Address PR reviews: rename directories and update launcher commands

Author: joeywan-google

…de-BF16-SEQ4096-GBS256/recipe/Chart.yaml …Q4096-GBS256-NEMO25.11/recipe/Chart.yamltraining/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256/recipe/Chart.yaml renamed to training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256-NEMO25.11/recipe/Chart.yaml training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256/recipe/Chart.yaml renamed to training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256-NEMO25.11/recipe/Chart.yaml

@@ -75,7 +75,7 @@ Clone the `gpu-recipes` repository and set a reference to the recipe folder.
 git clone https://github.com/ai-hypercomputer/gpu-recipes.git
 cd gpu-recipes
 export REPO_ROOT=`git rev-parse --show-toplevel`
-export RECIPE_ROOT=$REPO_ROOT/training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256/recipe
+export RECIPE_ROOT=$REPO_ROOT/training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256-NEMO25.11/recipe
 cd $RECIPE_ROOT
 ```
 

…ode-BF16-SEQ4096-GBS256/recipe/README.md …EQ4096-GBS256-NEMO25.11/recipe/README.mdtraining/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256/recipe/README.md renamed to training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256-NEMO25.11/recipe/README.md training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256/recipe/README.md renamed to training/a4/deepseek_v3/megatron-bridge-pretraining-gke/32node-BF16-SEQ4096-GBS256-NEMO25.11/recipe/README.md

@@ -112,7 +112,14 @@ worker_command=$(cat <<- EOM
     --model_recipe_name deepseek_v3 \
     --gpus_per_node 8 \
     --num_gpus 256 \
+    --global_batch_size 2048 \
+    --micro_batch_size 1 \
     --seq_length 4096 \
+    --tensor_model_parallel_size 1 \
+    --pipeline_model_parallel_size 16 \
+    --context_parallel_size 1 \
+    --virtual_pipeline_model_parallel_size None \
+    --expert_model_parallel_size 8 \
     --compute_dtype bf16 \
     --max_steps 30
 EOM