biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel] [Quantization] Add MXFP4 and bias support for marlin kernel (#22428)

Browse Source 
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
Signed-off-by: Huzaifa Sidhpurwala <huzaifas@redhat.com>
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Animesh Jain <anijain@umich.edu>
Signed-off-by: Rui Qiao <ruisearch42@gmail.com>
Signed-off-by: Xiongfei Wei <isaacwxf23@gmail.com>
Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: kf <kuanfu.liu@embeddedllm.com>
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Dipika Sikka <dipikasikka1@gmail.com>
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Signed-off-by: tjtanaavllm <tunjian.tan@amd.com>
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
Signed-off-by: Chih-Chieh-Yang <7364402+cyang49@users.noreply.github.com>
Signed-off-by: Roger Wang <hey@rogerw.me>
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
Signed-off-by: Isotr0py <2037008807@qq.com>
Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Signed-off-by: Chih-Chieh Yang <7364402+cyang49@users.noreply.github.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Signed-off-by: yan <yan.ma@intel.com>
Signed-off-by: Yan Ma <yan.ma@intel.com>
Signed-off-by: Xiao Liu <xiszishu@gmail.com>
Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Ye (Charlotte) Qi <yeq@meta.com>
Signed-off-by: LopezCastroRoberto <roberto.lopez.castro@udc.es>
Signed-off-by: Andy Xie <andy.xning@gmail.com>
Signed-off-by: Haibin Lin <haibin.lin@bytedance.com>
Signed-off-by: David Ben-David <davidb@pliops.com>
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Signed-off-by: zitian.zhao <zitian.zhao@tencentmusic.com>
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
Signed-off-by: Abirdcfly <fp544037857@gmail.com>
Signed-off-by: Giancarlo Delfin <gdelfin@meta.com>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Signed-off-by: huangweixiao <huangweixiao@msh.team>
Signed-off-by: alyosha-swamy <raghav@arcee.ai>
Signed-off-by: Eric Hanley <ericehanley@google.com>
Signed-off-by: Abatom <abzhonghua@gmail.com>
Signed-off-by: CLFutureX <775523362@qq.com>
Signed-off-by: Linkun Chen <github@lkchen.net>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Signed-off-by: tlipoca9 <tlipoca9@gmail.com>
Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Signed-off-by: zitian zhao <zitian.zhao@tencentmusic.com>
Signed-off-by: mgoin <michael@neuralmagic.com>
Signed-off-by: wang.yuqi <noooop@126.com>
Signed-off-by: Benji Beck <benjibeck@meta.com>
Signed-off-by: Siyuan Liu <lsiyuan@google.com>
Signed-off-by: Benjamin Chislett <benjamin.chislett@centml.ai>
Signed-off-by: isotr0py <2037008807@qq.com>
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Signed-off-by: simon-mo <xmo@berkeley.edu>
Signed-off-by: LucasWilkinson <lwilkinson@neuralmagic.com>
Signed-off-by: Zhang Jason <ning.zhang2@amd.com>
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Signed-off-by: asafg <asafg@ai21.com>
Signed-off-by: Siyuan Fu <siyuanf@nvidia.com>
Signed-off-by: Lain <fusiyuan2000@hotmail.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Signed-off-by: QscQ <qscqesze@gmail.com>
Signed-off-by: qingjun <qingjun@minimaxi.com>
Signed-off-by: Syed Muhammad Bin Asif <syedmba7@connect.hku.hk>
Signed-off-by: Lionel Villard <villard@us.ibm.com>
Signed-off-by: ycyaw66 <497410282@qq.com>
Signed-off-by: David Chen <530634352@qq.com>
Signed-off-by: Linkun <github@lkchen.net>
Signed-off-by: Moritz Sanft <58110325+msanft@users.noreply.github.com>
Signed-off-by: Ming Yang <minos.future@gmail.com>
Signed-off-by: Adrian Garcia <adrian.garcia@inceptionai.ai>
Signed-off-by: shaojunqi <shaojunqi.sjq@alibaba-inc.com>
Signed-off-by: Ricardo Decal <rdecal@anyscale.com>
Signed-off-by: Andrew Chan <andrewkchan.akc@gmail.com>
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: Andrew Sansom <andrew@protopia.ai>
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
Signed-off-by: Shu Wang <shuw@nvidia.com>
Signed-off-by: Po-Han Huang <pohanh@nvidia.com>
Signed-off-by: Shu Wang. <shuw@nvidia.com>
Signed-off-by: XIn Li <xinli@nvidia.com>
Signed-off-by: Junhao Li <junhao@ubicloud.com>
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Signed-off-by: iAmir97 <Amir.balwel@embeddedllm.com>
Signed-off-by: iAmir97 <71513472+iAmir97@users.noreply.github.com>
Signed-off-by: <zyy1102000@gmail.com>
Signed-off-by: Guy Stone <guys@spotify.com>
Signed-off-by: <yyweiss@gmail.com>
Signed-off-by: yyw <yyweiss@gmail.com>
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Signed-off-by: Pradyun Ramadorai <pradyunr@amazon.com>
Signed-off-by: Pradyun92 <142861237+Pradyun92@users.noreply.github.com>
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Co-authored-by: rongfu.leng <rongfu.leng@daocloud.io>
Co-authored-by: Huzaifa Sidhpurwala <huzaifas@redhat.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: Varun Sundar Rabindranath <varunsundar08@gmail.com>
Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Animesh Jain <jainanimesh2305@yahoo.com>
Co-authored-by: Rui Qiao <161574667+ruisearch42@users.noreply.github.com>
Co-authored-by: XiongfeiWei <isaacwxf23@gmail.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: JartX <sagformas@gmail.com>
Co-authored-by: fhl2000 <63384265+fhl2000@users.noreply.github.com>
Co-authored-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: kf <kuanfu.liu@embeddedllm.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>
Co-authored-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: tjtanaavllm <tunjian.tan@amd.com>
Co-authored-by: Yong Hoon Shin <48474650+sarckk@users.noreply.github.com>
Co-authored-by: Chih-Chieh Yang <7364402+cyang49@users.noreply.github.com>
Co-authored-by: Roger Wang <hey@rogerw.me>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Yuxuan Zhang <2448370773@qq.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Yan Ma <yan.ma@intel.com>
Co-authored-by: Xiao <xiszishu@gmail.com>
Co-authored-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Ye (Charlotte) Qi <yeq@meta.com>
Co-authored-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Ning Xie <andy.xning@gmail.com>
Co-authored-by: H <linhaibin.eric@gmail.com>
Co-authored-by: David Ben-David <sdavidbd@gmail.com>
Co-authored-by: David Ben-David <davidb@pliops.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
Co-authored-by: TankNee <nee@tanknee.cn>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Seiji Eicher <58963096+eicherseiji@users.noreply.github.com>
Co-authored-by: ZiTian.Zhao <zitian.zhao@tencentmusic.com>
Co-authored-by: 22quinn <33176974+22quinn@users.noreply.github.com>
Co-authored-by: Abirdcfly <fp544037857@gmail.com>
Co-authored-by: Giancarlo Delfin <32987265+TheEpicDolphin@users.noreply.github.com>
Co-authored-by: Chenxi Yang <cxyang@cs.utexas.edu>
Co-authored-by: Chenxi Yang <cxyang@meta.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Weixiao Huang <hwx.simle@gmail.com>
Co-authored-by: Raghav Ravishankar <113712354+alyosha-swamy@users.noreply.github.com>
Co-authored-by: ericehanley <ericehanley@google.com>
Co-authored-by: Zhonghua Deng <abzhonghua@gmail.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: PiteXChen <44110731+CLFutureX@users.noreply.github.com>
Co-authored-by: lkchen <github@lkchen.net>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com>
Co-authored-by: tlipoca9 <160737620+tlipoca9@users.noreply.github.com>
Co-authored-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: wang.yuqi <noooop@126.com>
Co-authored-by: Benji Beck <benjibeck@meta.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Siyuan Liu <lsiyuan@google.com>
Co-authored-by: Benjamin Chislett <chislett.ben@gmail.com>
Co-authored-by: LiuXiaoxuanPKU <lilyliupku@gmail.com>
Co-authored-by: simon-mo <xmo@berkeley.edu>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
Co-authored-by: Minseok Lee <47620120+minseokl@users.noreply.github.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: Zhang Jason <ning.zhang2@amd.com>
Co-authored-by: Asaf Joseph Gardin <39553475+Josephasafg@users.noreply.github.com>
Co-authored-by: asafg <asafg@ai21.com>
Co-authored-by: Lain <siyuanf@nvidia.com>
Co-authored-by: tc-mb <157115220+tc-mb@users.noreply.github.com>
Co-authored-by: imning3 <hbning@pku.edu.cn>
Co-authored-by: Maximilien de Bayser <mbayser@br.ibm.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: Tao He <linzhu.ht@alibaba-inc.com>
Co-authored-by: qscqesze <qingjun@minimaxi.com>
Co-authored-by: Syed Muhammad Bin Asif <92625830+syedmba@users.noreply.github.com>
Co-authored-by: Lionel Villard <villard@us.ibm.com>
Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>
Co-authored-by: ycyaw66 <497410282@qq.com>
Co-authored-by: Moritz Sanft <58110325+msanft@users.noreply.github.com>
Co-authored-by: Ming Yang <minos.future@gmail.com>
Co-authored-by: Adrián García García <adrigarvk8@gmail.com>
Co-authored-by: Michael Goin <mgoin@redhat.com>
Co-authored-by: JaceyShao <65159281+JaceyShao@users.noreply.github.com>
Co-authored-by: shaojunqi <shaojunqi.sjq@alibaba-inc.com>
Co-authored-by: Ricardo Decal <crypdick@users.noreply.github.com>
Co-authored-by: Andrew Chan <andrewkchan.akc@gmail.com>
Co-authored-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: Andrew Sansom <andrew@protopia.ai>
Co-authored-by: Zhiyu <zhiyuc@nvidia.com>
Co-authored-by: Shu Wang <shuw@nvidia.com>
Co-authored-by: XIn Li <xinli@nvidia.com>
Co-authored-by: Junhao Li <streaver91@gmail.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: iAmir97 <71513472+iAmir97@users.noreply.github.com>
Co-authored-by: iAmir97 <Amir.balwel@embeddedllm.com>
Co-authored-by: Hong Hanh <hanh.usth@gmail.com>
Co-authored-by: Daniel Serebrenik <74646983+pliops-daniels@users.noreply.github.com>
Co-authored-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: Guy Stone <guys@spotify.com>
Co-authored-by: yyweiss <70619747+yyweiss@users.noreply.github.com>
Co-authored-by: Pradyun92 <142861237+Pradyun92@users.noreply.github.com>
Co-authored-by: Pradyun Ramadorai <pradyunr@amazon.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
This commit is contained in:
Jinzhen Lin
2025-08-15 02:23:22 +08:00
committed by GitHub
parent ab9f2cfd19
commit 33c63e9547
34 changed files with 1128 additions and 324 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
vllm/model_executor/layers/quantization/mxfp4.py
View File

@@ -15,13 +15,17 @@ from vllm.model_executor.layers.linear import (LinearBase,
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
prepare_moe_fp4_layer_for_marlin)
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
_can_support_mxfp4, _swizzle_mxfp4)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
is_layer_skipped)
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.utils import next_power_of_2, round_up
from vllm.scalar_type import scalar_types
from vllm.utils import (has_triton_kernels, is_torch_equal_or_newer,
next_power_of_2, round_up)
if (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):
@@ -81,6 +85,21 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
super().__init__()
self.topk_indices_dtype = None
self.moe = moe
self.use_marlin = self._should_use_marlin()
def _should_use_marlin(self):
if envs.VLLM_MXFP4_USE_MARLIN is not None:
return envs.VLLM_MXFP4_USE_MARLIN
if current_platform.is_cuda() and \
not current_platform.has_device_capability(100):
if not current_platform.is_device_capability(90):
# marlin kernel has better performance on ampere
return True
if not has_triton_kernels():
return True
if not is_torch_equal_or_newer("2.8.0"):
return True
return False
def create_weights(self, layer: torch.nn.Module, num_experts: int,
hidden_size: int, intermediate_size_per_partition: int,
@@ -101,11 +120,29 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
intermediate_size_per_partition_after_pad = \
intermediate_size_per_partition
# pad the intermediate size to be a multiple of 2 * mxfp4_block
# for to hold non-uniform sharded tensor as well as swizzling
# other padding to increase performance
if (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):
if self.use_marlin:
# The moe marlin kernel requires that for each linear
# n % 256 == 0 and k % 128 == 0.
# In gate_up_proj:
# n = 2 * intermediate_size_per_partition_after_pad
# k = hidden_size
# In down_proj
# n = hidden_size
# k = intermediate_size_per_partition_after_pad
intermediate_size_per_partition_after_pad = round_up(
intermediate_size_per_partition, 128)
hidden_size = round_up(hidden_size, 256)
layer.params_dtype = params_dtype
layer.num_experts = num_experts
layer.hidden_size = hidden_size
layer.intermediate_size_per_partition = \
intermediate_size_per_partition_after_pad
elif (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):
# pad the intermediate size to be a multiple of 2 * mxfp4_block
# for to hold non-uniform sharded tensor as well as swizzling
# other padding to increase performance
intermediate_size_per_partition_after_pad = round_up(
intermediate_size_per_partition, 256)
hidden_size = round_up(hidden_size, 256)
@@ -191,8 +228,10 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
set_weight_attrs(w2_bias, extra_weight_attrs)
def process_weights_after_loading(self, layer):
if (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):
if self.use_marlin:
prepare_moe_fp4_layer_for_marlin(layer)
elif (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):
layer.gemm1_alpha = Parameter(torch.tensor(
[1.702] * self.num_experts, dtype=torch.float32).cuda(),
requires_grad=False)
@@ -399,13 +438,45 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
if enable_eplb:
raise NotImplementedError("EPLB is not supported for mxfp4")
if self.use_marlin:
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
router_logits=router_logits,
use_grouped_topk=use_grouped_topk,
top_k=top_k,
renormalize=renormalize,
topk_group=topk_group,
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
return torch.ops.vllm.fused_marlin_moe(
x,
layer.w13_weight,
layer.w2_weight,
layer.w13_bias,
layer.w2_bias,
layer.w13_weight_scale,
layer.w2_weight_scale,
router_logits,
topk_weights,
topk_ids,
global_scale1=None,
global_scale2=None,
quant_type_id=scalar_types.float4_e2m1f.id,
apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts,
activation=activation,
expert_map=expert_map)
assert _can_support_mxfp4(
use_grouped_topk, topk_group, num_expert_group, expert_map,
custom_routing_function, e_score_correction_bias,
apply_router_weight_on_input, scoring_func, activation,
expert_load_view, logical_to_physical_map,
logical_replica_count), ("MXFP4 are not supported\
with this configuration.")
logical_replica_count), (
"MXFP4 are not supported with this configuration.")
if (envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):