diff --git a/dsv4/layers/moe.py b/dsv4/layers/moe.py
index 6d7bf149..b9b327d7 100644
--- a/dsv4/layers/moe.py
+++ b/dsv4/layers/moe.py
@@ -164,6 +164,10 @@ class Nvfp4MoE:
         self._l1_gsa_buf = torch.zeros(self.num_experts, dtype=torch.float32, device=self.device)
         self._l2_gsa_buf = torch.zeros(self.num_experts, dtype=torch.float32, device=self.device)
         
+        # Pre-allocated tokens-per-expert buffer — replaces torch.bincount
+        # (bincount produces data-dependent shapes, breaks CUDA graph capture)
+        self._tokens_per_expert_buf = torch.zeros(self.num_experts, dtype=torch.int32, device=self.device)
+        
         # Row indices for scale assembly (max_num_tokens * top_k slots)
         self._row_indices_buf = torch.arange(
             self.max_num_tokens * self.top_k, device=self.device
@@ -466,7 +470,16 @@ class Nvfp4MoE:
             # Quantize slot_hidden for GEMM
             slot_x_fp4, slot_x_sf = quantize_activation_nvfp4(slot_hidden, l1_gs)
             
-            tokens_per_expert = torch.bincount(sorted_ids, minlength=self.num_experts)[:self.num_experts].int()
+            # Compute tokens_per_expert — CUDA-graph-safe alternative to torch.bincount.
+            # torch.bincount produces data-dependent shapes (violates graph capture).
+            # Instead, use scatter_add_ into a pre-allocated buffer (fixed shape, GPU-only).
+            self._tokens_per_expert_buf.zero_()
+            # Pre-allocated ones buffer — avoids per-call torch.ones() allocation
+            n_slots = sorted_ids.shape[0]
+            if not hasattr(self, '_ones_buf') or self._ones_buf.shape[0] < n_slots:
+                self._ones_buf = torch.ones(self.max_num_tokens * self.top_k, dtype=torch.int32, device=sorted_ids.device)
+            self._tokens_per_expert_buf.scatter_add_(0, sorted_ids, self._ones_buf[:n_slots])
+            tokens_per_expert = self._tokens_per_expert_buf[:self.num_experts]
             expert_offsets = self._expert_offsets_buf
             expert_offsets.zero_()
             expert_offsets[1:self.num_experts + 1] = tokens_per_expert.cumsum(0)
@@ -494,7 +507,9 @@ class Nvfp4MoE:
                 padded_expert_offsets,
                 self._padded_x_sf_buf_l1, self._per_expert_scale_bufs_l1
             )
-            l1_gsa = torch.full((self.num_experts,), l1_gs, dtype=torch.float32, device=device)
+            # l1_gsa: pre-allocated buffer, no per-call allocation
+            self._l1_gsa_buf.fill_(l1_gs)
+            l1_gsa = self._l1_gsa_buf
             
             l1_out = run_nvfp4_grouped_gemm(
                 mat_a=padded_x_fp4, mat_b=self._l1_mat_b,