fix(cambricon): stop index autotune from killing the process on untileable expand_shape

src/flag_gems/runtime/backend/_cambricon/tune_configs.yaml

@@ -1462,10 +1462,14 @@ index:
   - 1
   - 2
   - 4
+  # BLOCK_SIZE1=4096 excluded: with the default num_stages=2 a 4096-wide tile
+  # exceeds NRAM (always OutOfResources, can never win) and BLOCK_SIZE0=4 x
+  # BLOCK_SIZE1=4096 trips an MLU AutoTileForTritonPass failure
+  # ("PassManager::run failed" on an untileable tensor.expand_shape that
+  # collapses [1, 4, 4096]) in inductor-wrapped contexts with dynamic shapes.
   block_size1:
   - 1024
   - 2048
-  - 4096
   warps:
   - 1
   - 4

src/flag_gems/utils/libentry.py

@@ -995,7 +995,20 @@ def run(self, *args, **kwargs):
                 def bench(config: triton.Config) -> List[float]:
                     ret = cache.get(config)
                     if ret is None:
-                        ret = self._bench(*args, config=config, **kwargs)
+                        try:
+                            ret = self._bench(*args, config=config, **kwargs)
+                        except RuntimeError as e:
+                            # A config whose COMPILE raises a plain RuntimeError
+                            # is outside triton's autotuner catch list
+                            # (OutOfResources / CompileTimeAssertionFailure /
+                            # PTXASError) and would kill the whole process. Some
+                            # backend compilers do this — e.g. cambricon MLU
+                            # AutoTileForTritonPass raises "PassManager::run
+                            # failed" on a tensor.expand_shape it cannot tile.
+                            # Treat such a config as a non-candidate (inf) so
+                            # tuning continues and a working config is selected.
+                            print(f"[libentry] config {config} failed to compile: {e}")
+                            ret = (float("inf"), float("inf"), float("inf"))
                         # Some Triton backends (e.g. tsingmicro with
                         # use_cuda_graph=True) return a scalar float from
                         # _bench instead of the standard (p50, p20, p80) tuple.