tensorflow-metal-plugin 0.1.0__tar.gz

tensorflow_metal_plugin-0.1.0/LICENSE

@@ -0,0 +1,251 @@
+                                 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.
+
+## Some of TensorFlow's code is derived from Caffe, which is subject to the following copyright notice:
+
+COPYRIGHT
+
+All contributions by the University of California:
+
+Copyright (c) 2014, The Regents of the University of California (Regents)
+All rights reserved.
+
+All other contributions:
+
+Copyright (c) 2014, the respective contributors
+All rights reserved.
+
+Caffe uses a shared copyright model: each contributor holds copyright over
+their contributions to Caffe. The project versioning records all such
+contribution and copyright details. If a contributor wants to further mark
+their specific copyright on a particular contribution, they should indicate
+their copyright solely in the commit message of the change when it is
+committed.
+
+LICENSE
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+   ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+   WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+   DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+   ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+   LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+   ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+CONTRIBUTION AGREEMENT
+
+By contributing to the BVLC/caffe repository through pull-request, comment,
+or otherwise, the contributor releases their content to the
+license and copyright terms herein.

tensorflow_metal_plugin-0.1.0/MANIFEST.in

@@ -0,0 +1,9 @@
+# The wheel is built from source at install time, against whichever
+# TensorFlow the installing interpreter has, so the sdist has to carry
+# everything the build needs.
+include Makefile
+include LICENSE
+include README.md
+recursive-include src *.cc *.h *.mm
+recursive-include tools *.sh *.py
+recursive-include tensorflow-plugins *.py

tensorflow_metal_plugin-0.1.0/Makefile

@@ -0,0 +1,94 @@
+# Builds the Metal PluggableDevice plugin against an installed TensorFlow.
+#
+# The header and library paths come from the TensorFlow package itself rather
+# than from a checked-in copy, so the plugin is built against exactly the
+# TensorFlow it will be loaded into.
+
+# Whichever interpreter has the TensorFlow you intend to load the plugin into.
+# Override it when `python3` on PATH is not that one:
+#   make PYTHON=/path/to/venv/bin/python
+PYTHON ?= python3
+TF_INCLUDE := $(shell $(PYTHON) -c "import tensorflow as tf; print(tf.sysconfig.get_include())")
+TF_LIB := $(shell $(PYTHON) -c "import tensorflow as tf; print(tf.sysconfig.get_lib())")
+SDK := $(shell xcrun --sdk macosx --show-sdk-path 2>/dev/null)
+
+BUILD := build
+OUT := $(BUILD)/libmetal_plugin.dylib
+
+SOURCES := src/plugin_init.cc \
+           $(wildcard src/tensorflow/core/common_runtime/metal/*.mm) \
+           $(wildcard src/tensorflow/core/common_runtime/metal/kernels/*.mm)
+OBJECTS := $(patsubst src/%,$(BUILD)/%.o,$(SOURCES))
+
+# The sources use manual retain and release, matching how TensorFlow's own
+# objc_library targets compile them. Turning ARC on here would reject them.
+#
+# STREAM_OPTIONS probes the installed headers for a StreamExecutor C API
+# callback added after the last release, so one source tree serves both an
+# in-tree build and a build against whatever TensorFlow is installed.
+STREAM_OPTIONS := $(shell bash tools/probe_stream_options.sh $(TF_INCLUDE))
+ifeq ($(STREAM_OPTIONS),no)
+COMPAT := -DTF_METAL_NO_STREAM_OPTIONS
+endif
+
+# -MMD -MP writes a .d file per object listing the headers it included, so a
+# change to a header rebuilds everything that reads it. Without this, adding a
+# field to a struct recompiled only the file it was declared in and left the
+# rest reading the old layout, which shows up as a stream reporting a failure
+# that never happened.
+CXXFLAGS := -std=c++17 -O2 -fPIC -isysroot $(SDK) $(COMPAT) -MMD -MP \
+            -mmacosx-version-min=13.0 \
+            -Isrc -I$(TF_INCLUDE) \
+            -I$(TF_INCLUDE)/external/farmhash_archive/src \
+            -DNDEBUG -DTF_METAL_OUT_OF_TREE \
+            -DTF_CAPI_WEAK
+
+FRAMEWORKS := -framework Metal -framework MetalPerformanceShaders \
+              -framework MetalPerformanceShadersGraph -framework Foundation
+
+# Linking against the framework library rather than deferring every symbol to
+# load time: an undefined symbol should be a build failure here, not a crash
+# in someone's first import.
+# TF_CAPI_WEAK, which TensorFlow's own c_api_macros.h provides for exactly this
+# case, makes every C API reference a weak one. Part of the kernel C API is
+# declared in the headers a released TensorFlow ships without being exported by
+# any binary in it, and dyld on macOS 13 and later binds at load rather than at
+# first call, so an ordinary reference to one of those would make dlopen fail
+# outright. Weak references bind to null instead, and the kernels that need
+# them are not registered when they are null (see ResourceVariableApiAvailable).
+LDFLAGS := -dynamiclib $(FRAMEWORKS) \
+           -L$(TF_LIB) -ltensorflow_framework.2 \
+           -Wl,-undefined,dynamic_lookup \
+           -Wl,-rpath,$(TF_LIB)
+
+.PHONY: all clean test sweep install
+
+all: $(OUT)
+
+$(OUT): $(OBJECTS)
+	@mkdir -p $(dir $@)
+	$(CXX) $(OBJECTS) $(LDFLAGS) -o $@
+	@echo "built $@"
+
+$(BUILD)/%.o: src/%
+	@mkdir -p $(dir $@)
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+# Fails loudly if the library still has unresolved TensorFlow or Metal symbols.
+check-symbols: $(OUT)
+	@bash tools/check_symbols.sh $(OUT)
+
+test: $(OUT)
+	$(PYTHON) tests/run_tests.py
+
+# Every registered op, through TensorFlow's own dispatch, against the CPU.
+sweep: $(OUT)
+	PYTHONPATH=tools $(PYTHON) tools/op_sweep.py
+
+install: $(OUT)
+	$(PYTHON) -m pip install .
+
+clean:
+	rm -rf $(BUILD)
+
+-include $(OBJECTS:.o=.d)

tensorflow_metal_plugin-0.1.0/PKG-INFO

@@ -0,0 +1,273 @@
+Metadata-Version: 2.4
+Name: tensorflow-metal-plugin
+Version: 0.1.0
+Summary: Metal GPU backend for TensorFlow on Apple silicon
+License: Apache-2.0
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: GPU
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: MacOS :: MacOS X
+Classifier: Programming Language :: Python :: 3
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: tensorflow>=2.16
+Dynamic: classifier
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: license
+Dynamic: license-file
+Dynamic: requires-dist
+Dynamic: requires-python
+Dynamic: summary
+
+# tensorflow-metal-plugin
+
+A Metal GPU backend for TensorFlow on Apple silicon, built as an out-of-tree
+PluggableDevice. It loads into a stock TensorFlow wheel and adds
+`/physical_device:GPU:0`.
+
+This is the out-of-tree form of the backend proposed in
+[tensorflow/tensorflow#126254](https://github.com/tensorflow/tensorflow/pull/126254).
+The sources are the same; the only difference is this repository exports
+`SE_InitPlugin` and `TF_InitKernel` from a shared object, where the in-tree
+form hands the same function pointers to `RegisterPluggableDevicePlugin`.
+
+## Status
+
+Working, and every op it registers has been run on a real GPU and checked.
+One significant limitation is not this project's to fix: see
+[What a released TensorFlow cannot do](#what-a-released-tensorflow-cannot-do).
+
+`make sweep` calls all 356 registered ops through TensorFlow's own dispatch,
+once on the GPU and once on the CPU with identical inputs, with soft placement
+off so that a missing kernel raises rather than answering from the host:
+
+| | |
+| --- | --- |
+| Verified against the CPU kernel, or against a property where there is no CPU kernel | 323 |
+| Removed from TensorFlow, so no device can run them | 19 |
+| Need kernel C API entry points a released TensorFlow does not export | 14 |
+| **Unaccounted for** | **0** |
+
+Every op is also run twice and required to give the same answer, which is how
+an inverse transform that rewrote its own input was caught. The sweep
+separately enumerates every registration TensorFlow holds for these ops and
+rejects any that is duplicated or that constrains an attribute the op does not
+have, since either makes an op unusable while looking registered.
+
+Two of the nineteen announce themselves differently, complaining that a
+kernel constrains an attribute the node lacks: TensorFlow's own CPU
+registrations for `TopK` and `TileGrad` constrain `index_type` and
+`Tmultiples`, which their op defs do not have. That is true and is not why
+they cannot run. Both are deprecated in their op def, `TopK` from GraphDef
+version 7 and `TileGrad` from version 3, so nothing can call them either way.
+
+Verified on an Apple M4 Max, macOS 26.6, against the stock
+`tensorflow==2.20.0` wheel for Python 3.12:
+
+```
+before: ['/physical_device:CPU:0']
+after : ['/physical_device:CPU:0', '/physical_device:GPU:0']
+Executing op MatMul in device /job:localhost/replica:0/task:0/device:GPU:0
+```
+
+`MatMul`, `Conv2D`, `Softmax`, `Relu`, `MaxPool2D` and `ReduceSum` match the
+CPU kernels with soft placement disabled, so a missing GPU kernel raises
+instead of quietly producing a correct answer on the wrong device.
+
+## Install
+
+```
+pip install tensorflow
+pip install git+https://github.com/IPNP-BIPN/tensorflow-metal-plugin
+```
+
+That is all. The shared object is built at install time against the
+TensorFlow of the interpreter doing the installing, and lands in
+`site-packages/tensorflow-plugins`, which TensorFlow scans at import. Nothing
+has to be loaded by hand:
+
+```python
+>>> import tensorflow as tf
+>>> tf.config.list_physical_devices()
+[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'),
+ PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
+```
+
+Verified on a clean environment with `tensorflow==2.21.0`, Python 3.12, macOS
+26.6 on an M4 Max.
+
+## Training works, and what it costs today
+
+`model.fit(optimizer="adam")` trains and the loss goes down. Getting there
+needed a correction worth stating plainly, because it changes the speed.
+
+TensorFlow's own kernels for resource variables reach a tensor through its
+data pointer. On a unified memory device that pointer is host-addressable, so
+those kernels read and write device memory from the host with no idea that GPU
+work is in flight against it. A plugin is supposed to implement those ops
+itself, through `tensorflow/c/kernels_experimental.h`, and order them on its
+own stream. Since 2.20.0 no shipped binary defines those entry points
+([#126374](https://github.com/tensorflow/tensorflow/issues/126374)), so the
+ops fall back to the host and race.
+
+The symptom was the worst kind: an optimiser read a slot variable mid-write,
+took the square root of whatever was there, and produced `nan` weights with no
+error raised. `model.fit` reported `[nan, nan, nan]` and carried on.
+
+While those entry points are missing, every Metal kernel waits for the GPU
+before returning, which closes the window. It is announced in a warning at
+load, and `TF_METAL_SYNCHRONOUS` forces it either way. When the entry points
+come back the plugin returns to running asynchronously with no change here.
+
+## Is it faster than the CPU
+
+Sometimes, and by how much depends entirely on the shape of the work. Measured
+on an M4 Max against TensorFlow 2.21.0, median of ten runs each, both devices
+in the same process on the same data, waiting for the device before stopping
+the clock:
+
+Two columns of speedup, because the wait above costs most of it. "Today" is
+what you get from a released TensorFlow; "async" is the same machine with
+`TF_METAL_SYNCHRONOUS=0`, which is what the plugin does once
+[#126374](https://github.com/tensorflow/tensorflow/issues/126374) is fixed.
+
+| | GPU today | CPU | today | async |
+| --- | ---: | ---: | ---: | ---: |
+| MatMul 2048x2048 | 3.80 ms | 14.42 ms | **3.8x** | 6.5x |
+| Conv2D, batch 64, 64x64x32 to 64 | 4.58 ms | 10.91 ms | **2.4x** | 3.0x |
+| MatMul 1024x1024 | 1.12 ms | 2.17 ms | 1.9x | 1.5x |
+| CNN training step, SGD, batch 128 | 19.46 ms | 20.61 ms | 1.1x | 1.7x |
+| CNN forward, batch 128 | 5.53 ms | 6.00 ms | 1.1x | 1.5x |
+| MatMul 512x512 | 0.44 ms | 0.36 ms | 0.8x | 0.9x |
+| ReduceSum 4096x4096 | 0.44 ms | 0.34 ms | 0.8x | 0.8x |
+| Elementwise 4096x4096 | 2.88 ms | 1.54 ms | 0.5x | 0.5x |
+
+The pattern is the ordinary one and worth stating plainly: the GPU wins where
+there is arithmetic to do per byte moved, and loses where there is not. A
+4096x4096 elementwise chain moves 67 MB and does three floating point
+operations per element, so it is bound by memory on a machine whose CPU shares
+that same memory. Small matrices lose to the cost of getting work to the
+device at all.
+
+A training step is 1.7x the CPU when the plugin can run asynchronously, and
+barely ahead of it while it cannot. That gap is the cost of the missing entry
+points, not of the backend.
+
+`benchmarks/benchmark.py` reproduces the table.
+
+## Build
+
+Needs the macOS 15 SDK or later and a Python with TensorFlow installed. The
+backend aliases an `MTLBuffer` through `MPSNDArray` with packed rows, and both
+`initWithBuffer:offset:descriptor:` and `preferPackedRows` arrived in that SDK;
+an older one does not declare them and the build stops rather than degrading. The
+header and library paths come from that TensorFlow, so the plugin is built
+against exactly the one it will be loaded into.
+
+```
+make                                  # or: make PYTHON=/path/to/venv/bin/python
+make check-symbols
+make test
+```
+
+Then either point TensorFlow at it directly:
+
+```python
+from tensorflow.python.framework import load_library
+load_library.load_pluggable_device_library("build/libmetal_plugin.dylib")
+```
+
+or install it so that `import tensorflow` finds it:
+
+```
+make install
+```
+
+`TF_DISABLE_METAL=1` keeps the backend out of the process without
+uninstalling it.
+
+## What a released TensorFlow cannot do
+
+Six entry points of the kernel C API are declared in the headers a released
+TensorFlow ships and are exported by no binary in it:
+
+```
+TF_AssignRefVariable
+TF_AssignUpdateVariable
+TF_GetInputTensorFromVariable
+TF_MaybeLockVariableInputMutexesInOrder
+TF_ReleaseVariableInputLockHolder
+TF_OpKernelConstruction_GetAttrTensorShape
+TF_OpKernelContext_ForwardRefInputToRefOutput
+```
+
+Checked against `tensorflow==2.20.0` on macOS arm64: absent from
+`libtensorflow_framework.2.dylib`, from `libtensorflow_cc.2.dylib`, and from
+every pywrap module, and unresolvable by `dlsym` inside a live process.
+`TF_AllocateOutput` and `TF_NewKernelBuilder`, from the same header set, are
+exported normally, so this is not a matter of the whole C API being private.
+
+Fifteen ops need them, and the plugin does not register those when the symbols
+are missing, logging one warning instead:
+
+| Family | Ops |
+| --- | --- |
+| Optimisers | `ResourceApplyAdam`, `ResourceApplyGradientDescent`, `ResourceApplyMomentum`, `ResourceApplyKerasMomentum`, `ResourceApplyRMSProp` |
+| Resource gather and scatter | `ResourceGather`, `ResourceGatherNd`, `ResourceScatterUpdate`, `GatherNd` |
+| Reference variables | `Assign`, `AssignAdd`, `AssignSub` |
+| Parallel stacking | `ParallelConcat`, `_ParallelConcatStart`, `_ParallelConcatUpdate` |
+
+The optimisers are the whole of that list that matters: **without them there is
+no training on the GPU**, only inference and manual gradient work. They run on
+the host instead, which is correct and slow.
+
+This is a regression, not a standing limitation. All fourteen symbols of
+`tensorflow/c/kernels_experimental.cc` are exported by `libtensorflow_framework`
+in 2.19.1 and 2.18.1, and absent from every binary in the 2.20.0 wheel, with
+none added in exchange. The headers still declare them. Filed upstream as
+[tensorflow/tensorflow#126374](https://github.com/tensorflow/tensorflow/issues/126374).
+
+So this is not something the plugin can work around, and it is not permanent
+either: when those exports come back, the fifteen ops below start working here
+with no change to this repository.
+
+It is also the sharpest argument for the in-tree form, where the same code
+links these functions directly and all fifteen ops work. That trade is the
+subject of the discussion on
+[#126254](https://github.com/tensorflow/tensorflow/pull/126254).
+
+## Why this exists
+
+Apple's `tensorflow-metal` last shipped 1.2.0 on 2025-01-31, publishes no
+wheel past cp312, has no sdist, and its repository was archived in 2021. TF
+master requires Python 3.10 or later and classifies up to cp313, so on a
+current Python there is no GPU path for TensorFlow on a Mac at all.
+
+## Op coverage
+
+The backend registers every op TensorFlow registers for `DEVICE_GPU`, less the
+five TensorRT ops that `if_tensorrt` excludes from a macOS build, and less the
+fifteen above when the C API entry points they need are missing. The table of
+Metal kernels with their dtypes is in
+[docs/ops.md](docs/ops.md).
+
+## Layout
+
+```
+src/plugin_init.cc                          the two exported entry points
+src/tensorflow/core/common_runtime/metal/   the backend, verbatim from the
+                                            TensorFlow tree
+tools/                                      build probes and the symbol check
+tests/                                      on-device checks against CPU
+```
+
+The backend sources keep their TensorFlow paths so that syncing them from the
+tree is a copy rather than a patch. Two macros, `TF_METAL_OUT_OF_TREE` and
+`TF_METAL_NO_STREAM_OPTIONS`, are the whole of what the out-of-tree build
+turns on; both are no-ops in the tree.
+
+## Licence
+
+Apache 2.0, the same as TensorFlow.