small fix for inference roofline model (#3990)

Update

[ghstack-poisoned]

Author: vkuzo

benchmarks/float8/float8_inference_roofline.py

@@ -581,6 +581,7 @@ def run(
             M, K, N, torch.float8_e4m3fn, gemm_recipe_name
         )
     print("bf16_gemm_time_sympy", bf16_gemm_time_sympy)
+    print("bf16_ovhd_time_sympy", bf16_ovhd_time_sympy)
     print("fp8_gemm_time_sympy", fp8_gemm_time_sympy)
     print("fp8_ovhd_time_sympy", fp8_ovhd_time_sympy)
     print()

docs/source/workflows/inference.md

@@ -121,10 +121,10 @@ torch version           2.12.0.dev20260218+cu130
 torchao version         0.17.0+git3075bb624
 ...
    fwd_M  fwd_K  fwd_N  r_fp8_gemm_and_ovhd_spdp  b_fp8_e2e_spdp
-0   1024   1024   1024                      0.64            0.94
-1   2048   2048   2048                      1.75            1.21
-2   4096   4096   4096                      1.90            1.45
-3   8192   8192   8192                      1.94            1.75
+0   1024   1024   1024                      1.00            0.93
+1   2048   2048   2048                      1.75            1.20
+2   4096   4096   4096                      1.90            1.46
+3   8192   8192   8192                      1.94            1.76
 4  16384  16384  16384                      1.97            1.77
 
 #
@@ -137,11 +137,11 @@ torch version           2.12.0.dev20260218+cu130
 torchao version         0.17.0+git3075bb624
 ...
    fwd_M  fwd_K  fwd_N  r_fp8_gemm_and_ovhd_spdp  b_fp8_e2e_spdp
-0   1024   1024   1024                      0.64            0.37
-1   2048   2048   2048                      2.39            0.74
+0   1024   1024   1024                      1.00            0.38
+1   2048   2048   2048                      2.39            0.73
 2   4096   4096   4096                      2.92            1.19
-3   8192   8192   8192                      3.34            1.78
-4  16384  16384  16384                      3.63            2.57
+3   8192   8192   8192                      3.34            1.80
+4  16384  16384  16384                      3.63            2.56
 ```
 
 ## Other Available Quantization Techniques

torchao/testing/training/roofline_utils.py

@@ -561,6 +561,7 @@ def get_inference_bf16_activation_mem_sympy(M, K, N, gpu_name: Optional[str] = N
     kernel_rw = BYTES_PER_EL_BF16 * M * K * 2
     # convert from bytes to seconds
     res_s = kernel_rw / specs["peak_mem_bw_bytes_sec"] / specs["pct_achievable_mem_bw"]
+    res_s = sympy.Max(res_s, KERNEL_LAUNCH_OVERHEAD_SEC)
     return res_s