metaljax 0.1.0__tar.gz

metaljax-0.1.0/.gitignore

@@ -0,0 +1,11 @@
+# reference clones (not part of this repo)
+/jax/
+/llvm-project/
+
+.venv/
+__pycache__/
+*.egg-info/
+.pytest_cache/
+plugin/build/
+src/metaljax/lib/
+.DS_Store

metaljax-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.

metaljax-0.1.0/PKG-INFO

@@ -0,0 +1,263 @@
+Metadata-Version: 2.4
+Name: metaljax
+Version: 0.1.0
+Summary: Experimental Metal backend for JAX on Apple silicon: PJRT plugin executing StableHLO on MLX
+Project-URL: Repository, https://github.com/eterevsky/metaljax
+Project-URL: Issues, https://github.com/eterevsky/metaljax/issues
+Author-email: Oleg Eterevsky <oleg@eterevsky.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: apple-silicon,gpu,jax,metal,mlx,pjrt
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: GPU
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: MacOS
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.12
+Requires-Dist: jax<0.12,>=0.11
+Requires-Dist: ml-dtypes
+Requires-Dist: mlx>=0.32
+Requires-Dist: numpy
+Description-Content-Type: text/markdown
+
+# metaljax — a Metal backend for JAX
+
+metaljax lets unmodified JAX code run on Apple-silicon GPUs:
+
+```python
+$ JAX_PLATFORMS=metal python -c \
+    "import jax.numpy as jnp; a = jnp.array([1, 2, 3]); print(2 * a, (2*a).device)"
+[2 4 6] MetalDevice(id=0)
+```
+
+From JAX's point of view it is a regular PJRT backend: `jax.devices()`
+reports a `METAL` device, and `jit`, `grad`, `vmap`, `lax.scan`,
+`jax.random` (threefry), optax training loops, etc. all work. Under the
+hood the compiled StableHLO programs are interpreted onto
+[MLX](https://github.com/ml-explore/mlx) arrays, which execute on the GPU
+via Metal.
+
+**Status**: experimental, published as a demo. Real training runs work
+(the [texmo](https://github.com/eterevsky/texmo) project trains
+end-to-end) and transformer training steps run within a few percent of
+PyTorch's MPS backend, but expect gaps — unsupported ops fail with a
+clear `UnsupportedOpError`. If a Metal backend ever lands upstream in the
+JAX ecosystem, this package will be deprecated in its favor.
+
+## Install
+
+```bash
+pip install metaljax
+```
+
+Requirements: Apple-silicon Mac, macOS 14+, Python 3.12+, jax 0.11.x
+(installed automatically). Then select the backend per program:
+
+```bash
+JAX_PLATFORMS=metal python -c "import jax; print(jax.devices())"
+```
+
+CPU remains the default backend when `JAX_PLATFORMS` is unset, so
+installing metaljax does not change existing workflows. Installing from
+the source distribution (rather than the wheel) additionally requires the
+Xcode command-line tools, since the PJRT plugin compiles at build time.
+
+## How it works
+
+```
+jax.jit(f)(x)
+  │  StableHLO (serialized portable artifact)
+  ▼
+plugin/metal_pjrt.cc          ── PJRT C-API dylib loaded by jaxlib.
+  │                              No dependencies; trampolines every call
+  ▼                              back into Python (same process, GIL).
+src/metaljax/engine.py        ── compile: deserialize + wrap Interpreter
+  │                              execute: run on device buffers
+  ▼
+src/metaljax/interpreter.py   ── walks the StableHLO module op by op
+  │  + src/metaljax/ops/*     ── one handler per op family
+  ▼
+mlx.core                      ── lazy Metal arrays; unified memory
+```
+
+- The dylib implements PJRT API v0.114 (`plugin/vendor/pjrt_c_api.h`,
+  vendored at jaxlib's exact openxla/xla pin).
+- Registration happens through the `jax_plugins` namespace package
+  (`src/jax_plugins/metal/`), at **priority −1**: CPU stays the default
+  backend unless you opt in via `JAX_PLATFORMS`.
+
+## Requirements
+
+- Apple-silicon Mac (developed on an M5 Max, macOS 26.5, Xcode 26.6 —
+  any arm64 Mac with a recent Xcode/CLT should work).
+- [uv](https://docs.astral.sh/uv/) (only for creating the venv).
+- Python **3.14** and jax/jaxlib **0.11.x** (what the venv setup below
+  installs; the vendored PJRT header matches jaxlib 0.11.0).
+
+## Developing from source
+
+```bash
+git clone https://github.com/eterevsky/metaljax && cd metaljax
+uv venv --python 3.14 .venv
+uv pip install -p .venv/bin/python jax mlx numpy pytest
+uv pip install -p .venv/bin/python -e .
+./plugin/build.sh          # builds plugin/build/libmetal_pjrt.dylib (clang)
+```
+
+Verify:
+
+```bash
+JAX_PLATFORMS=metal .venv/bin/python -c "import jax; print(jax.devices())"
+```
+
+should print `[MetalDevice(id=0)]`.
+
+## Running the tests
+
+The pytest suite lowers each construct with `jax.jit(...).lower()`, runs
+the StableHLO module through the interpreter on the GPU, and compares
+against the JAX CPU backend (this exercises the interpreter directly and
+does not need the plugin dylib):
+
+```bash
+.venv/bin/python -m pytest tests/ -q
+```
+
+Current suite: 129 tests across elementwise/transcendental ops, shapes and
+broadcasting, `dot_general`/einsum, reductions and cumulative ops, control
+flow (`while`/`cond`/`scan`), gather/scatter, RNG, and bf16/f16/x64 dtype
+handling.
+
+End-to-end smoke test through the real plugin (device buffers, compile,
+execute, PJRT events):
+
+```bash
+JAX_PLATFORMS=metal .venv/bin/python -c "
+import jax, jax.numpy as jnp
+g = jax.jit(jax.grad(lambda x: jnp.sum(jnp.tanh(x) ** 2)))(jnp.arange(4.0))
+print(g, g.device)"
+```
+
+## Running texmo on Metal
+
+texmo pins its platform in its own `config.py`, so use the bundled driver,
+which imports texmo's `ManagerJax` directly (set `TEXMO_DIR` if your
+checkout is not `~/texmo`):
+
+```bash
+# extra deps texmo imports at module level (torch is never executed by JAX)
+uv pip install -p .venv/bin/python optax safetensors regex scipy scikit-learn torch
+
+# platform spec steps batch length precision
+.venv/bin/python scripts/texmo_train.py metal,cpu 'bits.1+bp|mgru.4-dense.4.gelu' 64 16 64 fp32
+```
+
+texmo's own GRU benchmark accepts a platform flag:
+
+```bash
+cd ~/texmo
+~/metaljax/.venv/bin/python scripts/bench_jax.py --platform metal --mode ram --steps 8 --batch 64
+```
+
+## Using metaljax from another project
+
+Add `metaljax` to your dependencies (it declares `jax` itself):
+
+```toml
+[project]
+dependencies = ["metaljax"]
+```
+
+and set `JAX_PLATFORMS=metal` (or
+`jax.config.update("jax_platforms", "metal")` before first use).
+
+To develop against a local checkout instead, use a path source:
+
+```toml
+[tool.uv.sources]
+metaljax = { path = "../metaljax", editable = true }
+```
+
+(with an editable install, build the plugin once in the checkout via
+`./plugin/build.sh`). A git source
+(`metaljax = { git = "https://github.com/eterevsky/metaljax" }`) works
+too; like sdist installs it compiles the plugin during the build, which
+needs the Xcode command-line tools.
+
+## Environment variables
+
+| Variable | Default | Meaning |
+|---|---|---|
+| `JAX_PLATFORMS` | *(unset)* | Set to `metal` (or `metal,cpu`) to select the backend; unset keeps CPU default. |
+| `METALJAX_MATMUL_PRECISION` | `highest` | On M5-class GPUs MLX routes f32 GEMM through the neural accelerators at ~bf16 input precision (~4e-3 error). `highest` pins MLX kernels to the previous GPU generation for exact f32; set `default` to allow the fast path. |
+| `METALJAX_F64` | `error` | Metal has no float64. Default (`error`): f64 values may pass **through** the device (x64 mode wraps Python scalars as f64 buffers that programs immediately convert to f32 — stored as f32, which rounds exactly once and stays bit-identical to CPU), but any op that **computes** in f64 fails at compile time, naming the op. `downcast`: emulate all f64 in f32 (one warning). Example: under `jax_enable_x64`, optax AdamW's `beta**step` bias correction is real f64 arithmetic — strict mode rejects it, so `scripts/texmo_train.py` sets `downcast` explicitly. |
+| `METALJAX_PLUGIN_PATH` | *(auto)* | Override the path to `libmetal_pjrt.dylib`. |
+
+## Repository layout
+
+```
+CLAUDE.md                  project decisions/status (kept current)
+pyproject.toml             python package + jax_plugins entry point
+src/metaljax/
+  interpreter.py           StableHLO walker (SSA env, blocks, funcs)
+  ops/                     op handlers: elementwise, shape, linalg,
+                           reduction, control, gather
+  engine.py                PJRT-facing compile/execute/buffer layer
+  dtypes.py, _ir.py        dtype tables, MLIR context & attr decoding
+src/jax_plugins/metal/     backend registration (priority -1)
+plugin/
+  metal_pjrt.cc            the PJRT C-API dylib (no deps, ~1100 lines)
+  vendor/pjrt_c_api.h      vendored PJRT header (API 0.114)
+  build.sh                 clang build → plugin/build/libmetal_pjrt.dylib
+tests/                     pytest suite (Metal vs CPU)
+scripts/texmo_train.py     texmo-on-Metal driver
+```
+
+## Benchmarks
+
+Full training steps (fwd + bwd + AdamW), f32, M5 Max, via
+`scripts/bench_compare.py` (16 timed steps after warmup):
+
+| workload | jax CPU | **metaljax** | torch MPS | torch CPU |
+|---|---:|---:|---:|---:|
+| transformer d256 L4 T256 b32 | 174.3 | **31.2** | 30.1 | 209.3 |
+| transformer d512 L4 T256 b64 | — | **157.6** | 154.6 | — |
+| GRU.256 T256 b256 (scan) | — | **112.0** | 50.7¹ | — |
+| texmo `bench_jax.py` GRU b256 | 273.3 | **147.7** | — | — |
+
+¹ torch uses its fused `nn.GRU` kernel; metaljax runs the scan as a
+compiled-body loop — sequential models still pay per-timestep dispatch.
+
+How: pure programs and counted-loop (`scan`/`fori_loop`) bodies are traced
+once through `mx.compile` and replayed as fused Metal graphs; small
+statically-counted loops are unrolled into the enclosing trace, so e.g. a
+whole texmo training step (forward scan + backward + AdamW) becomes a
+single graph replay. `METALJAX_COMPILE=0` disables compilation,
+`METALJAX_TRACE_BUDGET` (default 20000 ops) caps trace sizes, and
+`METALJAX_DEBUG=1` logs loop/compile decisions.
+
+On texmo itself: models of texmo's typical tiny size (a handful of units)
+remain CPU territory — thousands of microscopic kernels can't beat XLA:CPU
+compiling the whole scan natively — while at `mgru.256` metal already
+edges out CPU (5.1s vs 5.3s for 64 steps, batch 64, length 256).
+
+## Known limitations
+
+- Sequential (scan-heavy) models run each timestep as a compiled-graph
+  replay from Python; fused-RNN-kernel frameworks are ~2× faster there.
+- float64 is emulated in f32 (see `METALJAX_F64`); complex dtypes are
+  unsupported.
+- Not yet implemented (fail with a clear `UnsupportedOpError`):
+  argmax/argmin-style multi-result `reduce`, `sort`, general
+  `reduce_window` (only cumsum/cumprod/cummax/cummin patterns),
+  partial-window scatter, send/recv, multi-device anything.
+- Buffer donation is ignored (correct, but no memory savings).
+
+## License / provenance
+
+Experimental personal project; builds against public JAX/OpenXLA (PJRT
+header vendored from openxla/xla) and Apple's MLX.