mxfp8 inference roofline: add fusion to observed (#3223)

Summary:

Adds option to benchmark with relu -> linear to capture the impact
of fusing the activation to the quant kernel

Test Plan:

```bash
(pt_nightly_312_2) [[email protected] ~/local/ao
(20251021_inference_fusion_modeling)]$ python
benchmarks/float8/float8_inference_roofline.py ~/local/tmp/test.csv
--recipe_name mxfp8_cublas --shape_gen_name pow2_extended
--enable_fusion_modeling True
```

Reviewers:

Subscribers:

Tasks:

Tags:

Author: vkuzo

benchmarks/float8/float8_inference_roofline.py

@@ -168,6 +168,7 @@ def run(
     shape_gen_name: str = "pow2",
     n_limit: Optional[int] = None,
     save_profile_traces: bool = False,
+    enable_fusion_modeling: bool = False,
 ):
     """
     Args:
@@ -176,6 +177,7 @@ def run(
     * `shape_gen_name`: `llama`, `pow2`, `pow2_extended`, or `sweep`
     * `n_limit (optional)`: if specified, only runs `n_limit` iterations
     # `save_profile_traces (optional)`: if True, saves profiling traces
+    # `enable_fusion_modeling`: if True, models activation -> gemm instead of just gemm
     """
     config_table = [
         ["GPU", torch.cuda.get_device_name(0)],
@@ -184,6 +186,7 @@ def run(
         ["recipe_name", recipe_name],
         ["do_benchmarks", do_benchmarks],
         ["shape_gen_name", shape_gen_name],
+        ["enable_fusion_modeling", enable_fusion_modeling],
     ]
     print(tabulate(config_table, headers=["Parameter", "Value"], tablefmt="simple"))
 
@@ -194,6 +197,7 @@ def run(
         K,
         N,
         recipe_name,
+        # TODO(future): also enable fusion modeling here
     )
     bf16_gemm_time_sympy = get_inference_gemm_time_sympy(M, K, N, torch.bfloat16, None)
 
@@ -287,9 +291,11 @@ def run(
         b_bf16_e2e_time_s, b_fp8_e2e_time_s = 0, 0
         if do_benchmarks:
             # create the model
-            m_orig = (
-                nn.Sequential(nn.Linear(K_val, N_val, bias=False)).cuda().bfloat16()
-            )
+            if not enable_fusion_modeling:
+                m_orig = nn.Sequential(nn.Linear(K_val, N_val, bias=False))
+            else:
+                m_orig = nn.Sequential(nn.ReLU(), nn.Linear(K_val, N_val, bias=False))
+            m_orig = m_orig.cuda().bfloat16()
             x = torch.randn(
                 M_val, K_val, dtype=torch.bfloat16, device="cuda"
             ).requires_grad_()