[Compile] Conditional compilation. Introduce compile_ranges (#24252)

Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Luka Govedič <luka.govedic@gmail.com>
Signed-off-by: ProExpertProg <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Luka Govedič <luka.govedic@gmail.com>
This commit is contained in:
Ilya Markov
2025-12-05 18:17:32 +00:00
committed by GitHub
co-authored by Luka Govedič Luka Govedič Robert Shaw Luka Govedič
parent 66e674cdd5
commit 4e26d3b09e
15 changed files with 582 additions and 268 deletions
+3 -2
View File
@@ -9,6 +9,7 @@ import torch.fx as fx
from torch._inductor.pattern_matcher import PatternMatcherPass
from vllm.config import VllmConfig
from vllm.config.compilation import Range
from vllm.distributed import get_tp_group, tensor_model_parallel_all_reduce
from vllm.distributed.parallel_state import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
@@ -333,7 +334,7 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
self.dump_patterns(config, self.patterns)
def is_applicable(self, shape: int | None) -> bool:
def is_applicable_for_range(self, compile_range: Range) -> bool:
# When sequence parallelism is enabled, the residual tensor from RMSNorm
# needs to be split along the sequence dimension. However, this dimension
# is symbolic during piecewise compilation, and splitting symbolic shapes
@@ -353,7 +354,7 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
):
return True
tp_size = get_tensor_model_parallel_world_size()
return shape is not None and shape % tp_size == 0
return (compile_range.is_single_size()) and (compile_range.end % tp_size == 0)
@VllmInductorPass.time_and_log
def __call__(self, graph: fx.Graph):