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

[Quantization] Enable BNB support for more MoE models (#21370)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2025-07-25 18:57:34 +08:00
committed by GitHub
parent 7311f74468
commit 5c3f2628d5
2 changed files with 93 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
vllm/model_executor/models/glm4_moe.py
View File

@@ -53,7 +53,7 @@ from vllm.model_executor.model_loader.weight_utils import (
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
@@ -461,6 +461,15 @@ class Glm4MoeModel(nn.Module):
device=device),
})
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
# Params for weights, fp8 weight scales, fp8 activation scales
# (param_name, weight_name, expert_id, shard_id)
return FusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts)
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
stacked_params_mapping = [
@@ -472,16 +481,9 @@ class Glm4MoeModel(nn.Module):
("gate_up_proj", "up_proj", 1),
]
# Params for weights, fp8 weight scales, fp8 activation scales
# (param_name, weight_name, expert_id, shard_id)
expert_params_mapping = FusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts)
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
for name, loaded_weight in weights:
spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
if spec_layer is not None:
@@ -570,7 +572,7 @@ class Glm4MoeModel(nn.Module):
return loaded_params
class Glm4MoeForCausalLM(nn.Module, SupportsPP):
class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -677,6 +679,9 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP):
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()
def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
weight_name: str) -> Optional[int]: