[PERF]MiniMax-M2 gate kernel (#38445)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com> Signed-off-by: qianlihuang <91178480+qianlihuang@users.noreply.github.com> Co-authored-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
This commit is contained in:
@@ -43,10 +43,10 @@ from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.router.gate_linear import GateLinear
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
@@ -113,12 +113,12 @@ class MiniMaxM2MoE(nn.Module):
|
||||
router_logits_dtype=torch.float32,
|
||||
)
|
||||
|
||||
self.gate = ReplicatedLinear(
|
||||
self.gate = GateLinear(
|
||||
config.hidden_size,
|
||||
config.num_local_experts,
|
||||
bias=False,
|
||||
params_dtype=torch.float32,
|
||||
quant_config=None,
|
||||
out_dtype=torch.float32,
|
||||
prefix=f"{prefix}.gate",
|
||||
)
|
||||
|
||||
@@ -132,7 +132,7 @@ class MiniMaxM2MoE(nn.Module):
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
# router_logits: (num_tokens, n_experts)
|
||||
router_logits, _ = self.gate(hidden_states.to(torch.float32))
|
||||
router_logits, _ = self.gate(hidden_states)
|
||||
final_hidden_states = self.experts(
|
||||
hidden_states=hidden_states, router_logits=router_logits
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user