[Misc][BE] Type coverage for vllm/compilation [3/3] (#31748)

Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
This commit is contained in:
Lucas Kabela
2026-01-12 19:24:38 +00:00
committed by GitHub
parent 08e8e99ce7
commit ad8818bb5e
11 changed files with 331 additions and 278 deletions
+22 -18
View File
@@ -2,6 +2,8 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import functools
from collections.abc import Callable, Sequence
from typing import Any
import torch
import torch._inductor.pattern_matcher as pm
@@ -26,9 +28,11 @@ from .vllm_inductor_pass import VllmInductorPass, VllmPatternMatcherPass
logger = init_logger(__name__)
def get_first_out_wrapper(fn):
def get_first_out_wrapper(
fn: Callable[..., Sequence[torch.Tensor]],
) -> Callable[..., torch.Tensor]:
@functools.wraps(fn)
def wrapper(*args):
def wrapper(*args: Any) -> torch.Tensor:
return fn(*args)[0]
return wrapper
@@ -68,7 +72,7 @@ class FirstAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
super().__init__(epsilon, dtype, device)
self.rmsnorm_matcher = MatcherRMSNorm(epsilon)
def get_inputs(self):
def get_inputs(self) -> list[torch.Tensor]:
input = torch.empty([1, 8, 4], device=self.device, dtype=self.dtype)
arg3_1 = torch.empty([4], device=self.device, dtype=self.dtype)
@@ -78,7 +82,7 @@ class FirstAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
def pattern(
input: torch.Tensor,
arg3_1: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
all_reduce = self._all_reduce(input)
rmsnorm = self.rmsnorm_matcher(all_reduce, arg3_1)
@@ -87,7 +91,7 @@ class FirstAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
def replacement(
input: torch.Tensor,
arg3_1: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
reduce_scatter = self._reduce_scatter(input)
rmsnorm = self.rmsnorm_matcher(reduce_scatter, arg3_1)
@@ -100,11 +104,11 @@ class FirstAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
class MiddleAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
def __init__(self, epsilon: float, dtype: torch.dtype, device: str | None):
def __init__(self, epsilon: float, dtype: torch.dtype, device: str | None) -> None:
super().__init__(epsilon, dtype, device)
self.rmsnorm_matcher = MatcherFusedAddRMSNorm(epsilon)
def get_inputs(self):
def get_inputs(self) -> list[torch.Tensor]:
mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)
residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
@@ -116,7 +120,7 @@ class MiddleAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
rms_norm_weights,
]
def register(self, pm_pass: PatternMatcherPass):
def register(self, pm_pass: PatternMatcherPass) -> None:
def pattern(
residual: torch.Tensor,
mm_1: torch.Tensor,
@@ -163,23 +167,23 @@ class FirstAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
epsilon: float,
dtype: torch.dtype,
device: str | None,
):
) -> None:
super().__init__(epsilon, dtype, device)
self.rmsnorm_matcher = MatcherRMSNorm(epsilon)
self.quant_matcher = MatcherQuantFP8(kFp8StaticTensorSym)
def get_inputs(self):
def get_inputs(self) -> list[torch.Tensor]:
input = torch.zeros([1, 8, 4], device=self.device, dtype=self.dtype)
weight = torch.empty([4], device=self.device, dtype=self.dtype)
scale = torch.tensor(1.0, device=self.device, dtype=torch.float32)
return [input, weight, scale]
def register(self, pm_pass: PatternMatcherPass):
def register(self, pm_pass: PatternMatcherPass) -> None:
def pattern(
input: torch.Tensor,
weight: torch.Tensor,
scale: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
all_reduce = self._all_reduce(input)
rms = self.rmsnorm_matcher(all_reduce, weight)
quant, _ = self.quant_matcher(rms, scale)
@@ -189,7 +193,7 @@ class FirstAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
input: torch.Tensor,
weight: torch.Tensor,
scale: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
reduce_scatter = self._reduce_scatter(input)
rms = self.rmsnorm_matcher(reduce_scatter, weight)
quant, _ = self.quant_matcher(rms, scale)
@@ -203,12 +207,12 @@ class FirstAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
class MiddleAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
def __init__(self, epsilon: float, dtype: torch.dtype, device: str | None):
def __init__(self, epsilon: float, dtype: torch.dtype, device: str | None) -> None:
super().__init__(epsilon, dtype, device)
self.rmsnorm_matcher = MatcherFusedAddRMSNorm(epsilon)
self.quant_matcher = MatcherQuantFP8(kFp8StaticTensorSym)
def get_inputs(self):
def get_inputs(self) -> list[torch.Tensor]:
mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)
residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
rms_norm_weights = torch.empty([4, 4], device=self.device, dtype=self.dtype)
@@ -216,7 +220,7 @@ class MiddleAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
return [residual, mm_1, rms_norm_weights, scale]
def register(self, pm_pass: PatternMatcherPass):
def register(self, pm_pass: PatternMatcherPass) -> None:
def pattern(
residual: torch.Tensor,
mm_1: torch.Tensor,
@@ -302,7 +306,7 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
"""
@enable_fake_mode
def __init__(self, config: VllmConfig):
def __init__(self, config: VllmConfig) -> None:
super().__init__(config)
# Used to clean up redundant views created temporarily
@@ -357,7 +361,7 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
return (compile_range.is_single_size()) and (compile_range.end % tp_size == 0)
@VllmInductorPass.time_and_log
def __call__(self, graph: fx.Graph):
def __call__(self, graph: fx.Graph) -> None:
self.matched_count = self.patterns.apply(graph)
logger.debug("Replaced %s patterns", self.matched_count)
# Clean up reshape nodes