File size: 24,714 Bytes
e94400c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 | # Copyright 2025 starVLA community. All rights reserved.
# Licensed under the MIT License, Version 1.0 (the "License");
# Implemented by Jinhui YE / HKUST University] in [2025].
"""
Qwen-GROOT Framework
A lightweight implementation that Qwen2.5-vl + Flow-matching head to directly predict continuous actions
Flow-matching header is copyright from GR00T N1.5, but a sample MoE inspired by PI_0
"""
import sys
sys.path.append("/mnt/data/fangyu/code/rewardmodel")
from typing import List
from tqdm import tqdm
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from PIL import Image
import copy
from starVLA.training.trainer_utils import initialize_overwatch
from deployment.model_server.tools.image_tools import to_pil_preserve
from transformers import AutoImageProcessor, AutoModel
from omegaconf import OmegaConf
logger = initialize_overwatch(__name__)
# HuggingFace Default / LLaMa-2 IGNORE_INDEX (for labels)
IGNORE_INDEX = -100
from starVLA.model.framework.base_framework import baseframework
from starVLA.model.modules.vlm import get_vlm_model
from starVLA.model.modules.action_model.ActionModel_FM import ActionModelFM
from starVLA.model.modules.action_model.configuration_actionmodel import ActionModelConfig
from starVLA.dataloader.gr00t_lerobot.datasets import ACTION_REPRESENTATION_SLICES
from starVLA.training.trainer_utils.trainer_tools import resize_images
from starVLA.model.tools import FRAMEWORK_REGISTRY
####################################################
# ⚠️ Warning: This framework has been restructured and is NOT compatible with checkpoints created before 2025-10-20.
####################################################
@FRAMEWORK_REGISTRY.register("QwenLatent")
class QwenLatent(baseframework):
"""
Multimodal vision-language-action model.
Components:
- Qwen2.5 VL interface for fused language/vision token embeddings
- Layer-wise cross DiT diffusion head
Focus: Predict future continuous actions conditioned on images + instruction.
"""
@staticmethod
def _get_last_nonpad_indices(attention_mask: torch.Tensor) -> torch.Tensor:
"""
Return the index of the last non-padding token for each sequence.
Works for both tokenizer.padding_side == "left" and "right".
attention_mask: [B, T] with 1/True for real tokens and 0/False for pads.
"""
if attention_mask is None:
raise ValueError("attention_mask cannot be None")
if attention_mask.dim() != 2:
raise ValueError(f"attention_mask must be 2D [B,T], got shape {tuple(attention_mask.shape)}")
# Find distance-from-end to last 1 by reversing sequence dimension.
# Example:
# - left pad: [0,0,1,1,1] -> flip -> [1,1,1,0,0] -> argmax = 0 -> last = T-1
# - right pad: [1,1,1,0,0] -> flip -> [0,0,1,1,1] -> argmax = 2 -> last = T-1-2 = 2
mask = attention_mask.to(dtype=torch.long)
rev_first_one = torch.flip(mask, dims=[1]).argmax(dim=1)
last_nonpad = mask.size(1) - 1 - rev_first_one
return last_nonpad
#
def __init__(
self,
config: Optional[dict] = None,
**kwargs,
) -> None:
"""
Construct all submodules and cache key configuration values.
Args:
config: Hierarchical configuration (OmegaConf/dict) containing framework + trainer sections.
**kwargs: Reserved for future overrides (unused).
"""
super().__init__()
self.config = config
self.qwen_vl_interface = get_vlm_model(config=self.config)
# dynamic get llm config
num_vl_layers, llm_hidden_size = 36, self.qwen_vl_interface.model.config.hidden_size
self.llm_hidden_size = llm_hidden_size
self.config.framework.qwenvl.vl_hidden_dim = llm_hidden_size
self.config.framework.qwenvl.num_vl_layers = num_vl_layers
action_model_cfg = getattr(self.config.framework, "action_model", None)
if action_model_cfg is not None:
action_model_kwargs = OmegaConf.to_container(action_model_cfg, resolve=True)
print(f"{action_model_kwargs=}")
self.action_model = ActionModelFM(ActionModelConfig(**action_model_kwargs))
else:
self.action_model = ActionModelFM(ActionModelConfig())
ckpt_path = getattr(self.config.framework.action_model, "ckpt_path", None)
if ckpt_path:
self.action_model.load_state_dict(torch.load(ckpt_path, map_location="cpu"), strict=True)
print(f"✅ loaded action model from {ckpt_path}")
print(f"action model loss mode: {self.action_model.config.loss_mode}")
# Dataset soft prompt for QwenVL (conditioning on dataset_id)
self.dataset_vocab_size = getattr(self.config.framework.action_model, "dataset_vocab_size", 256)
self.num_data_tokens = getattr(self.config.framework.qwenvl, "num_data_tokens", 32)
self.dataset_embed = nn.Embedding(
self.dataset_vocab_size,
llm_hidden_size * self.num_data_tokens,
)
# Learnable query token appended to VLM inputs (for action embedding)
self.query_token = nn.Parameter(torch.randn(1, 1, llm_hidden_size))
# 使用 MLP 投影器,增加表达能力(2048 → 2048 → 1024)
action_hidden_size = self.action_model.config.hidden_size
self.action_embed_projector = nn.Sequential(
nn.Linear(llm_hidden_size, llm_hidden_size),
nn.GELU(),
nn.Linear(llm_hidden_size, action_hidden_size),
)
self.chunk_size = self.config.datasets.vla_data.chunk_size
self.num_history_steps = 0
self.use_state = self.action_model.use_state
# Multi-t sampling trick: sample K different t per example for the FM head
# to enlarge effective batch size without re-running the expensive VLM.
self.num_t_samples = getattr(self.config.framework.action_model, "num_t_samples", 1)
print(f"num_t_samples: {self.num_t_samples}")
def _maybe_log_align_stats(
self,
predicted_action_embeddings: torch.Tensor,
gt_action_embeddings: torch.Tensor,
) -> None:
if getattr(self, "_align_stats_logged", False):
return
if torch.distributed.is_available() and torch.distributed.is_initialized():
if torch.distributed.get_rank() != 0:
return
with torch.no_grad():
pred = predicted_action_embeddings.float()
gt = gt_action_embeddings.float()
pred_norm = pred.norm(dim=-1).mean().item()
gt_norm = gt.norm(dim=-1).mean().item()
logger.info(
"Align stats: pred(mean=%.4f,std=%.4f,avg_norm=%.4f) "
"gt(mean=%.4f,std=%.4f,avg_norm=%.4f)",
pred.mean().item(),
pred.std().item(),
pred_norm,
gt.mean().item(),
gt.std().item(),
gt_norm,
)
self._align_stats_logged = True
def forward(
self,
examples: List[dict] = None,
**kwargs,
):
"""
Args:
examples: List[dict], each dict requires:
- image: List[PIL.Image] (multi-view)
- lang: str instruction
- action: np.ndarray or list shaped [T, action_dim]
Returns:
dict:
action_loss (torch.Tensor): Scalar diffusion noise prediction loss.
"""
batch_images = [example["image"] for example in examples] # [B,[PLT]]
instructions = [example["lang"] for example in examples] # [B, str]
actions = [example["action"] for example in examples] # label [B, L, action_dim]
states = [example["state"] for example in examples] if self.use_state else None # [B, L, state_dim] when state_use_action_chunk
dataset_ids = [example.get("dataset_id", 0) for example in examples]
# Step 1: QWenVL input format
qwen_inputs = self.qwen_vl_interface.build_qwenvl_inputs(
images=batch_images,
instructions=instructions,
chunk_size=self.chunk_size,
)
# Prepend dataset soft prompt tokens to VLM inputs
if "input_ids" in qwen_inputs:
dataset_ids_tensor = torch.tensor(
dataset_ids, device=qwen_inputs["input_ids"].device, dtype=torch.long
)
ds_embeds = self.dataset_embed(dataset_ids_tensor).view(
len(dataset_ids), self.num_data_tokens, self.llm_hidden_size
)
token_embeds = self.qwen_vl_interface.model.get_input_embeddings()(qwen_inputs["input_ids"])
query_embeds = self.query_token.expand(len(dataset_ids), -1, -1)
qwen_inputs["inputs_embeds"] = torch.cat((ds_embeds, token_embeds, query_embeds), dim=1)
qwen_inputs.pop("input_ids")
if "attention_mask" in qwen_inputs:
prefix_mask = torch.ones(
(qwen_inputs["attention_mask"].shape[0], self.num_data_tokens),
device=qwen_inputs["attention_mask"].device,
dtype=qwen_inputs["attention_mask"].dtype,
)
query_mask = torch.ones(
(qwen_inputs["attention_mask"].shape[0], 1),
device=qwen_inputs["attention_mask"].device,
dtype=qwen_inputs["attention_mask"].dtype,
)
qwen_inputs["attention_mask"] = torch.cat(
(prefix_mask, qwen_inputs["attention_mask"], query_mask), dim=1
)
if "position_ids" in qwen_inputs:
prefix_pos = torch.arange(
self.num_data_tokens,
device=qwen_inputs["position_ids"].device,
dtype=qwen_inputs["position_ids"].dtype,
).unsqueeze(0).expand(qwen_inputs["position_ids"].shape[0], -1)
query_pos = (
torch.full(
(qwen_inputs["position_ids"].shape[0], 1),
qwen_inputs["position_ids"].shape[1] + self.num_data_tokens,
device=qwen_inputs["position_ids"].device,
dtype=qwen_inputs["position_ids"].dtype,
)
)
qwen_inputs["position_ids"] = torch.cat(
(prefix_pos, qwen_inputs["position_ids"] + self.num_data_tokens, query_pos), dim=1
)
with torch.autocast("cuda", dtype=torch.bfloat16):
qwenvl_outputs = self.qwen_vl_interface(
**qwen_inputs,
output_attentions=False,
output_hidden_states=True,
return_dict=True,
)
last_hidden_states = qwenvl_outputs.hidden_states[-1]
if "attention_mask" in qwen_inputs:
# 找到非 padding 的最后一个 token index(兼容 left/right padding)
last_token_indices = self._get_last_nonpad_indices(qwen_inputs["attention_mask"])
batch_indices = torch.arange(last_hidden_states.shape[0], device=last_hidden_states.device)
action_token_hidden = last_hidden_states[batch_indices, last_token_indices]
else:
action_token_hidden = last_hidden_states[:, -1, :]
predicted_action_embeddings = self.action_embed_projector(action_token_hidden).float() # [B, Action_Hidden]
predicted_action_embeddings = F.normalize(predicted_action_embeddings, p=2, dim=-1)
# Step 2: Action Expert Forward and Loss
loss_mode = getattr(self.action_model.config, "loss_mode", "full")
with torch.autocast("cuda", dtype=torch.float32):
actions_target = torch.as_tensor(np.array(actions), device=last_hidden_states.device, dtype=torch.float32)
# Multi-t sampling trick: expand the FM-head batch K times by sampling K
# independent t values per example. The expensive VLM embedding is computed
# only once and then tiled, so the extra cost is only in the lightweight FM head.
K = self.num_t_samples
def tile_batch(x: torch.Tensor, k: int) -> torch.Tensor:
"""Repeat tensor k times along dim-0, keeping all other dims intact."""
return x.repeat(k, *([1] * (x.dim() - 1)))
if K > 1:
actions_target_fm = tile_batch(actions_target, K) # [K*B, T, D]
predicted_embeddings_fm = tile_batch(predicted_action_embeddings, K)
else:
actions_target_fm = actions_target
predicted_embeddings_fm = predicted_action_embeddings
B_fm = actions_target_fm.shape[0]
t = self.action_model._sample_fm_time(B_fm, device=actions_target.device, dtype=actions_target.dtype)
noise = torch.randn_like(actions_target_fm)
if loss_mode == "predict_only":
# Only predict_loss: skip align_loss and recon_loss
align_loss = None
recon_loss = None
predict_loss = self.action_model.recon_loss_from_embedding(
actions=actions_target_fm,
action_embedding=predicted_embeddings_fm,
t=t,
noise=noise,
)
else:
# Full mode: align + recon + predict
# state chunk 与 action chunk 对齐(同长度)
states_target = None
if self.use_state:
states_target = torch.as_tensor(np.array(states), device=last_hidden_states.device, dtype=torch.float32)
gt_action_embeddings = self.action_model.encode_actions(
actions=actions_target,
dataset_ids=dataset_ids,
state=states_target,
)
self._maybe_log_align_stats(predicted_action_embeddings, gt_action_embeddings)
# align_loss only needs the original (non-expanded) embeddings
align_loss = F.l1_loss(predicted_action_embeddings, gt_action_embeddings.float().detach())
gt_embeddings_fm = tile_batch(gt_action_embeddings, K) if K > 1 else gt_action_embeddings
recon_loss = self.action_model.recon_loss_from_embedding(
actions=actions_target_fm,
action_embedding=gt_embeddings_fm,
t=t,
noise=noise,
)
predict_loss = self.action_model.recon_loss_from_embedding(
actions=actions_target_fm,
action_embedding=predicted_embeddings_fm,
t=t,
noise=noise,
)
return {
"align_loss": align_loss,
"recon_loss": recon_loss,
"predict_loss": predict_loss,
}
@torch.inference_mode()
def predict_action( # TODO align predict_action with forward, make api more flexible
self,
examples: List[dict] = None,
embodiment_tag: Optional[str] = None,
**kwargs: str,
) -> np.ndarray:
"""
推理:单次前向直接回归未来动作(无扩散采样)。
Steps:
1. Resize images to training resolution (if specified)
2. Encode with QwenVL (hidden states retained)
Args:
examples: List of example dicts containing image, lang, etc.
embodiment_tag: Optional embodiment tag (e.g., "franka", "oxe_rt1", "oxe_bridge").
If provided, will extract valid action dimensions based on ACTION_REPRESENTATION_SLICES.
If None, returns full unified action representation.
Returns:
dict:
normalized_actions (np.ndarray): Shape [B, T, action_dim], diffusion-sampled normalized actions.
If embodiment_tag is provided, shape is [B, T, valid_dim] where
valid_dim is determined by ACTION_REPRESENTATION_SLICES[embodiment_tag].
"""
from deployment.model_server.tools.image_tools import to_pil_preserve
batch_images = [to_pil_preserve(example["image"]) for example in examples] # [B,[PLT]]
instructions = [example["lang"] for example in examples] # [B, str]
dataset_ids = [example.get("dataset_id") for example in examples]
train_obs_image_size = getattr(self.config.datasets.vla_data, "image_size", None)
if train_obs_image_size:
batch_images = resize_images(batch_images, target_size=train_obs_image_size)
# Step 1: QWenVL input format
qwen_inputs = self.qwen_vl_interface.build_qwenvl_inputs(
images=batch_images,
instructions=instructions,
)
# Prepend dataset soft prompt tokens to VLM inputs
if "input_ids" in qwen_inputs:
dataset_ids_tensor = torch.tensor(
dataset_ids, device=qwen_inputs["input_ids"].device, dtype=torch.long
)
ds_embeds = self.dataset_embed(dataset_ids_tensor).view(
len(dataset_ids), self.num_data_tokens, self.llm_hidden_size
)
token_embeds = self.qwen_vl_interface.model.get_input_embeddings()(qwen_inputs["input_ids"])
query_embeds = self.query_token.expand(len(dataset_ids), -1, -1)
qwen_inputs["inputs_embeds"] = torch.cat((ds_embeds, token_embeds, query_embeds), dim=1)
qwen_inputs.pop("input_ids")
if "attention_mask" in qwen_inputs:
prefix_mask = torch.ones(
(qwen_inputs["attention_mask"].shape[0], self.num_data_tokens),
device=qwen_inputs["attention_mask"].device,
dtype=qwen_inputs["attention_mask"].dtype,
)
query_mask = torch.ones(
(qwen_inputs["attention_mask"].shape[0], 1),
device=qwen_inputs["attention_mask"].device,
dtype=qwen_inputs["attention_mask"].dtype,
)
qwen_inputs["attention_mask"] = torch.cat(
(prefix_mask, qwen_inputs["attention_mask"], query_mask), dim=1
)
if "position_ids" in qwen_inputs:
prefix_pos = torch.arange(
self.num_data_tokens,
device=qwen_inputs["position_ids"].device,
dtype=qwen_inputs["position_ids"].dtype,
).unsqueeze(0).expand(qwen_inputs["position_ids"].shape[0], -1)
query_pos = (
torch.full(
(qwen_inputs["position_ids"].shape[0], 1),
qwen_inputs["position_ids"].shape[1] + self.num_data_tokens,
device=qwen_inputs["position_ids"].device,
dtype=qwen_inputs["position_ids"].dtype,
)
)
qwen_inputs["position_ids"] = torch.cat(
(prefix_pos, qwen_inputs["position_ids"] + self.num_data_tokens, query_pos), dim=1
)
with torch.autocast("cuda", dtype=torch.bfloat16):
qwenvl_outputs = self.qwen_vl_interface(
**qwen_inputs,
output_attentions=False,
output_hidden_states=True,
return_dict=True,
)
last_hidden_states = qwenvl_outputs.hidden_states[-1]
if "attention_mask" in qwen_inputs:
# 找到非 padding 的最后一个 token index(兼容 left/right padding)
last_token_indices = self._get_last_nonpad_indices(qwen_inputs["attention_mask"])
batch_indices = torch.arange(last_hidden_states.shape[0], device=last_hidden_states.device)
action_token_hidden = last_hidden_states[batch_indices, last_token_indices]
else:
action_token_hidden = last_hidden_states[:, -1, :]
predicted_action_embeddings = self.action_embed_projector(action_token_hidden).float() # [B, Action_Hidden]
# L2 normalize before sending to decoder (consistent with training)
predicted_action_embeddings = F.normalize(predicted_action_embeddings, p=2, dim=-1)
# Step 4: 选择 decoder 进行推理
with torch.autocast("cuda", dtype=torch.float32):
pred_actions = self.action_model.decode_actions(
predicted_action_embeddings,
chunk_size=self.chunk_size
)
normalized_actions = pred_actions.detach().cpu().numpy()
# 如果提供了 embodiment_tag,根据 tag 提取有效的动作维度
if embodiment_tag is not None:
if embodiment_tag not in ACTION_REPRESENTATION_SLICES:
raise ValueError(
f"Unknown embodiment tag '{embodiment_tag}'. "
f"Known tags: {sorted(ACTION_REPRESENTATION_SLICES.keys())}"
)
# 获取对应的 slice
target_slice = ACTION_REPRESENTATION_SLICES[embodiment_tag]
# 从统一表示中提取对应的维度
normalized_actions = normalized_actions[..., target_slice]
return {"normalized_actions": normalized_actions}
if __name__ == "__main__":
from omegaconf import OmegaConf
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--config_yaml", type=str,
default="/fsx/home/yfang/projects/LearnLatent/starVLA/config/training/starvla_train_qwenlatent_oxe.yaml",
help="Path to YAML config")
args, clipargs = parser.parse_known_args()
cfg = OmegaConf.load(args.config_yaml)
# try get model
model = QwenLatent(cfg)
# ckpt="/mnt/petrelfs/yejinhui/Projects/llavavla/results/Checkpoints/1011_qwenpi/checkpoints/need_steps_10000_pytorch_model.pt"
# model = Qwen_PI.from_pretrained(ckpt)
print(model)
# fake sample
image = Image.fromarray(np.random.randint(0, 255, (224, 224, 3), dtype=np.uint8))
# Create a sample
sample = {
"action": np.random.uniform(-1, 1, size=(15, 7)).astype(np.float16), # action_chunk, action_dim
"image": [image], # two views
"image_past_half": [image],
"image_past_one": [image],
"image_future": [image],
"lang": "put the ball on the table",
"state": np.random.uniform(-1, 1, size=(1, 8)).astype(np.float16), # chunk, state_dim
}
batch = [sample, sample] # batch size 2
device = torch.device("cuda:7" if torch.cuda.is_available() else "cpu")
model = model.to(device)
forward_output = model(batch)
align_loss = forward_output['align_loss']
recon_loss = forward_output['recon_loss']
print(f"Align Loss: {align_loss.item()}")
print(f"Recon Loss: {recon_loss.item()}")
# # test predict action
# predict_output = model.predict_action([sample])
# normalized_actions = predict_output['normalized_actions']
# print(f"Unnormalized Action: {normalized_actions}")
# # Advance: try forward model with dataloader
# # can be fake sample, but here get from dataloader for simpler
# from starVLA.dataloader.lerobot_datasets import get_vla_dataset, collate_fn
# vla_dataset_cfg = cfg.datasets.vla_data
# dataset = get_vla_dataset(data_cfg=vla_dataset_cfg)
# from torch.utils.data import DataLoader
# train_dataloader = DataLoader(
# dataset,
# batch_size=2,
# num_workers=1, # For Debug
# collate_fn=collate_fn,
# )
# #
# for batch in tqdm(train_dataloader, desc="Processing Batches"):
# batch
# break
# # try get model
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model = model.to(device)
# model(batch)
# action = model.predict_action(batch_images=[batch[0]["image"]], instructions=[batch[0]["lang"]])
# # fake state
# for ba in batch:
# ba["state"] = ba["action"][0][None]
# model(batch)
# action = model.predict_action(batch_images=[batch[0]["image"]], instructions=[batch[0]["lang"]], state=[batch[0]["state"]]) |