[wip] one less allocation/copy in Pipe

torchstore/storage_volume.py

@@ -11,8 +11,10 @@
 import torch
 from monarch.actor import Actor, endpoint
 
-from torchstore.transport.buffers import TransportBuffer
-
+from torchstore.transport.buffers import (
+    create_default_transport_buffer,
+    TransportBuffer,
+)
 from torchstore.transport.pipe import Request, TensorSlice
 from torchstore.utils import assemble_global_tensor, spawn_actors
 
@@ -59,10 +61,8 @@ async def put(
         await self.store.put(key, transport_buffer, request)
 
     @endpoint
-    async def get(
-        self, key: str, transport_buffer: TransportBuffer, request: Request
-    ) -> TransportBuffer:
-        return await self.store.get(key, transport_buffer, request)
+    async def get(self, key: str, request: Request) -> TransportBuffer:
+        return await self.store.get(key, request)
 
     @endpoint
     async def get_meta(
@@ -86,9 +86,7 @@ async def put(
         """Store data in the storage backend."""
         raise NotImplementedError()
 
-    async def get(
-        self, key: str, transport_buffer: TransportBuffer, request: Request
-    ) -> TransportBuffer:
+    async def get(self, key: str, request: Request) -> TransportBuffer:
         """Retrieve data from the storage backend."""
         raise NotImplementedError()
 
@@ -201,13 +199,13 @@ async def put(
 
         self.kv[key] = tensor
 
-    async def get(
-        self, key: str, transport_buffer: TransportBuffer, request: Request
-    ) -> TransportBuffer:
+    async def get(self, key: str, request: Request) -> TransportBuffer:
 
         if key not in self.kv:
             raise KeyError(f"Key '{key}' not found. {list(self.kv.keys())=}")
 
+        transport_buffer = create_default_transport_buffer()
+
         # TODO: clean up
         val = self.kv[key]
         if isinstance(val, dict) and "obj" in val:
@@ -216,7 +214,7 @@ async def get(
             return transport_buffer
 
         if request.tensor_slice is None:
-            await transport_buffer.write_from(self.kv[key])
+            transport_buffer.from_contiguous_tensor(self.kv[key])
             return transport_buffer
 
         # TODO:
@@ -227,7 +225,7 @@ async def get(
 
         for shard in self.kv[key].values():
             if shard["slice"] == request.tensor_slice:
-                await transport_buffer.write_from(shard["tensor"])
+                transport_buffer.from_contiguous_tensor(shard["tensor"])
                 return transport_buffer
 
         raise RuntimeError(f"Tensor slice {request.tensor_slice} not found in {key}")

torchstore/transport/buffers.py

@@ -4,6 +4,10 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from __future__ import annotations
+
+import functools
+
 import logging
 import os
 from typing import Any, Dict, List, Optional, Tuple, Union
@@ -30,13 +34,21 @@ def RDMABuffer(*args: Any, **kwargs: Any) -> Any:
 # assert RDMA_CHUNK_SIZE_MB <= 1024, "Monarch does not support 1gb chunks via rdma"
 
 
+@functools.cache
 def rdma_available() -> bool:
     rdma_enabled = (
-        os.environ.get("TORCHSTORE_RDMA_ENABLED", "0") == "1"
+        os.environ.get("TORCHSTORE_RDMA_ENABLED", "1") == "1"
     )  # TODO: enable on this build
     return rdma_enabled and monarch_rdma_available()
 
 
+def create_default_transport_buffer() -> TransportBuffer:
+    if rdma_available():
+        return RDMATransportBuffer()
+    else:
+        return MonarchTransportBuffer()
+
+
 class TransportBuffer:
     finalize: bool = False
     is_object: bool = False
@@ -49,10 +61,7 @@ def update(self, other_buffer: "TransportBuffer") -> None:
         self.objects = other_buffer.objects
         self.requires_meta = other_buffer.requires_meta
 
-    def allocate(self, tensor_like: Union[torch.Tensor, Tuple]) -> None:
-        """Allocates internal buffers based on either an existing tensor
-        or a Tuple of (shape, dtype)
-        """
+    def from_contiguous_tensor(self, tensor: torch.Tensor) -> None:
         raise NotImplementedError()
 
     async def read_into(self, tensor: Optional[torch.Tensor]) -> torch.Tensor:
@@ -190,6 +199,15 @@ async def write_from(self, tensor: Optional[torch.Tensor]) -> None:
         for idx, chunk in enumerate(chunked_byte_view):
             await self.rdma_buffers[idx].write_from(chunk)
 
+    def from_contiguous_tensor(self, tensor: torch.Tensor) -> None:
+        assert tensor.is_contiguous(), "Tensor must be contiguous"
+        self.shape = tensor.shape
+        self.dtype = tensor.dtype
+        self.dim = tensor.dim()
+        byte_view_chunks = self._create_byte_views_from_tensor(tensor)
+        self.tensor_refs = [torch.empty_like(chunk) for chunk in byte_view_chunks]
+        self.rdma_buffers = [RDMABuffer(chunk) for chunk in self.tensor_refs]
+
 
 class MonarchTransportBuffer(TransportBuffer):
     """This interface is mostly a noop, intended to be used with Monarch's regular RPC.
@@ -224,3 +242,6 @@ async def write_from(self, tensor: Optional[torch.Tensor]) -> None:
     def update(self, other_buffer: "TransportBuffer") -> None:
         super().update(other_buffer)
         self.tensor = other_buffer.tensor
+
+    def from_contiguous_tensor(self, tensor: torch.Tensor) -> None:
+        self.tensor = tensor

torchstore/transport/pipe.py

@@ -145,9 +145,9 @@ def create_transport_buffer(self) -> TransportBuffer:
     async def put_to_storage_volume(self, key, request: Request):
         transport_buffer = self.create_transport_buffer()
         tensor = request.tensor_val
-
-        transport_buffer.allocate(tensor)
-        await transport_buffer.write_from(tensor)
+        if not tensor.is_contiguous():
+            tensor = tensor.contiguous()
+        transport_buffer.from_contiguous_tensor(tensor)
 
         # transporting tensors is handled by the buffer, so we don't want to send it
         # via monarch RPC since that would generate considerable overhead
@@ -156,24 +156,8 @@ async def put_to_storage_volume(self, key, request: Request):
         )
 
     async def get_from_storage_volume(self, key, request: Request):
-
-        transport_buffer = self.create_transport_buffer()
-
-        # Certain buffers (RDMA) need to know the size of the tensor
-        # so we can allocate the right amount of memory locally.
-        # This can be avoided if the request contains a tensor slice.
-        # Could likely be optimized away in the future.
-        if transport_buffer.requires_meta and request.tensor_val is None:
-            meta = await self.storage_volume.get_meta.call_one(key, request.meta_only())
-            transport_buffer.allocate(meta)
-        else:
-            transport_buffer.allocate(request.tensor_val)
-
-        # TODO: consider placing the buffer inside the request or vice versa
-        transport_buffer.update(
-            await self.storage_volume.get.call_one(
-                key, transport_buffer, request.meta_only()
-            )
+        transport_buffer = await self.storage_volume.get.call_one(
+            key, request.meta_only()
         )
 
         if transport_buffer.is_object: