[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
+33 -28
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from abc import ABC, abstractmethod
from typing import Any
import torch
from torch._higher_order_ops.auto_functionalize import auto_functionalized
@@ -52,7 +53,7 @@ class ActivationQuantPattern(ABC):
def __init__(
self,
quant_key: QuantKey,
):
) -> None:
self.quant_key = quant_key
self.quant_dtype = quant_key.dtype
@@ -68,12 +69,12 @@ class ActivationQuantPattern(ABC):
self.silu_and_mul_matcher = MatcherSiluAndMul()
def empty_quant(self, *args, **kwargs):
def empty_quant(self, *args: Any, **kwargs: Any) -> torch.Tensor:
kwargs = {"dtype": self.quant_dtype, "device": "cuda", **kwargs}
return torch.empty(*args, **kwargs)
@abstractmethod
def register(self, pm_pass: PatternMatcherPass):
def register(self, pm_pass: PatternMatcherPass) -> None:
raise NotImplementedError
@@ -82,15 +83,22 @@ class SiluMulFp8StaticQuantPattern(ActivationQuantPattern):
Fusion for SiluMul+Fp8StaticQuant Pattern
"""
def __init__(self):
def __init__(self) -> None:
super().__init__(kFp8StaticTensorSym)
self.quant_matcher = MatcherQuantFP8(kFp8StaticTensorSym)
def register(self, pm_pass: PatternMatcherPass):
def get_inputs(self) -> list[torch.Tensor]:
scale = self.quant_matcher.inputs()[1]
return [
*self.silu_and_mul_matcher.inputs(), # input
scale,
]
def register(self, pm_pass: PatternMatcherPass) -> None:
def pattern(
input: torch.Tensor,
scale: torch.Tensor,
):
) -> torch.Tensor:
result_silu_mul = self.silu_and_mul_matcher(input)
result_quant = self.quant_matcher(result_silu_mul, scale)
return result_quant[0]
@@ -98,7 +106,7 @@ class SiluMulFp8StaticQuantPattern(ActivationQuantPattern):
def replacement(
input: torch.Tensor,
scale: torch.Tensor,
):
) -> torch.Tensor:
d = input.shape[-1] // 2
output_shape = input.shape[:-1] + (d,)
result = torch.empty(
@@ -109,13 +117,10 @@ class SiluMulFp8StaticQuantPattern(ActivationQuantPattern):
)
return at[1]
inputs = [
*self.silu_and_mul_matcher.inputs(), # input
self.quant_matcher.inputs()[1], # scale
]
pattern(*inputs)
inps = self.get_inputs()
pattern(*inps)
register_replacement(pattern, replacement, inputs, fwd_only, pm_pass)
register_replacement(pattern, replacement, inps, fwd_only, pm_pass)
class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
@@ -123,16 +128,23 @@ class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
Fusion for SiluMul+Nvfp4Quant Pattern
"""
def __init__(self):
def __init__(self) -> None:
super().__init__(kNvfp4Quant)
def register(self, pm_pass: PatternMatcherPass):
def get_inputs(self) -> list[torch.Tensor]:
result = self.empty_quant(5, 32)
output_scale = empty_i32(128, 4)
input_ = empty_bf16(5, 64)
scale = empty_fp32(1, 1)
return [result, output_scale, input_, scale]
def register(self, pm_pass: PatternMatcherPass) -> None:
def pattern(
result: torch.Tensor,
output_scale: torch.Tensor,
input: torch.Tensor,
scale: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
result_silu_mul = self.silu_and_mul_matcher(input)
at = auto_functionalized(
self.QUANT_OP,
@@ -148,7 +160,7 @@ class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
output_scale: torch.Tensor,
input: torch.Tensor,
scale: torch.Tensor,
):
) -> tuple[torch.Tensor, torch.Tensor]:
at = auto_functionalized(
self.FUSED_OP,
result=result,
@@ -158,14 +170,7 @@ class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
)
return at[1], at[2]
inputs = [
self.empty_quant(5, 32), # result
empty_i32(128, 4), # output_scale
empty_bf16(5, 64), # input
empty_fp32(1, 1), # scale
]
register_replacement(pattern, replacement, inputs, fwd_only, pm_pass)
register_replacement(pattern, replacement, self.get_inputs(), fwd_only, pm_pass)
class ActivationQuantFusionPass(VllmPatternMatcherPass):
@@ -179,7 +184,7 @@ class ActivationQuantFusionPass(VllmPatternMatcherPass):
"""
@enable_fake_mode
def __init__(self, config: VllmConfig):
def __init__(self, config: VllmConfig) -> None:
super().__init__(config)
self.patterns: PatternMatcherPass = PatternMatcherPass(
@@ -196,11 +201,11 @@ class ActivationQuantFusionPass(VllmPatternMatcherPass):
self.dump_patterns(config, self.patterns)
@VllmInductorPass.time_and_log
def __call__(self, graph: torch.fx.Graph):
def __call__(self, graph: torch.fx.Graph) -> None:
self.matched_count = self.patterns.apply(graph)
logger.debug("Replaced %s patterns", self.matched_count)
def uuid(self):
def uuid(self) -> str:
return VllmInductorPass.hash_source(
self,
ActivationQuantPattern,