernelOptError):
-      k.apply_opt(Opt(OptOps.TC, 0, (-1, 1, 1)))
-
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   @unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
                       "test requires tensor cores with accumulation in half") # testing with half suffices.
@@ -2198,23 +1477,6 @@ class TestKernelOpts(unittest.TestCase):
       [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
     ])
 
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  def test_padto_group(self):
-    Tensor.manual_seed(0)
-    g0, g1, g2 = [UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=i) for i in range(3)]
-    ld0 = UOp(Ops.LOAD, dtypes.float, src=(g1.view(ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=False),))),)) # noqa: E501
-    ld1 = UOp(Ops.LOAD, dtypes.float, src=(g2.view(ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),))),)) # noqa: E501
-    store = UOp(Ops.STORE, src=(g0.view(ShapeTracker(views=(View(shape=(1, 1, 1, 1, 1, 4, 1, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=True),))), UOp(Ops.REDUCE_AXIS, dtypes.float, (ld0*ld1,), (Ops.ADD, (0, 2, 4, 6)),))) # noqa: E501
-    sink = UOp(Ops.SINK, src=(store,))
-    data1 = Tensor.randn(2, 1, 4, 1, 3, 4, 2, 6, 1, 3).realize()
-    data2 = Tensor.randn(2, 1, 4, 1, 3, 4, 2, 6, 1, 3).realize()
-    helper_linearizer_ast(sink, [data1, data2], opts=[
-      #[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.GROUP, 0, 4)],
-      #[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8)],
-      #[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8), Opt(OptOps.GROUP, 0, 4)]
-    ])
-
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
   def test_color_shapes_with_local(self):
diff --git a/test/test_linearizer_failures.py b/test/test_linearizer_failures.py
deleted file mode 100644
index a5e5a1a0ef..0000000000
--- a/test/test_linearizer_failures.py
+++ /dev/null
@@ -1,1706 +0,0 @@
-# ruff: noqa: E501
-import unittest, random
-import numpy as np
-from tinygrad.opt.kernel import Kernel, KernelOptError
-from tinygrad.device import is_dtype_supported
-from tinygrad.uop.ops import UOp, Ops
-from tinygrad.opt.search import Opt, OptOps
-from tinygrad import Device, dtypes, Tensor
-from tinygrad.helpers import CI
-from test.external.fuzz_linearizer import compare_linearizer
-
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.shape.view import View
-
-def helper_test_lin(lin: Kernel, opts, failed_platforms, rtol=1e-2, atol=1e-2):
-  if any(b.dtype.base == dtypes.half for b in lin.membufs) and not is_dtype_supported(dtypes.half): return
-  if any(b.dtype.base == dtypes.bfloat16 for b in lin.membufs) and not is_dtype_supported(dtypes.bfloat16): return
-
-  try:
-    lin.apply_opts(opts)
-  except KernelOptError:
-    # it's considered fixed if we invalidated the opts
-    assert Device.DEFAULT not in failed_platforms, f"unexpected success on {Device.DEFAULT}"
-    return
-
-  compare_result = compare_linearizer(lin, rtol=rtol, atol=atol)
-  if compare_result[0] in ["PASS", "KernelOptError"]:
-    # it's considered fixed if we invalidated the opts
-    assert Device.DEFAULT not in failed_platforms, f"unexpected success on {Device.DEFAULT}"
-  else:
-    assert Device.DEFAULT in failed_platforms, f"failed on {Device.DEFAULT} with {compare_result[0]}"
-  return lin
-
-@unittest.skipIf(CI and Device.DEFAULT in {"CUDA", "NV"}, "failed on CUDA CI")
-class TestLinearizerFailures(unittest.TestCase):
-  def setUp(self):
-    random.seed(42)
-    np.random.seed(42)
-    Tensor.manual_seed(42)
-
-  def test_failure_1(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(512), arg=ShapeTracker(views=(View(shape=(32, 16, 1), strides=(16, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(512), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(512), arg=ShapeTracker(views=(View(shape=(32, 16, 16), strides=(16, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  x8:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(512), arg=1, src=()),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(32, 16, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=2, src=()),)),)),)),
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(512), arg=ShapeTracker(views=(View(shape=(32, 16, 1), strides=(16, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-               x8,)),)),)),)),))
-    helper_test_lin(Kernel(ast), [], failed_platforms=[])
-
-  def test_failure_2(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(21312), arg=ShapeTracker(views=(View(shape=(32, 2, 37, 9, 1, 1), strides=(666, 333, 9, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(21312), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.MAX, (4, 5)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(197119), arg=ShapeTracker(views=(View(shape=(32, 2, 111, 27), strides=(6160, 3080, 28, 1), offset=0, mask=((0, 32), (0, 2), (0, 110), (0, 27)), contiguous=False), View(shape=(32, 2, 37, 9, 2, 2), strides=(5994, 2997, 81, 3, 27, 1), offset=0, mask=None, contiguous=False))), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(197119), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=0, arg=32)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_3(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(4096), arg=ShapeTracker(views=(View(shape=(32, 8, 16, 1), strides=(128, 16, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4096), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(65536), arg=ShapeTracker(views=(View(shape=(32, 8, 16, 16), strides=(2048, 256, 16, 1), offset=0, mask=None, contiguous=True),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(65536), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=2), Opt(op=OptOps.UNROLL, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.LOCAL, axis=0, arg=32)]
-    # METAL: AssertionError: Error Domain=AGXMetalG13X Code=3 "Threadgroup memory size (65536) exceeds the maximum threadgroup memory allowed (32768)" UserInfo={NSLocalizedDescription=Threadgroup memory size (65536) exceeds the maximum threadgroup memory allowed (32768)}
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_5(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(1, 1, 1, 1, 1, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 4, 6)), src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            x5:=UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.CONST, dtypes.float, arg=0.1464405059814453, src=(
-                  x8:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 1, 4, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                   x8,)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 1, 4, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=1, src=()),)),)),)),
-             x5,)),)),)),))
-    opts = [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-    # EXEC_ERROR, it has no global_size
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_6(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(10), arg=ShapeTracker(views=(View(shape=(10, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(10), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-              UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(11, 19), strides=(0, 0), offset=0, mask=((0, 11), (9, 19)), contiguous=False), View(shape=(10, 10), strides=(1, 20), offset=0, mask=None, contiguous=False))), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                x9:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(10, 10), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                 x9,)),)),)),
-          UOp(Ops.CONST, dtypes.int, arg=10, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(10, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=2), Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    # COMPILE FAILED, KeyError: Ops.CONST
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_7(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(18939904), arg=ShapeTracker(views=(View(shape=(512, 32, 1, 34, 1, 34), strides=(36992, 1156, 0, 34, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(18939904), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2, 4)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(37748736), arg=ShapeTracker(views=(View(shape=(512, 32, 6, 8, 4, 6, 8, 4), strides=(2048, 64, 6291456, 8, 0, 1048576, 1, 0), offset=0, mask=((0, 512), (0, 32), (0, 6), (0, 8), (0, 1), (0, 6), (0, 8), (0, 1)), contiguous=False), View(shape=(512, 32, 6, 35, 6, 35), strides=(1179648, 36864, 6144, 192, 32, 1), offset=0, mask=((0, 512), (0, 32), (0, 6), (0, 32), (0, 6), (0, 32)), contiguous=False), View(shape=(512, 32, 238, 238), strides=(1411200, 44100, 210, 1), offset=0, mask=((0, 512), (0, 32), (0, 210), (0, 210)), contiguous=False), View(shape=(512, 32, 7, 34, 7, 34), strides=(1812608, 56644, 8092, 238, 34, 1), offset=0, mask=None, contiguous=True))), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(37748736), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=4)]
-    # test/test_linearizer_failures.py Fatal Python error: Segmentation fault
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_8(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(1, 1, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=0, src=()),)),
-        UOp(Ops.SQRT, dtypes.float, arg=None, src=(
-          UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-            UOp(Ops.ADD, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-                  UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                    x9:=UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(4096), arg=ShapeTracker(views=(View(shape=(1, 1, 4096), strides=(0, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4096), arg=1, src=()),)),)),
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(4096), arg=ShapeTracker(views=(View(shape=(1, 1, 4096), strides=(0, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4096), arg=2, src=()),)),)),)),
-                     x9,)),)),
-                UOp(Ops.CONST, dtypes.float, arg=0.000244140625, src=(
-                  x17:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 1, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()),)),)),
-              UOp(Ops.CONST, dtypes.float, arg=1e-06, src=(
-                 x17,)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4)]
-    # fatal error: bracket nesting level exceeded maximum of 256
-    # note: use -fbracket-depth=N to increase maximum nesting level
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_9(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(13500), arg=ShapeTracker(views=(View(shape=(1, 1, 1, 3, 1, 1, 1, 1, 5, 15, 5, 3, 4), strides=(0, 0, 0, 4500, 0, 0, 0, 0, 900, 60, 12, 4, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(13500), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(6), arg=ShapeTracker(views=(View(shape=(1, 2, 1, 3, 1, 1, 1, 1, 5, 15, 5, 3, 4), strides=(0, 3, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(6), arg=1, src=()),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(9000), arg=ShapeTracker(views=(View(shape=(1, 2, 1, 3, 1, 1, 1, 1, 5, 15, 5, 3, 4), strides=(0, 4500, 0, 0, 0, 0, 0, 0, 900, 60, 12, 4, 1), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9000), arg=2, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_10(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(1024), arg=ShapeTracker(views=(View(shape=(1, 1, 1024, 1), strides=(0, 0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1024), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.half, arg=(Ops.ADD, (3,)), src=(
-            UOp(Ops.MUL, dtypes.half, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(50257), arg=ShapeTracker(views=(View(shape=(1, 1, 1024, 50257), strides=(0, 0, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(50257), arg=1, src=()),)),)),
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(51463168), arg=ShapeTracker(views=(View(shape=(1, 1, 1024, 50257), strides=(0, 0, 1, 1024), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(51463168), arg=2, src=()),)),)),)),)),
-          UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.half.ptr(1024), arg=ShapeTracker(views=(View(shape=(1, 1, 1024, 1), strides=(0, 0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1024), arg=3, src=()),)),)),)),)),))
-    helper_test_lin(Kernel(ast), [], failed_platforms=[])
-
-  def test_failure_11(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(1, 64, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=0, src=()),)),
-        UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 3)), src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                  UOp(Ops.MAX, dtypes.float, arg=None, src=(
-                    UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(1179648), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True),)), src=(
-                          x12:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1179648), arg=1, src=()),)),)),
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                          x15:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=2, src=()),)),)),)),
-                    UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-                      x17:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-                  UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      x20:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=3, src=()),)),)),)),
-                UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                   x17,)),)),
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                      UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(512, 64, 3, 3, 2, 2), strides=(0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=()),)),
-                    UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                      UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                UOp(Ops.MAX, dtypes.float, arg=None, src=(
-                                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                      UOp(Ops.VIEW, dtypes.float.ptr(1179648), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                                         x12,)),)),
-                                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                      UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                                         x15,)),)),)),
-                                  x39:=UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-                                    x40:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=()),)),)),
-                                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                  UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                                     x20,)),)),)),
-                              UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                                 x40,)),)),
-                            UOp(Ops.SQRT, dtypes.float, arg=None, src=(
-                              UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                                UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-                                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                        UOp(Ops.VIEW, dtypes.float.ptr(64), arg=ShapeTracker(views=(View(shape=(64,), strides=(1,), offset=0, mask=None, contiguous=True), View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=4, src=()),)),)),
-                                      UOp(Ops.CONST, dtypes.float, arg=5.425347222222222e-05, src=(
-                                        x53:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(64,), strides=(0,), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 3, 2, 2), strides=(2304, 36, 12, 2, 6, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=()),)),)),
-                                    UOp(Ops.CONST, dtypes.float, arg=1e-05, src=(
-                                       x53,)),)),)),)),)),)),
-                           x39,)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(294912), arg=ShapeTracker(views=(View(shape=(512, 64, 3, 3, 2, 2), strides=(576, 9, 3, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(294912), arg=5, src=()),)),)),)),)),)),
-                  UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(294912), arg=ShapeTracker(views=(View(shape=(512, 64, 3, 3, 2, 2), strides=(576, 9, 3, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(294912), arg=6, src=()),)),)),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(294912), arg=ShapeTracker(views=(View(shape=(512, 64, 3, 3, 2, 2), strides=(576, 9, 3, 1, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 3, 2, 3, 2), strides=(2304, 36, 12, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(512, 64, 6, 6), strides=(2304, 36, 6, 1), offset=0, mask=None, contiguous=True))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(294912), arg=7, src=()),)),)),)),)),)),)),)),))
-    helper_test_lin(Kernel(ast), [], failed_platforms=[])
-
-  def test_failure_12(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(72), arg=ShapeTracker(views=(View(shape=(1, 1, 1, 1, 1, 4, 1, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(72), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 4, 6)), src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            x5:=UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(72), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(72), arg=1, src=()),)),)),
-                UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=2, src=()),)),)),)),
-             x5,)),)),)),))
-    opts = [Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.GROUP, axis=0, arg=4)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  # both kernels are correct from a code standpoint, but generate different results due to precision errors (switching to float results in output matches)
-  def test_failure_13(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(768), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1), strides=(384, 0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(768), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.half, arg=(Ops.ADD, (3,)), src=(
-            UOp(Ops.MUL, dtypes.half, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(103728), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 51864), strides=(51864, 0, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(103728), arg=1, src=()),)),)),
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(19915776), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 51864), strides=(0, 0, 1, 384), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(19915776), arg=2, src=()),)),)),)),)),
-          UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.half.ptr(19968), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1), strides=(0, 0, 1, 0), offset=19584, mask=None, contiguous=False),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(19968), arg=3, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=4)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=["METAL", "GPU", "CUDA"])
-
-  def test_failure_14(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(72), arg=ShapeTracker(views=(View(shape=(1, 1, 1, 1, 1, 4, 1, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(72), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 4, 6)), src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            x5:=UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(72), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 18, 0, 3, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(72), arg=1, src=()),)),)),
-                UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 3, 4, 2, 6, 1, 3), strides=(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=2, src=()),)),)),)),
-             x5,)),)),)),))
-    opts = [Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4)]
-    # COMPILE_ERROR on METAL in fuzz_linearizer: unused variables and undeclared variables
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_15(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(21952), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 196, 14, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(21952), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5,)), src=(
-                  UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(94080), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 480, 1, 1), strides=(0, 0, 0, 14, 1, 196, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(94080), arg=1, src=()),)),)),
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(53760), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 480, 1, 1), strides=(0, 0, 480, 0, 0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(53760), arg=2, src=()),)),)),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(112), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(112), arg=3, src=()),)),)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(112), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(112), arg=4, src=()),)),)),)),
-            UOp(Ops.SQRT, dtypes.float, arg=None, src=(
-              UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-                UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.float.ptr(112), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(112), arg=5, src=()),)),)),
-                  UOp(Ops.CONST, dtypes.float, arg=1e-05, src=(
-                    UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)),
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(112), arg=ShapeTracker(views=(View(shape=(1, 1, 112, 14, 14, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(112), arg=6, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=2), Opt(op=OptOps.PADTO, axis=1, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.UPCAST, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=3, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=1, arg=2), Opt(op=OptOps.LOCAL, axis=1, arg=16)]
-    # COMPILE_ERROR on METAL in fuzz_linearizer ast 115: Error Domain=AGXMetalG14X Code=3 "Compiler encountered an internal error"
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_16(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(13), arg=ShapeTracker(views=(View(shape=(1, 13, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(13), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(13312), arg=ShapeTracker(views=(View(shape=(1, 13, 1024), strides=(0, 1024, 1), offset=0, mask=None, contiguous=True),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(13312), arg=1, src=()),)),)),)),
-          UOp(Ops.CONST, dtypes.float, arg=0.0009765625, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 13, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=1, arg=4)]
-    # COMPILE_ERROR on METAL/GPU (probably HIP/CUDA too) in fuzz_linearizer ast 154: bracket nesting level exceeded maximum of 256
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_17(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(62720), arg=ShapeTracker(views=(View(shape=(2, 1, 40, 1, 28, 28, 1, 1), strides=(31360, 0, 784, 0, 28, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(62720), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(9600), arg=ShapeTracker(views=(View(shape=(2, 1, 40, 240, 28, 28, 1, 1), strides=(0, 0, 1, 40, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9600), arg=1, src=()),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(376320), arg=ShapeTracker(views=(View(shape=(2, 1, 40, 240, 28, 28, 1, 1), strides=(188160, 0, 0, 784, 28, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(376320), arg=2, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=1, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=2), Opt(op=OptOps.GROUPTOP, axis=0, arg=16), Opt(op=OptOps.PADTO, axis=1, arg=32), Opt(op=OptOps.LOCAL, axis=1, arg=4)]
-    # COMPILE_ERROR on METAL in fuzz_linearizer ast 178: Error Domain=AGXMetalG14X Code=3 "Compiler encountered an internal error"
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_18(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(768), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1), strides=(384, 0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(768), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(768), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1), strides=(384, 0, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(768), arg=1, src=()),)),)),
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(3072), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1536), strides=(1536, 0, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(3072), arg=2, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(589824), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1536), strides=(0, 0, 1536, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(589824), arg=3, src=()),)),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(384), arg=ShapeTracker(views=(View(shape=(2, 1, 384, 1), strides=(0, 0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(384), arg=4, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUPTOP, axis=0, arg=256), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=3)]
-    # COMPILE_ERROR on METAL in fuzz_linearizer ast 239: Error Domain=AGXMetalG14X Code=3 "Compiler encountered an internal error"
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_19(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(4536), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 1, 9, 7, 3, 3), strides=(2268, 0, 567, 0, 63, 9, 3, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4536), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(144), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 4, 9, 7, 3, 3), strides=(0, 0, 36, 9, 0, 0, -3, -1), offset=8, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(144), arg=1, src=()),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(504), arg=ShapeTracker(views=(View(shape=(2, 1, 4, 4, 9, 7, 3, 3), strides=(252, 0, 0, 63, 7, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(504), arg=2, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=2, arg=3), Opt(op=OptOps.UPCAST, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=7), Opt(op=OptOps.UPCAST, axis=2, arg=3), Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.LOCAL, axis=0, arg=3)]
-    # COMPILE_ERROR on METAL in fuzz_linearizer ast 379: Error Domain=AGXMetalG14X Code=3 "Compiler encountered an internal error"
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_20(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(4, 4), strides=(4, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(4), arg=ShapeTracker(views=(View(shape=(4, 4), strides=(0, 1), offset=0, mask=None, contiguous=False),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4), arg=1, src=()),)),)),
-          UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 4), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_21(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(2925), arg=ShapeTracker(views=(View(shape=(45, 65), strides=(65, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(2925), arg=0, src=()),)),
-        UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(45, 65), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),))
-    opts = [Opt(op=OptOps.PADTO, axis=0, arg=32)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  #@unittest.skipIf(Device.DEFAULT in ("LLVM", "METAL", "CPU"), "flaky")
-  @unittest.skip("flaky everywhere")
-  def test_failure_22(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(1, 96, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          x4:=UOp(Ops.CONST, dtypes.float, arg=0.000244140625, src=(
-            x5:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 96, 1, 1), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 3)), src=(
-                  UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                      UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                  UOp(Ops.VIEW, dtypes.float.ptr(393216), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True),)), src=(
-                                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(393216), arg=1, src=()),)),)),
-                                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                  UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=2, src=()),)),)),)),
-                              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=3, src=()),)),)),)),
-                            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=4, src=()),)),)),)),
-                          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=5, src=()),)),)),)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=6, src=()),)),)),)),
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(32, 96, 8, 16), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=7, src=()),)),)),)),
-                    UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=8, src=()),)),)),
-                      UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=9, src=()),)),)),
-                        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=10, src=()),)),)),
-                          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=11, src=()),)),)),
-                            UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=12, src=()),)),)),
-                              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                  UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=13, src=()),)),)),
-                                UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                  UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                    UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                      UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=14, src=()),)),)),
-                                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                      UOp(Ops.VIEW, dtypes.float.ptr(276461), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 8640, 180, 18, 1), offset=19, mask=((1, 2), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(276461), arg=15, src=()),)),)),
-                                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                                      UOp(Ops.VIEW, dtypes.float.ptr(544301), arg=ShapeTracker(views=(View(shape=(2, 32, 48, 8, 16), strides=(0, 17280, 180, 18, 1), offset=19, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(2, 32, 48, 8, 16), strides=(0, 12288, 128, 16, 1), offset=0, mask=((0, 1), (0, 32), (0, 48), (0, 8), (0, 16)), contiguous=False), View(shape=(1536, 2, 128), strides=(128, 196608, 1), offset=0, mask=None, contiguous=False), View(shape=(32, 96, 8, 16), strides=(12288, 128, 16, 1), offset=0, mask=None, contiguous=True))), src=(
-                                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(544301), arg=16, src=()),)),)),)),)),)),)),)),)),)),)),)),)),
-                UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-                  UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(1, 96, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=17, src=()),)),)),
-                    UOp(Ops.CONST, dtypes.float, arg=2.0, src=(
-                       x5,)),)),)),)),
-              x79:=UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-                UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                  UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(96), arg=ShapeTracker(views=(View(shape=(1, 96, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96), arg=18, src=()),)),)),
-                     x4,)),
-                  UOp(Ops.CONST, dtypes.float, arg=1e-05, src=(
-                     x5,)),)),)),)),
-             x79,)),)),)),))
-    opts = []
-    helper_test_lin(Kernel(ast), opts, failed_platforms=["METAL", "CUDA"])
-
-  def test_failure_23(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(9600), arg=ShapeTracker(views=(View(shape=(240, 40, 1, 1), strides=(40, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9600), arg=0, src=()),)),
-        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.VIEW, dtypes.float.ptr(9600), arg=ShapeTracker(views=(View(shape=(240, 40, 1, 1), strides=(1, 240, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9600), arg=1, src=()),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.LOCAL, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=3, arg=2)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_24(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(256), arg=ShapeTracker(views=(View(shape=(8, 32, 1, 1), strides=(32, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(256), arg=0, src=()),)),
-        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.VIEW, dtypes.float.ptr(256), arg=ShapeTracker(views=(View(shape=(8, 32, 1, 1), strides=(1, 8, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(256), arg=1, src=()),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=2, arg=2), Opt(op=OptOps.LOCAL, axis=1, arg=8), Opt(op=OptOps.UPCAST, axis=2, arg=0), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=2)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  # this is the cause of the GPT2 BEAM instability. bisects to PR#3530 O(n) arange attempt
-  def test_failure_25(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1024), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-              UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1025, 2047), strides=(0, 0), offset=0, mask=((0, 1025), (1023, 2047)), contiguous=False), View(shape=(1024, 1024), strides=(1, 2048), offset=0, mask=None, contiguous=False))), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                x9:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1024), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                 x9,)),)),)),
-          UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=16), Opt(op=OptOps.UNROLL, axis=0, arg=4)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  # COMPARE_ERROR from GPT2 kernel - stems from uops.py self.simplify_phi_loops
-  def test_failure_26(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(128), arg=ShapeTracker(views=(View(shape=(128, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(128), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-              UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(129, 255), strides=(0, 0), offset=0, mask=((0, 129), (127, 255)), contiguous=False), View(shape=(128, 128), strides=(1, 256), offset=0, mask=None, contiguous=False))), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                x9:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(128, 128), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                 x9,)),)),)),
-          UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(128, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    all_failing_opts = [
-      [Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.GROUPTOP, axis=0, arg=32), Opt(op=OptOps.UNROLL, axis=0, arg=0)],
-      [Opt(op=OptOps.GROUPTOP, axis=0, arg=32), Opt(op=OptOps.UNROLL, axis=0, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=4)],
-      [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-      [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-      [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=4)],
-      [Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-      [Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4)],
-      [Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=1, arg=4)],
-      [Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.GROUP, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=1, arg=4)],
-      [Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.GROUP, axis=0, arg=16), Opt(op=OptOps.UNROLL, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-      [Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-    ]
-    for opts in all_failing_opts:
-      helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  # COMPARE_ERROR from GPT2 kernel - just the first element off
-  # testing ast 41
-  # 0 ━┳ STORE MemBuffer(idx=0, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 16, 13, 1), strides=(0, 13, 1, 0), offset=0, mask=None, contiguous=True),)))
-  # 1  ┗━┳ MAX (3,)
-  # 2    ┗━━ LOAD MemBuffer(idx=1, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 16, 13, 13), strides=(0, 169, 13, 1), offset=0, mask=None, contiguous=True),)))
-  # 208   13
-  # ...
-  # Mismatched elements: 1 / 1232 (0.0812%)
-  # Max absolute difference: 0.8687
-  # Max relative difference: 1.
-  #  x: array([0.   , 0.996, 0.829, ..., 0.   , 0.   , 0.   ], dtype=float16)
-  #  y: array([0.8687, 0.996 , 0.829 , ..., 0.    , 0.    , 0.    ], dtype=float16)
-  # COMPARE FAILED!!
-  def test_failure_27(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(208), arg=ShapeTracker(views=(View(shape=(1, 16, 13, 1), strides=(0, 13, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(208), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.half, arg=(Ops.MAX, (3,)), src=(
-          UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.half.ptr(2704), arg=ShapeTracker(views=(View(shape=(1, 16, 13, 13), strides=(0, 169, 13, 1), offset=0, mask=None, contiguous=True),)), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(2704), arg=1, src=()),)),)),)),)),))
-    all_failing_opts = [
-      [Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=7), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
-    ]
-    for opts in all_failing_opts:
-      helper_test_lin(Kernel(ast), opts, failed_platforms=[])
-
-  def test_failure_28(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.bfloat16.ptr(1), arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.bfloat16.ptr(1), arg=0, src=()),)),
-        UOp(Ops.WHERE, dtypes.bfloat16, arg=None, src=(
-          UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-            x5:=UOp(Ops.CAST, dtypes.bfloat16, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.int.ptr(1), arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=1, src=()),)),)),)),
-            x9:=UOp(Ops.CONST, dtypes.bfloat16, arg=230.0, src=(
-              x10:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=()),)),)),
-          UOp(Ops.ADD, dtypes.bfloat16, arg=None, src=(
-            UOp(Ops.MUL, dtypes.bfloat16, arg=None, src=(
-              UOp(Ops.MUL, dtypes.bfloat16, arg=None, src=(
-                 x5,
-                UOp(Ops.CONST, dtypes.bfloat16, arg=0.004347826086956522, src=(
-                   x10,)),)),
-              UOp(Ops.CONST, dtypes.bfloat16, arg=0.199374800625, src=(
-                 x10,)),)),
-            UOp(Ops.CONST, dtypes.bfloat16, arg=1.99375e-07, src=(
-               x10,)),)),
-          UOp(Ops.ADD, dtypes.bfloat16, arg=None, src=(
-            UOp(Ops.MUL, dtypes.bfloat16, arg=None, src=(
-              UOp(Ops.MUL, dtypes.bfloat16, arg=None, src=(
-                UOp(Ops.ADD, dtypes.bfloat16, arg=None, src=(
-                   x5,
-                   x9,)),
-                UOp(Ops.CONST, dtypes.bfloat16, arg=0.0012987012987012987, src=(
-                   x10,)),)),
-              UOp(Ops.CONST, dtypes.bfloat16, arg=-0.19439062499999998, src=(
-                 x10,)),)),
-            UOp(Ops.CONST, dtypes.bfloat16, arg=0.199375, src=(
-               x10,)),)),)),)),))
-    helper_test_lin(Kernel(ast), opts=[], failed_platforms=[])
-
-  def test_failure_29(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(25690112), arg=ShapeTracker(views=(View(shape=(128, 1, 64, 56, 56, 1, 1, 1), strides=(200704, 0, 3136, 56, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(25690112), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (7, 6, 5)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(25690112), arg=ShapeTracker(views=(View(shape=(1, 128, 1, 64, 4, 58, 4, 58), strides=(0, 200704, 0, 3136, 0, 56, 0, 1), offset=-57, mask=((0, 1), (0, 128), (0, 1), (0, 64), (0, 4), (1, 57), (0, 4), (1, 57)), contiguous=False), View(shape=(128, 1, 64, 56, 56, 64, 3, 3), strides=(3444736, 0, 0, 232, 1, 53824, 13688, 59), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(25690112), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(36864), arg=ShapeTracker(views=(View(shape=(128, 1, 64, 56, 56, 64, 3, 3), strides=(0, 0, 576, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(36864), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 1, 1)), Opt(op=OptOps.PADTO, axis=2, arg=32)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=[], atol=1.0)
-
-  def test_failure_30(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(2952192), arg=ShapeTracker(views=(View(shape=(256, 1, 12, 31, 31, 1, 1, 1), strides=(11532, 0, 961, 31, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(2952192), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (7, 6, 5)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(786432), arg=ShapeTracker(views=(View(shape=(256, 1, 12, 31, 31, 3, 2, 2), strides=(3072, 0, 0, 32, 1, 1024, 32, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(786432), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(144), arg=ShapeTracker(views=(View(shape=(256, 1, 12, 31, 31, 3, 2, 2), strides=(0, 0, 12, 0, 0, 4, 2, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(144), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.PADTO, axis=3, arg=32), Opt(op=OptOps.LOCAL, axis=3, arg=32), Opt(op=OptOps.UPCAST, axis=3, arg=4), Opt(op=OptOps.UPCAST, axis=3, arg=0)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  # from METAL=1 fuzz_linearizer command in test.yml
-  def test_failure_31(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(208), arg=ShapeTracker(views=(View(shape=(1, 16, 13, 1), strides=(0, 13, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(208), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-          UOp(Ops.EXP2, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(2704), arg=ShapeTracker(views=(View(shape=(1, 16, 13, 13), strides=(0, 169, 13, 1), offset=0, mask=None, contiguous=True),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(2704), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(208), arg=ShapeTracker(views=(View(shape=(1, 16, 13, 13), strides=(0, 13, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(208), arg=2, src=()),)),)),)),
-              UOp(Ops.CONST, dtypes.float, arg=1.4426950408889634, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 16, 13, 13), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UNROLL, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=1, arg=32)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  @unittest.skipIf(CI, "for real AMD GPU")
-  def test_failure_32(self):
-    # kernel from beaming resnet
-    # Memory access fault on tinybox red
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(12845056), arg=ShapeTracker(views=(View(shape=(256, 1, 256, 14, 14, 1, 1, 1), strides=(50176, 0, 196, 14, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(12845056), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (7, 6, 5)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(12845056), arg=ShapeTracker(views=(View(shape=(1, 256, 1, 256, 4, 16, 4, 16), strides=(0, 50176, 0, 196, 0, 14, 0, 1), offset=-15, mask=((0, 1), (0, 256), (0, 1), (0, 256), (0, 4), (1, 15), (0, 4), (1, 15)), contiguous=False), View(shape=(256, 1, 256, 14, 14, 256, 3, 3), strides=(1048576, 0, 0, 64, 1, 4096, 1088, 17), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(12845056), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(589824), arg=ShapeTracker(views=(View(shape=(256, 1, 256, 14, 14, 256, 3, 3), strides=(0, 0, 2304, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(589824), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=2, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=2, arg=7), Opt(op=OptOps.UNROLL, axis=1, arg=0), Opt(op=OptOps.LOCAL, axis=1, arg=16)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[], atol=0.1, rtol=0.05)
-
-  def test_failure_33(self):
-    # Ops.UNMUL left after linearize
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            x5:=UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(26040), arg=ShapeTracker(views=(View(shape=(32640,), strides=(1,), offset=0, mask=((0, 26040),), contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(26040), arg=1, src=()),)),)),
-            UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-              UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                 x5,
-                x10:=UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-                  x11:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(32640,), strides=(0,), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-              UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                      UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                        UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                          x18:=UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(32640,), strides=(0,), offset=0, mask=((0, 26040),), contiguous=False),)), src=()),)),
-                          UOp(Ops.CONST, dtypes.float, arg=0.06788442333021306, src=(
-                             x11,)),
-                           x10,)),
-                         x5,)),
-                      UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                         x18,
-                        UOp(Ops.CONST, dtypes.float, arg=-0.03394221166510653, src=(
-                           x11,)),
-                         x10,)),)),
-                    UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(6600), arg=ShapeTracker(views=(View(shape=(32640,), strides=(1,), offset=-26040, mask=((26040, 32640),), contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(6600), arg=2, src=()),)),)),
-                      UOp(Ops.WHERE, dtypes.float, arg=None, src=(
-                        UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(32640,), strides=(0,), offset=0, mask=((26040, 32640),), contiguous=False),)), src=()),)),
-                        UOp(Ops.CONST, dtypes.float, arg=-0.18257418583505536, src=(
-                           x11,)),
-                         x10,)),)),)),
-                   x10,)),
-                UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                   x11,)),
-                UOp(Ops.CONST, dtypes.float, arg=1.0, src=(
-                   x11,)),)),
-               x10,)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUPTOP, axis=0, arg=16)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  # from fuzzing on metal
-  def test_failure_34(self, unroll=False):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(720), arg=ShapeTracker(views=(View(shape=(4, 1, 6, 10, 3, 1, 1, 1), strides=(180, 0, 30, 3, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(720), arg=0, src=()),)),
-        UOp(Ops.MAX, dtypes.float, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (6, 7)), src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(308), arg=ShapeTracker(views=(View(shape=(4, 1, 6, 10, 3, 1, 2, 5), strides=(77, 0, 0, 7, 1, 0, 7, 1), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(308), arg=1, src=()),)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(60), arg=ShapeTracker(views=(View(shape=(4, 1, 6, 10, 3, 1, 2, 5), strides=(0, 0, 10, 0, 0, 0, 5, 1), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(60), arg=2, src=()),)),)),)),)),
-          UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1, 6, 10, 3, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1)), Opt(op=OptOps.UNROLL, axis=0, arg=0)] if unroll else [Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  def test_failure_35(self): self.test_failure_34(True)
-
-  # from world fuzz_linearizer: PYTHONPATH=. METAL=1 FUZZ_ALL_ACTIONS=1 DEPTH=1 FUZZ_N=100 FUZZ_NTH=84 python3 ./test/external/fuzz_linearizer.py
-  def test_failure_36(self):
-    # Ops.UNMUL left after linearize
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.uchar.ptr(5), arg=ShapeTracker(views=(View(shape=(5, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(5), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.uchar, arg=None, src=(
-          UOp(Ops.ADD, dtypes.uint, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.uint, arg=(Ops.ADD, (1,)), src=(
-              UOp(Ops.CAST, dtypes.uint, arg=None, src=(
-                UOp(Ops.WHERE, dtypes.uchar, arg=None, src=(
-                  UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(6, 9), strides=(0, 0), offset=0, mask=((0, 6), (4, 9)), contiguous=False), View(shape=(5, 5), strides=(1, 10), offset=0, mask=None, contiguous=False))), src=()),)),
-                  UOp(Ops.CONST, dtypes.uchar, arg=1, src=(
-                    x11:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(5, 5), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                  UOp(Ops.CONST, dtypes.uchar, arg=0, src=(
-                     x11,)),)),)),)),
-            UOp(Ops.CONST, dtypes.uint, arg=-1, src=(
-              UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(5, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  # BEGIN METAL=1 ./examples/beautiful_mnist.py failures
-  # log : PYTHONPATH=. LOGKERNS=/tmp/beautiful_mnist.kernels.txt METAL=1 python3 ./examples/beautiful_mnist.py
-  def test_failure_37(self):
-    # beautiful mnist kernel number 28: 6 possible TC axis_choices (3 for axis_buf1 and 2 reduce) and all fail
-    # fuzz: PYTHONPATH=. METAL=1 FUZZ_ALL_ACTIONS=1 DEPTH=1 FUZZ_NTH=28 DEBUG=2 python3 ./test/external/fuzz_linearizer.py --logfile /tmp/beautiful_mnist.kernels.txt
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(9437184), arg=ShapeTracker(views=(View(shape=(512, 1, 32, 24, 24, 1, 1, 1), strides=(18432, 0, 576, 24, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9437184), arg=0, src=()),)),
-        UOp(Ops.MAX, dtypes.float, arg=None, src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (6, 7)), src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.uchar, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.uchar.ptr(401408), arg=ShapeTracker(views=(View(shape=(512, 1, 32, 24, 24, 1, 5, 5), strides=(784, 0, 0, 28, 1, 0, 28, 1), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(401408), arg=1, src=()),)),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(800), arg=ShapeTracker(views=(View(shape=(512, 1, 32, 24, 24, 1, 5, 5), strides=(0, 0, 25, 0, 0, 0, 5, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(800), arg=2, src=()),)),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(32), arg=ShapeTracker(views=(View(shape=(512, 1, 32, 24, 24, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(32), arg=3, src=()),)),)),)),
-          UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(512, 1, 32, 24, 24, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    for axis in [0,1,2,3,4,5]:
-      opts = [Opt(op=OptOps.TC, axis=axis, arg=(-1, 2, 1))]
-      helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  def test_failure_38(self):
-    # beautiful mnist kernel number 87: 6 possible TC axis_choices (2 for axis_buf1 and 3 reduce) and first/second reduce axis fail for both axis_buf1 choices
-    # fuzz: PYTHONPATH=. METAL=1 FUZZ_ALL_ACTIONS=1 DEPTH=1 FUZZ_NTH=87 DEBUG=2 python3 ./test/external/fuzz_linearizer.py --logfile /tmp/beautiful_mnist.kernels.txt
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(204800), arg=ShapeTracker(views=(View(shape=(1, 1, 32, 1, 1, 1, 5, 5, 256), strides=(0, 0, 6400, 0, 0, 0, 1280, 256, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(204800), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 3, 4)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.uchar, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.uchar.ptr(401408), arg=ShapeTracker(views=(View(shape=(2, 1, 32, 24, 24, 1, 5, 5, 256), strides=(784, 0, 0, 28, 1, 0, 28, 1, 1568), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(401408), arg=1, src=()),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(9437184), arg=ShapeTracker(views=(View(shape=(2, 1, 32, 24, 24, 1, 5, 5, 256), strides=(18432, 0, 576, 24, 1, 0, 0, 0, 36864), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9437184), arg=2, src=()),)),)),)),)),)),))
-    for axis in [0,1,3,4]:
-      opts = [Opt(op=OptOps.TC, axis=axis, arg=(-1, 2, 1))]
-      helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  @unittest.skip("very slow, similar to test_failure_37")
-  def test_failure_39(self):
-    # beautiful mnist kernel number 127: 6 possible TC axis_choices (3 for axis_buf1 and 2 reduce) and all fail
-    # fuzz: PYTHONPATH=. METAL=1 FUZZ_ALL_ACTIONS=1 DEPTH=1 FUZZ_NTH=127 DEBUG=2 python3 ./test/external/fuzz_linearizer.py --logfile /tmp/beautiful_mnist.kernels.txt
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(184320000), arg=ShapeTracker(views=(View(shape=(10000, 1, 32, 24, 24, 1, 1, 1), strides=(18432, 0, 576, 24, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(184320000), arg=0, src=()),)),
-        UOp(Ops.MAX, dtypes.float, arg=None, src=(
-          UOp(Ops.ADD, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (6, 7)), src=(
-              UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.uchar, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.uchar.ptr(7840000), arg=ShapeTracker(views=(View(shape=(10000, 1, 32, 24, 24, 1, 5, 5), strides=(784, 0, 0, 28, 1, 0, 28, 1), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(7840000), arg=1, src=()),)),)),)),
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(800), arg=ShapeTracker(views=(View(shape=(10000, 1, 32, 24, 24, 1, 5, 5), strides=(0, 0, 25, 0, 0, 0, 5, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(800), arg=2, src=()),)),)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(32), arg=ShapeTracker(views=(View(shape=(10000, 1, 32, 24, 24, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(32), arg=3, src=()),)),)),)),
-          UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(10000, 1, 32, 24, 24, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    for axis in [0,1,2,3,4,5]:
-      opts = [Opt(op=OptOps.TC, axis=axis, arg=(-1, 2, 1))]
-      helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  def test_failure_40(self):
-    # beautiful mnist kernel number 3:
-    # fuzz: PYTHONPATH=. METAL=1 FUZZ_ALL_ACTIONS=1 DEPTH=2 DEBUG=2 FUZZ_NTH=3 python3 ./test/external/fuzz_linearizer.py --logfile /tmp/beautiful_mnist.kernels.txt
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(60000), arg=ShapeTracker(views=(View(shape=(60000, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(60000), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-              UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60001, 119999), strides=(0, 0), offset=0, mask=((0, 60001), (59999, 119999)), contiguous=False), View(shape=(60000, 60000), strides=(1, 120000), offset=0, mask=None, contiguous=False))), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                x9:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60000, 60000), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                 x9,)),)),)),
-          UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60000, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    for amt in [16,32]:
-      opts = [Opt(op=OptOps.GROUPTOP, axis=0, arg=amt), Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-      helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-  # END METAL=1 ./examples/beautiful_mnist.py failures
-
-  @unittest.skipIf(CI, "for real AMD GPU")
-  def test_failure_41(self):
-    # One more resnet crash with a page fault on AMD. Checked on rocm6.1.3, -O1 works, -O2 fails
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(25690112), arg=ShapeTracker(views=(View(shape=(256, 1, 128, 28, 28, 1, 1, 1), strides=(100352, 0, 784, 28, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(25690112), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(102760448), arg=ShapeTracker(views=(View(shape=(1, 256, 1, 128, 4, 58, 4, 58), strides=(0, 401408, 0, 3136, 0, 56, 0, 1), offset=-57, mask=((0, 1), (0, 256), (0, 1), (0, 128), (0, 4), (1, 57), (0, 4), (1, 57)), contiguous=False), View(shape=(256, 1, 128, 28, 28, 128, 3, 3), strides=(6889472, 0, 0, 464, 2, 53824, 13688, 59), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(102760448), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(147456), arg=ShapeTracker(views=(View(shape=(256, 1, 128, 28, 28, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(147456), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts=[Opt(op=OptOps.TC, axis=5, arg=(-1, 2, 1)), Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=["AMD", "HIP"], atol=0.02)
-
-  # llama3 8B failure with BEAM=2 https://github.com/tinygrad/tinygrad/actions/runs/10150118124/job/28066519425#step:14:1, these don't compile
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test needs local")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test needs shared")
-  def test_failure_42(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(25, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(26, 49), strides=(0, -1), offset=48, mask=((0, 26), (24, 49)), contiguous=False), View(shape=(25, 25), strides=(1, 50), offset=0, mask=None, contiguous=False))), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.UPCAST, axis=0, arg=2), Opt(op=OptOps.PADTO, axis=0, arg=32)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test needs local")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test needs shared")
-  def test_failure_43(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(25, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(26, 49), strides=(0, -1), offset=48, mask=((0, 26), (24, 49)), contiguous=False), View(shape=(25, 25), strides=(1, 50), offset=0, mask=None, contiguous=False))), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test needs local")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test needs shared")
-  def test_failure_44(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(25, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-            UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(26, 49), strides=(0, -1), offset=48, mask=((0, 26), (24, 49)), contiguous=False), View(shape=(25, 25), strides=(1, 50), offset=0, mask=None, contiguous=False))), src=(
-              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=1, src=()),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUP, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4)]
-    k = helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-    assert k is not None
-    ifs = [u for u in k.uops if u.op is Ops.IF]
-    self.assertEqual(len(ifs), 3)
-    #for st in k.uops.sink.src: self.assertEqual(len(st.src), 4)
-    self.assertLessEqual(len(ifs[0].src[0].toposort()), 17)
-
-  def test_failure_45(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(6), arg=ShapeTracker(views=(View(shape=(2, 3, 1, 1, 1), strides=(3, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(6), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2, 3)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(6), arg=ShapeTracker(views=(View(shape=(2, 3, 2, 3, 1), strides=(0, 0, 3, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(6), arg=1, src=()),)),)),
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.bool, arg=None, src=(
-                UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                  UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.int.ptr(1), arg=ShapeTracker(views=(View(shape=(2, 3, 2, 3, 1), strides=(0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=2, src=()),)),)),
-                    UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                      UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (4,)), src=(
-                        UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                          UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(3, 3), strides=(0, 0), offset=0, mask=((0, 3), (1, 3)), contiguous=False), View(shape=(2, 3, 2, 3, 3), strides=(0, 0, 1, 0, 4), offset=0, mask=((0, 2), (0, 3), (0, 2), (0, 3), (0, 2)), contiguous=False))), src=()),)),
-                          x20:=UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                            x21:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 3, 2, 3, 3), strides=(0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                          x22:=UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                             x21,)),)),)),
-                      x23:=UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                        x24:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 3, 2, 3, 1), strides=(0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),
-                  x25:=UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                     x24,)),)),
-                UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                  UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.int.ptr(6), arg=ShapeTracker(views=(View(shape=(2, 3, 2, 3, 1), strides=(3, 1, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(6), arg=3, src=()),)),)),
-                    UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                      UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (4,)), src=(
-                        UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                          UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 5), strides=(0, 0), offset=0, mask=((0, 4), (2, 5)), contiguous=False), View(shape=(2, 3, 2, 3, 3), strides=(0, 0, 0, 1, 6), offset=0, mask=None, contiguous=False))), src=()),)),
-                           x20,
-                           x22,)),)),
-                       x23,)),)),
-                   x25,)),)),)),)),)),)),))
-    # ValueError: size mismatched, can't reshape self.shape=(6, 2, 3, 3) -> new_shape=(6, 2, 3, 1, 2)
-    opts = [Opt(op=OptOps.UNROLL, axis=2, arg=0)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  def test_failure_46(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(512), arg=ShapeTracker(views=(View(shape=(512, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(512), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                UOp(Ops.MUL, dtypes.bool, arg=None, src=(
-                  UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                    UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.int.ptr(10), arg=ShapeTracker(views=(View(shape=(512, 10), strides=(0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(10), arg=1, src=()),)),)),
-                      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.int.ptr(512), arg=ShapeTracker(views=(View(shape=(512, 10), strides=(1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(512), arg=2, src=()),)),)),)),
-                    UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                      UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(512, 10), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-                  UOp(Ops.LOAD, dtypes.bool, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.bool.ptr(512), arg=ShapeTracker(views=(View(shape=(512, 10), strides=(1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(512), arg=3, src=()),)),)),)),)),
-              UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.float.ptr(1), arg=ShapeTracker(views=(View(shape=(512, 10), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=4, src=()),)),)),)),)),
-          UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(512), arg=ShapeTracker(views=(View(shape=(512, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(512), arg=5, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=2)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  def test_failure_47(self):
-    # upcast an arange, failed with UOP_IS_SYMBOLIC=1 (fixed!)
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.int.ptr(60000), arg=ShapeTracker(views=(View(shape=(60000, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(60000), arg=0, src=()),)),
-        UOp(Ops.ADD, dtypes.int, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (1,)), src=(
-            UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-              UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60001, 119999), strides=(0, 0), offset=0, mask=((0, 60001), (59999, 119999)), contiguous=False), View(shape=(60000, 60000), strides=(1, 120000), offset=0, mask=None, contiguous=False))), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                x9:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60000, 60000), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-              UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                 x9,)),)),)),
-          UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(60000, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=3)]
-    helper_test_lin(Kernel(ast), opts=opts, failed_platforms=[])
-
-  @unittest.skipUnless(not CI and Device.DEFAULT in ("NV", "CUDA"), "for real NV")
-  def test_failure_48(self):
-    # with UOP_IS_SYMBOLIC=1, generates the wrong IDIV (fixed!)
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(4194304), arg=ShapeTracker(views=(View(shape=(1, 1, 64, 1, 1, 256, 1, 1, 256), strides=(0, 0, 65536, 0, 0, 256, 0, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(4194304), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3, 4)), src=(
-          UOp(Ops.CAST, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.half, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(205520896), arg=ShapeTracker(views=(View(shape=(1, 1, 64, 56, 56, 256, 1, 1, 256), strides=(0, 0, 0, 56, 1, 3136, 0, 0, 802816), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(205520896), arg=1, src=()),)),)),
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(51380224), arg=ShapeTracker(views=(View(shape=(1, 1, 64, 56, 56, 256, 1, 1, 256), strides=(0, 0, 3136, 56, 1, 0, 0, 0, 200704), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(51380224), arg=2, src=()),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 0, 1)), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=2)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_49(self):
-    # with UOP_IS_SYMBOLIC=1, on METAL it breaks store fusion and has A+B and B+A being two different UOp
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(60), arg=ShapeTracker(views=(View(shape=(10, 6, 1), strides=(6, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(60), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(100), arg=ShapeTracker(views=(View(shape=(10, 6, 10), strides=(10, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(100), arg=1, src=()),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(60), arg=ShapeTracker(views=(View(shape=(10, 6, 10), strides=(0, 1, 6), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(60), arg=2, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=0, arg=2)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_50(self):
-    # from BEAM_COMPARE=2 running tinyphysics.onnx model
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.bool.ptr(400), arg=ShapeTracker(views=(View(shape=(1, 1, 20, 1, 20), strides=(0, 0, 20, 0, 1), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(400), arg=0, src=()),)),
-        UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.bool, arg=(Ops.ADD, (3,)), src=(
-            UOp(Ops.MUL, dtypes.bool, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.bool, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.bool.ptr(400), arg=ShapeTracker(views=(View(shape=(1, 1, 20, 20, 20), strides=(0, 0, 0, 20, 1), offset=0, mask=None, contiguous=False),)), src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(400), arg=1, src=()),)),)),
-              UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.int.ptr(20), arg=ShapeTracker(views=(View(shape=(1, 1, 20, 20, 20), strides=(0, 0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(20), arg=2, src=()),)),)),
-                  UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.int.ptr(20), arg=ShapeTracker(views=(View(shape=(1, 1, 20, 20, 20), strides=(0, 0, 0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(20), arg=3, src=()),)),)),)),
-                UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 1, 20, 20, 20), strides=(0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),
-          UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-            UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 1, 20, 1, 20), strides=(0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=2)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_51(self):
-    # regression test for #7019, training bert on tinybox red
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(12288), arg=ShapeTracker(views=(View(shape=(12, 1024, 1), strides=(1024, 1, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(12288), arg=0, src=()),)),
-        UOp(Ops.RECIP, dtypes.half, arg=None, src=(
-          UOp(Ops.ADD, dtypes.half, arg=None, src=(
-            UOp(Ops.CONST, dtypes.half, arg=1.0, src=(
-              x6:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(12, 1024, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-            UOp(Ops.EXP2, dtypes.half, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                  UOp(Ops.CONST, dtypes.half, arg=2.0, src=(
-                     x6,)),
-                  UOp(Ops.ADD, dtypes.half, arg=None, src=(
-                    UOp(Ops.CAST, dtypes.half, arg=None, src=(
-                      UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-                        UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                          UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                            UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.half.ptr(5768192), arg=ShapeTracker(views=(View(shape=(12, 1024, 1024), strides=(524288, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(5768192), arg=1, src=()),)),)),
-                            UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.half.ptr(1048576), arg=ShapeTracker(views=(View(shape=(12, 1024, 1024), strides=(0, 1024, 1), offset=0, mask=None, contiguous=False),)), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1048576), arg=2, src=()),)),)),)),)),)),)),
-                    UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.half.ptr(1024), arg=ShapeTracker(views=(View(shape=(12, 1024, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1024), arg=3, src=()),)),)),)),)),
-                UOp(Ops.CONST, dtypes.half, arg=-1.4426950408889634, src=(
-                   x6,)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  @unittest.skip("allocating over 200MB buffer")
-  @unittest.skipIf(CI and Device.DEFAULT in {"METAL"}, "hangs metal gpu CI")
-  def test_failure_52(self):
-    # resnet beam.
-    # NV also fails with a pf.
-    # CUDA Error 700, an illegal memory access was encountered
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(205520896), arg=ShapeTracker(views=(View(shape=(256, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(205520896), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(38535168), arg=ShapeTracker(views=(View(shape=(1, 256, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, 256), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), View(shape=(256, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(38535168), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(9408), arg=ShapeTracker(views=(View(shape=(256, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(9408), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=16)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_53(self):
-    # COMPILE_ERROR, val scope issue
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.uchar.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 1, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(1024), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.uchar, arg=(Ops.ADD, (1,)), src=(
-          UOp(Ops.MUL, dtypes.uchar, arg=None, src=(
-            UOp(Ops.LOAD, dtypes.uchar, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.uchar.ptr(50000), arg=ShapeTracker(views=(View(shape=(1024, 50000, 1), strides=(0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(50000), arg=1, src=()),)),)),
-            UOp(Ops.CAST, dtypes.uchar, arg=None, src=(
-              UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.int.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 50000, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1024), arg=2, src=()),)),)),
-                  UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                    UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (2,)), src=(
-                      UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                        UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(50001, 99999), strides=(0, 0), offset=0, mask=((0, 50001), (49999, 99999)), contiguous=False), View(shape=(1024, 50000, 50000), strides=(0, 1, 100000), offset=0, mask=None, contiguous=False))), src=()),)),
-                        UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                          x20:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 50000, 50000), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                        UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                           x20,)),)),)),
-                    UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                      x23:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 50000, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),
-                UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                   x23,)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.GROUPTOP, axis=1, arg=16)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=["AMD", "GPU", "METAL", "NV", "CUDA"])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"METAL"}, "hangs metal gpu CI")
-  def test_failure_54(self):
-    # resnet beam
-    # HIP: Memory access fault by GPU node-1 (Agent handle: 0x56c21f1d1480) on address 0x730cc242e000. Reason: Page not present or supervisor privilege.
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(51380224), arg=ShapeTracker(views=(View(shape=(256, 1, 64, 56, 56, 1, 1, 1), strides=(200704, 0, 3136, 56, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(51380224), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(51380224), arg=ShapeTracker(views=(View(shape=(1, 256, 1, 64, 4, 58, 4, 58), strides=(0, 200704, 0, 3136, 0, 56, 0, 1), offset=-57, mask=((0, 1), (0, 256), (0, 1), (0, 64), (0, 4), (1, 57), (0, 4), (1, 57)), contiguous=False), View(shape=(256, 1, 64, 56, 56, 64, 3, 3), strides=(3444736, 0, 0, 232, 1, 53824, 13688, 59), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(51380224), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(36864), arg=ShapeTracker(views=(View(shape=(256, 1, 64, 56, 56, 64, 3, 3), strides=(0, 0, 576, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(36864), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.TC, axis=2, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=2, arg=7), Opt(op=OptOps.UPCAST, axis=1, arg=2)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=["HIP", "AMD"])
-
-  @unittest.skipIf(CI and Device.DEFAULT in {"METAL"}, "hangs metal gpu CI")
-  def test_failure_55(self):
-    W = 2
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(W * 200704), arg=ShapeTracker(views=(View(shape=(W, 1, 64, 56, 56, 1, 1, 1), strides=(200704, 0, 3136, 56, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(W * 200704), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(W * 200704), arg=ShapeTracker(views=(View(shape=(1, W, 1, 64, 4, 58, 4, 58), strides=(0, 200704, 0, 3136, 0, 56, 0, 1), offset=-57, mask=((0, 1), (0, W), (0, 1), (0, 64), (0, 4), (1, 57), (0, 4), (1, 57)), contiguous=False), View(shape=(W, 1, 64, 56, 56, 64, 3, 3), strides=(3444736, 0, 0, 232, 1, 53824, 13688, 59), offset=0, mask=None, contiguous=False))), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(W * 200704), arg=1, src=()),)),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.half.ptr(W * 18432), arg=ShapeTracker(views=(View(shape=(W, 1, 64, 56, 56, 64, 3, 3), strides=(0, 0, 576, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(W * 18432), arg=2, src=()),)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.SWAP, axis=1, arg=2)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_56(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(1, 16, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 3)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                x7:=UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-                  x8:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                UOp(Ops.MAX, dtypes.float, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                      UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.float.ptr(247808), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(1936, 121, 11, 1), offset=0, mask=None, contiguous=True),)), src=(
-                              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(247808), arg=1, src=()),)),)),
-                          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=2, src=()),)),)),
-                            UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                               x8,)),)),)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=3, src=()),)),)),)),
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=4, src=()),)),)),)),
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=5, src=()),)),)),)),
-                   x7,)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(204800), arg=ShapeTracker(views=(View(shape=(128, 16, 5, 2, 5, 2), strides=(1600, 100, 20, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(128, 16, 11, 11), strides=(1600, 100, 10, 1), offset=0, mask=((0, 128), (0, 16), (0, 10), (0, 10)), contiguous=False))), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(204800), arg=6, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=2, arg=32)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  def test_failure_57(self):
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(1, 16, 1, 1), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (0, 2, 3)), src=(
-          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.CMPLT, dtypes.bool, arg=None, src=(
-                x7:=UOp(Ops.CONST, dtypes.float, arg=0.0, src=(
-                  x8:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                UOp(Ops.MAX, dtypes.float, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                      UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                        UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                            UOp(Ops.VIEW, dtypes.float.ptr(247808), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(1936, 121, 11, 1), offset=0, mask=None, contiguous=True),)), src=(
-                              UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(247808), arg=1, src=()),)),)),
-                          UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                              UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=2, src=()),)),)),
-                            UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                               x8,)),)),)),
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                            UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=3, src=()),)),)),)),
-                      UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                        UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=4, src=()),)),)),)),
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(16), arg=ShapeTracker(views=(View(shape=(128, 16, 11, 11), strides=(0, 1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(16), arg=5, src=()),)),)),)),
-                   x7,)),)),)),
-            UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-              UOp(Ops.VIEW, dtypes.float.ptr(204800), arg=ShapeTracker(views=(View(shape=(128, 16, 5, 2, 5, 2), strides=(1600, 100, 20, 2, 4, 1), offset=0, mask=None, contiguous=False), View(shape=(128, 16, 11, 11), strides=(1600, 100, 10, 1), offset=0, mask=((0, 128), (0, 16), (0, 10), (0, 10)), contiguous=False))), src=(
-                UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(204800), arg=6, src=()),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=1, arg=32)]
-    helper_test_lin(Kernel(ast, opts=Device[Device.DEFAULT].renderer), opts=opts, failed_platforms=[])
-
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test needs local")
-  def test_failure_59(self):
-    # stable diffusion with SINGLE_KERNEL_SOFTMAX=1
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(268435456), arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 1, 1), strides=(134217728, 16777216, 4096, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(268435456), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.EXP2, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(268435456), arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 1, 1), strides=(134217728, 16777216, 4096, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-                    x9:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(268435456), arg=1, src=()),)),)),
-                UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                  UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.MAX, (5,), True), src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(268435456), arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 1, 4096), strides=(134217728, 16777216, 4096, 0, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                         x9,)),)),)),
-                  UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                    x15:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 1, 1), strides=(0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),
-              UOp(Ops.CONST, dtypes.float, arg=1.4426950408889634, src=(
-                 x15,)),)),)),
-          UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (4,)), src=(
-              UOp(Ops.EXP2, dtypes.float, arg=None, src=(
-                UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(268435456), arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 4096, 1), strides=(134217728, 16777216, 4096, 0, 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                         x9,)),)),
-                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                      UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.MAX, (5,), True), src=(
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(268435456), arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 4096, 4096), strides=(134217728, 16777216, 4096, 0, 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                             x9,)),)),)),
-                      UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                        x29:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 8, 4096, 4096, 4096, 1), strides=(0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),
-                  UOp(Ops.CONST, dtypes.float, arg=1.4426950408889634, src=(
-                     x29,)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UNROLL, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.LOCAL, axis=1, arg=16)]
-    # NOTE: this is slow to run, just confirm it can generate the program without Exception
-    Kernel(ast, opts=Device[Device.DEFAULT].renderer).apply_opts(opts).to_program()
-
-  @unittest.expectedFailure
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test needs local")
-  def test_failure_60(self):
-    # TestSymbolicOps.test_attention
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(80), arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1, 1), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x2:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),
-         x2,)),
-      UOp(Ops.CONST, dtypes.int, arg=4, src=()),)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),
-       x1,)), 0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(80), arg=0, src=()),)),
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.EXP2, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.float, arg=None, src=(
-              UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                  UOp(Ops.VIEW, dtypes.float.ptr(80), arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1, 1), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x2:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),
-         x2,)),
-      UOp(Ops.CONST, dtypes.int, arg=4, src=()),)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),
-       x1,)), 0, 1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-                    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(80), arg=1, src=()),)),)),
-                UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                  UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.MAX, (5,), True), src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(80), arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=())), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=4, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-           x1,
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x1,)), 0, UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=0, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), 0, 1), offset=0, mask=None, contiguous=False),)), src=(
-                        x14:=UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(80), arg=2, src=()),)),)),)),
-                  UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                    x16:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1, 1), strides=(0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),
-              UOp(Ops.CONST, dtypes.float, arg=1.4426950408889634, src=(
-                 x16,)),)),)),
-          UOp(Ops.RECIP, dtypes.float, arg=None, src=(
-            UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (4,)), src=(
-              UOp(Ops.EXP2, dtypes.float, arg=None, src=(
-                UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                  UOp(Ops.ADD, dtypes.float, arg=None, src=(
-                    UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                      UOp(Ops.VIEW, dtypes.float.ptr(80), arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=4, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-           x1,
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x1,)), 0, UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=0, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), 1, 0), offset=0, mask=None, contiguous=False),)), src=(
-                         x14,)),)),
-                    UOp(Ops.MUL, dtypes.float, arg=None, src=(
-                      UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.MAX, (5,), True), src=(
-                        UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-                          UOp(Ops.VIEW, dtypes.float.ptr(80), arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=())), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.CONST, dtypes.int, arg=4, src=()),
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      x0:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-         x0,
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)), 0, UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.CONST, dtypes.int, arg=0, src=()),
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)), 1), offset=0, mask=None, contiguous=False), View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=())), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=4, src=()),
-        x2:=UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x2,)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        x2:=UOp(Ops.MUL, dtypes.int, arg=None, src=(
-           x1,
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x2,)), 0, UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=0, src=()),
-        x2:=UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x2,)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      x0:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-         x0,
-        UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)), 1), offset=0, mask=None, contiguous=False))), src=(
-                             x14,)),)),)),
-                      UOp(Ops.CONST, dtypes.float, arg=-1.0, src=(
-                        x30:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=())), strides=(0, 0, 0, 0), offset=0, mask=None, contiguous=False), View(shape=(2, 4, 1, UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()), 1), strides=(UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=4, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-           x1,
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x1,)), 0, UOp(Ops.MUL, dtypes.int, arg=None, src=(
-      UOp(Ops.MUL, dtypes.int, arg=None, src=(
-        UOp(Ops.CONST, dtypes.int, arg=0, src=()),
-        UOp(Ops.MUL, dtypes.int, arg=None, src=(
-          x3:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),
-          UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10), src=()),)),)),
-       x3,)), 1, 0), offset=0, mask=None, contiguous=False))), src=()),)),)),)),
-                  UOp(Ops.CONST, dtypes.float, arg=1.4426950408889634, src=(
-                     x30,)),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=0, arg=2), Opt(op=OptOps.LOCAL, axis=0, arg=4)]
-    # NOTE: this is slow to run, just confirm it can generate the program without Exception
-    Kernel(ast, opts=Device[Device.DEFAULT].renderer).apply_opts(opts).to_program()
-
-  def test_failure_61(self):
-    # WINO=1 JITBEAM=4 python3 examples/beautiful_cifar.py
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.half.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 1, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1024), arg=0, src=()),)),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.CAST, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.half, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                x7:=UOp(Ops.CONST, dtypes.half, arg=0.6931471805599453, src=(
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                UOp(Ops.CAST, dtypes.half, arg=None, src=(
-                  UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (1,)), src=(
-                    UOp(Ops.CAST, dtypes.float, arg=None, src=(
-                      UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                        UOp(Ops.CONST, dtypes.half, arg=-1.0, src=(
-                          x14:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 10, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                        UOp(Ops.ADD, dtypes.half, arg=None, src=(
-                          UOp(Ops.CONST, dtypes.half, arg=-0.010000000000000002, src=(
-                             x14,)),
-                          UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                            UOp(Ops.CAST, dtypes.half, arg=None, src=(
-                              UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                                UOp(Ops.CMPNE, dtypes.bool, arg=None, src=(
-                                  UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-                                    UOp(Ops.VIEW, dtypes.int.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 10, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
-                                      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1024), arg=1, src=()),)),)),
-                                  UOp(Ops.ADD, dtypes.int, arg=None, src=(
-                                    UOp(Ops.REDUCE_AXIS, dtypes.int, arg=(Ops.ADD, (2,), True), src=(
-                                      UOp(Ops.WHERE, dtypes.int, arg=None, src=(
-                                        UOp(Ops.VALID, dtypes.bool, arg=None, src=(
-                                          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(11, 19), strides=(0, 0), offset=0, mask=((0, 11), (9, 19)), contiguous=False), View(shape=(1024, 10, 10), strides=(0, 1, 20), offset=0, mask=None, contiguous=False))), src=()),)),
-                                        UOp(Ops.CONST, dtypes.int, arg=1, src=(
-                                          x30:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 10, 10), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),
-                                        UOp(Ops.CONST, dtypes.int, arg=0, src=(
-                                           x30,)),)),)),
-                                    UOp(Ops.CONST, dtypes.int, arg=-1, src=(
-                                       x14,)),)),)),
-                                UOp(Ops.CONST, dtypes.bool, arg=True, src=(
-                                   x14,)),)),)),
-                            UOp(Ops.CONST, dtypes.half, arg=-0.4, src=(
-                               x14,)),)),)),)),)),)),)),)),
-              UOp(Ops.RECIP, dtypes.half, arg=None, src=(
-                UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                  UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                    UOp(Ops.VIEW, dtypes.half.ptr(1024), arg=ShapeTracker(views=(View(shape=(1024, 1, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=True),)), src=(
-                      UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(1024), arg=2, src=()),)),)),
-                   x7,)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=0, arg=32), Opt(op=OptOps.GROUP, axis=1, arg=0)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=["AMD", "METAL", "CUDA", "NV"])
-
-  def test_failure_62(self):
-    # WINO=1 DEFAULT_FLOAT=HALF FUSE_ARANGE=1 JITBEAM=4 BS=1024 STEPS=500 python examples/hlb_cifar10.py
-    # RuntimeError: UOp verification failed at 4 on Ops.LOAD dtypes.half 2 [<Ops.DEFINE_GLOBAL: 24>, <Ops.VIEW: 23>] None
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(11808768), arg=0, src=()),
-        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1, 12, 31, 31, 1, 1, 1), strides=(11532, 0, 961, 31, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()),
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(3145728), arg=1, src=()),
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1, 12, 31, 31, 3, 2, 2), strides=(3072, 0, 0, 32, 1, 1024, 32, 1), offset=0, mask=None, contiguous=False),)), src=()),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(144), arg=2, src=()),
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1024, 1, 12, 31, 31, 3, 2, 2), strides=(0, 0, 12, 0, 0, 4, 2, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)),)),))
-    opts = [Opt(op=OptOps.LOCAL, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=1, arg=3), Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.PADTO, axis=2, arg=32), Opt(op=OptOps.UPCAST, axis=2, arg=4), Opt(op=OptOps.UPCAST, axis=2, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=0)]
-    helper_test_lin(Kernel(ast), opts, failed_platforms=["AMD", "HIP", "NV", "CUDA"])
-
-if __name__ == '__main__':
-  unittest.main()
diff --git a/test/test_search.py b/test/test_search.py
index af2c4f0a39..8faa25b823 100644
--- a/test/test_search.py
+++ b/test/test_search.py
@@ -83,6 +83,7 @@ class TestBEAM(unittest.TestCase):
     actions_after = actions.copy()
     assert actions_after == actions_before, "actions state was not preserved"
 
+  @unittest.skip("invalid reduce now")
   def test_filter_global_buffer(self):
     # taken from https://github.com/tinygrad/tinygrad/issues/4612
     ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
diff --git a/tinygrad/opt/kernel.py b/tinygrad/opt/kernel.py
index 581b165e9b..4dd0a8ad15 100644
--- a/tinygrad/opt/kernel.py
+++ b/tinygrad/opt/kernel.py
@@ -58,11 +58,6 @@ class Kernel:
     # add a shapetracker to the end to track the full shape, with 0 strides so it can merge
     self.sts.append(ShapeTracker.from_shape(tuple([smax(*s) for s in zip(*[x.shape for x in self.sts])]), (0,)*self.shape_len))
 
-    # move all reduce axes to the end
-    reduce = list(enumerate(zip(self.full_shape, self.output_shape)))
-    permute = tuple([i for i,(s,n) in reduce if not resolve(s != n)] + [i for i,(s,n) in reduce if resolve(s != n)])
-    self.reshape_and_permute(None, permute)
-
     # parameters for optimization
     self.applied_opts: list[Opt] = []
     self.group_for_reduces: int = 0
@@ -77,6 +72,11 @@ class Kernel:
     self.simplify_ones()
     self.simplify_merge_adjacent()
 
+    # confirm all reduce axes are at the end
+    final_reduces = [i for i,(s,n) in enumerate(zip(self.full_shape, self.output_shape)) if resolve(s != n)]
+    if final_reduces != list(range(len(self.full_shape)-len(final_reduces), len(self.full_shape))):
+      raise RuntimeError(f"reduces are not at the end of the shape {self.full_shape} -> {self.output_shape}")
+
   def copy(self):
     ret = type(self).__new__(type(self))