Expert Parallelism (EP) Support for DeepSeek V2 (#12583)
This commit is contained in:
@@ -670,6 +670,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
|
||||
use_grouped_topk: bool = False,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
global_num_experts: int = -1,
|
||||
expert_map: Optional[torch.Tensor] = None,
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
scoring_func: str = "softmax",
|
||||
e_score_correction_bias: Optional[torch.Tensor] = None,
|
||||
@@ -697,6 +699,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
|
||||
topk_ids=topk_ids,
|
||||
inplace=True,
|
||||
use_fp8_w8a8=True,
|
||||
global_num_experts=global_num_experts,
|
||||
expert_map=expert_map,
|
||||
w1_scale=(layer.w13_weight_scale_inv
|
||||
if self.block_quant else layer.w13_weight_scale),
|
||||
w2_scale=(layer.w2_weight_scale_inv
|
||||
|
||||
Reference in New Issue
Block a user