210 lines
9.0 KiB
Python
210 lines
9.0 KiB
Python
"""j2t 转换入口 (design_v2 §8, D8/R-10)。
|
|
|
|
JAX checkpoint (契约管线产物, 含 model_structure.json) -> torch model.safetensors。
|
|
|
|
改造 (相对 examples/convert_jax_model_to_pytorch.py):
|
|
- 结构判定读 model_structure.json (契约快照), 不再用 config 名 / 目录名字符串
|
|
(R-10: examples 内三处 "pi05" in checkpoint_dir 已改为 pi05 形参, 本入口以
|
|
契约字段驱动);
|
|
- 转换前结构校验 (params/ 存在 + 契约可恢复);
|
|
- 转换后抽样数值一致性自检 (JAX vs torch 前向对拍: 同一 fake_obs 与噪声,
|
|
actions 余弦相似度 / 最大绝对误差), 结果写进产物 metadata;
|
|
- 产物附加: model_structure.json 透传 + paligemma_tokenizer.model 拷贝
|
|
(下游消费: model.safetensors + tokenizer 同目录)。
|
|
|
|
超参: CHECKPOINT_PATH (与上游端口 checkpoint 二选一上游优先), OUTPUT 精度
|
|
(bfloat16 默认)。
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import importlib.util
|
|
import json
|
|
import logging
|
|
import os
|
|
import shutil
|
|
import sys
|
|
import types
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
|
from pi_pipeline import contract as _contract # noqa: E402
|
|
from ws_common import report_utils # noqa: E402
|
|
|
|
|
|
def locate_checkpoint() -> Path:
|
|
"""上游端口 (portId=checkpoint) 优先, CHECKPOINT_PATH 兜底。
|
|
|
|
与 serve_xpolicylab.locate_checkpoint 同语义; 此处独立实现避免顶层引入
|
|
XPolicyLab client_server 依赖 (convert 场景无需 serve 栈)。
|
|
"""
|
|
import json
|
|
|
|
artifacts_path = os.environ.get("INPUT_ARTIFACTS_PATH", "")
|
|
if artifacts_path and os.path.isfile(artifacts_path):
|
|
with open(artifacts_path, "r", encoding="utf-8") as f:
|
|
entries = json.load(f) if artifacts_path else []
|
|
for entry in entries or []:
|
|
if entry.get("portId") == "checkpoint" and not str(entry.get("path", "")).startswith("bos://"):
|
|
return Path(entry["path"])
|
|
value = os.environ.get("CHECKPOINT_PATH", "").strip()
|
|
if value:
|
|
return Path(value)
|
|
sys.exit("[input] neither upstream checkpoint port nor CHECKPOINT_PATH is set")
|
|
|
|
logger = logging.getLogger("openpi.convert")
|
|
|
|
_CONVERT_SCRIPT = Path(__file__).resolve().parents[1] / "openpi" / "examples" / "convert_jax_model_to_pytorch.py"
|
|
|
|
|
|
def _load_convert_module():
|
|
spec = importlib.util.spec_from_file_location("openpi_convert", _CONVERT_SCRIPT)
|
|
module = importlib.util.module_from_spec(spec)
|
|
spec.loader.exec_module(module)
|
|
return module
|
|
|
|
|
|
def numeric_consistency_check(checkpoint_dir: Path, contract: _contract.Contract, torch_output: Path) -> dict:
|
|
"""JAX vs torch 前向对拍: 同一 obs + 噪声, 比较 sample_actions。"""
|
|
import jax
|
|
import torch
|
|
|
|
import openpi.models.model as _model
|
|
import dataclasses
|
|
from openpi.models.pi0_config import Pi0Config
|
|
|
|
# 自检单次前向, 关闭 torch.compile (max-autotune 首编 10min+ 且对拍数值无益)
|
|
model_cfg = dataclasses.replace(
|
|
Pi0Config(**contract.pi0_config_kwargs()), pytorch_compile_mode=None
|
|
)
|
|
|
|
jax_model = model_cfg.load(_model.restore_params(checkpoint_dir / "params", dtype="bfloat16"))
|
|
# load_pytorch 按 train_config.model 取模型配置 (官方 scripts 惯例), 此处以
|
|
# shim 适配独立调用
|
|
train_cfg_shim = types.SimpleNamespace(model=model_cfg)
|
|
torch_model = model_cfg.load_pytorch(train_cfg_shim, str(torch_output / "model.safetensors"))
|
|
torch_model = torch_model.to("cuda" if torch.cuda.is_available() else "cpu").eval()
|
|
device = next(torch_model.parameters()).device
|
|
|
|
obs_spec, _ = model_cfg.inputs_spec()
|
|
obs = jax.tree.map(
|
|
lambda s: np.random.uniform(-1.0, 1.0, size=s.shape).astype(np.float32)
|
|
if s.dtype == np.float32 or str(s.dtype) == "float32"
|
|
else (np.random.uniform(0, 2, size=s.shape).astype(np.int32) if "int" in str(s.dtype) else np.zeros(s.shape, bool)),
|
|
obs_spec,
|
|
)
|
|
rng = np.random.default_rng(0)
|
|
noise = rng.normal(size=(1, contract.action_horizon, model_cfg.action_dim)).astype(np.float32)
|
|
|
|
# jaxtyping>=0.3 的 Array hint 不含 numpy 数组, jax 侧统一转 jnp
|
|
import jax.numpy as jnp
|
|
|
|
jax_obs = jax.tree.map(jnp.asarray, obs)
|
|
jax_actions = np.asarray(jax_model.sample_actions(jax.random.key(0), jax_obs, noise=jnp.asarray(noise)))
|
|
|
|
torch_obs = jax.tree.map(lambda x: torch.from_numpy(np.asarray(x)).to(device), obs)
|
|
# torch 栈图像约定 CHW (官方 train_pytorch 数据管线/transformers get_image_features),
|
|
# jax 侧 inputs_spec 为 HWC — 图像叶 permute
|
|
for cam in list(torch_obs.images.keys()):
|
|
torch_obs.images[cam] = torch.permute(torch_obs.images[cam], (0, 3, 1, 2))
|
|
torch_noise = torch.from_numpy(noise).to(device)
|
|
with torch.no_grad():
|
|
torch_actions = np.asarray(torch_model.sample_actions(device, torch_obs, noise=torch_noise).cpu())
|
|
|
|
a, b = jax_actions.reshape(-1), torch_actions.reshape(-1)
|
|
cos = float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-9))
|
|
max_abs = float(np.abs(a - b).max())
|
|
return {"cosine": cos, "max_abs_diff": max_abs, "shape": list(jax_actions.shape)}
|
|
|
|
|
|
def main() -> None:
|
|
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
|
|
output_dir = Path(os.environ.get("OUTPUT_DIR", "").strip())
|
|
if not output_dir:
|
|
sys.exit("[hyperparam] OUTPUT_DIR not set")
|
|
|
|
checkpoint_dir = locate_checkpoint()
|
|
contract = load_contract_checked(checkpoint_dir)
|
|
precision = os.environ.get("PRECISION", "bfloat16").strip() or "bfloat16"
|
|
|
|
logger.info("[convert] JAX %s -> torch (%s)", checkpoint_dir, precision)
|
|
convert = _load_convert_module()
|
|
from openpi.models.pi0_config import Pi0Config
|
|
|
|
model_cfg = Pi0Config(**contract.pi0_config_kwargs())
|
|
torch_output = output_dir / "checkpoint_torch"
|
|
convert.convert_pi0_checkpoint(str(checkpoint_dir), precision, str(torch_output), model_cfg)
|
|
|
|
# 契约快照透传 (checkpoint 自带或 CONTRACT_JSON 注入时由当前契约生成)
|
|
source_structure = checkpoint_dir / "model_structure.json"
|
|
if source_structure.is_file():
|
|
shutil.copy2(source_structure, torch_output / "model_structure.json")
|
|
else:
|
|
(torch_output / "model_structure.json").write_text(contract.to_json(), encoding="utf-8")
|
|
logger.info("[convert] contract snapshot generated from CONTRACT_JSON injection")
|
|
|
|
# tokenizer 随行 (量化交接)
|
|
tokenizer_src = _find_tokenizer()
|
|
if tokenizer_src:
|
|
shutil.copy2(tokenizer_src, torch_output / "paligemma_tokenizer.model")
|
|
logger.info("[convert] tokenizer copied: %s", tokenizer_src)
|
|
else:
|
|
logger.warning(
|
|
"[convert] paligemma_tokenizer.model not found in cache; quantization handoff "
|
|
"requires it next to model.safetensors (run serve/train once or place manually)"
|
|
)
|
|
|
|
# 数值一致性自检
|
|
check = numeric_consistency_check(checkpoint_dir, contract, torch_output)
|
|
verdict = check["cosine"] >= 0.95
|
|
check["verdict"] = "pass" if verdict else "fail"
|
|
logger.info("[convert] numeric check: cos=%.4f max_abs=%.4f -> %s", check["cosine"], check["max_abs_diff"], check["verdict"])
|
|
(torch_output / "convert_metadata.json").write_text(
|
|
json.dumps({"numeric_check": check, "precision": precision, "source_checkpoint": str(checkpoint_dir)},
|
|
ensure_ascii=False, indent=2),
|
|
encoding="utf-8",
|
|
)
|
|
|
|
report_utils.write_artifacts(
|
|
output_dir,
|
|
[{"name": "checkpoint_torch", "type": "file", "portId": "checkpoint", "path": str(torch_output)}],
|
|
)
|
|
if not verdict:
|
|
sys.exit("[convert] numeric consistency check FAILED (cosine < 0.95)")
|
|
|
|
|
|
def load_contract_checked(checkpoint_dir: Path) -> _contract.Contract:
|
|
"""契约来源优先级: checkpoint 内 model_structure.json > CONTRACT_JSON 超参注入
|
|
(官方权重/外部权重无契约快照时, 如 RoboTwin cotrain 权重, T5)。"""
|
|
structure = checkpoint_dir / "model_structure.json"
|
|
if structure.is_file():
|
|
return _contract.Contract.from_model_structure(str(structure))
|
|
raw = os.environ.get("CONTRACT_JSON", "").strip()
|
|
if raw:
|
|
logger.info("[convert] model_structure.json not found; using CONTRACT_JSON injection")
|
|
# 快照只落产物侧 (main 透传), 不回写源 checkpoint 目录 (外部资产可能只读)
|
|
return _contract.Contract.from_json(raw)
|
|
sys.exit(
|
|
f"[convert] {structure} not found and CONTRACT_JSON not set; provide either a "
|
|
"contract-pipeline checkpoint or an explicit CONTRACT_JSON"
|
|
)
|
|
|
|
|
|
def _find_tokenizer() -> Path | None:
|
|
"""openpi 缓存中的 paligemma tokenizer (量化交接)。"""
|
|
candidates = []
|
|
data_home = os.environ.get("OPENPI_DATA_HOME", "")
|
|
if data_home:
|
|
candidates.append(Path(data_home) / "big_vision" / "paligemma_tokenizer.model")
|
|
candidates.append(Path.home() / ".cache" / "openpi" / "big_vision" / "paligemma_tokenizer.model")
|
|
for cand in candidates:
|
|
if cand.is_file():
|
|
return cand
|
|
return None
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|