Upstream Llama4 Support to Main (#16113)

Signed-off-by: Aston Zhang <22279212+astonzhang@users.noreply.github.com>
Signed-off-by: Chris Thi <chris.c.thi@gmail.com>
Signed-off-by: drisspg <drisspguessous@gmail.com>
Signed-off-by: Jon Swenson <jmswen@gmail.com>
Signed-off-by: Keyun Tong <tongkeyun@gmail.com>
Signed-off-by: Lu Fang <fanglu@meta.com>
Signed-off-by: Xiaodong Wang <xdwang@meta.com>
Signed-off-by: Yang Chen <yangche@fb.com>
Signed-off-by: Ye (Charlotte) Qi <yeq@meta.com>
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>
Signed-off-by: Lu Fang <lufang@fb.com>
Signed-off-by: Lu Fang <fanglu@fb.com>
Signed-off-by: Lucia Fang <fanglu@fb.com>
Signed-off-by: Roger Wang <ywang@roblox.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Lu Fang <fanglu@fb.com>
Co-authored-by: Roger Wang <ywang@roblox.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Lu Fang
2025-04-07 08:06:27 -07:00
committed by GitHub
co-authored by Lu Fang Roger Wang DarkLight1337
parent 8017c8db7f
commit 55dcce91df
43 changed files with 2436 additions and 155 deletions
@@ -113,6 +113,7 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -129,18 +130,20 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
return fused_experts(x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
use_int8_w8a16=True,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_scale,
w2_scale=layer.w2_scale)
return fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
use_int8_w8a16=True,
global_num_experts=global_num_experts,
apply_router_weight_on_input=apply_router_weight_on_input,
expert_map=expert_map,
w1_scale=layer.w13_scale,
w2_scale=layer.w2_scale)
@staticmethod
def quantizing_weight_loader(layer, weight_loader):