paddle.zeros对于paddle.CUDAPinnedPlace()不生效

[CSWYF3634076]

bug描述 Describe the Bug

paddle.zeros对于paddle.CUDAPinnedPlace()不生效
paddle.to_tensor指定dtype后paddle.CUDAPinnedPlace()不生效

import os
import time
import psutil
import numpy as np
import paddle
from pynvml import (
    nvmlInit,
    nvmlShutdown,
    nvmlDeviceGetHandleByIndex,
    nvmlDeviceGetMemoryInfo,
)

GPU_ID = 0  # 物理 2 号卡

# 初始化 NVML
nvmlInit()
handle = nvmlDeviceGetHandleByIndex(GPU_ID)

def cpu_mem_gb():
    return psutil.virtual_memory().used / 1024**3

def gpu_mem_gb():
    info = nvmlDeviceGetMemoryInfo(handle)
    return info.used / 1024**3

print(f"Before CPU RAM: {cpu_mem_gb():.2f} GB")
print(f"Before GPU{GPU_ID} VRAM: {gpu_mem_gb():.2f} GB")

# 建议同时限制 Paddle 只用 2 号卡，避免起别的卡 context
paddle.device.set_device(f"gpu:{GPU_ID}")

# 分配普通 CPU 内存
arr = np.empty([1024, 1024, 256, 5], dtype="float32")

# 拷到 pinned memory
x = paddle.to_tensor(arr, place=paddle.CUDAPinnedPlace())
x_bf16 = paddle.to_tensor(arr, place=paddle.CUDAPinnedPlace(), dtype=paddle.bfloat16)
cpu_data = paddle.zeros(arr.shape, device=paddle.CUDAPinnedPlace())

print(f"x.place:",x.place)
print(f"x_bf16.place:",x_bf16.place)
print(f"cpu_data.place:",cpu_data.place)
print(f"After  CPU RAM: {cpu_mem_gb():.2f} GB")
print(f"After  GPU{GPU_ID} VRAM: {gpu_mem_gb():.2f} GB")

# print("Sleeping 60s... You can also run: nvidia-smi -i 2 -l 1")
# time.sleep(60)

nvmlShutdown()

其他补充信息 Additional Supplementary Information

No response

[CSWYF3634076]

@YqGe585 麻烦看一下这个？

[YqGe585]

cpu_data = paddle.zeros(arr.shape, device=paddle.CUDAPinnedPlace(), pin_memory=True)
这样使用即可。