feat: consolidate the step 5 (checkpoint conversion back to HF)

Author: YixuanWang-99

tests/end_to_end/tpu/gemma3/4b/test_gemma3.sh

@@ -6,7 +6,6 @@
 # 1. Run inference on the pre-converted checkpoint.
 # 2. Run pre-training starting from the pre-converted checkpoint.
 # 3. Run inference on the checkpoint produced by the pre-training run.
-# 4. Convert the checkpoint produced by the pre-training run back to HuggingFace format.
 
 # Usage:
 # export HF_TOKEN=<your Hugging Face access token>
@@ -85,10 +84,4 @@ else
     scan_layers=false prompt='I love to' attention=\'dot_product\'
 fi
 
-# Step 5: Convert the checkpoint from MaxText format to Hugging Face format
-python3 -m maxtext.checkpoint_conversion.to_huggingface \
-    model_name=${MODEL_NAME} tokenizer_type="huggingface" \
-    load_parameters_path=${BASE_OUTPUT_DIRECTORY}/train/${run_id}/checkpoints/4/items \
-    base_output_directory=${BASE_OUTPUT_DIRECTORY}/to_huggingface/unscanned/${run_id} \
-    use_multimodal=${USE_MULTIMODAL} \
-    scan_layers=false
+

tests/end_to_end/tpu/gemma3/4b/test_gemma3_multimodal_sft.sh

@@ -6,7 +6,6 @@
 # 1. Run inference on the pre-converted checkpoint.
 # 2. Run SFT of Gemma3-4B on ChartQA dataset with the converted checkpoint.
 # 3. Run inference on the checkpoint produced by the SFT run.
-# 4. Convert the checkpoint produced by the SFT run back to HuggingFace format.
 
 # Usage:
 # export HF_TOKEN=<your Hugging Face access token>
@@ -79,9 +78,3 @@ python3 -m maxtext.inference.decode \
     image_path=\'tests/assets/test_image.jpg\' \
     attention=\'dot_product\'
 
-# Step 5: Convert the SFT checkpoint back to HuggingFace format
-python3 -m maxtext.checkpoint_conversion.to_huggingface \
-    model_name=${MODEL_NAME} \
-    load_parameters_path=${BASE_OUTPUT_DIRECTORY}/multimodal/sft/${run_id}/checkpoints/4/items \
-    base_output_directory=${BASE_OUTPUT_DIRECTORY}/to_huggingface/unscanned/${run_id} \
-    use_multimodal=true scan_layers=false

tests/end_to_end/tpu/gemma3/4b/test_gemma3_rl.sh

@@ -6,7 +6,6 @@
 # 1. Run inference on the pre-converted checkpoint.
 # 2. Run RL starting from the pre-converted checkpoint.
 # 3. Run inference on the checkpoint produced by the RL run.
-# 4. Convert the checkpoint produced by the RL run back to HuggingFace format.
 
 # Usage:
 # export HF_TOKEN=<your Hugging Face access token>
@@ -58,10 +57,4 @@ python3 -m maxtext.inference.vllm_decode \
     hbm_utilization_vllm=0.5 \
     prompt='Suggest some famous landmarks in London.' \
     use_chat_template=True scan_layers=true
-
-# Step 5: Convert the checkpoint from MaxText format to Hugging Face format
-python3 -m maxtext.checkpoint_conversion.to_huggingface \
-    model_name=${MODEL_NAME} \
-    load_parameters_path=${BASE_OUTPUT_DIRECTORY}/rl/${run_id}/checkpoints/actor/5/model_params \
-    base_output_directory=${BASE_OUTPUT_DIRECTORY}/to_huggingface/unscanned/${run_id} \
-    use_multimodal=false scan_layers=true
+    

tests/end_to_end/tpu/gemma3/4b/test_gemma3_sft.sh

@@ -6,7 +6,6 @@
 # 1. Run inference on the pre-converted checkpoint.
 # 2. Run SFT starting from the pre-converted checkpoint.
 # 3. Run inference on the checkpoint produced by the SFT run.
-# 4. Convert the checkpoint produced by the SFT run back to HuggingFace format.
 
 # Usage:
 # export HF_TOKEN=<your Hugging Face access token>
@@ -55,9 +54,4 @@ python3 -m maxtext.inference.vllm_decode \
     prompt="Suggest some famous landmarks in London." \
     use_chat_template=True scan_layers=true
 
-# Step 5: Convert the checkpoint from MaxText format to Hugging Face format
-python3 -m maxtext.checkpoint_conversion.to_huggingface \
-    model_name=${MODEL_NAME} \
-    load_parameters_path=${BASE_OUTPUT_DIRECTORY}/sft/${run_id}/checkpoints/5/model_params \
-    base_output_directory=${BASE_OUTPUT_DIRECTORY}/to_huggingface/unscanned/${run_id} \
-    use_multimodal=false scan_layers=true
+

tests/end_to_end/tpu/gemma3/4b/test_gemma3_to_hf.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+# Converts a MaxText checkpoint to a Hugging Face model checkpoint.
+
+# Usage:
+# export RUN_ID=$(date +%Y-%m-%d-%H-%M-%S)
+# bash test_gemma3_to_hf.sh $RUN_ID $CHECKPOINT_PATH $USE_MULTIMODAL $SCAN_LAYERS
+
+set -ex
+
+run_id=$1
+CKPT_PATH=$2
+USE_MULTIMODAL=${3:-false}
+SCAN_LAYERS=${4:-false}
+
+MODEL_NAME='gemma3-4b'
+BASE_OUTPUT_DIRECTORY="gs://runner-maxtext-logs/${MODEL_NAME}"
+
+if [ "${SCAN_LAYERS,,}" = "true" ]; then
+    scan_status="scanned"
+else
+    scan_status="unscanned"
+fi
+
+python3 -m maxtext.checkpoint_conversion.to_huggingface \
+    model_name=${MODEL_NAME} \
+    tokenizer_type="huggingface" \
+    load_parameters_path=${CKPT_PATH} \
+    base_output_directory=${BASE_OUTPUT_DIRECTORY}/to_huggingface/${scan_status}/${run_id} \
+    use_multimodal=${USE_MULTIMODAL} \
+    scan_layers=$SCAN_LAYERS