[ROCm][FEAT] Fuse DeepSeek shared experts into AITER fused_moe ops (#24097)

Signed-off-by: chenjun <junchen2@amd.com>
Signed-off-by: kliuae <kuanfu.liu@embeddedllm.com>
Co-authored-by: valarLip <103567126+valarLip@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
kliuae
2025-10-16 10:41:34 +08:00
committed by GitHub
co-authored by valarLip TJian
parent 0ecc553ee6
commit 1317034379
8 changed files with 347 additions and 83 deletions
@@ -1056,6 +1056,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
routed_scaling_factor=routed_scaling_factor,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
num_fused_shared_experts=layer.num_fused_shared_experts,
)
per_act_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
@@ -1169,6 +1169,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
global_num_experts=global_num_experts,
zero_expert_num=zero_expert_num,
zero_expert_type=zero_expert_type,
num_fused_shared_experts=layer.num_fused_shared_experts,
)
#