[torch.compile] Undo the fast_moe_cold_start hack in torch>=2.11 (#35475)

Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
Richard Zou
2026-03-01 21:44:22 +00:00
committed by GitHub
parent 6290470843
commit e82fbeec7b
4 changed files with 109 additions and 10 deletions
+35
View File
@@ -740,6 +740,41 @@ def is_torch_equal(target: str) -> bool:
return Version(importlib.metadata.version("torch")) == Version(target)
HAS_OPAQUE_TYPE = is_torch_equal_or_newer("2.11.0.dev")
if HAS_OPAQUE_TYPE:
from torch._opaque_base import OpaqueBase
else:
OpaqueBase = object # type: ignore[misc, assignment]
class ModuleName(OpaqueBase): # type: ignore[misc]
"""Wraps a module name string for use as a torch opaque type.
When torch >= 2.11, this is registered as a hoisted value-type opaque
object so that torch.compile lifts it as a graph input instead of baking
it as a constant. This avoids per-layer recompilation for MOE ops.
"""
def __init__(self, value: str):
self.value = value
def __eq__(self, other):
return isinstance(other, ModuleName) and self.value == other.value
def __hash__(self):
return hash(self.value)
def __fx_repr__(self):
return (f"ModuleName({self.value!r})", {ModuleName})
if HAS_OPAQUE_TYPE:
from torch._library.opaque_object import register_opaque_type
register_opaque_type(ModuleName, typ="value", hoist=True)
# Supports xccl with PyTorch versions >= 2.8.0.dev for XPU platform
def supports_xccl() -> bool:
return torch.distributed.is_xccl_available()