镜像自地址
https://github.com/MINT-SJTU/Evo-RL.git
已同步 2026-11-20 18:17:06 +00:00
fix(evo1): support stage2 checkpoint resume
这个提交包含在:
@ -161,8 +161,25 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC): # type: igno
|
|||||||
return None
|
return None
|
||||||
|
|
||||||
def _save_pretrained(self, save_directory: Path) -> None:
|
def _save_pretrained(self, save_directory: Path) -> None:
|
||||||
with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
|
with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as tmp, draccus.config_type("json"):
|
||||||
draccus.dump(self, f, indent=4)
|
draccus.dump(self, tmp, indent=4)
|
||||||
|
tmp_path = tmp.name
|
||||||
|
|
||||||
|
with open(tmp_path) as f:
|
||||||
|
config = json.load(f)
|
||||||
|
|
||||||
|
if "type" not in config:
|
||||||
|
matching_choices = [
|
||||||
|
name for name in self.get_known_choices() if self.get_choice_class(name) is self.__class__
|
||||||
|
]
|
||||||
|
if len(matching_choices) != 1:
|
||||||
|
raise ValueError(
|
||||||
|
f"Expected exactly one registered choice for {self.__class__.__name__}, got {matching_choices}"
|
||||||
|
)
|
||||||
|
config["type"] = matching_choices[0]
|
||||||
|
|
||||||
|
with open(save_directory / CONFIG_NAME, "w") as f:
|
||||||
|
json.dump(config, f, indent=4)
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
def from_pretrained(
|
def from_pretrained(
|
||||||
@ -207,16 +224,34 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC): # type: igno
|
|||||||
# apply cli overrides.
|
# apply cli overrides.
|
||||||
# This is very ugly, ideally we'd like to be able to do that natively with draccus
|
# This is very ugly, ideally we'd like to be able to do that natively with draccus
|
||||||
# something like --policy.path (in addition to --policy.type)
|
# something like --policy.path (in addition to --policy.type)
|
||||||
with draccus.config_type("json"):
|
|
||||||
orig_config = draccus.parse(cls, config_file, args=[])
|
|
||||||
|
|
||||||
if config_file is None:
|
if config_file is None:
|
||||||
raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
|
raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
|
||||||
|
|
||||||
with open(config_file) as f:
|
with open(config_file) as f:
|
||||||
config = json.load(f)
|
config = json.load(f)
|
||||||
|
|
||||||
config.pop("type")
|
if "type" in config:
|
||||||
|
with draccus.config_type("json"):
|
||||||
|
orig_config = draccus.parse(cls, config_file, args=[])
|
||||||
|
else:
|
||||||
|
parse_successes = []
|
||||||
|
for choice_name in cls.get_known_choices():
|
||||||
|
config_cls = cls.get_choice_class(choice_name)
|
||||||
|
try:
|
||||||
|
with draccus.config_type("json"):
|
||||||
|
parsed_config = draccus.parse(config_cls, config_file, args=[])
|
||||||
|
except Exception:
|
||||||
|
continue
|
||||||
|
parse_successes.append((choice_name, parsed_config))
|
||||||
|
|
||||||
|
if len(parse_successes) != 1:
|
||||||
|
raise ValueError(
|
||||||
|
f"Expected to infer exactly one policy type from {config_file}, got "
|
||||||
|
f"{[choice for choice, _ in parse_successes]}"
|
||||||
|
)
|
||||||
|
orig_config = parse_successes[0][1]
|
||||||
|
|
||||||
|
config.pop("type", None)
|
||||||
with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
|
with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
|
||||||
json.dump(config, f)
|
json.dump(config, f)
|
||||||
config_file = f.name
|
config_file = f.name
|
||||||
|
|||||||
@ -410,6 +410,14 @@ def train(
|
|||||||
policy, optimizer, dataloader, lr_scheduler = accelerator.prepare(
|
policy, optimizer, dataloader, lr_scheduler = accelerator.prepare(
|
||||||
policy, optimizer, dataloader, lr_scheduler
|
policy, optimizer, dataloader, lr_scheduler
|
||||||
)
|
)
|
||||||
|
if (
|
||||||
|
accelerator.num_processes > 1
|
||||||
|
and cfg.policy.type == "evo1"
|
||||||
|
and getattr(cfg.policy, "training_stage", None) == "stage2"
|
||||||
|
and hasattr(policy, "_set_static_graph")
|
||||||
|
):
|
||||||
|
logging.info("Enabling DDP static graph for Evo1 stage2.")
|
||||||
|
policy._set_static_graph()
|
||||||
dl_iter = cycle(dataloader)
|
dl_iter = cycle(dataloader)
|
||||||
|
|
||||||
policy.train()
|
policy.train()
|
||||||
|
|||||||
在新工单中引用
屏蔽一个用户