diff --git a/benchmark/scripts/benchmark_llama4_rope.py b/benchmark/scripts/benchmark_llama4_rope.py
index 3e71dceb..a3b03239 100644
--- a/benchmark/scripts/benchmark_llama4_rope.py
+++ b/benchmark/scripts/benchmark_llama4_rope.py
@@ -52,7 +52,7 @@ def setup_llama4_rope(input: SingleBenchmarkRunInput):
         Llama4TextRotaryEmbedding,
         Llama4TextRotaryEmbedding,
         before_kwargs={"config": config, "device": device},
-        after_kwargs={"config": config, "device": device},
+        after_kwargs={"config": config},
     )
 
     q = torch.randn(
diff --git a/benchmark/scripts/benchmark_rope.py b/benchmark/scripts/benchmark_rope.py
index d7e1868c..b788cb21 100644
--- a/benchmark/scripts/benchmark_rope.py
+++ b/benchmark/scripts/benchmark_rope.py
@@ -48,7 +48,7 @@ def setup_rope(input: SingleBenchmarkRunInput):
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
     q = torch.randn(
         (1, seq_len, num_q_heads, head_dim),
diff --git a/src/liger_kernel/utils.py b/src/liger_kernel/utils.py
index 8c11ac08..6421762f 100644
--- a/src/liger_kernel/utils.py
+++ b/src/liger_kernel/utils.py
@@ -183,7 +183,7 @@ def transformers_version_dispatch(
         ...     before_args=(head_dim,),
         ...     after_args=(LlamaConfig(head_dim=head_dim),),
         ...     before_kwargs={'device': device},
-        ...     after_kwargs={'device': device}
+        ...     after_kwargs={}  # config-based constructor derives device from input; newer transformers rejects `device`
         ... )
     """
     from packaging import version
diff --git a/test/cutedsl/test_rope.py b/test/cutedsl/test_rope.py
index 64d02a85..52002a0e 100644
--- a/test/cutedsl/test_rope.py
+++ b/test/cutedsl/test_rope.py
@@ -70,7 +70,7 @@ def test_cutedsl_matches_triton_and_hf(
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((bsz, seq_len, num_q_heads, head_dim), device=device).transpose(1, 2).to(dtype)
@@ -148,7 +148,7 @@ def test_cutedsl_backward_contiguous_grad(bsz, seq_len, num_q_heads, num_kv_head
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((bsz, seq_len, num_q_heads, head_dim), device=device).transpose(1, 2).to(dtype)
@@ -211,7 +211,7 @@ def test_cutedsl_vision_2d_cos_sin(seq_len, num_heads, head_dim, dtype, atol, rt
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((seq_len, num_heads, head_dim), device=device).to(dtype)
diff --git a/test/cutile/test_rope.py b/test/cutile/test_rope.py
index 87afa085..5d2204e3 100644
--- a/test/cutile/test_rope.py
+++ b/test/cutile/test_rope.py
@@ -80,7 +80,7 @@ def test_cutile_matches_triton_and_hf(
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((bsz, seq_len, num_q_heads, head_dim), device=device).transpose(1, 2).to(dtype)
@@ -162,7 +162,7 @@ def test_cutile_backward_contiguous_grad(bsz, seq_len, num_q_heads, num_kv_heads
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((bsz, seq_len, num_q_heads, head_dim), device=device).transpose(1, 2).to(dtype)
@@ -247,7 +247,7 @@ def test_cutile_vision_2d_cos_sin(seq_len, num_heads, head_dim, dtype, atol, rto
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim)},
     )
 
     # Vision layout: q, k arrive as (seq_len, num_heads, head_dim); the wrapper
diff --git a/test/transformers/test_rope.py b/test/transformers/test_rope.py
index 132ceb95..9b0a34a8 100644
--- a/test/transformers/test_rope.py
+++ b/test/transformers/test_rope.py
@@ -67,7 +67,7 @@ def test_correctness(
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     _tensor_q = torch.randn((bsz, seq_len, num_q_heads, head_dim), device=device).transpose(1, 2).to(dtype)
@@ -152,7 +152,7 @@ def test_functional_correctness(
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_kv_heads, head_dim=head_dim)},
     )
 
     pos_ids = torch.arange(seq_len, device=device, dtype=torch.long).unsqueeze(0)
@@ -224,7 +224,7 @@ def test_vision_2d_cos_sin(seq_len, num_heads, head_dim, dtype, atol, rtol):
         LlamaRotaryEmbedding,
         LlamaRotaryEmbedding,
         before_kwargs={"dim": head_dim, "device": device},
-        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim), "device": device},
+        after_kwargs={"config": LlamaConfig(num_kv_heads=num_heads, head_dim=head_dim)},
     )
 
     _q = torch.randn((seq_len, num_heads, head_dim), device=device).to(dtype)