[Fix] Clear stale FlashInfer BF16 MoE index cache (#33016)

2026-07-31 00:35:15 -07:00
parent 585a7d05e3
commit 937c77cf50
1 changed files with 5 additions and 0 deletions
@@ -374,6 +374,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):

        # Reorder rows of W1 for fused gated activation
        if self.use_flashinfer_trtllm_moe:
+            # The cached indices are GPU tensors. Colocated weight offloading
+            # can release their backing memory between reloads, so rebuild them
+            # once per post-processing cycle.
+            self._cache_permute_indices.clear()
+
            from flashinfer.fused_moe.core import (
                _maybe_get_cached_w3_w1_permute_indices,
                convert_to_block_layout,