feat(runtime): implement device runtime layer (Python)#277
feat(runtime): implement device runtime layer (Python)#277
Conversation
|
This PR is intended to show the impact of RFC : Device and runtime layer (aligned with pluggable compile backends) |
| # single GPU stack per process; validate against FLYDSL_RUNTIME_KIND. | ||
| from ...runtime.device_runtime import get_device_runtime | ||
|
|
||
| get_device_runtime() |
There was a problem hiding this comment.
Good point — get_backend() is the compile-backend resolver, and calling get_device_runtime() here couples compile resolution to the device-runtime singleton. We did it to fail fast on FLYDSL_COMPILE_BACKEND vs FLYDSL_RUNTIME_KIND mismatch (RFC).
If we’d rather keep compile resolution free of runtime, we can move this check to the first JIT compile / launch path (e.g. jit_function / CompiledArtifact) and leave get_backend() pure. Happy to adjust in the next push.
| n = 0 | ||
| self._torch_device_count = max(n, 1) | ||
| return self._torch_device_count | ||
|
|
There was a problem hiding this comment.
So we add those Jit runner funcs here after done?
There was a problem hiding this comment.
RocmDeviceRuntime in this PR is intentionally minimal — it only exists so we have a process-wide runtime kind (rocm) and can validate it against FLYDSL_COMPILE_BACKEND.
device_count() is a small probe via PyTorch for now; we are not planning to add JIT runner / ExecutionEngine glue here. That stays in jit_executor / compiler as today.
Follow-up work (per RFC) could add stream/event helpers on DeviceRuntime, but still not duplicate the JIT launch path into this module.
We can change device_count() to return 0 when no CUDA/ROCm device is visible instead of max(n, 1) to make it clear.
Peter9606
force-pushed
the
fujun.han/pluggable-compile-backend
branch
2 times, most recently
from
0e746c2 to
d47854a
Compare
Peter9606
changed the base branch from
refactor/pluggable-compile-backend
to
main
- Add flydsl.runtime.device_runtime: Device/DeviceRuntime/Event, RocmDeviceRuntime - FLYDSL_RUNTIME_KIND (env.runtime.kind), compile/runtime pairing validation - get_backend() triggers get_device_runtime() for single-stack invariant - register_device_runtime, register_compile_runtime_mapping for extensions - Unit tests. Signed-off-by: Fujun Han <fujun.han@iluvatar.com>
Peter9606
force-pushed
the
fujun.han/pluggable-compile-backend
branch
from
01cd244 to
0a4c4a8
Compare
|
Hi @coderfeli , I've updated the base branch of this PR from refactor/pluggable-compile-backend to main. The current diff now correctly reflects the changes against the main branch. Thanks. |
2 participants
code with cursor