Skip to content

[BugFix] [RL] Fix cpu cache for rl #7764

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Jiang-Jia-Jun merged 4 commits into from
May 11, 2026
Merged

[BugFix] [RL] Fix cpu cache for rl #7764

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b8d5bb1
[BugFix] [RL] Fix cpu cache for rl
liyonghua0910 May 9, 2026
08e0884
[fix] fix stable ci
liyonghua0910 May 9, 2026
303ff3a
[fix] fix stable ci
liyonghua0910 May 9, 2026
f56f8b3
[fix] fix stable ci
liyonghua0910 May 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 13 additions & 9 deletions fastdeploy/cache_manager/cache_transfer_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -540,24 +540,28 @@ def _init_gpu_cache(self):
logger.info("GPU KV cache is initialized")

def _clear_gpu_cache(self):

if self.create_cache_tensor:

This comment was marked as outdated.

Sign in to view

logger.debug("Waiting for gpu runner to unlink cuda ipc")
while self.cache_ready_signal.value[self.rank] != 0:
time.sleep(0.1)
logger.debug("Stop waiting! gpu runner has unlinked cuda ipc")
self.gpu_cache_kvs.clear()
self.gpu_cache_k_tensors.clear()
self.gpu_cache_v_tensors.clear()
if hasattr(self, "gpu_cache_scales_k_tensors"):
self.gpu_cache_scales_k_tensors.clear()
if hasattr(self, "gpu_cache_scales_v_tensors"):
self.gpu_cache_scales_v_tensors.clear()
paddle.set_flags({"FLAGS_selected_gpus": f"{self.device}"})
paddle.device.cuda.empty_cache()
else:
for name, tensor in self.gpu_cache_kvs.items():
unset_data_ipc(tensor, name, True, False)
logger.debug("Successfully unlinked gpu caches cuda ipc")

self.gpu_cache_kvs.clear()
self.gpu_cache_k_tensors.clear()
self.gpu_cache_v_tensors.clear()
if hasattr(self, "gpu_cache_scales_k_tensors"):
self.gpu_cache_scales_k_tensors.clear()
if hasattr(self, "gpu_cache_scales_v_tensors"):
self.gpu_cache_scales_v_tensors.clear()
paddle.set_flags({"FLAGS_selected_gpus": f"{self.device}"})
paddle.device.cuda.empty_cache()
Copy link
Copy Markdown

@PaddlePaddle-bot PaddlePaddle-bot May 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

❓ 疑问 当 create_cache_tensor=True 时：

  • _clear_gpu_cache（transfer manager 侧）等待 cache_ready_signal.value[rank] == 0，日志为 "Waiting for gpu runner to unlink"
  • gpu_model_runner.clear_cache（GPU runner 侧）同样等待 cache_ready_signal.value[local_rank] == 0，日志为 "Waiting for cache transfer manager to unlink"

两侧均声称在等待对方，但此分支下均未将 signal 置 0。请确认：

  1. create_cache_tensor=True 时 signal 的初始值是否始终为 0（即等待为空操作）？
  2. 若 signal 在 IPC 建立阶段被置为非零，谁负责将其清零？是否存在两侧互相等待的死锁风险？


if not self.create_cache_tensor:
self.cache_ready_signal.value[self.rank] = 0

while np.sum(self.cache_ready_signal.value) != 0:
Expand Down
15 changes: 11 additions & 4 deletions fastdeploy/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3010,10 +3010,17 @@ def clear_cache(self, profile=False):
)
local_rank = self.local_rank % self.parallel_config.tensor_parallel_size

if not create_cache_tensor:
for name, tensor in self.cache_kvs_map.items():
unset_data_ipc(tensor, name, True, False)
self.cache_ready_signal.value[local_rank] = 0
if not profile:

This comment was marked as outdated.

Sign in to view

Copy link
Copy Markdown

@PaddlePaddle-bot PaddlePaddle-bot May 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 建议 gpu_model_runner.py 属于 GPU 通用路径，architecture.md §A6 指出此类变更通常需同步到其他硬件 ModelRunner。

clear_cache 中原有的 if not create_cache_tensor: 条件反转 Bug，是否在 xpu_model_runner.py、dcu_model_runner.py、gcu_model_runner.py、hpu_model_runner.py、iluvatar_model_runner.py、metax_model_runner.py 中也存在相同问题？若存在，建议一并修复，避免其他硬件 CI 在 RL 场景下出现类似的 CPU cache 清理异常。

if create_cache_tensor:
if self.fd_config.cache_config.num_cpu_blocks > 0:
logger.info("Waiting for cache transfer manager to unlink cuda ipc")
while self.cache_ready_signal.value[local_rank] != 0:
time.sleep(0.1)
logger.info("Stop waiting! cache transfer manager has unlinked cuda ipc")
else:
for name, tensor in self.cache_kvs_map.items():
unset_data_ipc(tensor, name, True, False)
self.cache_ready_signal.value[local_rank] = 0

self.cache_kvs_map.clear()
self.share_inputs.pop("caches", None)
Expand Down
Loading