diff --git a/extra/onnx.py b/extra/onnx.py
index 265a3dc505..43f0f25aa1 100644
--- a/extra/onnx.py
+++ b/extra/onnx.py
@@ -812,7 +812,7 @@ def get_onnx_ops():
       else:
         # DSP swizzle memory
         x = x.reshape(x.shape[0], x.shape[1]//WEIGHT_SHIFT, WEIGHT_SHIFT).permute(1,0,2).contiguous().permute(1,0,2).reshape(x.shape)
-    if False and getenv("NHWC") and x.shape == (1000, 1280):
+    if getenv("NHWC") and x.shape == (1000, 1280):
       x = x.reshape(-1, 320, 4)
       order = (1,0,2)
       x = x.permute(*order).contiguous().permute(*argsort(order))
diff --git a/extra/replay_pkl.py b/extra/replay_pkl.py
index 1a6af06de5..d63154fb5d 100644
--- a/extra/replay_pkl.py
+++ b/extra/replay_pkl.py
@@ -32,7 +32,8 @@ if __name__ == "__main__":
             k.apply_opt(Opt(OptOps.UPCAST, 1, 4))
           elif knum == 66:
             k.apply_opt(Opt(OptOps.UNROLL, 0, 4))
-            k.apply_opt(Opt(OptOps.UPCAST, 0, 8))
+            k.apply_opt(Opt(OptOps.PADTO, 0, 128))
+            k.apply_opt(Opt(OptOps.UPCAST, 0, 128))
           elif k.full_shape[-3:] == (32,3,3):
             # NOTE: there's an issue with the 7 -> 8 upcast and the valid removal
             if knum not in [52,56,60] and k.full_shape[-4]%4 != 0: k.apply_opt(Opt(OptOps.PADTO, len(k.full_shape)-4, 4))
@@ -55,6 +56,8 @@ if __name__ == "__main__":
             k.apply_opt(Opt(OptOps.UNROLL, 0, 8))
             k.apply_opt(Opt(OptOps.UPCAST, 2, 32))
             if k.full_shape[1]%4 == 0: k.apply_opt(Opt(OptOps.UPCAST, 1, 4))
+          elif len(k.full_shape) == 2:
+            if k.full_shape[0]%128 == 0: k.apply_opt(Opt(OptOps.UPCAST, 0, 128))
           elif len(k.full_shape) == 1:
             for sz in [128,64,32]:
               if k.full_shape[0]%sz == 0:
diff --git a/tinygrad/codegen/kernel.py b/tinygrad/codegen/kernel.py
index 34a81e80b4..b1dd926c5b 100644
--- a/tinygrad/codegen/kernel.py
+++ b/tinygrad/codegen/kernel.py
@@ -387,7 +387,7 @@ class Kernel:
       self.group_for_reduces += 1
     elif opt.op is OptOps.UNROLL:                     # purple
       check(axis < self.first_upcast, "can't upcasted already upcasted")
-      check(amt <= 32, "don't unroll more than 32")
+      #check(amt <= 32, "don't unroll more than 32")
       # TODO: fix upcast_count to put purples before yellows. broken because of METAL tensor cores
       #upcast_count = sum(x == y for x,y in zip(self.full_shape[-self.upcasted:], self.output_shape[-self.upcasted:])) if self.upcasted else 0
       #self.shift_to(axis, amt, insert_before=None if upcast_count == 0 else self.shape_len-upcast_count)