Update cudagraphs.md

Signed-off-by: JimmyZhang12 <67203904+JimmyZhang12@users.noreply.github.com>

docs/source/performance/cudagraphs.md

@@ -8,20 +8,19 @@ Currently, cudagraphs increases the memory usage of activations by ~20%. End to
 
 ### When to use
 Cudagraphs are recommended to be used when host overheads are expected to be significant, for instance systems with weak single-threaded CPU performance.
-As a demonstration we show that with a GPT3 20B model, cudagraphs can improve performance by 11.4%. The increase in memory usage due to cudagraphs was observed to be 15GB.
+As a demonstration we show that with a GPT3 20B model, cudagraphs can improve performance by 11.4%. The increase in memory usage due to cudagraphs was observed to be 15GB per gpu.
 | Setting         | TFLOP/s |
 | -----         | ------ |
 | No Cudagraphs      | 750     |
 | Cudagraphs     | 836     |
 - Setup: 16x GH200, MBS=2, TP=4, FP8 precision
 
-
-
-
 ### Usage
 As of the NeMo 24.09 release, Cudagraphs is currently supported only for the pretraining of dense models. 
 To enable please add the following configs:
-`model.enable_cuda_graph=True `
-`model.use_te_rng_tracker=True `
+`model.enable_cuda_graph=True`
+
+`model.use_te_rng_tracker=True`
+
 `model.get_attention_mask_from_fusion=True`