[AMD][CI][BugFix] Fix Distributed Compile Unit Tests (2xH100-2xMI300) group (#43120)
Signed-off-by: Randall Smith <Randall.Smith@amd.com>
This commit is contained in:
@@ -1101,6 +1101,14 @@ class RocmAiterAllReduceFusionPass(VllmFusionPatternMatcherPass):
|
||||
return False
|
||||
return bool(compile_range.end <= self.max_token_num)
|
||||
|
||||
@VllmInductorPass.time_and_log
|
||||
def __call__(self, graph: fx.Graph) -> None:
|
||||
self.matched_count = self.pm_pass.apply(graph)
|
||||
VllmPatternMatcherPass.match_table[self.pass_name] += self.matched_count
|
||||
logger.debug(
|
||||
"%s Replaced %s patterns", self.__class__.__name__, self.matched_count
|
||||
)
|
||||
|
||||
def __del__(self) -> None:
|
||||
if getattr(self, "disabled", True):
|
||||
return
|
||||
|
||||
@@ -917,13 +917,13 @@ class AsyncTPPass(VllmFusionPatternMatcherPass):
|
||||
AllGatherScaledMMPattern(self.model_dtype, self.device).register(
|
||||
self.pm_pass
|
||||
)
|
||||
|
||||
CutlassScaledMMReduceScatterPattern(self.model_dtype, self.device).register(
|
||||
self.pm_pass
|
||||
)
|
||||
AllGatherCutlassScaledMMPattern(self.model_dtype, self.device).register(
|
||||
self.pm_pass
|
||||
)
|
||||
if hasattr(torch.ops._C, "cutlass_scaled_mm"):
|
||||
CutlassScaledMMReduceScatterPattern(
|
||||
self.model_dtype, self.device
|
||||
).register(self.pm_pass)
|
||||
AllGatherCutlassScaledMMPattern(self.model_dtype, self.device).register(
|
||||
self.pm_pass
|
||||
)
|
||||
with suppress(ImportError):
|
||||
import vllm.utils.flashinfer # noqa: F401
|
||||
if hasattr(torch.ops.vllm, "bmm_fp8"):
|
||||
|
||||
Reference in New Issue
Block a user