Parvulus / c_engine /infer_c.py
aasivas's picture
Upload folder using huggingface_hub
67c4f10 verified
Raw History Blame Contribute Delete
17.8 kB
"""C++ and CUDA inference backends for Parvulus."""
import ctypes
import os
import time
from pathlib import Path
from typing import Dict, List, Optional, Sequence, Tuple, Union
import numpy as np
import torch
C_ENGINE_DIR = Path(__file__).resolve().parent
LIB_PATH = C_ENGINE_DIR / "libparvulus.so"
_lib = None
if LIB_PATH.exists():
try:
_lib = ctypes.CDLL(str(LIB_PATH))
_lib.parvulus_create.argtypes = [ctypes.c_int]
_lib.parvulus_create.restype = ctypes.c_void_p
_lib.parvulus_load.argtypes = [
ctypes.c_void_p,
ctypes.c_char_p,
ctypes.c_char_p,
ctypes.c_char_p,
]
_lib.parvulus_load.restype = ctypes.c_bool
_lib.parvulus_set_task_vector.argtypes = [
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_float),
]
_lib.parvulus_set_task_vector.restype = ctypes.c_bool
_lib.parvulus_set_task_string.argtypes = [
ctypes.c_void_p,
ctypes.c_char_p,
]
_lib.parvulus_set_task_string.restype = ctypes.c_bool
_lib.parvulus_predict_chunk.argtypes = [
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_uint8),
ctypes.POINTER(ctypes.c_uint8),
ctypes.c_char_p,
ctypes.POINTER(ctypes.c_float),
ctypes.c_int,
ctypes.POINTER(ctypes.c_float),
]
_lib.parvulus_predict_chunk.restype = ctypes.c_bool
_lib.parvulus_predict_chunk_with_vector.argtypes = [
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_uint8),
ctypes.POINTER(ctypes.c_uint8),
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.c_int,
ctypes.POINTER(ctypes.c_float),
]
_lib.parvulus_predict_chunk_with_vector.restype = ctypes.c_bool
_lib.parvulus_predict_chunk_with_emb.argtypes = [
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_uint8),
ctypes.POINTER(ctypes.c_uint8),
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.c_int,
ctypes.POINTER(ctypes.c_float),
]
_lib.parvulus_predict_chunk_with_emb.restype = ctypes.c_bool
_lib.parvulus_register_task_embedding.argtypes = [
ctypes.c_void_p,
ctypes.c_char_p,
ctypes.POINTER(ctypes.c_float),
]
_lib.parvulus_register_task_embedding.restype = ctypes.c_bool
_lib.parvulus_destroy.argtypes = [ctypes.c_void_p]
_lib.parvulus_destroy.restype = None
except Exception as e:
print(f"[Warning] Could not load libparvulus.so: {e}")
_lib = None
class ParvulusEngine:
"""Native Parvulus CUDA/cuBLAS runtime engine (Zero LibTorch, Zero PyTorch)."""
def __init__(
self,
device_id: int = 0,
weights_path: Optional[Union[str, Path]] = None,
embeddings_path: Optional[Union[str, Path]] = None,
scale_path: Optional[Union[str, Path]] = None,
chunk_size: int = 8,
replan_stride: int = 8,
num_ode_steps: int = 10,
):
if _lib is None:
raise RuntimeError(f"libparvulus.so not loaded. Check: {LIB_PATH}")
self.device_id = device_id
self.chunk_size = chunk_size
self.replan_stride = replan_stride
self.num_ode_steps = num_ode_steps
if weights_path is None:
weights_path = C_ENGINE_DIR / "parvulus_model_weights.bin"
if embeddings_path is None:
embeddings_path = C_ENGINE_DIR / "task_embeddings.bin"
if scale_path is None:
scale_path = C_ENGINE_DIR / "action_scale.bin"
weights_path = Path(weights_path)
embeddings_path = Path(embeddings_path)
scale_path = Path(scale_path)
if not weights_path.exists():
print(f"[ParvulusEngine] Generating {weights_path.name} from SafeTensors...")
import sys
if str(C_ENGINE_DIR) not in sys.path:
sys.path.insert(0, str(C_ENGINE_DIR))
from export_from_safetensors import export_binary_weights, export_action_scale, export_task_embeddings
st_path = C_ENGINE_DIR.parent / "parvulus.safetensors"
if not st_path.exists():
st_path = C_ENGINE_DIR.parent / "model.safetensors"
export_binary_weights(str(st_path), str(weights_path))
if not scale_path.exists():
export_action_scale(str(C_ENGINE_DIR.parent / "action_scale.json"), str(scale_path))
if not embeddings_path.exists() and (C_ENGINE_DIR / "canonical_tasks.json").exists():
export_task_embeddings(str(C_ENGINE_DIR / "canonical_tasks.json"), str(st_path), str(embeddings_path))
self.handle = _lib.parvulus_create(self.device_id)
if not self.handle:
raise RuntimeError(f"Failed to initialize Parvulus engine on device {self.device_id}")
ok = _lib.parvulus_load(
self.handle,
str(weights_path).encode("utf-8"),
str(embeddings_path).encode("utf-8"),
str(scale_path).encode("utf-8"),
)
if not ok:
_lib.parvulus_destroy(self.handle)
self.handle = None
raise RuntimeError("Failed to load Parvulus engine weights!")
self.current_task_desc: Optional[str] = None
self.action_queue = []
self.stride_counter: int = 0
self._actions_buffer = np.zeros(56, dtype=np.float32)
self._onnx_session = None
self._tokenizer = None
self._torch_text_encoder = None
def __del__(self):
if hasattr(self, "handle") and self.handle is not None and _lib is not None:
_lib.parvulus_destroy(self.handle)
self.handle = None
def set_task_vector(self, task_vector: np.ndarray) -> bool:
"""Sets the active 256-dim task conditioning vector on the GPU."""
vec_c = np.ascontiguousarray(task_vector.flatten()[:256], dtype=np.float32)
vec_ptr = vec_c.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
return bool(_lib.parvulus_set_task_vector(self.handle, vec_ptr))
def encode_text(self, text: str) -> np.ndarray:
"""Encodes arbitrary natural language prompts or perturbations into 256-dim vector."""
onnx_path = C_ENGINE_DIR / "parvulus_text_encoder.onnx"
if not onnx_path.exists() and os.environ.get("PARVULUS_HF_REPO"):
try:
from huggingface_hub import hf_hub_download
hf_hub_download(
repo_id=os.environ["PARVULUS_HF_REPO"],
filename="parvulus_text_encoder.onnx",
subfolder="c_engine",
local_dir=str(C_ENGINE_DIR.parent),
)
except Exception:
pass
if onnx_path.exists():
try:
import onnxruntime as ort
from transformers import AutoTokenizer
if self._onnx_session is None:
available = ort.get_available_providers()
providers = [p for p in ["CUDAExecutionProvider", "CPUExecutionProvider"] if p in available]
self._onnx_session = ort.InferenceSession(str(onnx_path), providers=providers)
self._tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
tokens = self._tokenizer([text], padding=True, truncation=True, return_tensors="np")
ort_out = self._onnx_session.run(
None,
{
"input_ids": tokens["input_ids"].astype(np.int64),
"attention_mask": tokens["attention_mask"].astype(np.int64),
},
)[0]
return ort_out[0].astype(np.float32)
except Exception:
pass
try:
from text_encoder import ParvulusTextEncoder
from safetensors.torch import load_file
if self._torch_text_encoder is None:
te = ParvulusTextEncoder().cpu().eval()
st_path = C_ENGINE_DIR.parent / "parvulus.safetensors"
if st_path.exists():
weights = load_file(str(st_path))
proj_dict = {
"0.weight": weights["text_encoder.proj.0.weight"],
"0.bias": weights["text_encoder.proj.0.bias"],
"2.weight": weights["text_encoder.proj.2.weight"],
"2.bias": weights["text_encoder.proj.2.bias"],
}
te.proj.load_state_dict(proj_dict)
self._torch_text_encoder = te
with torch.no_grad():
emb = self._torch_text_encoder([text]).cpu().numpy()[0]
return emb.astype(np.float32)
except Exception as e:
raise RuntimeError(
f"Failed to dynamically encode text prompt '{text}': {e}. "
"Ensure onnxruntime or transformers/torch is installed."
)
def reset(
self,
task_description: Optional[str] = None,
task_vector: Optional[np.ndarray] = None,
):
if task_vector is not None:
self.set_task_vector(task_vector)
self.current_task_desc = task_description or "<vector>"
elif task_description is not None:
self.current_task_desc = task_description
task_bytes = task_description.encode("utf-8")
if not _lib.parvulus_set_task_string(self.handle, task_bytes):
encoded_vec = self.encode_text(task_description)
self.set_task_vector(encoded_vec)
self.action_queue.clear()
self.stride_counter = 0
def predict_chunk(
self,
agentview_img: np.ndarray,
wrist_img: np.ndarray,
task_description: Optional[str] = None,
task_vector: Optional[np.ndarray] = None,
custom_z0: Optional[np.ndarray] = None,
) -> np.ndarray:
if task_vector is not None:
self.set_task_vector(task_vector)
elif task_description is not None:
self.reset(task_description=task_description)
agent_c = np.ascontiguousarray(agentview_img, dtype=np.uint8)
wrist_c = np.ascontiguousarray(wrist_img, dtype=np.uint8)
agent_ptr = agent_c.ctypes.data_as(ctypes.POINTER(ctypes.c_uint8))
wrist_ptr = wrist_c.ctypes.data_as(ctypes.POINTER(ctypes.c_uint8))
out_ptr = self._actions_buffer.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
z0_ptr = None
if custom_z0 is not None:
z0_c = np.ascontiguousarray(custom_z0, dtype=np.float32)
z0_ptr = z0_c.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
ok = _lib.parvulus_predict_chunk(
self.handle,
agent_ptr,
wrist_ptr,
None,
out_ptr,
self.num_ode_steps,
z0_ptr,
)
if not ok:
raise RuntimeError("Parvulus predict_chunk failed!")
return self._actions_buffer.reshape(8, 7).copy()
def predict_chunk_with_emb(
self,
agentview_img: np.ndarray,
wrist_img: np.ndarray,
text_embedding: np.ndarray,
custom_z0: Optional[np.ndarray] = None,
) -> np.ndarray:
"""Predicts action chunk using a directly provided 256-dim text embedding vector."""
return self.predict_chunk(
agentview_img,
wrist_img,
task_vector=text_embedding,
custom_z0=custom_z0,
)
def register_task_embedding(self, task_description: str, text_embedding: np.ndarray) -> bool:
"""Dynamically registers or updates a task embedding vector in engine memory."""
task_bytes = task_description.encode("utf-8")
emb_c = np.ascontiguousarray(text_embedding.flatten()[:256], dtype=np.float32)
emb_ptr = emb_c.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
return bool(_lib.parvulus_register_task_embedding(self.handle, task_bytes, emb_ptr))
def step(self, agentview_img: np.ndarray, wrist_img: np.ndarray) -> np.ndarray:
if not self.action_queue or self.stride_counter >= self.replan_stride:
chunk = self.predict_chunk(agentview_img, wrist_img)
self.action_queue = list(chunk[: self.replan_stride])
self.stride_counter = 0
action = self.action_queue.pop(0)
self.stride_counter += 1
return action
# Aliases for backward compatibility
ParvulusPureCUDAEngine = ParvulusEngine
ParvulusCUDAEngine = ParvulusEngine
class ParvulusCUDAGraphEngine:
"""CUDA Graph replay engine for PyTorch models.
Pre-allocates static GPU buffers, captures the complete ODE Transformer loop into
a CUDA execution graph, and replays it with zero host kernel launch overhead.
"""
def __init__(
self,
device_id: int = 0,
model_path: Optional[Union[str, Path]] = None,
embeddings_path: Optional[Union[str, Path]] = None,
scale_path: Optional[Union[str, Path]] = None,
chunk_size: int = 8,
replan_stride: int = 8,
num_ode_steps: int = 10,
):
self.device = torch.device(f"cuda:{device_id}")
self.chunk_size = chunk_size
self.replan_stride = replan_stride
self.num_ode_steps = num_ode_steps
if model_path is None:
model_path = C_ENGINE_DIR / "parvulus_traced_model.pt"
if embeddings_path is None:
embeddings_path = C_ENGINE_DIR / "task_embeddings.bin"
if scale_path is None:
scale_path = C_ENGINE_DIR / "action_scale.bin"
import struct
with open(scale_path, "rb") as f:
min_vals = struct.unpack("<7f", f.read(28))
max_vals = struct.unpack("<7f", f.read(28))
self.action_min = torch.tensor(min_vals, device=self.device, dtype=torch.float32).view(1, 1, 7)
self.action_max = torch.tensor(max_vals, device=self.device, dtype=torch.float32).view(1, 1, 7)
self.task_embeddings: Dict[str, torch.Tensor] = {}
with open(embeddings_path, "rb") as f:
num_tasks = struct.unpack("<I", f.read(4))[0]
for _ in range(num_tasks):
s_len = struct.unpack("<H", f.read(2))[0]
task_str = f.read(s_len).decode("utf-8")
emb = struct.unpack("<256f", f.read(1024))
self.task_embeddings[task_str] = torch.tensor(emb, device=self.device, dtype=torch.float32).view(1, 256)
if Path(model_path).exists():
self.model = torch.jit.load(str(model_path), map_location=self.device).eval()
self.static_obs = torch.zeros(1, 6, 256, 256, device=self.device, dtype=torch.float32)
self.static_text = torch.zeros(1, 256, device=self.device, dtype=torch.float32)
self.static_z0 = torch.zeros(1, 8, 7, device=self.device, dtype=torch.float32)
self.static_out = torch.zeros(1, 8, 7, device=self.device, dtype=torch.float32)
s = torch.cuda.Stream(device=self.device)
s.wait_stream(torch.cuda.current_stream(device=self.device))
with torch.cuda.stream(s):
for _ in range(5):
_ = self.model(self.static_obs, self.static_text, self.static_z0)
torch.cuda.current_stream(device=self.device).wait_stream(s)
self.graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(self.graph, stream=s):
self.static_out = self.model(self.static_obs, self.static_text, self.static_z0)
else:
self.model = None
self.graph = None
self.current_task_desc: Optional[str] = None
self.action_queue = []
self.stride_counter: int = 0
def reset(self, task_description: Optional[str] = None):
if task_description is not None:
self.current_task_desc = task_description
if self.graph is not None and task_description in self.task_embeddings:
self.static_text.copy_(self.task_embeddings[task_description])
self.action_queue.clear()
self.stride_counter = 0
def predict_chunk(
self,
agentview_img: np.ndarray,
wrist_img: np.ndarray,
task_description: Optional[str] = None,
) -> np.ndarray:
if self.graph is None:
raise RuntimeError("ParvulusCUDAGraphEngine requires a valid TorchScript model.")
if task_description is not None:
self.reset(task_description)
combined_np = np.stack([agentview_img, wrist_img], axis=0)
combined_t = torch.from_numpy(combined_np).to(self.device, non_blocking=True).float()
combined_t = combined_t.flip(1, 2)
combined_t = (combined_t / 127.5) - 1.0
self.static_obs.copy_(combined_t.permute(0, 3, 1, 2).reshape(1, 6, 256, 256))
self.static_z0.normal_()
self.graph.replay()
actions = (self.static_out + 1.0) * 0.5 * (self.action_max - self.action_min) + self.action_min
return actions.squeeze(0).detach().cpu().numpy()
def step(self, agentview_img: np.ndarray, wrist_img: np.ndarray) -> np.ndarray:
if not self.action_queue or self.stride_counter >= self.replan_stride:
chunk = self.predict_chunk(agentview_img, wrist_img)
self.action_queue = list(chunk[: self.replan_stride])
self.stride_counter = 0
action = self.action_queue.pop(0)
self.stride_counter += 1
return action