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

[Quantization][1/N] MoE support BNB-Inflight Quantization (#20061)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2025-07-11 16:01:13 +08:00
committed by GitHub
parent 762be26a8e
commit 8020e98c9f
8 changed files with 561 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/model_executor/models/phimoe.py
View File

@@ -516,6 +516,14 @@ class PhiMoEModel(nn.Module):
hidden_states = self.norm(hidden_states)
return hidden_states
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return FusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="w1",
ckpt_down_proj_name="w2",
ckpt_up_proj_name="w3",
num_experts=self.config.num_local_experts,
)
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
stacked_params_mapping = [
@@ -672,3 +680,6 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)
return loader.load_weights(weights)
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return self.model.get_expert_mapping()