Update on "compiled RMSNorm"

On Llama3 8B model, no AC
`compiled_rmsnorm` is ~9% faster than `rmsnorm`, but ~2% slower than `fused_rmsnorm`.
Please see below for details.

rmsnorm
<img width="757" alt="image" src="https://github.com/pytorch/torchtitan/assets/150487191/79645518-e38b-4ddb-b01d-b0c93ec27dd4">

compiled_rmsnorm
<img width="754" alt="image" src="https://github.com/pytorch/torchtitan/assets/150487191/c457b388-793f-452b-9bce-17bc1823df66">

fused_rmsnorm
<img width="753" alt="image" src="https://github.com/pytorch/torchtitan/assets/150487191/ea1db7ad-5887-4efa-9788-e708e4b40428">



[ghstack-poisoned]

estimation.py

@@ -57,8 +57,12 @@ def estimate_memory(job_config: JobConfig):
         )
         job_config.model.norm_type = "rmsnorm"
 
+    if job_config.model.norm_type == "compiled_rmsnorm":
+        logger.info("Compiled RMSNorm is not supported yet. Switching to RMSNorm.")
+        job_config.model.norm_type = "rmsnorm"
+
     if job_config.training.compile:
-        logger.info("Compile mode is not supported yet. " "Switching to Eager mode.")
+        logger.info("Compile mode is not supported yet. Switching to eager mode.")
         job_config.training.compile = False
 
     parallel_dims = ParallelDims(

test_runner.py

@@ -266,7 +266,7 @@ def build_test_list():
         OverrideDefinitions(
             [
                 [
-                    "--memory_estimation.enabled",
+                    "--memory_estimation.enabled --model.norm_type rmsnorm",
                 ]
             ],
             "FSDP2 Memory Tracking and Estimation",