GitLab

移除 forward 中对 experts.use_overlapped/_shared_experts 的状态改写，避免 torch.compile 启动期 shared/non-shared 路径不一致
FusedMoE.forward_impl 仅在 shared_output 为空时计算 shared experts，防止透传值被本地重算覆盖

  新增环境变量 VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD 用于控制
  fused sum+mul+add 开关。
  在 DeepseekV2MoE 中增加 fused 路径，预计算 shared_output，并下传 iqis 与 routed_scaling_factor。
  扩展 FusedMoE/SharedFusedMoE 及相关 custom op 接口，统一透传 i_q/i_s/shared_output/routed_scaling_factor。
  同步适配 Triton、Marlin W16A16、SlimQuant W4A8、CompressedTensors W8A8 等实现，支持在内核侧完成 sum+mul+add。

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Signed-off-by: ganyi <ygan@amd.com>

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

Signed-off-by: Kebe <mail@kebe7jun.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

Signed-off-by: Hollow Man <hollowman@opensuse.org>

Signed-off-by: ganyi <ygan@amd.com>

Signed-off-by: yewentao256 <zhyanwentao@126.com>

Signed-off-by: yewentao256 <zhyanwentao@126.com>

Signed-off-by: baoqian <1354987947@qq.com>
Signed-off-by: baoqian426 <1354987947@qq.com>

修复CompressedTensorsLinearMethod中的w4a16的冲突问题
feat(moe): add Marlin W16A16 fused MoE behind VLLM_USE_MARLIN_W16A16_MOE
replace the fp8_mqa_logits and fp8_paged_mqa_logits interfaces in deepgemm with mqa_logits and paged_mqa_logits from lightop

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

Signed-off-by: juliendenize <julien.denize@mistral.ai>
(cherry picked from commit 5e5646e2)

Signed-off-by: juliendenize <julien.denize@mistral.ai>

Signed-off-by: Julien Denize <julien.denize@mistral.ai>
Signed-off-by: Julien Denize <40604584+juliendenize@users.noreply.github.com>
Signed-off-by: Mickael Seznec <mickael@mistral.ai>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Mickael Seznec <mickael@mistral.ai>