WIP

benchmarks/python/test_rope.py

@@ -262,22 +262,10 @@ def grads():
 
 # { 'name_benchmark' : (fn, [[sizes0, optional_strides0, dtype0], [sizes1, dtype1], ...]) }
 rope_setup = {
-    "llama_2_7b_hf_rope": (
+    "llama_2_7b_hf_rope":
         partial(llama_hf_rope, config_str="llama_2_7b_hf_rope"),
-        [
-            ((2, 4096, 12288), torch.bfloat16),
-            ((4096, 128), torch.bfloat16),
-            ((4096, 128), torch.bfloat16),
-        ],
-    ),
-    "llama_3_8B_rope": (
+    "llama_3_8B_rope":
         partial(llama_hf_rope, config_str="llama_3_8B_rope"),
-        [
-            ((2, 8192, 6144), torch.bfloat16),
-            ((8192, 128), torch.bfloat16),
-            ((8192, 128), torch.bfloat16),
-        ],
-    ),
     # "hf_qwen2_rope": (
     #     hf_qwen2_rope,
     #     [
@@ -328,9 +316,7 @@ def test_rope_variations_fwd_benchmark(
     if executor == "torchcompile":
         clear_dynamo_cache()
 
-    config = rope_setup[rope_variation]
-
-    model, inputs, _ = config[0]()
+    model, inputs, _ = rope_setup[rope_variation]
 
     def fwd_call(inp):
         return model(*inp)
@@ -358,8 +344,8 @@ def test_rope_variations_bwd_benchmark(
     if executor == "torchcompile":
         clear_dynamo_cache()
 
-    config = rope_setup[rope_variation]
-    model, fwd_inputs, grad = config[0]()
+    # TODO why not just a random like for grad on output instead of returning a grad function
+    model, fwd_inputs, grad = rope_setup[rope_variation]
 
     def fwd_call(inp):
         return model(*inp)
@@ -368,11 +354,10 @@ def fwd_call(inp):
     fwd_fn = with_executor(executor, fwd_call)
     outputs = fwd_fn(fwd_inputs())
 
+    # NOTE does this look about right?
     output = outputs[0]
     for i in range(1, len(outputs)):
         output += outputs[i]
 
-    print(f"{output.shape=}")
-    print(f"{grad().shape=}")
     benchmark_fn = with_executor(executor, fwd_call)
     run_benchmark(benchmark, unary_bwd_torch, [output, grad()])