pointwise_dynamic.patch

diff --git a/src/flag_gems/utils/pointwise_dynamic.py b/src/flag_gems/utils/pointwise_dynamic.py
index b499310..28efc08 100644
--- a/src/flag_gems/utils/pointwise_dynamic.py
+++ b/src/flag_gems/utils/pointwise_dynamic.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import dataclasses
 import importlib
 import os
 from dataclasses import dataclass
@@ -1318,9 +1319,20 @@ class PointwiseDynamicFunction:
 
     def _call_real_impl(self, *args, **kwargs):
         """Single entry point for real kernel invocation."""
-        ndim, args, kwargs = self.prepare_args(*args, **kwargs)
-        overload = self.instantiate(ndim)
-        out = overload(*args, **kwargs)
+        ndim, args, kwargs, call_config = self.prepare_args(*args, **kwargs)
+        # wt-2026-09-11-fix (#5739): prepare_args may return a per-call config
+        # (block pointer disabled for broadcasted operands). Use it only for
+        # this call's instantiation so the shared vendor config is never
+        # polluted, and always restore afterwards (thread-safety best effort).
+        # wt <wangt635@ustc.edu.cn>
+        saved_config = self.config
+        if call_config is not None:
+            self.config = call_config
+        try:
+            overload = self.instantiate(ndim)
+            out = overload(*args, **kwargs)
+        finally:
+            self.config = saved_config
         return self._unwrap(out)
 
     # -------------------- complex helpers --------------------
@@ -1529,8 +1541,13 @@ class PointwiseDynamicFunction:
 
         tensors = out_tensors + in_tensors
         INT32_MAX = torch.iinfo(torch.int32).max
-        if tensors[0].numel() > INT32_MAX:
-            self.config.prefer_block_pointer = False
+        # wt-2026-09-11-fix (#5739): the old code mutated the *shared* vendor
+        # config in place (self.config.prefer_block_pointer = False), silently
+        # disabling block pointers for every other pointwise operator after one
+        # big call (singleton pollution, also called out by upstream PR #5668).
+        # wt <wangt635@ustc.edu.cn>
+        call_config = None
+        need_no_block_pointer = tensors[0].numel() > INT32_MAX
         if self.use_fast_path(tensors):  # dimension collapse & use physical ordering
             allocated_outputs = [
                 torch.empty_like(tensors[0], dtype=dtype)
@@ -1563,6 +1580,23 @@ class PointwiseDynamicFunction:
 
             task_shape = broadcast_shapes(shapes)
 
+            # wt-2026-09-11-fix (#5739, Ascend): broadcasted operands get a zero
+            # stride on expanded axes (broadcasted_stride), which block pointers
+            # cannot represent — the TritonToLinalg pipeline rejects them with
+            # "strides must not be zero" (ConvertTritonIRToLinalgIR on
+            # triton-ascend 3.5.1; ConvertLinalgRToBinary on the issue's 3.2).
+            # Reproduced: with prefer_block_pointer=True, non-broadcast add
+            # compiles fine while broadcast (128,1)+(1,128) always fails; the
+            # zero stride is the only variable. Disable block pointers for this
+            # call only (linear addressing handles stride 0 natively) and route
+            # through a copied config so the shared vendor singleton is untouched.
+            # wt <wangt635@ustc.edu.cn>
+            if not need_no_block_pointer:
+                for item in tensors:
+                    if item.shape != task_shape:
+                        need_no_block_pointer = True
+                        break
+
             if out_tensors:
                 for index, item in enumerate(out_tensors):
                     if list(item.shape) != list(task_shape):
@@ -1616,7 +1650,13 @@ class PointwiseDynamicFunction:
                     task_shape,
                     broadcasted_stride(item.shape, item.stride(), task_shape),
                 )
-        return (ndim, args, kwargs)
+        # wt-2026-09-11-fix (#5739): materialize the per-call config once here;
+        # dataclasses.replace copies the shared vendor config so flipping
+        # prefer_block_pointer off never leaks to other operators.
+        # wt <wangt635@ustc.edu.cn>
+        if need_no_block_pointer and self.config.prefer_block_pointer:
+            call_config = dataclasses.replace(self.config, prefer_block_pointer=False)
+        return (ndim, args, kwargs, call_config)
 
     def _unwrap(self, tensors):
         # unwrap StridedBuffer to get Tensor