Fix BackendImplementationMissing: repair codegen registration skipped by circular-import partial cache - #3286
Open
mengluy0125 wants to merge 2 commits into
Open
Fix BackendImplementationMissing: repair codegen registration skipped by circular-import partial cache#3286mengluy0125 wants to merge 2 commits into
mengluy0125 wants to merge 2 commits into
Conversation
Summary: `_find_device` scans kernel arguments for a `torch.device`/`torch.Tensor` to determine the launch device, recursing into nested `tuple`/`list`/`dict` containers. The recursion passed each element directly as `_find_device(item)`. `_find_device` treats its argument as a tuple of args and iterates it (`for arg in args`). Passing a bare `item` that is itself a `torch.Tensor` made the recursive call iterate over the tensor (`for arg in tensor`), materializing the rows of the tensor on the host before the first row matched the `isinstance(arg, torch.Tensor)` check -- a real per-call host overhead on the hot argument-processing path. A 0-dim scalar tensor also fails to iterate at all. Wrap the recursed element in a single-element tuple, `_find_device((item,))`, so the recursion sees exactly one arg and hits the `torch.Tensor` fast path directly. Behavior is unchanged for every previously-working case; the tensor is no longer iterated. Differential Revision: D114780871
… by circular-import partial cache
Summary:
Under torch.compile in packaged runtimes (e.g. aps_ads_vm training), Helion kernels
could fail at pt2_warmup with `BackendImplementationMissing: Backend 'triton' is
missing required implementation: codegen for API function load`.
Root cause: `import_backend_codegen()` registers each backend's per-op codegen by
`importlib.import_module("<pkg>._codegen_modules")`. If that codegen module (or one
of its leaf modules such as `triton/memory_ops.py`) is already present in
`sys.modules` in a partially-initialized state -- a circular import during package
init cached it before its module-scope `_decorators.codegen` / `register_codegen`
handlers ran -- `import_module` returns the incomplete module and the registrations
are silently skipped, leaving `APIFunc._codegen` empty for that backend. This is
import-order dependent (nondeterministic) and surfaces only at codegen time.
Fix:
- `CodegenDict.__missing__`: when a backend codegen key is absent, attempt a
one-time global repair (imports are settled by codegen time) and retry before
raising KeyError.
- `backend_registry.repair_backend_codegen()`: reload each backend's codegen leaf
modules so their registration decorators re-run.
- Make codegen registration idempotent (`_decorators.codegen`,
`aten_lowering.register_codegen`) so the repair reload cannot trip the
"already registered" assert.
Differential Revision: D114782445
meta-cla
Bot
added
the
CLA Signed
|
@mengluy0125 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114782445. |
ethche
requested changes
Aug 6, 2026
ethche
left a comment
ethche
left a comment
Contributor
There was a problem hiding this comment.
Looks good but marked some areas of concern. Could you take a look?
I asked codex, and it suggested something like this:
- Add a locked ensure_backend_codegen() in backend_registry.
- Set its completion flag only after successful reload.
- Call it before falling back to "common" in CodegenDict.
- Call it when AtenLowering cannot find a backend handler, then retry.
| def __missing__(self, key: str) -> Callable[[CodegenState], object]: | ||
| if key != "common" and "common" in self: | ||
| return self["common"] | ||
| # A backend codegen can be absent because its codegen module was cached |
Contributor
There was a problem hiding this comment.
Could repair happen before falling back to common? Otherwise a missing backend-specific registration silently uses the common implementation and never triggers repair. Maybe double check
| assert backend not in self.codegen_impls, ( | ||
| f"codegen already registered for backend {backend!r}" | ||
| ) | ||
| # Idempotent: repair_backend_codegen() may reload this codegen module |
Contributor
There was a problem hiding this comment.
What triggers repair when an Aten codegen registration is missing?
| # this dict empty). By codegen time imports are settled, so force-complete | ||
| # the registrations once and retry before giving up. | ||
| if key != "common" and not CodegenDict._repaired: | ||
| CodegenDict._repaired = True |
Contributor
There was a problem hiding this comment.
Should _repaired be set only after repair succeeds, with synchronization? As written, an exception or concurrent lookup permanently skips repair.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Under torch.compile in packaged runtimes (e.g. aps_ads_vm training), Helion kernels
could fail at pt2_warmup with
BackendImplementationMissing: Backend 'triton' is missing required implementation: codegen for API function load.Root cause:
import_backend_codegen()registers each backend's per-op codegen byimportlib.import_module("<pkg>._codegen_modules"). If that codegen module (or oneof its leaf modules such as
triton/memory_ops.py) is already present insys.modulesin a partially-initialized state -- a circular import during packageinit cached it before its module-scope
_decorators.codegen/register_codegenhandlers ran --
import_modulereturns the incomplete module and the registrationsare silently skipped, leaving
APIFunc._codegenempty for that backend. This isimport-order dependent (nondeterministic) and surfaces only at codegen time.
Fix:
CodegenDict.__missing__: when a backend codegen key is absent, attempt aone-time global repair (imports are settled by codegen time) and retry before
raising KeyError.
backend_registry.repair_backend_codegen(): reload each backend's codegen leafmodules so their registration decorators re-run.
_decorators.codegen,aten_lowering.register_codegen) so the repair reload cannot trip the"already registered" assert.
Differential Revision: D114782445