镜像自地址
https://github.com/MINT-SJTU/Evo-RL.git
已同步 2026-11-20 18:17:06 +00:00
feat(evo1): merge stage2 training optimizations
这个提交包含在:
@ -33,6 +33,17 @@ Validated in this repo / recent experiments:
|
||||
- `training_stage=stage2` finetunes both the VLM and the action head.
|
||||
- In this repo, `resume_pretrain=true` is the Evo1-style stage handoff: load weights only, but start a fresh optimizer/scheduler state.
|
||||
- We currently recommend `dataset.image_transforms.enable=false` in the validated training commands because the image augmentation pipeline was not stable in the tested environment.
|
||||
- Stage2 now defaults to `embedder_tensor_fastpath=true`, `enable_vlm_gradient_checkpointing=true`, and
|
||||
`gradient_checkpointing_use_reentrant=false`.
|
||||
|
||||
## Optimization Summary
|
||||
|
||||
The current Evo1 integration includes a small set of stage2-oriented training optimizations merged from the Evo_2 branch while preserving the LeRobot/Evo-RL structure:
|
||||
|
||||
- Runtime config switches were added for `embedder_tensor_fastpath`, `enable_vlm_gradient_checkpointing`, and `gradient_checkpointing_use_reentrant`.
|
||||
- Evo1 image collection during training no longer forces a `GPU -> CPU -> GPU` round trip before the VLM embedder.
|
||||
- InternVL3 stage2 now enables a more memory-friendly checkpointing path and disables `use_cache` when checkpointing is active.
|
||||
- The main `lerobot_train` entrypoint was kept intact; only lightweight Evo1 stage2 runtime logging was added.
|
||||
|
||||
## Stage 1 Command
|
||||
|
||||
|
||||
@ -56,6 +56,9 @@ class Evo1Config(PreTrainedConfig):
|
||||
vlm_num_layers: int | None = 14
|
||||
vlm_dtype: str = "bfloat16"
|
||||
use_flash_attn: bool = True
|
||||
embedder_tensor_fastpath: bool = True
|
||||
enable_vlm_gradient_checkpointing: bool | None = None
|
||||
gradient_checkpointing_use_reentrant: bool = False
|
||||
action_head: str = "flowmatching"
|
||||
embed_dim: int = 896
|
||||
hidden_dim: int = 1024
|
||||
@ -97,6 +100,9 @@ class Evo1Config(PreTrainedConfig):
|
||||
self.finetune_vlm = True
|
||||
self.finetune_action_head = True
|
||||
|
||||
if self.enable_vlm_gradient_checkpointing is None:
|
||||
self.enable_vlm_gradient_checkpointing = self.training_stage == "stage2"
|
||||
|
||||
if self.n_action_steps > self.chunk_size:
|
||||
raise ValueError(
|
||||
f"n_action_steps ({self.n_action_steps}) must be <= chunk_size ({self.chunk_size})"
|
||||
|
||||
@ -22,6 +22,9 @@ class EVO1(nn.Module):
|
||||
num_language_layers=config.get("vlm_num_layers", 14),
|
||||
model_dtype=config.get("vlm_dtype", "bfloat16"),
|
||||
use_flash_attn=config.get("use_flash_attn", True),
|
||||
enable_gradient_checkpointing=config.get("enable_vlm_gradient_checkpointing", False),
|
||||
enable_tensor_fastpath=config.get("embedder_tensor_fastpath", True),
|
||||
gradient_checkpointing_use_reentrant=config.get("gradient_checkpointing_use_reentrant", False),
|
||||
)
|
||||
|
||||
action_head_type = config.get("action_head", "flowmatching").lower()
|
||||
|
||||
@ -88,12 +88,18 @@ class InternVL3Embedder(nn.Module):
|
||||
num_language_layers: int | None = 14,
|
||||
model_dtype: str | torch.dtype = "bfloat16",
|
||||
use_flash_attn: bool = True,
|
||||
enable_gradient_checkpointing: bool = False,
|
||||
enable_tensor_fastpath: bool = True,
|
||||
gradient_checkpointing_use_reentrant: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
self._requested_device = device
|
||||
self.image_size = image_size
|
||||
self.num_language_layers = num_language_layers
|
||||
self.max_text_length = 1024 # InternVL3 supports up to 1024 tokens
|
||||
self.enable_gradient_checkpointing = bool(enable_gradient_checkpointing)
|
||||
self.enable_tensor_fastpath = bool(enable_tensor_fastpath)
|
||||
self.gradient_checkpointing_use_reentrant = bool(gradient_checkpointing_use_reentrant)
|
||||
self.transform = build_transform(image_size)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, use_fast=False)
|
||||
if isinstance(model_dtype, str):
|
||||
@ -123,12 +129,53 @@ class InternVL3Embedder(nn.Module):
|
||||
else:
|
||||
self.model.language_model.layers = torch.nn.ModuleList(layers)
|
||||
self.model.language_model.lm_head = torch.nn.Identity()
|
||||
|
||||
if hasattr(self.model, "vision_model") and hasattr(self.model.vision_model, "encoder"):
|
||||
self.model.vision_model.encoder.gradient_checkpointing = False
|
||||
self._configure_memory_features()
|
||||
|
||||
self.img_context_token_id = self.tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
||||
|
||||
def _configure_memory_features(self) -> None:
|
||||
checkpoint_kwargs = {"use_reentrant": self.gradient_checkpointing_use_reentrant}
|
||||
|
||||
def _enable_ckpt(module) -> bool:
|
||||
if module is None or not hasattr(module, "gradient_checkpointing_enable"):
|
||||
return False
|
||||
module.gradient_checkpointing_enable(gradient_checkpointing_kwargs=checkpoint_kwargs)
|
||||
return True
|
||||
|
||||
if not self.enable_gradient_checkpointing:
|
||||
if hasattr(self.model, "vision_model") and hasattr(self.model.vision_model, "encoder"):
|
||||
self.model.vision_model.encoder.gradient_checkpointing = False
|
||||
return
|
||||
|
||||
enabled_any = False
|
||||
enabled_any = _enable_ckpt(self.model) or enabled_any
|
||||
|
||||
if hasattr(self.model, "vision_model") and hasattr(self.model.vision_model, "encoder"):
|
||||
self.model.vision_model.encoder.gradient_checkpointing = True
|
||||
enabled_any = True
|
||||
|
||||
if hasattr(self.model, "language_model"):
|
||||
language_model = self.model.language_model
|
||||
enabled_any = _enable_ckpt(language_model) or enabled_any
|
||||
if hasattr(language_model, "model"):
|
||||
enabled_any = _enable_ckpt(language_model.model) or enabled_any
|
||||
if hasattr(language_model, "config"):
|
||||
language_model.config.use_cache = False
|
||||
|
||||
if hasattr(self.model, "config"):
|
||||
self.model.config.use_cache = False
|
||||
|
||||
if hasattr(self.model, "enable_input_require_grads"):
|
||||
self.model.enable_input_require_grads()
|
||||
|
||||
if enabled_any:
|
||||
logging.info(
|
||||
"Enabled InternVL3 gradient checkpointing (use_reentrant=%s).",
|
||||
self.gradient_checkpointing_use_reentrant,
|
||||
)
|
||||
else:
|
||||
logging.warning("Requested InternVL3 gradient checkpointing, but no checkpointable module was found.")
|
||||
|
||||
def _dynamic_preprocess_tensor(
|
||||
self, image_t, min_num=1, max_num=1, use_thumbnail=False
|
||||
):
|
||||
@ -191,7 +238,7 @@ class InternVL3Embedder(nn.Module):
|
||||
for image_tensors in image_tensors_batch:
|
||||
num_tiles_list = []
|
||||
for image in image_tensors:
|
||||
if isinstance(image, torch.Tensor):
|
||||
if isinstance(image, torch.Tensor) and self.enable_tensor_fastpath:
|
||||
if (
|
||||
image.ndim == 3
|
||||
and image.shape[0] == 3
|
||||
@ -216,6 +263,8 @@ class InternVL3Embedder(nn.Module):
|
||||
tiles = self._dynamic_preprocess_tensor(image_t)
|
||||
tile_tensors = (tiles - mean) / std
|
||||
else:
|
||||
if isinstance(image, torch.Tensor):
|
||||
image = to_pil_image(image.detach().to(device="cpu"))
|
||||
image_t = TF.to_tensor(image).to(device=self.device, dtype=torch.float32)
|
||||
tiles = self._dynamic_preprocess_tensor(image_t)
|
||||
tile_tensors = (tiles - mean) / std
|
||||
|
||||
@ -70,6 +70,9 @@ class EVO1Policy(PreTrainedPolicy):
|
||||
"vlm_num_layers": config.vlm_num_layers,
|
||||
"vlm_dtype": config.vlm_dtype,
|
||||
"use_flash_attn": config.use_flash_attn,
|
||||
"embedder_tensor_fastpath": config.embedder_tensor_fastpath,
|
||||
"enable_vlm_gradient_checkpointing": config.enable_vlm_gradient_checkpointing,
|
||||
"gradient_checkpointing_use_reentrant": config.gradient_checkpointing_use_reentrant,
|
||||
"action_head": config.action_head,
|
||||
"action_horizon": config.chunk_size,
|
||||
"per_action_dim": config.max_action_dim,
|
||||
@ -300,7 +303,7 @@ class EVO1Policy(PreTrainedPolicy):
|
||||
raise ValueError(
|
||||
f"EVO1 batch mismatch across cameras: key={camera_key} batch={image.shape[0]} expected={batch_size}"
|
||||
)
|
||||
sample_images.append(image[batch_index].detach().cpu())
|
||||
sample_images.append(image[batch_index])
|
||||
if not sample_images:
|
||||
raise ValueError("EVO1 received a batch without any image tensor.")
|
||||
while len(sample_images) < self.config.max_views:
|
||||
|
||||
@ -273,6 +273,14 @@ def train(
|
||||
" scheduler, and global step will start fresh.",
|
||||
cfg.policy.pretrained_path,
|
||||
)
|
||||
if cfg.policy.type == "evo1" and getattr(cfg.policy, "training_stage", None) == "stage2":
|
||||
logging.info(
|
||||
"Evo1 stage2 runtime settings: embedder_tensor_fastpath=%s "
|
||||
"enable_vlm_gradient_checkpointing=%s gradient_checkpointing_use_reentrant=%s",
|
||||
getattr(cfg.policy, "embedder_tensor_fastpath", None),
|
||||
getattr(cfg.policy, "enable_vlm_gradient_checkpointing", None),
|
||||
getattr(cfg.policy, "gradient_checkpointing_use_reentrant", None),
|
||||
)
|
||||
policy = make_policy(
|
||||
cfg=cfg.policy,
|
||||
ds_meta=dataset.meta,
|
||||
|
||||
在新工单中引用
屏蔽一个用户