diff --git a/python/sglang/srt/layers/attention/triton_backend.py b/python/sglang/srt/layers/attention/triton_backend.py
index c1f921ce1..2daae2f82 100644
--- a/python/sglang/srt/layers/attention/triton_backend.py
+++ b/python/sglang/srt/layers/attention/triton_backend.py
@@ -205,9 +205,11 @@ class TritonAttnBackend(AttentionBackend):
             self.v_head_dim = model_runner.token_to_kv_pool.get_v_head_dim()
             self.swa_v_head_dim = None
         else:
-            self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[
-                -1
-            ]
+            # Use start_layer instead of 0 to handle pipeline parallelism.
+            # In PP, start_layer may be > 0, so layer 0 isn't in this stage's buffer.
+            self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(
+                model_runner.token_to_kv_pool.start_layer
+            ).shape[-1]
             self.swa_v_head_dim = None
         self.max_context_len = model_runner.model_config.context_len
         self.device = model_runner.device
diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py
index 0aa942076..3da4022d9 100644
--- a/python/sglang/srt/mem_cache/memory_pool.py
+++ b/python/sglang/srt/mem_cache/memory_pool.py
@@ -3869,7 +3869,11 @@ class HybridLinearKVPool(KVCache):
             )
 
     def get_v_head_dim(self):
-        return self.full_kv_pool.get_value_buffer(0).shape[-1]
+        # Use start_layer to handle pipeline parallelism where layer 0
+        # may not be present in this stage's buffer.
+        return self.full_kv_pool.get_value_buffer(self.full_kv_pool.start_layer).shape[
+            -1
+        ]
 
     def set_mla_kv_buffer(
         self,
@@ -4991,4 +4995,6 @@ class MiniMaxSparseKVPool(KVCache):
         )
 
     def get_v_head_dim(self):
-        return self.main_pool.get_value_buffer(0).shape[-1]
+        # Use start_layer to handle pipeline parallelism where layer 0
+        # may not be present in this stage's buffer.
+        return self.main_pool.get_value_buffer(self.main_pool.start_layer).shape[-1]