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

[Model] Apply shared experts overlap optimization to all models with shared experts (#26145)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-10-09 11:31:04 -04:00
committed by GitHub
parent 3b736e1c38
commit 47e66c24e2
15 changed files with 285 additions and 297 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/llama4.py
View File

@@ -33,7 +33,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
QKVParallelLinear,
@@ -42,7 +42,6 @@ from vllm.model_executor.layers.linear import (
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.shared_fused_moe import SharedFusedMoE
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
@@ -399,7 +398,7 @@ class Llama4Model(LlamaModel):
params_dict: The dictionary of module parameters.
loaded_params: The set of already loaded parameters.
expert_params_mapping: The mapping of expert parameters. Must be
generated by FusedMoE.make_expert_params_mapping().
generated by SharedFusedMoE.make_expert_params_mapping().
fused: Whether the expert weights are fused into a single weight
tensor or are separate weight tensors for each expert.
When fused is True, loaded_weight should have shape of:
@@ -522,7 +521,7 @@ class Llama4Model(LlamaModel):
fused_experts_params = False
# Expert parameter mapping for the case where the expert weights are
# not fused into a single weight tensor.
expert_params_mapping = FusedMoE.make_expert_params_mapping(
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
@@ -530,7 +529,7 @@ class Llama4Model(LlamaModel):
)
# Expert parameter mapping for the case where the expert weights are
# fused into a single weight tensor.
expert_params_mapping_fused = FusedMoE.make_expert_params_mapping(
expert_params_mapping_fused = SharedFusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_up_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="gate_up_proj",