[TRITON_KERNELS] tweak matmul_ogs heuristics (#7664)

Author: ptillet

python/triton_kernels/triton_kernels/matmul_ogs_details/opt_flags.py

@@ -157,7 +157,8 @@ def make_default_opt_flags_nvidia(
     elif enforce_bitwise_invariance:
         block_m = 128
     else:
-        block_m = max(16, min(triton.next_power_of_2(tokens_per_expt), 128))
+        min_block_m = 64 if torch.cuda.get_device_capability()[0] == 10 else 16
+        block_m = max(min_block_m, min(triton.next_power_of_2(tokens_per_expt), 128))
     # block n
     arch = None
     block_n = opt_flags_nvidia.compute_block_n(n, arch, precision_config)