updating gptj-config (#109)

* updating gptj-config * added distributed config logging to wandb * update * black fix * adding check for ds_plugin * removing wandb entity name from default config Co-authored-by: Louis Castricato <[email protected]>
CarperAI · Dec 4, 2022 · b229288 · b229288
1 parent b60f05e
commit b229288
Show file tree

Hide file tree

Showing 4 changed files with 36 additions and 5 deletions.
diff --git a/configs/ppo_gptj.yml b/configs/ppo_gptj.yml
@@ -35,9 +35,9 @@ method:
   cliprange: 0.2  # clip range
   cliprange_value: 0.2  # clip range
   vf_coef: 0.2  # value term weight
-  scale_reward: False  # False | "ref" | "running" estimate against which to scale rewards
+  scale_reward: False # False | "ref" | "running" estimate against which to scale rewards
   ref_mean: null
-  ref_std: null  # rescale rewards with this deviation
+  ref_std: null # rescale rewards with this deviation
   cliprange_reward: 10
   gen_kwargs:
     max_length: 48  # LM max sample gen length

diff --git a/examples/ppo_sentiments.py b/examples/ppo_sentiments.py
@@ -46,7 +46,6 @@ def reward_fn(samples: List[str]) -> List[float]:
     prompts = [" ".join(review.split()[:4]) for review in imdb["text"]]
 
     model = trlx.train(
-        "lvwerra/gpt2-imdb",
         reward_fn=reward_fn,
         prompts=prompts,
         eval_prompts=["I don't know much about Hungarian underground"] * 64,

diff --git a/trlx/model/accelerate_base_model.py b/trlx/model/accelerate_base_model.py
@@ -23,7 +23,7 @@
 import ray
 from ray.air import session
 from ray.air.checkpoint import Checkpoint
-from trlx.utils import filter_non_scalars, get_git_tag
+from trlx.utils import filter_non_scalars, get_distributed_config, get_git_tag
 
 
 @register_model
@@ -76,9 +76,12 @@ def __init__(self, config, train_mode=True):
         run_name = f"{script_name}/{model_name}"
 
         if self.accelerator.is_main_process and not ray.is_initialized():
+            config_dict = self.config.to_dict()
+            dist_config = get_distributed_config(self.accelerator)
+            config_dict["distributed"] = dist_config
             self.accelerator.init_trackers(
                 project_name=self.config.train.project_name,
-                config=self.config.to_dict(),
+                config=config_dict,
                 init_kwargs={
                     "wandb": {
                         "name": run_name,

diff --git a/trlx/utils/__init__.py b/trlx/utils/__init__.py
@@ -11,6 +11,9 @@
 from torch.optim.lr_scheduler import ChainedScheduler, LinearLR
 from torchtyping import TensorType
 
+import accelerate
+from accelerate import Accelerator
+
 
 def set_seed(seed: int):
     """
@@ -57,6 +60,32 @@ def safe_mkdir(path: str):
     os.mkdir(path)
 
 
+def get_distributed_config(accelerator: Accelerator):
+    """
+    Return accelerator distributed config
+    """
+
+    accelerate_config = accelerator.state
+    dist_config = {
+        "mixed_precision": accelerate_config.mixed_precision,
+        "num_gpus": accelerate_config.num_processes,
+    }
+
+    if hasattr(accelerator.state, "deepspeed_plugin"):
+        ds_plugin = accelerator.state.deepspeed_plugin
+        dist_config.upate(
+            {
+                "gradient_accumulation_steps": ds_plugin.gradient_accumulation_steps,
+                "gradient_clipping": ds_plugin.gradient_clipping,
+                "zero_stage": ds_plugin.zero_stage,
+                "offload_optimizer_device": ds_plugin.offload_optimizer_device,
+                "offload_param_device": ds_plugin.offload_param_device,
+            }
+        )
+
+    return dist_config
+
+
 # Stats