diff --git a/benchmarks/python/test_dropout_layernorm_bwd.py b/benchmarks/python/test_dropout_layernorm_bwd.py
index 7182d2326aa..ff5de768ba5 100644
--- a/benchmarks/python/test_dropout_layernorm_bwd.py
+++ b/benchmarks/python/test_dropout_layernorm_bwd.py
@@ -220,6 +220,6 @@ def test_dropout_layernorm_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=dropout_layernorm_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_dropout_rmsnorm_bwd.py b/benchmarks/python/test_dropout_rmsnorm_bwd.py
index 332cacbb243..5182a9f4e6a 100644
--- a/benchmarks/python/test_dropout_rmsnorm_bwd.py
+++ b/benchmarks/python/test_dropout_rmsnorm_bwd.py
@@ -196,6 +196,6 @@ def test_dropout_rmsnorm_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=dropout_rmsnorm_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_gelu_bwd.py b/benchmarks/python/test_gelu_bwd.py
index 90ec6bc2ad9..04b6f705255 100644
--- a/benchmarks/python/test_gelu_bwd.py
+++ b/benchmarks/python/test_gelu_bwd.py
@@ -117,6 +117,6 @@ def test_gelu_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=gelu_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_huggingface_attn_bwd.py b/benchmarks/python/test_huggingface_attn_bwd.py
index c886204d5ee..258d98d1a37 100644
--- a/benchmarks/python/test_huggingface_attn_bwd.py
+++ b/benchmarks/python/test_huggingface_attn_bwd.py
@@ -144,6 +144,6 @@ def test_huggingface_attn_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=huggingface_attn_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_layernorm_bwd.py b/benchmarks/python/test_layernorm_bwd.py
index 165afb7967f..08be8e6a7b6 100644
--- a/benchmarks/python/test_layernorm_bwd.py
+++ b/benchmarks/python/test_layernorm_bwd.py
@@ -178,6 +178,6 @@ def test_layernorm_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=layernorm_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_nanogpt_attn_bwd.py b/benchmarks/python/test_nanogpt_attn_bwd.py
index d4884893436..54d19fd7093 100644
--- a/benchmarks/python/test_nanogpt_attn_bwd.py
+++ b/benchmarks/python/test_nanogpt_attn_bwd.py
@@ -162,6 +162,6 @@ def test_nanogpt_attn_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=nanogpt_attn_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_rmsnorm_bwd.py b/benchmarks/python/test_rmsnorm_bwd.py
index 3fe7ed8d653..3f7a15fa1a1 100644
--- a/benchmarks/python/test_rmsnorm_bwd.py
+++ b/benchmarks/python/test_rmsnorm_bwd.py
@@ -143,6 +143,6 @@ def test_rmsnorm_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=rmsnorm_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_scale_bias_relu_bwd.py b/benchmarks/python/test_scale_bias_relu_bwd.py
index 6f4b0498480..219247fc8e7 100644
--- a/benchmarks/python/test_scale_bias_relu_bwd.py
+++ b/benchmarks/python/test_scale_bias_relu_bwd.py
@@ -110,6 +110,6 @@ def test_sbr_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=sbr_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_silu_mul_bwd.py b/benchmarks/python/test_silu_mul_bwd.py
index f3da0cbdb9f..b5aaa805161 100644
--- a/benchmarks/python/test_silu_mul_bwd.py
+++ b/benchmarks/python/test_silu_mul_bwd.py
@@ -109,6 +109,6 @@ def test_silu_mul_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=silu_mul_bwd_iobytes(size, dtype),
     )
diff --git a/benchmarks/python/test_softmax_bwd.py b/benchmarks/python/test_softmax_bwd.py
index 885fba8b75a..685f9ad4b3e 100644
--- a/benchmarks/python/test_softmax_bwd.py
+++ b/benchmarks/python/test_softmax_bwd.py
@@ -121,6 +121,6 @@ def test_softmax_bwd_baseline_benchmark(
     run_benchmark(
         benchmark,
         unary_bwd_torch,
-        [outputs, grads],
+        [outputs, grads, *fwd_inputs],
         iobytes=softmax_bwd_iobytes(size, dtype),
     )