Handle enable CUDA graph param in SD example (microsoft#246)

This PR updates how the enable_cuda_graph param is set depending on the world_size i.e. CUDA graphs should only be enabled when world_size==1.

inference/huggingface/stable-diffusion/README.md

@@ -11,6 +11,8 @@ pip install -r requirements.txt
 Examples can be run as follows:
 <pre>deepspeed --num_gpus [number of GPUs] test-[model].py</pre>
 
+NOTE: Local CUDA graphs for replaced SD modules will only be enabled when `mp_size==1`.
+
 # Example Output
 Command:
 <pre>

inference/huggingface/stable-diffusion/test-stable-diffusion.py

@@ -9,7 +9,7 @@
 model = "prompthero/midjourney-v4-diffusion"
 local_rank = int(os.getenv("LOCAL_RANK", "0"))
 device = torch.device(f"cuda:{local_rank}")
-world_size = int(os.getenv('WORLD_SIZE', '4'))
+world_size = int(os.getenv('WORLD_SIZE', '1'))
 generator = torch.Generator(device=torch.cuda.current_device())
 
 pipe = DiffusionPipeline.from_pretrained(model, torch_dtype=torch.half)
@@ -19,12 +19,14 @@
 baseline_image = pipe(prompt, guidance_scale=7.5, generator=generator).images[0]
 baseline_image.save(f"baseline.png")
 
-# NOTE: DeepSpeed inference supports local CUDA graphs for replaced SD modules
+# NOTE: DeepSpeed inference supports local CUDA graphs for replaced SD modules.
+#       Local CUDA graphs for replaced SD modules will only be enabled when `mp_size==1`
 pipe = deepspeed.init_inference(
     pipe,
+    mp_size=world_size,
     dtype=torch.half,
     replace_with_kernel_inject=True,
-    enable_cuda_graph=True,
+    enable_cuda_graph=True if world_size==1 else False,
 )
 
 generator.manual_seed(0xABEDABE7)