[vLLM IR] 1/N Implement IR skeleton and rms_norm op (#33825)

Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
Signed-off-by: chzhang <chaojun.zhang@intel.com>
Signed-off-by: Luka Govedic <luka.govedic@gmail.com>
Co-authored-by: Xinyu Chen <xinyu1.chen@intel.com>
Co-authored-by: Chaojun Zhang <chaojun.zhang@intel.com>
Co-authored-by: Luka Govedič <ProExpertProg@h100-01.nemg-001.lab.rdu2.dc.redhat.com>
This commit is contained in:
Luka Govedič
2026-03-31 22:15:05 -04:00
committed by GitHub
co-authored by Xinyu Chen Chaojun Zhang Luka Govedič
parent 0fab52f0aa
commit 40bb175027
49 changed files with 2177 additions and 265 deletions
+23 -2
View File
@@ -8,7 +8,7 @@ import functools
import json
import sys
from collections.abc import Callable
from dataclasses import MISSING, dataclass, fields, is_dataclass
from dataclasses import MISSING, asdict, dataclass, fields, is_dataclass
from itertools import permutations
from types import UnionType
from typing import (
@@ -70,7 +70,7 @@ from vllm.config.cache import (
PrefixCachingHashAlgo,
)
from vllm.config.device import Device
from vllm.config.kernel import MoEBackend
from vllm.config.kernel import IrOpPriorityConfig, MoEBackend
from vllm.config.lora import MaxLoRARanks
from vllm.config.model import (
ConvertOption,
@@ -401,6 +401,7 @@ class EngineArgs:
max_cudagraph_capture_size: int | None = get_field(
CompilationConfig, "max_cudagraph_capture_size"
)
ir_op_priority: IrOpPriorityConfig = get_field(KernelConfig, "ir_op_priority")
# Note: Specifying a custom executor backend by passing a class
# is intended for expert use only. The API may change without
# notice.
@@ -657,6 +658,9 @@ class EngineArgs:
self.weight_transfer_config = WeightTransferConfig(
**self.weight_transfer_config
)
if isinstance(self.ir_op_priority, dict):
self.ir_op_priority = IrOpPriorityConfig(**self.ir_op_priority)
# Setup plugins
from vllm.plugins import load_general_plugins
@@ -1293,6 +1297,7 @@ class EngineArgs:
title="KernelConfig",
description=KernelConfig.__doc__,
)
kernel_group.add_argument("--ir-op-priority", **kernel_kwargs["ir_op_priority"])
kernel_group.add_argument(
"--enable-flashinfer-autotune",
**kernel_kwargs["enable_flashinfer_autotune"],
@@ -1917,6 +1922,22 @@ class EngineArgs:
if self.moe_backend != "auto":
kernel_config.moe_backend = self.moe_backend
# Transfer top-level ir_op_priority into KernelConfig.ir_op_priority
for op_name, op_priority in asdict(self.ir_op_priority).items():
# Empty means unset
if not op_priority:
continue
# Priority cannot be set 2x for the same op
if getattr(kernel_config.ir_op_priority, op_name):
raise ValueError(
f"Op priority for {op_name} specified via both ir_op_priority "
f"and KernelConfig.ir_op_priority, only one allowed at a time."
)
# Set the attribute
setattr(kernel_config.ir_op_priority, op_name, op_priority)
load_config = self.create_load_config()
# Pass reasoning_parser into StructuredOutputsConfig