checkpoint-engine 0.3.2__tar.gz → 0.3.4__tar.gz

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: checkpoint-engine
-Version: 0.3.2
+Version: 0.3.4
 Summary: checkpoint-engine is a lightweight, decoupling and efficient weight update middleware
 Project-URL: Homepage, https://github.com/MoonshotAI/checkpoint-engine
 Project-URL: Repository, https://github.com/MoonshotAI/checkpoint-engine

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/checkpoint_engine/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.3.2'
-__version_tuple__ = version_tuple = (0, 3, 2)
+__version__ = version = '0.3.4'
+__version_tuple__ = version_tuple = (0, 3, 4)
 
-__commit_id__ = commit_id = 'g4a73109a3'
+__commit_id__ = commit_id = 'g15446dd22'

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/checkpoint_engine/pin_memory.py

@@ -209,7 +209,9 @@ def _inplace_pin_memory(files: list[str], rank: int | None = None) -> list[Memor
             torch.cuda.set_device(device_index)
             cudart = torch.cuda.cudart()
             r = cudart.cudaHostRegister(t.data_ptr(), t.numel() * t.element_size(), 0)
-            assert r == 0, f"pin memory error, error code: {r}"
+            if r != 0:
+                error_msg = cudart.cudaGetErrorString(r)
+                raise RuntimeError(f"pin memory error, error code: {r}, error message: {error_msg}")
 
         # TODO: should only support /dev/shm? but we found files in disk also work?
         size = os.stat(file_path).st_size
@@ -254,6 +256,12 @@ def _inplace_pin_memory(files: list[str], rank: int | None = None) -> list[Memor
         # Remove the file after successfully loading. This will avoid doubling the memory usage.
         # We assume files in /dev/shm/ are temporary files. So it's safe to remove them after loading.
         os.remove(file_path)
+        if not metas:
+            # TODO: should we still return this buffer?
+            assert buffer.nbytes == 0, f"buffer nbytes {buffer.nbytes} should be 0"
+            logger.warning(f"[rank{rank}] no metas found in {file_path}, skip pin memory")
+            return MemoryBuffer(buffer=buffer, size=buffer.nbytes, metas=[], manually_pinned=False)
+
         _pin(buffer)
         logger.info(
             f"[rank{rank}] inplace pin memory for file {file_path} finished, size {buffer.nbytes / 1024 / 1024:.2f}MiB"

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/checkpoint_engine/ps.py

@@ -391,7 +391,11 @@ class ParameterServer:
                 )
                 cudart = torch.cuda.cudart()
                 r = cudart.cudaHostUnregister(t.data_ptr())
-                assert r == 0, f"unpin memory error, error code: {r}"
+                if r != 0:
+                    error_msg = cudart.cudaGetErrorString(r)
+                    raise RuntimeError(
+                        f"unpin memory error, error code: {r}, error message: {error_msg}"
+                    )
 
             # if the checkpoint is pinned by cudaHostRegister manually, we need to unpin it manually
             try:
@@ -407,7 +411,13 @@ class ParameterServer:
             del self._memory_pool[checkpoint_name]
         # see https://github.com/pytorch/pytorch/blob/31d5c675394705f8a6bc767f80ae14bf4f01246b/torch/csrc/cuda/Module.cpp#L2018
         # this works by using torch>=2.5.0
-        torch._C._host_emptyCache()
+        if self.device_manager.device_type == "cuda":
+            torch._C._host_emptyCache()
+        else:
+            # torch._C._host_emptyCache() is not supported on NPU, so we call gc.collect() to empty host cache.
+            import gc
+
+            gc.collect()
 
     def gather_metas(self, checkpoint_name: str):
         """

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/checkpoint_engine/worker.py

@@ -10,6 +10,9 @@ import zmq
 from checkpoint_engine.device_utils import DeviceManager, npu_generate_uuid
 
 
+_WEIGHTS_TYPE = list[tuple[str, torch.Tensor]]
+
+
 def _rebuild_ipc(handle: tuple[Callable, tuple], device_id: int | None = None) -> torch.Tensor:
     func, args = handle
     list_args = list(args)
@@ -29,11 +32,9 @@ class FlattenedTensorMetadata(TypedDict):
     offset: int
 
 
-def _extract_weights(
-    payload: list[FlattenedTensorMetadata], buffer: torch.Tensor
-) -> list[tuple[str, torch.Tensor]]:
+def _extract_weights(payload: list[FlattenedTensorMetadata], buffer: torch.Tensor) -> _WEIGHTS_TYPE:
     assert buffer is not None
-    weights: list[tuple[str, torch.Tensor]] = []
+    weights: _WEIGHTS_TYPE = []
     for item in payload:
         shape = item["shape"]
         if isinstance(shape, list | tuple):
@@ -166,12 +167,31 @@ class VllmColocateWorkerExtension:
             self.device = torch.device(f"npu:{self.local_rank}")
         assert self.device is not None
 
+        def _load_weights(weights: _WEIGHTS_TYPE):
+            # Load main model weights
+            self.model_runner.model.load_weights(weights)
+            # Load drafter model weights if MTP/speculative decoding is enabled
+            if (
+                getattr(self.model_runner, "drafter", None) is not None
+                and getattr(self.model_runner.drafter, "model", None) is not None
+            ):
+                self.model_runner.drafter.model.load_weights(weights=weights)
+
+        def _post_hook():
+            process_weights_after_loading(self.model_runner.model, self.model_config, self.device)
+            # Also trigger drafter model's post processing if MTP is enabled
+            if (
+                getattr(self.model_runner, "drafter", None) is not None
+                and getattr(self.model_runner.drafter, "model", None) is not None
+            ):
+                process_weights_after_loading(
+                    self.model_runner.drafter.model, self.model_config, self.device
+                )
+
         update_weights_from_ipc(
             self._zmq_ctx,
             zmq_handles[self._device_uuid],
             device_id=self.device.index,
-            run=self.model_runner.model.load_weights,
-            post_hook=lambda: process_weights_after_loading(
-                self.model_runner.model, self.model_config, self.device
-            ),
+            run=_load_weights,
+            post_hook=_post_hook,
         )

{checkpoint_engine-0.3.2 → checkpoint_engine-0.3.4}/checkpoint_engine.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: checkpoint-engine
-Version: 0.3.2
+Version: 0.3.4
 Summary: checkpoint-engine is a lightweight, decoupling and efficient weight update middleware
 Project-URL: Homepage, https://github.com/MoonshotAI/checkpoint-engine
 Project-URL: Repository, https://github.com/MoonshotAI/checkpoint-engine