skip some PTX IGNORE_OOB validation (#13927)

.github/workflows/test.yml

@@ -822,6 +822,7 @@ jobs:
         NV_PTX: 1
         NV: 1
         FORWARD_ONLY: 1
+        IGNORE_OOB: 0
       run: |
         python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
     - name: Run process replay tests

tinygrad/uop/validate.py

@@ -1,6 +1,6 @@
 from typing import Callable, cast
 from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, python_alu
-from tinygrad.dtype import ImageDType, dtypes, Invalid
+from tinygrad.dtype import ImageDType, dtypes, Invalid, PtrDType
 from tinygrad.helpers import IGNORE_OOB, cpu_profile
 
 try:
@@ -70,6 +70,9 @@ def validate_index(buf:UOp, idx:UOp, gate:UOp|None=None):
   # WEBGPU has a BITCAST in the index. TODO: fix
   if any(x.op is Ops.BITCAST for x in idx.toposort() | gate.toposort()): return True
 
+  # PTX uses absolute addresses (pointer cast to long), skip validation
+  if any(x.op is Ops.CAST and isinstance(x.src[0].dtype, PtrDType) for x in idx.toposort()): return True
+
   if not z3_imported: raise ImportError("bounds checking requires z3 >= 4.12.4, use IGNORE_OOB=1 to disable, or \"pip install 'z3-solver>=4.12.4\"")
   solver = z3.Solver(ctx=z3.Context())
   z3_idx, z3_mask = uops_to_z3(solver, idx, gate)