3 files changed, 168 insertions, 0 deletions
diff --git a/configs/deepspeed_zero2.json b/configs/deepspeed_zero2.json
new file mode 100644
index 0000000..bb7f7aa
--- /dev/null
+++ b/configs/deepspeed_zero2.json
@@ -0,0 +1,31 @@
+{
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "gradient_accumulation_steps": "auto",
+
+  "zero_optimization": {
+    "stage": 2,
+    "offload_optimizer": {
+      "device": "none"
+    },
+    "contiguous_gradients": true,
+    "overlap_comm": true,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 5e8,
+    "allgather_bucket_size": 5e8
+  },
+
+  "bf16": {
+    "enabled": false
+  },
+
+  "fp16": {
+    "enabled": false
+  },
+
+  "gradient_clipping": 1.0,
+
+  "zero_allow_untested_optimizer": true,
+
+  "wall_clock_breakdown": false
+}
diff --git a/configs/deepspeed_zero3.json b/configs/deepspeed_zero3.json
new file mode 100644
index 0000000..6e68c8f
--- /dev/null
+++ b/configs/deepspeed_zero3.json
@@ -0,0 +1,38 @@
+{
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "gradient_accumulation_steps": "auto",
+
+  "zero_optimization": {
+    "stage": 3,
+    "offload_optimizer": {
+      "device": "none"
+    },
+    "offload_param": {
+      "device": "none"
+    },
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 1e9,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 1e9,
+    "stage3_max_reuse_distance": 1e9,
+    "stage3_gather_16bit_weights_on_model_save": true
+  },
+
+  "bf16": {
+    "enabled": false
+  },
+
+  "fp16": {
+    "enabled": false
+  },
+
+  "gradient_clipping": 1.0,
+
+  "zero_allow_untested_optimizer": true,
+
+  "wall_clock_breakdown": false
+}
diff --git a/configs/eval_tasks_config.json b/configs/eval_tasks_config.json
new file mode 100644
index 0000000..e0dda43
--- /dev/null
+++ b/configs/eval_tasks_config.json
@@ -0,0 +1,99 @@
+[
+  {
+    "name": "dm_val",
+    "task_type": "math",
+    "dataset_path": "./data/dm_val.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": -1,
+    "max_gen_len": 2048,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "aime24",
+    "task_type": "math",
+    "dataset_path": "./data/aime24.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": -1,
+    "max_gen_len": 4096,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "aime25",
+    "task_type": "math",
+    "dataset_path": "./data/aime25.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": -1,
+    "max_gen_len": 4096,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "amc23",
+    "task_type": "math",
+    "dataset_path": "./data/amc23.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": -1,
+    "max_gen_len": 2048,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "math500",
+    "task_type": "math",
+    "dataset_path": "./data/math500.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": 500,
+    "max_gen_len": 2048,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "gsm8k",
+    "task_type": "math",
+    "dataset_path": "./data/gsm8k.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": 500,
+    "max_gen_len": 1024,
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "mmlu_stem",
+    "task_type": "qa",
+    "dataset_path": "./data/mmlu_stem.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": 500,
+    "max_gen_len": 512,
+    "temperature": 0.3,
+    "top_p": 0.9,
+    "num_samples_per_prompt": 1
+  },
+  {
+    "name": "humaneval",
+    "task_type": "code",
+    "dataset_path": "./data/humaneval.json",
+    "is_verifiable": true,
+    "metric_type": "accuracy",
+    "num_samples": 164,
+    "max_gen_len": 1024,
+    "temperature": 0.2,
+    "top_p": 0.95,
+    "num_samples_per_prompt": 1
+  }
+]
+