| # LEADER: tinygrad |
|
|
| **Chosen leader:** tinygrad |
| **Repository URL:** https://github.com/tinygrad/tinygrad |
| **Commit SHA:** 1b779a9058b4a9acb12387133e7ded436ae3c0ca |
| **License:** MIT |
| **License file URL:** https://github.com/tinygrad/tinygrad/blob/1b779a9058b4a9acb12387133e7ded436ae3c0ca/LICENSE |
| **License holder:** the tiny corp |
|
|
| ## Absorbed Source |
|
|
| **File:** `tinygrad/device.py` |
| **Line range absorbed:** 39–54 |
|
|
| ```python |
| # Lines 39-54 of tinygrad/device.py (MIT licensed) |
| def get_available_devices(self) -> Iterator[str]: |
| for device in ALL_DEVICES: |
| with contextlib.suppress(Exception): yield self[device].device |
| @property |
| def DEFAULT(self) -> str: return DEV.device or self._select_device |
| @DEFAULT.setter |
| def DEFAULT(self, v): raise AttributeError(...) |
| @functools.cached_property |
| def _select_device(self) -> str: |
| assert (dev:=next((d for d in self._devices if d not in ["DISK","TINYFS","NPY"] and getenv(d)==1), None)) is None, ... |
| try: |
| device = next(self.get_available_devices()) |
| os.environ["DEV"] = device |
| return device |
| except StopIteration as exc: raise RuntimeError("no usable devices") from exc |
| ``` |
|
|
| ## Why tinygrad was chosen |
|
|
| - **Minimal LOC:** 420-line `device.py` vs BitNet's complex C++ CUDA kernel infrastructure. |
| - **Cleanest dispatch primitive:** The `get_available_devices` + `_select_device` pattern is a pure iterator-based priority dispatch over a static ordered list — directly translatable to our `PATHS = [...]` + `min(costs)` pattern. |
| - **MIT license:** Confirmed by reading LICENSE file directly via GitHub API. |
| - **No transitive complexity:** The core loop (lines 39–54) has zero external deps in our kernel. |
|
|
| ## Selection rationale over other candidates |
|
|
| - **BitNet** (MIT): Dispatch lives in C++/CUDA kernel launch wrappers — not cleanly Python-extractable to ≤10 lines without hallucinating function signatures. Evaluated and rejected (see REJECTED.md). |
| - **vLLM** (Apache-2.0): MoE expert routing is valid license-wise but dispatch logic spans multiple files (`moe_align_block_size`, `topk_softmax`). Could not be honestly minimized to ≤10 lines. Evaluated and rejected (see REJECTED.md). |
|
|