[MOE] Standardize and abstract fused shared expert optimization selection (#51695)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-08-18 12:56:56 -04:00
committed by GitHub
parent f9f066d195
commit 88b2bff2c6
20 changed files with 1288 additions and 281 deletions
+22 -10
View File
@@ -54,6 +54,10 @@ from vllm.model_executor.layers.fused_moe import (
GateLinear,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.fused_moe.utils import (
is_model_fused_shared_expert_compatible,
resolve_layer_fused_shared_expert,
)
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
@@ -329,9 +333,13 @@ class DeepseekV2MoE(nn.Module):
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
self.is_fusion_moe_shared_experts_enabled = (
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
)
self.is_fused_shared_expert_enabled = False
if config.n_shared_experts is not None:
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
quant_config, prefix
)
if (
self.is_rocm_aiter_moe_enabled
and self.gate.e_score_correction_bias is not None
@@ -340,7 +348,7 @@ class DeepseekV2MoE(nn.Module):
# Accumulates in fp32; avoids bf16->fp32 cast.
self.gate.set_out_dtype(self.gate.weight.dtype)
if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
if config.n_shared_experts is None or self.is_fused_shared_expert_enabled:
self.shared_experts = None
else:
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
@@ -377,8 +385,9 @@ class DeepseekV2MoE(nn.Module):
is_sequence_parallel=self.is_sequence_parallel,
reduce_results=reduce_results,
n_shared_experts=config.n_shared_experts
if self.is_fusion_moe_shared_experts_enabled
if self.is_fused_shared_expert_enabled
else None,
fuse_shared_experts=self.is_fused_shared_expert_enabled,
router_logits_dtype=self.gate.out_dtype,
)
@@ -1401,6 +1410,12 @@ class DeepseekV2Model(nn.Module):
prefix=f"{prefix}.layers",
)
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
self.layers,
DeepseekV2MoE,
"mlp",
)
if get_pp_group().is_last_rank:
self.norm = RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
else:
@@ -1511,9 +1526,6 @@ class DeepseekV2Model(nn.Module):
return hidden_states
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
rocm_aiter_moe_shared_expert_enabled = (
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
)
stacked_params_mapping: list[tuple[str, str, int | str]] = [
# (param_name, shard_name, shard_id)
("gate_up_proj", "gate_proj", 0),
@@ -1554,7 +1566,7 @@ class DeepseekV2Model(nn.Module):
num_experts=self.config.n_routed_experts
+ (
self.config.n_shared_experts
if rocm_aiter_moe_shared_expert_enabled
if self.is_fused_shared_expert_enabled
else 0
),
num_redundant_experts=self.num_redundant_experts,
@@ -1582,7 +1594,7 @@ class DeepseekV2Model(nn.Module):
continue # this layer has no indexer; drop its checkpoint weights
is_fusion_moe_shared_experts_layer = (
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
)
if _try_load_fp8_indexer_wk(