rgpu 0.1.0__tar.gz

rgpu-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

rgpu-0.1.0/PKG-INFO

@@ -0,0 +1,35 @@
+Metadata-Version: 2.4
+Name: rgpu
+Version: 0.1.0
+Summary: A PyTorch device whose tensors live on a remote GPU
+Author: Yan Michalevsky
+License-Expression: Apache-2.0
+Project-URL: Documentation, https://rgpu.pages.dev/docs/
+Project-URL: Source, https://github.com/ymcrcat/rgpu
+Classifier: Programming Language :: Python :: 3
+Classifier: Operating System :: OS Independent
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.14
+Provides-Extra: test
+Requires-Dist: pytest>=8; extra == "test"
+Requires-Dist: torchvision; extra == "test"
+Dynamic: license-file
+
+# rgpu
+
+A PyTorch device whose tensors live on a remote GPU. Train on a laptop with
+`device="rgpu"` while the work runs on a GPU host you reach over SSH.
+
+```sh
+pip install rgpu
+```
+
+The GPU host runs `rgpu-opserver`; the laptop runs your script through
+`rgpu-run --server 127.0.0.1:9720 python train.py`. The protocol has no
+authentication, so keep the server on localhost and reach it through an SSH
+tunnel.
+
+Documentation: https://rgpu.pages.dev/docs/

rgpu-0.1.0/README.md

@@ -0,0 +1,15 @@
+# rgpu
+
+A PyTorch device whose tensors live on a remote GPU. Train on a laptop with
+`device="rgpu"` while the work runs on a GPU host you reach over SSH.
+
+```sh
+pip install rgpu
+```
+
+The GPU host runs `rgpu-opserver`; the laptop runs your script through
+`rgpu-run --server 127.0.0.1:9720 python train.py`. The protocol has no
+authentication, so keep the server on localhost and reach it through an SSH
+tunnel.
+
+Documentation: https://rgpu.pages.dev/docs/

rgpu-0.1.0/pyproject.toml

@@ -0,0 +1,39 @@
+[build-system]
+requires = ["setuptools>=77"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "rgpu"
+version = "0.1.0"
+description = "A PyTorch device whose tensors live on a remote GPU"
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+authors = [{ name = "Yan Michalevsky" }]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Operating System :: OS Independent",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+requires-python = ">=3.10"
+dependencies = ["torch>=2.14"]
+
+[project.urls]
+Documentation = "https://rgpu.pages.dev/docs/"
+Source = "https://github.com/ymcrcat/rgpu"
+
+[project.optional-dependencies]
+test = ["pytest>=8", "torchvision"]
+
+[project.scripts]
+rgpu-opserver = "rgpu.server.__main__:main"
+rgpu-run = "rgpu_run:main"
+
+[tool.setuptools]
+# A top-level module on purpose: the launcher only opens an ssh tunnel and
+# spawns a child, so importing the rgpu package - which registers the device
+# and pulls in torch, about three seconds - would be pure startup cost.
+py-modules = ["rgpu_run"]
+
+[tool.setuptools.packages.find]
+include = ["rgpu*"]

rgpu-0.1.0/rgpu/__init__.py

@@ -0,0 +1,18 @@
+"""A PyTorch device whose tensors live on a remote GPU.
+
+    import rgpu, torch
+    x = torch.randn(1024, 1024, device="rgpu")   # the data is on the server
+    y = (x @ x).relu().sum().item()               # one round trip, at .item()
+
+Set RGPU_OPSERVER=host:port to reach rgpu-opserver (default 127.0.0.1:9720),
+normally through an ssh tunnel: the server has no authentication.
+"""
+
+from .device import register_device as _register_device
+
+_register_device()
+
+from .compile import compile_backend  # noqa: E402
+from .session import RemoteError, SessionLost, stats  # noqa: E402
+
+__all__ = ["RemoteError", "SessionLost", "compile_backend", "stats"]

rgpu-0.1.0/rgpu/autocast.py

@@ -0,0 +1,140 @@
+"""Autocast for rgpu: which ops run in half precision and which stay float32.
+
+PyTorch ships an autocast policy only for its own devices. A new device gets
+none, and under autocast every op fails until something is registered at its
+Autocast key. This registers a fallthrough for everything, then kernels for a
+documented subset of what PyTorch registers for AutocastCUDA: LOWER runs in
+the autocast dtype, FP32 widens half inputs back to float32, and BANNED
+raises the way CUDA does. Everything else falls through uncast.
+
+Falling through is right for most of what is missing, because much of CUDA's
+list is CompositeImplicitAutograd and decomposes into ops that are covered -
+einsum and layer_norm were checked and come out as they do on CUDA. These are
+the differences that remain, where rgpu is not CUDA:
+
+  - grid_sampler and the RNN cells (lstm_cell, gru_cell, rnn_tanh_cell,
+    rnn_relu_cell, and the fused _thnn_* forms), which CUDA runs in the
+    autocast dtype and rgpu leaves alone.
+  - dot, vdot, bilinear and conv_tbc, likewise on CUDA's lower-precision list
+    and not here.
+  - CUDA's whole "promote" category, which runs an op at the widest dtype
+    among its inputs: addcmul, addcdiv, atan2, cross, index_put, scatter_add
+    and the upsample_* family. On rgpu a mix of float16 and float32 inputs
+    reaches them unpromoted.
+"""
+
+import torch
+from torch.utils._pytree import tree_map
+
+# Ops that run in the autocast dtype (PyTorch's CUDA "lower precision" list).
+LOWER = [
+    "mm", "addmm", "bmm", "baddbmm", "addbmm", "addmv", "addr", "matmul", "mv",
+    "linear", "conv1d", "conv2d", "conv3d", "conv_transpose1d", "conv_transpose2d",
+    "conv_transpose3d", "convolution", "_convolution", "prelu", "chain_matmul",
+    "linalg_multi_dot", "scaled_dot_product_attention",
+]
+
+# Ops whose half-precision inputs are widened to float32 (the CUDA fp32 list).
+FP32 = [
+    "acos", "asin", "cosh", "erfinv", "exp", "expm1", "log", "log10", "log2",
+    "log1p", "reciprocal", "rsqrt", "sinh", "tan", "pow", "softplus", "layer_norm",
+    "group_norm", "norm", "nuclear_norm", "cosine_similarity", "poisson_nll_loss",
+    "cosine_embedding_loss", "nll_loss", "nll_loss2d", "hinge_embedding_loss",
+    "kl_div", "l1_loss", "smooth_l1_loss", "huber_loss", "mse_loss",
+    "margin_ranking_loss", "multilabel_margin_loss", "soft_margin_loss",
+    "triplet_margin_loss", "multi_margin_loss", "binary_cross_entropy_with_logits",
+    "dist", "pdist", "cdist", "renorm", "logsumexp", "softmax", "log_softmax",
+    "cross_entropy_loss", "cumprod", "cumsum", "prod", "sum",
+]
+
+# Ops CUDA refuses under autocast rather than casting, because they are
+# unsafe in half precision. Running one anyway would diverge silently from a
+# local GPU, so rgpu refuses them with the same advice.
+BANNED = {
+    "binary_cross_entropy": (
+        "torch.nn.functional.binary_cross_entropy and torch.nn.BCELoss are unsafe to "
+        "autocast. Many models use a sigmoid layer right before the binary cross "
+        "entropy layer. In this case, combine the two layers using "
+        "torch.nn.functional.binary_cross_entropy_with_logits or "
+        "torch.nn.BCEWithLogitsLoss. binary_cross_entropy_with_logits and "
+        "BCEWithLogits are safe to autocast."),
+}
+
+_KEY = torch._C.DispatchKey.AutocastPrivateUse1
+_libs = []
+
+
+def _kernel(op, target, eligible):
+    def kernel(*args, **kwargs):
+        dtype = target()
+
+        def cast(x):
+            if (isinstance(x, torch.Tensor) and x.device.type == "rgpu"
+                    and eligible(x.dtype) and x.dtype != dtype):
+                return x.to(dtype)
+            return x
+
+        with torch._C._ExcludeDispatchKeyGuard(torch._C.DispatchKeySet(_KEY)):
+            return op(*tree_map(cast, args), **tree_map(cast, kwargs))
+
+    return kernel
+
+
+def _refusing_kernel(message):
+    def kernel(*args, **kwargs):
+        raise RuntimeError(message)
+
+    return kernel
+
+
+def _has_out_argument(schema):
+    """True if any argument writes into a tensor the caller passed in.
+
+    An out= kernel's cast() would allocate a fresh cast tensor, compute into
+    that, and leave the caller's out tensor untouched - silently wrong
+    values, no exception. PyTorch's own CUDA autocast policy does not cover
+    out= variants either, so - like CUDA - they fall through to the uncast
+    path instead of getting a cast kernel here.
+    """
+    return any(a.alias_info is not None and a.alias_info.is_write for a in schema.arguments)
+
+
+def install():
+    if _libs:
+        return
+    everything = torch.library.Library("_", "IMPL")
+    everything.fallback(torch.library.fallthrough_kernel, "AutocastPrivateUse1")
+    lib = torch.library.Library("aten", "IMPL")
+
+    def lower_eligible(dt):
+        return dt.is_floating_point and dt != torch.float64
+
+    def fp32_eligible(dt):
+        return dt in (torch.float16, torch.bfloat16)
+
+    for names, target, eligible in [
+        (LOWER, lambda: torch.get_autocast_dtype("rgpu"), lower_eligible),
+        (FP32, lambda: torch.float32, fp32_eligible),
+    ]:
+        for name in names:
+            packet = getattr(torch.ops.aten, name, None)
+            if packet is None:
+                continue
+            for overload in packet.overloads():
+                op = getattr(packet, overload)
+                if _has_out_argument(op._schema):
+                    continue
+                qualified = name if overload == "default" else f"{name}.{overload}"
+                lib.impl(qualified, _kernel(op, target, eligible), "AutocastPrivateUse1")
+
+    for name, message in BANNED.items():
+        packet = getattr(torch.ops.aten, name, None)
+        if packet is None:
+            continue
+        # Every overload, out= included: this kernel computes nothing, so the
+        # reason out= variants are skipped above does not apply.
+        for overload in packet.overloads():
+            qualified = name if overload == "default" else f"{name}.{overload}"
+            lib.impl(qualified, _refusing_kernel(message), "AutocastPrivateUse1")
+
+    _libs.extend([everything, lib])

rgpu-0.1.0/rgpu/compile.py

@@ -0,0 +1,169 @@
+"""torch.compile for rgpu: ship the graph once, then one message per call.
+
+The backend hooks aot_autograd, which hands over forward and backward graphs
+of aten ops. Each is serialized with the same op names and argument encoding
+as eager ops, sent once, and compiled on the server. After that a call is a
+single CALL message, and it streams: output shapes come from the traced
+graph, so nothing has to wait for them.
+
+Compiled code works on the unwrapped meta tensors, never on RemoteTensor
+itself, which is why each tensor's id lives on its meta tensor: inputs are
+looked up by it and outputs are given one.
+"""
+
+import itertools
+import logging
+import operator
+
+import torch
+from torch._dynamo.backends.common import aot_autograd
+from torch.utils._pytree import tree_map
+
+from . import session, wire
+from .tensor import RemoteTensor, id_of, meta_like, register
+
+log = logging.getLogger("rgpu")
+_graph_ids = itertools.count(1)
+
+
+class Unshippable(Exception):
+    """A graph containing something the wire cannot carry."""
+
+
+class UnsupportedOp(Unshippable):
+    """An op in the graph that rgpu-opserver would refuse to run.
+
+    Unlike the other refusals this one has no eager fallback. Running the
+    graph through the ordinary dispatch path posts the very same op as a RUN,
+    and the server resolves ops by name and runs only aten ops, so it fails
+    there instead - later, and with a worse message. Raising at compile time
+    says the true thing: this cannot run on rgpu at all.
+    """
+
+
+def serialize(gm):
+    nodes, index = [], {}
+
+    def ref(a):
+        if isinstance(a, torch.fx.Node):
+            return wire.NodeRef(index[a])
+        if isinstance(a, torch.Tensor):
+            raise Unshippable("a constant tensor inside the graph")
+        if isinstance(a, torch.device):
+            # An rgpu tensor is a meta tensor on this side, so a graph traced
+            # through one allocates on "meta". On the server that means rgpu.
+            return wire.Dev("rgpu" if a.type in ("rgpu", "meta") else str(a))
+        if isinstance(a, (torch.SymInt, torch.SymFloat)):
+            raise Unshippable("dynamic shapes; compile with dynamic=False")
+        try:
+            wire.encode(a)
+        except wire.EncodeError as e:
+            raise Unshippable(str(e)) from None
+        return a
+
+    for n in gm.graph.nodes:
+        if n.op == "placeholder":
+            nodes.append(["placeholder"])
+        elif n.op == "call_function":
+            t = n.target
+            if t is operator.getitem:
+                # Not an op: how a graph picks one output of an op that has
+                # several, which aot_autograd writes for max_pool2d_with_indices,
+                # native_batch_norm and the rest. It travels as its own kind of
+                # node, so the server still runs nothing but aten ops.
+                src, i = n.args
+                if not isinstance(src, torch.fx.Node) or not isinstance(i, int):
+                    raise Unshippable(f"getitem with a {type(i).__name__} index")
+                nodes.append(["getitem", wire.NodeRef(index[src]), i])
+                index[n] = len(nodes) - 1
+                continue
+            if not isinstance(t, torch._ops.OpOverload):
+                # A plain Python callable: not shippable as part of a graph,
+                # but the ops it calls are, so eager still works.
+                raise Unshippable(f"{t} is not an aten op")
+            if t.namespace != "aten":
+                raise UnsupportedOp(
+                    f"{t._schema.name}.{t._overloadname} cannot run on rgpu: "
+                    "rgpu-opserver runs aten ops only, so a custom op has nowhere "
+                    "to run, compiled or eager")
+            nodes.append(["call", t._schema.name, t._overloadname,
+                          tree_map(ref, list(n.args)),
+                          {k: tree_map(ref, v) for k, v in n.kwargs.items()}])
+        elif n.op == "output":
+            outs = n.args[0] if isinstance(n.args[0], (list, tuple)) else [n.args[0]]
+            nodes.append(["output", [tree_map(ref, o) for o in outs]])
+        else:
+            raise Unshippable(f"a {n.op} node")
+        index[n] = len(nodes) - 1
+    return nodes
+
+
+def _output_values(gm):
+    out = next(n for n in gm.graph.nodes if n.op == "output")
+    outs = out.args[0] if isinstance(out.args[0], (list, tuple)) else [out.args[0]]
+    return [o.meta.get("val") if isinstance(o, torch.fx.Node) else o for o in outs]
+
+
+def _boxed(fn):
+    def run(args):
+        return fn(*args)
+    run._boxed_call = True
+    return run
+
+
+def _ship(gid, nodes, compiler, mode):
+    session.get().post(wire.COMPILE, gid, nodes, compiler, mode)
+
+
+def _eager(gm):
+    """Run a graph that cannot be shipped op by op, through the normal path."""
+    def run(*inner):
+        wrapped = [RemoteTensor(x) if isinstance(x, torch.Tensor) else x for x in inner]
+        outs = gm(*wrapped)
+        return [o._rgpu_meta if isinstance(o, RemoteTensor) else o for o in outs]
+    return _boxed(run)
+
+
+def _compiler(mode, compiler):
+    def compile_graph(gm, example_inputs):
+        try:
+            nodes = serialize(gm)
+            values = _output_values(gm)
+            if any(isinstance(s, torch.SymInt) for v in values
+                   if isinstance(v, torch.Tensor) for s in v.shape):
+                raise Unshippable("dynamic shapes; compile with dynamic=False")
+        except UnsupportedOp:
+            raise   # no eager fallback exists for this one; see UnsupportedOp
+        except Unshippable as e:
+            log.warning("rgpu: running a graph eagerly because it has %s", e)
+            return _eager(gm)
+        gid = next(_graph_ids)
+        _ship(gid, nodes, compiler, mode)
+
+        def run(*inputs):
+            in_ids = [id_of(x) for x in inputs]
+            outs, out_ids = [], []
+            for v in values:
+                if isinstance(v, torch.Tensor):
+                    m = meta_like(v.dtype, list(v.shape), list(v.stride()), v.storage_offset(),
+                                  v.is_conj(), v.is_neg())
+                    out_ids.append(register(m))
+                    outs.append(m)
+                else:
+                    out_ids.append(None)
+                    outs.append(v)
+            session.get().post(wire.CALL, gid, in_ids, out_ids)
+            return outs
+
+        return _boxed(run)
+
+    return compile_graph
+
+
+def compile_backend(mode=None, compiler="inductor"):
+    """A torch.compile backend that compiles on the rgpu server."""
+    comp = _compiler(mode, compiler)
+    return aot_autograd(fw_compiler=comp, bw_compiler=comp)
+
+
+torch._dynamo.register_backend(name="rgpu")(compile_backend())