[Kernel] Delegate construction of FusedMoEQuantConfig to FusedMoEMethodBase subclasses (#22537)
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
@@ -15,8 +15,8 @@ from vllm.distributed import (divide, get_tensor_model_parallel_rank,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.model_executor.layers.activation import (get_act_and_mul_fn,
|
||||
get_act_fn)
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
fused_topk, torch_vllm_outplace_fused_experts)
|
||||
from vllm.model_executor.layers.fused_moe import (activation_without_mul,
|
||||
fused_topk)
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
@@ -230,7 +230,7 @@ class NomicMoE(nn.Module):
|
||||
self.hidden_size = hidden_size
|
||||
self.total_intermediate_size = intermediate_size
|
||||
self.intermediate_size = divide(intermediate_size, self.tp_size)
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_act = activation_without_mul(hidden_act)
|
||||
|
||||
if params_dtype is None:
|
||||
params_dtype = torch.get_default_dtype()
|
||||
@@ -297,14 +297,14 @@ class NomicMoE(nn.Module):
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=False)
|
||||
final_hidden_states = torch_vllm_outplace_fused_experts(
|
||||
|
||||
final_hidden_states = torch.ops.vllm.outplace_fused_experts(
|
||||
hidden_states=hidden_states,
|
||||
w1=self.w1,
|
||||
w2=self.w2,
|
||||
topk_weights=topk_weights,
|
||||
topk_ids=topk_ids,
|
||||
activation=self.hidden_act,
|
||||
is_act_and_mul=False,
|
||||
)
|
||||
|
||||
if self.tp_size > 1:
|
||||
|
||||
@@ -37,7 +37,7 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import fused_moe
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
@@ -163,13 +163,19 @@ class DeepseekMoE(nn.Module):
|
||||
shared_output = self.shared_experts(hidden_states)
|
||||
# router_logits: (num_tokens, n_experts)
|
||||
router_logits, _ = self.gate(hidden_states)
|
||||
final_hidden_states = fused_moe(hidden_states,
|
||||
self.w1,
|
||||
self.w2,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=self.config.norm_topk_prob,
|
||||
inplace=True)
|
||||
|
||||
topk_weights, topk_ids, _ = fused_topk(
|
||||
hidden_states,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=self.config.norm_topk_prob)
|
||||
|
||||
final_hidden_states = fused_experts(hidden_states,
|
||||
self.w1,
|
||||
self.w2,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
inplace=True)
|
||||
|
||||
if self.config.n_shared_experts is not None:
|
||||
final_hidden_states = final_hidden_states + shared_output
|
||||
|
||||
@@ -39,7 +39,7 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.model_executor.layers.activation import FatreluAndMul, SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import fused_moe
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
@@ -136,13 +136,18 @@ class MiniCPMMoE(nn.Module):
|
||||
hidden_states = hidden_states.view(-1, self.hidden_size)
|
||||
# router_logits: (num_tokens, n_experts)
|
||||
router_logits, _ = self.gate(hidden_states)
|
||||
final_hidden_states = fused_moe(hidden_states,
|
||||
self.ws,
|
||||
self.w2s,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=True,
|
||||
inplace=True)
|
||||
|
||||
topk_weights, topk_ids, _ = fused_topk(hidden_states,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=True)
|
||||
|
||||
final_hidden_states = fused_experts(hidden_states,
|
||||
self.ws,
|
||||
self.w2s,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
inplace=True)
|
||||
|
||||
if self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(
|
||||
|
||||
@@ -702,4 +702,4 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
|
||||
return loader.load_weights(weights)
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
return self.model.get_expert_mapping()
|
||||
return self.model.get_expert_mapping()
|
||||
Reference in New Issue
Block a user