feat: default disable hugepage

Author: niushengxiao

lightllm/server/multi_level_kv_cache/disk_cache_worker.py

@@ -76,8 +76,7 @@ def __init__(
         )
 
     def _prepare_tensor(self, tensor: torch.Tensor) -> torch.Tensor:
-        num_page, num_layer = tensor.shape[0], tensor.shape[1]
-        return tensor.reshape(num_page, num_layer, -1)
+        return tensor.flatten(1).view(dtype=torch.uint8)
 
     def run(self) -> None:
         while True:

lightllm/utils/envs_utils.py

@@ -199,3 +199,15 @@ def get_radix_tree_merge_update_delta() -> int:
 @lru_cache(maxsize=None)
 def get_disk_cache_prompt_limit_length():
     return int(os.getenv("LIGHTLLM_DISK_CACHE_PROMPT_LIMIT_LENGTH", 10000))
+
+
+@lru_cache(maxsize=None)
+def enable_huge_page():
+    """
+    大页模式：启动后可大幅缩短cpu kv cache加载时间
+    "sudo sed -i 's/^GRUB_CMDLINE_LINUX=\"/& default_hugepagesz=1G \
+        hugepagesz=1G hugepages={需要启用的大页容量}/' /etc/default/grub"
+    "sudo update-grub"
+    "sudo reboot"
+    """
+    return enable_env_vars("LIGHTLLM_HUGE_PAGE_ENABLE")

lightllm/utils/kv_cache_utils.py

@@ -8,7 +8,7 @@
 import numpy as np
 import triton
 from functools import lru_cache
-from lightllm.utils.envs_utils import get_env_start_args
+from lightllm.utils.envs_utils import get_env_start_args, enable_huge_page
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.config_utils import get_num_key_value_heads, get_head_dim, get_layer_num, get_model_type
 from typing import List, Tuple, Optional
@@ -93,7 +93,7 @@ def create_shm_kv_cache_ptr() -> int:
     args = get_env_start_args()
     key = args.cpu_kv_cache_shm_id
     requested_size = calcu_cpu_cache_meta().calcu_size()
-    use_hugetlb = True
+    use_hugetlb = enable_huge_page()
 
     # 计算大页大小（默认从 /proc/meminfo 读取 Hugepagesize）
     def _get_default_hugepage_size() -> int:
@@ -109,16 +109,19 @@ def _get_default_hugepage_size() -> int:
             pass
         return 2 * 1024 * 1024  # fallback 2MB
 
-    # 向上对齐到大页大小
-    huge_sz = _get_default_hugepage_size()
-    size_to_alloc = triton.cdiv(requested_size, huge_sz) * huge_sz
     shmflg = 0o666 | 0o1000  # 权限和 IPC_CREAT 标志
     if use_hugetlb:
+        # 向上对齐到大页大小
+        huge_sz = _get_default_hugepage_size()
+        size_to_alloc = triton.cdiv(requested_size, huge_sz) * huge_sz
         SHM_HUGETLB = 0o4000
         shmflg |= SHM_HUGETLB
         logger.info(
             f"Using SHM_HUGETLB, hugepage_size={huge_sz} bytes, requested={requested_size}, alloc={size_to_alloc}"
         )
+    else:
+        size_to_alloc = requested_size
+        logger.info(f"Using regular pages, requested={requested_size}, alloc={size_to_alloc}")
 
     # 优先尝试 HugeTLB 分配，失败则回退到普通页
     shmid = libc.shmget(key, size_to_alloc, shmflg)