diff --git a/PROJECT_ISSUES.md b/PROJECT_ISSUES.md
new file mode 100644
index 0000000..657a8eb
--- /dev/null
+++ b/PROJECT_ISSUES.md
@@ -0,0 +1,178 @@
+# CodeGeeX Project Issues List
+
+This document lists all identified issues in the CodeGeeX project.
+
+## 1. Hardcoded Paths (High Priority) ✅ FIXED
+
+### MindSpore Scripts
+**Status: All hardcoded paths have been replaced with configurable options**
+
+All hardcoded paths have been fixed:
+
+- **`codegeex/mindspore/generation_values.py`** ✅
+  - Now uses `--output_path` argument (defaults to `./output`)
+  - Output file: `output_values.npy`
+
+- **`codegeex/mindspore/generation_humaneval.py`** ✅
+  - Now uses `--input_path` for dataset (with smart fallback to repository-relative paths)
+  - Now uses `--output_path` for save directory (defaults to `./output`)
+  - Language parameter is now properly used (no longer hardcoded to C++)
+
+- **`codegeex/mindspore/generation_finetune.py`** ✅
+  - Now uses `--output_path` argument (defaults to `./output`)
+
+- **`codegeex/mindspore/generation_batch.py`** ✅
+  - Now uses `--output_path` argument (defaults to `./output`)
+
+- **`codegeex/mindspore/src/dataset.py`** ✅
+  - Now uses `eval_data_url` argument (if provided)
+
+- **`codegeex/mindspore/train.py`** ✅
+  - Cache paths now use `CODEGEEX_CACHE_BASE` environment variable (defaults to `/home/work/sfs/cache`)
+  - `BATCH_JOB_ID` now uses `.get()` with fallback
+
+- **`codegeex/mindspore/scripts/run_modelarts*.py`** ✅
+  - Temp directory now uses `MODELARTS_TEMP_DIR` environment variable (defaults to `/home/work/sfs/xx`)
+  - Added file existence checks before copying
+
+**New command-line arguments added:**
+- `--output_path`: Output directory for generated files (default: `./output`)
+- `--input_path`: Input path for data files (optional)
+
+**New environment variables:**
+- `CODEGEEX_CACHE_BASE`: Base directory for cache files (default: `/home/work/sfs/cache`)
+- `MODELARTS_TEMP_DIR`: Temp directory for ModelArts scripts (default: `/home/work/sfs/xx`)
+
+## 2. Configuration Placeholders (Medium Priority)
+
+Multiple configuration files contain placeholder values that must be set:
+
+- **`configs/codegeex_13b.sh`**: `CHECKPOINT_PATH` placeholder
+- **`configs/codegeex_13b_parallel.sh`**: `CHECKPOINT_PATH` placeholder  
+- **`configs/codegeex_13b_paddle.sh`**: `CHECKPOINT_PATH` placeholder
+- **`scripts/pretrain_codegeex.sh`**: 
+  - `HOSTFILE` placeholder
+  - `DATA_PATH` placeholder
+  - `CKPT_PATH` placeholder
+  - `OUTPUT_DIR` placeholder
+- **`scripts/finetune_codegeex.sh`**: Same placeholders as pretrain script
+- **`codegeex/mindspore/configs/*.sh`**: Multiple config files with `CODE_DATA_DIR` and `<TODO>` placeholders
+
+## 3. TODO Comments / Incomplete Code (Medium Priority)
+
+Multiple TODO comments indicate incomplete work:
+
+- **`codegeex/mindspore/train.py`** (lines 214, 216):
+  - TODO: remove after warming-up!
+  - TODO: add them back if not for the 1st run!
+
+- **`codegeex/mindspore/src/sat_dataset.py`** (line 81):
+  - TODO ARGS comment
+
+- **`codegeex/mindspore/src/dataset.py`** (line 122):
+  - TODO: set as current validation set path
+
+- **`codegeex/mindspore/generation_values_1p.py`** (line 166):
+  - TODO: add them back if not for the 1st run!
+
+- **`codegeex/mindspore/finetune.py`** (lines 216, 218):
+  - TODO: remove after warming-up!
+  - TODO: add them back if not for the 1st run!
+
+- **`codegeex/mindspore/generation_1p.py`** (line 166):
+  - TODO: add them back if not for the 1st run!
+
+- **`codegeex/mindspore/convertion_1p.py`** (lines 154, 160, 180):
+  - Multiple TODOs for checkpoint names and paths
+
+- All generation scripts have TODO comments for setting paths
+
+## 4. Security Issues (High Priority)
+
+- **`codegeex/benchmark/execution.py`** (line 347):
+  - Java execution code is commented out with security warning
+  - Warning states: "This program exists to execute untrusted model-generated code"
+  - Code execution should be sandboxed
+  - Currently the `exec_result` is None but code tries to access `.returncode` which will cause AttributeError
+
+- **`codegeex/benchmark/execution.py`** (lines 477-546):
+  - `reliability_guard()` function has explicit warning: "This function is NOT a security sandbox"
+  - Users should not blindly execute untrusted code
+
+## 5. Known Bugs (Documented)
+
+- **VS Code Extension** (mentioned in README):
+  - Bug: If cursor is moved before generation finishes, it may cause issues
+  - Location: `vscode-extension/README.md` and `README_zh.md`
+  - Status: Acknowledged, team working on making generation faster
+
+## 6. Debug/Test Code Left in Production (Low Priority)
+
+- **`scripts/evaluate_humaneval_x.py`** (lines 47-50):
+  - Hardcoded debug values left in code:
+    ```python
+    #Debugging
+    INPUT_FILE='/home/rog0d/Escritorio/CodeGeeX/generations/humaneval_rust_generations.jsonl.gz'
+    LANGUAGE='rust'
+    ```
+  - These override command-line arguments
+
+## 7. Incomplete Implementation (Medium Priority)
+
+- **`tests/test_inference_paddle.py`** (line 149):
+  - `raise NotImplementedError("quantize")` - quantization not implemented for Paddle backend
+
+## 8. Path Construction Bug (Low Priority)
+
+- **`scripts/evaluate_humaneval_x.py`** (line 44):
+  - Incorrect path join: `os.path.join(MAIN_DIR, "/codegeex/benchmark/humaneval-x/")`
+  - Leading slash makes it an absolute path, ignoring MAIN_DIR
+
+## 9. Deprecated/Outdated Information
+
+- **`README.md`** (line 12):
+  - Notes that CodeGeeX4 is newer and released
+  - Current codebase may be considered legacy version
+
+## 10. Missing Error Handling
+
+- **`codegeex/benchmark/execution.py`** (line 348):
+  - Code accesses `exec_result.returncode` when `exec_result` is `None` (line 336)
+  - Will cause `AttributeError` - Java execution path is broken
+
+## 11. Hardcoded CUDA Path (Low Priority)
+
+- **`scripts/generate_humaneval_x.sh`** (line 13):
+  - `export CUDA_HOME=/usr/local/cuda-11.1/` - hardcoded CUDA version
+- **`scripts/translate_humaneval_x.sh`** (line 14):
+  - Same hardcoded CUDA path
+
+## 12. Configuration Dependencies
+
+- Scripts require specific environment variables:
+  - `BATCH_JOB_ID` (used in train.py and scripts)
+  - Various NCCL environment variables
+  - Platform-specific paths for Ascend/MindSpore
+
+## Summary by Priority
+
+### Critical (Must Fix Before Production Use)
+1. ~~Hardcoded paths in generation scripts~~ ✅ **FIXED**
+2. Security issue: Java execution code broken (None.returncode error)
+3. Configuration placeholders not set
+
+### High Priority
+4. Security warnings for code execution
+5. Debug code left in evaluate script
+6. Known VS Code extension cursor bug
+
+### Medium Priority
+7. Multiple TODO comments indicating incomplete work
+8. Hardcoded CUDA paths
+9. Path construction bug
+
+### Low Priority
+10. Missing quantization implementation for Paddle
+11. Deprecated version notice (CodeGeeX4 available)
+12. Platform-specific hardcoded paths
+
diff --git a/codegeex/mindspore/generation_batch.py b/codegeex/mindspore/generation_batch.py
index 06798fc..dccd5a6 100644
--- a/codegeex/mindspore/generation_batch.py
+++ b/codegeex/mindspore/generation_batch.py
@@ -226,15 +226,19 @@ def run_predict(model_predict, config, args_opt, rank):
     generations = []
     batch_size = config.batch_size
     verbose = (rank % 8 == 0)
-    save_path = f'/home/work/sfs/xx/pangu_alpha_code/generation_batch/{args_opt.temperature}.txt'  # TODO: set as current save path
-    save_dir = os.path.split(save_path)[0]
+    
+    # Use configurable output path
+    output_dir = getattr(args_opt, 'output_path', './output')
+    save_dir = os.path.join(output_dir, 'generation_batch')
+    save_path = os.path.join(save_dir, f'temp_{args_opt.temperature}.txt')
+    
     if rank == 0:
-        if not os.path.exists(save_dir):
-            os.makedirs(save_dir)
+        os.makedirs(save_dir, exist_ok=True)
         if not os.path.exists(save_path):
-            f = open(save_path, 'w')
-            f.close()
-        os.system(f'sudo chmod 777 -R {save_dir}')
+            with open(save_path, 'w') as f:
+                pass  # Create empty file
+        if os.name != 'nt':  # Only on Unix-like systems
+            os.system(f'chmod 777 -R {save_dir}')
     batch = []
     input_length = []
     sample_ids = []
diff --git a/codegeex/mindspore/generation_finetune.py b/codegeex/mindspore/generation_finetune.py
index 9f2a96e..360e515 100644
--- a/codegeex/mindspore/generation_finetune.py
+++ b/codegeex/mindspore/generation_finetune.py
@@ -210,11 +210,14 @@ def run_predict(model_predict, config, args_opt, rank):
     generations = []
     batch_size = config.batch_size
     verbose = (rank % 8 == 0)
-    save_path = f'/home/work/sfs/xx/pangu_alpha_code/generation_finetune/code_translation/{lang}/temp_{args_opt.temperature}.txt'  # TODO: set as current save path
-    save_dir = os.path.split(save_path)[0]
+    
+    # Use configurable output path
+    output_dir = getattr(args_opt, 'output_path', './output')
+    save_dir = os.path.join(output_dir, 'generation_finetune', 'code_translation', lang)
+    save_path = os.path.join(save_dir, f'temp_{args_opt.temperature}.txt')
+    
     if rank == 0:
-        if not os.path.exists(save_dir):
-            os.makedirs(save_dir)
+        os.makedirs(save_dir, exist_ok=True)
         if not os.path.exists(save_path):
             f = open(save_path, 'w')
             f.close()
diff --git a/codegeex/mindspore/generation_humaneval.py b/codegeex/mindspore/generation_humaneval.py
index 030ff47..ecba456 100644
--- a/codegeex/mindspore/generation_humaneval.py
+++ b/codegeex/mindspore/generation_humaneval.py
@@ -15,6 +15,7 @@
 """
 PanGu predict run
 """
+import gzip
 import json
 import os
 import time
@@ -198,25 +199,74 @@ def run_predict(model_predict, config, args_opt, rank):
     # Define tokenizer
     tokenizer = CodeTokenizer(mode='6b')
 
-    # Tokenize input sentence to ids
-    humaneval_path = '/home/work/sfs/xx/human_eval_x/data/humaneval_cpp.jsonl'  # TODO: set as current humaneval path
-    humaneval = open(humaneval_path, 'r').readlines()
-    humaneval = [json.loads(task) for task in humaneval if len(task) != 0]
+    # Determine language (default to cpp for backward compatibility)
+    lang = getattr(args_opt, 'language', 'cpp') or 'cpp'
+    lang_lower = lang.lower()
+    
+    # Language tag mapping
+    lang_tags = {
+        'cpp': '// language: C++\n',
+        'c++': '// language: C++\n',
+        'python': '# language: Python\n',
+        'java': '// language: Java\n',
+        'javascript': '// language: JavaScript\n',
+        'js': '// language: JavaScript\n',
+        'go': '// language: Go\n',
+    }
+    tag = lang_tags.get(lang_lower, f'// language: {lang}\n')
+
+    # Determine input path
+    if hasattr(args_opt, 'input_path') and args_opt.input_path:
+        humaneval_path = args_opt.input_path
+    else:
+        # Try relative path from current script location
+        script_dir = os.path.dirname(os.path.abspath(__file__))
+        repo_root = os.path.dirname(os.path.dirname(os.path.dirname(script_dir)))
+        default_path = os.path.join(repo_root, 'codegeex', 'benchmark', 'humaneval-x', 
+                                   lang_lower, 'data', f'humaneval_{lang_lower}.jsonl.gz')
+        # Check if .gz file exists, otherwise try .jsonl
+        if os.path.exists(default_path):
+            humaneval_path = default_path
+        else:
+            default_path_jsonl = default_path.replace('.jsonl.gz', '.jsonl')
+            if os.path.exists(default_path_jsonl):
+                humaneval_path = default_path_jsonl
+            else:
+                # Fallback: use input_path or raise error
+                humaneval_path = default_path_jsonl
+                if rank == 0:
+                    print(f"Warning: Default path {humaneval_path} does not exist. Please set --input_path")
+    
+    # Open file (handle .gz files)
+    if humaneval_path.endswith('.gz'):
+        with gzip.open(humaneval_path, 'rt') as f:
+            humaneval = [json.loads(line) for line in f if line.strip()]
+    else:
+        with open(humaneval_path, 'r') as f:
+            humaneval = [json.loads(line) for line in f if line.strip()]
+    
     samples = [task['prompt'] for task in humaneval]
     generations = []
     batch_size = config.batch_size
     verbose = (rank % 8 == 0)
-    part = int(args_opt.part)
+    part = int(args_opt.part) if args_opt.part else 0
     gen_times = 12 # TODO: set as generation times of current task
     print(f"gen times: {gen_times}, part: {part}")
-    save_path = f'/home/work/sfs/xx/pangu_alpha_code/generation_humanevalx/cpp/temp_{args_opt.temperature}/samples_{args_opt.load_ckpt_epoch}_part_{part}.jsonl'  # TODO: set as current save path
+    
+    # Determine output path
+    output_dir = getattr(args_opt, 'output_path', './output')
+    os.makedirs(output_dir, exist_ok=True)
+    save_path = os.path.join(output_dir, 
+                            f'humaneval_{lang_lower}_temp_{args_opt.temperature}_samples_{args_opt.load_ckpt_epoch}_part_{part}.jsonl')
+    
     if rank == 0 and not os.path.exists(save_path):
         os.makedirs(os.path.split(save_path)[0], exist_ok=True)
-        f = open(save_path, 'w')
-        f.close()
-        os.system(f'sudo chmod 777 {save_path}')
+        with open(save_path, 'w') as f:
+            pass  # Create empty file
+        if os.name != 'nt':  # Only on Unix-like systems
+            os.system(f'chmod 777 {save_path}')
+    
     for i, sample in enumerate(samples):
-        tag = "// language: C++\n"
         sample = tag + sample
         if rank % 8 == 0:
             print(f"=================== prompt {i} ====================")
diff --git a/codegeex/mindspore/generation_values.py b/codegeex/mindspore/generation_values.py
index 7c81035..2209e07 100644
--- a/codegeex/mindspore/generation_values.py
+++ b/codegeex/mindspore/generation_values.py
@@ -199,9 +199,15 @@ def run_predict(model_predict, config, args_opt, rank):
                                           init, batch_valid_length)
     output = output_logits.asnumpy()
     if rank == 0:
-        np.save("/home/work/sfs/xx/pangu_alpha_code/output_6_7375_8.13.npy", output)  # TODO: set as current save path
-        os.system(
-            "chmod 777 /home/work/sfs/xx/pangu_alpha_code/output_6_7375_8.13.npy")  # TODO: set as current save path
+        # Use configurable output path
+        output_dir = getattr(args_opt, 'output_path', './output')
+        os.makedirs(output_dir, exist_ok=True)
+        output_file = os.path.join(output_dir, "output_values.npy")
+        np.save(output_file, output)
+        # Only try to chmod if on Unix-like system
+        if os.name != 'nt':
+            os.system(f"chmod 777 {output_file}")
+        print(f"== Output saved to: {output_file}")
         print("== Output shape: ", output.shape)
 
 
diff --git a/codegeex/mindspore/scripts/run_modelarts.py b/codegeex/mindspore/scripts/run_modelarts.py
index 05715a0..a7e25c5 100644
--- a/codegeex/mindspore/scripts/run_modelarts.py
+++ b/codegeex/mindspore/scripts/run_modelarts.py
@@ -21,11 +21,16 @@
 
 os.environ["LOG_PATH"] = tb_path
 
-print("=================RANK_TABLE_FILE: ", os.environ["RANK_TABLE_FILE"], flush=True)
+print("=================RANK_TABLE_FILE: ", os.environ.get("RANK_TABLE_FILE", "not set"), flush=True)
 print("=================ms import done", flush=True)
 time.sleep(10)
-os.system(
-    "cp /home/work/rank_table/jobstart_hccl.json /home/work/sfs/xx; sudo chmod +777 /home/work/rank_table/jobstart_hccl.json")
+# Use configurable temp directory (platform-specific for ModelArts)
+temp_dir = os.environ.get("MODELARTS_TEMP_DIR", "/home/work/sfs/xx")
+rank_table_source = "/home/work/rank_table/jobstart_hccl.json"
+if os.path.exists(rank_table_source):
+    os.system(f"cp {rank_table_source} {temp_dir}; sudo chmod +777 {rank_table_source}")
+else:
+    print(f"Warning: {rank_table_source} does not exist. Skipping copy.")
 ret = os.system(f"cd {log_path} && bash {args.script} 2>&1 | tee output.log")
 if os.environ.get("RANK_ID") == 0:
     log_dir = os.path.join(args.work_dir, "logs", os.environ.get("JOB_ID"))
diff --git a/codegeex/mindspore/scripts/run_modelarts_gen_finetune.py b/codegeex/mindspore/scripts/run_modelarts_gen_finetune.py
index 2e5f1ec..0945dd5 100644
--- a/codegeex/mindspore/scripts/run_modelarts_gen_finetune.py
+++ b/codegeex/mindspore/scripts/run_modelarts_gen_finetune.py
@@ -26,11 +26,16 @@
 else:
     os.environ["LANGUAGE"] = "Null"
 
-print("=================RANK_TABLE_FILE: ", os.environ["RANK_TABLE_FILE"], flush=True)
+print("=================RANK_TABLE_FILE: ", os.environ.get("RANK_TABLE_FILE", "not set"), flush=True)
 print("=================ms import done", flush=True)
 time.sleep(10)
-os.system(
-    "cp /home/work/rank_table/jobstart_hccl.json /home/work/sfs/xx; sudo chmod +777 /home/work/rank_table/jobstart_hccl.json")
+# Use configurable temp directory (platform-specific for ModelArts)
+temp_dir = os.environ.get("MODELARTS_TEMP_DIR", "/home/work/sfs/xx")
+rank_table_source = "/home/work/rank_table/jobstart_hccl.json"
+if os.path.exists(rank_table_source):
+    os.system(f"cp {rank_table_source} {temp_dir}; sudo chmod +777 {rank_table_source}")
+else:
+    print(f"Warning: {rank_table_source} does not exist. Skipping copy.")
 ret = os.system(f"cd {log_path} && bash {args.script} 2>&1 | tee output.log")
 if os.environ.get("RANK_ID") == 0:
     log_dir = os.path.join(args.work_dir, "logs", os.environ.get("JOB_ID"))
diff --git a/codegeex/mindspore/scripts/run_modelarts_gen_humaneval_x.py b/codegeex/mindspore/scripts/run_modelarts_gen_humaneval_x.py
index 256d7d1..6bae2d8 100644
--- a/codegeex/mindspore/scripts/run_modelarts_gen_humaneval_x.py
+++ b/codegeex/mindspore/scripts/run_modelarts_gen_humaneval_x.py
@@ -26,11 +26,16 @@
 else:
     os.environ["PART"] = "-1"
 
-print("=================RANK_TABLE_FILE: ", os.environ["RANK_TABLE_FILE"], flush=True)
+print("=================RANK_TABLE_FILE: ", os.environ.get("RANK_TABLE_FILE", "not set"), flush=True)
 print("=================ms import done", flush=True)
 time.sleep(10)
-os.system(
-    "cp /home/work/rank_table/jobstart_hccl.json /home/work/sfs/xx; sudo chmod +777 /home/work/rank_table/jobstart_hccl.json")
+# Use configurable temp directory (platform-specific for ModelArts)
+temp_dir = os.environ.get("MODELARTS_TEMP_DIR", "/home/work/sfs/xx")
+rank_table_source = "/home/work/rank_table/jobstart_hccl.json"
+if os.path.exists(rank_table_source):
+    os.system(f"cp {rank_table_source} {temp_dir}; sudo chmod +777 {rank_table_source}")
+else:
+    print(f"Warning: {rank_table_source} does not exist. Skipping copy.")
 ret = os.system(f"cd {log_path} && bash {args.script} 2>&1 | tee output.log")
 if os.environ.get("RANK_ID") == 0:
     log_dir = os.path.join(args.work_dir, "logs", os.environ.get("JOB_ID"))
diff --git a/codegeex/mindspore/src/dataset.py b/codegeex/mindspore/src/dataset.py
index 725ecea..45f5737 100644
--- a/codegeex/mindspore/src/dataset.py
+++ b/codegeex/mindspore/src/dataset.py
@@ -118,8 +118,14 @@ def create_dataset(batch_size, data_path, args_opt, device_num=1, rank=0, drop=T
     num_parallel_workers = 4
     train_data = get_code_data_train(data_path, args_opt, skip_num=(skip_num // num_parallel_workers))
     if train_and_eval:
-        val_data = get_code_data_eval("/home/work/sfs/xx/data_valid",
-                                      args_opt)  # TODO: set as current validation set path
+        # Use eval_data_url if provided, otherwise skip validation
+        val_data_path = getattr(args_opt, 'eval_data_url', None)
+        if val_data_path:
+            val_data = get_code_data_eval(val_data_path, args_opt)
+        else:
+            if rank == 0:
+                print("Warning: train_and_eval is True but eval_data_url is not set. Skipping validation data.")
+            val_data = None
     else:
         val_data = None
 
diff --git a/codegeex/mindspore/src/utils.py b/codegeex/mindspore/src/utils.py
index 767fc5d..6bb5ac4 100644
--- a/codegeex/mindspore/src/utils.py
+++ b/codegeex/mindspore/src/utils.py
@@ -310,6 +310,14 @@ def add_inference_params(opt):
                      default="true",
                      choices=["true", "false"],
                      help="Whether enable state reuse")
+    opt.add_argument("--output_path",
+                     type=str,
+                     default="./output",
+                     help="Output path for generated files")
+    opt.add_argument("--input_path",
+                     type=str,
+                     default=None,
+                     help="Input path for data files (e.g., HumanEval dataset)")
 
 
 def add_training_params(opt):
diff --git a/codegeex/mindspore/train.py b/codegeex/mindspore/train.py
index 9ccd461..e5853e6 100644
--- a/codegeex/mindspore/train.py
+++ b/codegeex/mindspore/train.py
@@ -219,9 +219,13 @@ def run_train(args_opt):
         #     args_opt.has_trained_steps = int(param_dict["step_num"].data.asnumpy())
         # args_opt.has_trained_steps = 9000
 
-        os.mkdir(f'/home/work/sfs/cache/{os.environ["BATCH_JOB_ID"]}/1/rank_{rank}')
+        # Use configurable cache directory (platform-specific for checkpoint synchronization)
+        cache_base = os.environ.get('CODEGEEX_CACHE_BASE', '/home/work/sfs/cache')
+        batch_job_id = os.environ.get('BATCH_JOB_ID', 'default_job')
+        cache_dir_1 = os.path.join(cache_base, batch_job_id, '1')
+        os.makedirs(os.path.join(cache_dir_1, f'rank_{rank}'), exist_ok=True)
         while True:
-            num = len(os.listdir(f'/home/work/sfs/cache/{os.environ["BATCH_JOB_ID"]}/1'))
+            num = len([d for d in os.listdir(cache_dir_1) if os.path.isdir(os.path.join(cache_dir_1, d)) and d.startswith('rank_')])
             if num == device_num:
                 break
             if rank % 64 == 0:
@@ -280,9 +284,13 @@ def run_train(args_opt):
         net_not_load = load_param_into_net(pangu_alpha_with_loss, param_dict)
         opt_not_load = load_param_into_net(optimizer, param_dict)
 
-        os.mkdir(f'/home/work/sfs/cache/{os.environ["BATCH_JOB_ID"]}/2/rank_{rank}')
+        # Use configurable cache directory (platform-specific for checkpoint synchronization)
+        cache_base = os.environ.get('CODEGEEX_CACHE_BASE', '/home/work/sfs/cache')
+        batch_job_id = os.environ.get('BATCH_JOB_ID', 'default_job')
+        cache_dir_2 = os.path.join(cache_base, batch_job_id, '2')
+        os.makedirs(os.path.join(cache_dir_2, f'rank_{rank}'), exist_ok=True)
         while True:
-            num = len(os.listdir(f'/home/work/sfs/cache/{os.environ["BATCH_JOB_ID"]}/2'))
+            num = len([d for d in os.listdir(cache_dir_2) if os.path.isdir(os.path.join(cache_dir_2, d)) and d.startswith('rank_')])
             if num == device_num:
                 break
             if rank % 64 == 0:
@@ -317,7 +325,7 @@ def run_train(args_opt):
                     sink_size=args_opt.sink_size, dataset_sink_mode=True)
     finally:
         if args_opt.profiling:
-            jobid = os.environ["BATCH_JOB_ID"]
+            jobid = os.environ.get("BATCH_JOB_ID", "default_job")
             profiler.analyse()
             rank_id = rank
             if context.get_context("save_graphs"):