apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-24 01:29:57 +08:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

[Models] Add forward_meta to moe models' forward function (#5138)

Browse Source 
* [Models] Add forward_meta to moe models' forward function

* fix missing param

* fix

* fix

* fix forward_meta

* fix test and remove chunked MoE releated in config

* fix test

* fix

* fix
This commit is contained in:
Longzhi Wang
2025-12-04 13:26:58 +08:00
committed by GitHub
parent f5bdb36e9b
commit 5cd17fd662
21 changed files with 131 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/models/qwen3moe.py
+6 -6
View File
@@ -79,8 +79,8 @@ class Qwen3MoeBlock(nn.Layer):
weight_dtype="float32",
)
def forward(self, x):
return self.experts(x, self.gate)
def forward(self, x, forward_meta):
return self.experts(x, self.gate, forward_meta)
def load_state_dict(self, state_dict):
""" """
@@ -125,7 +125,7 @@ class Qwen3MLP(nn.Layer):
self.up_gate_proj.load_state_dict(state_dict)
self.down_proj.load_state_dict(state_dict)
def forward(self, x):
def forward(self, x, forward_meta):
""" """
gate_up_out = self.up_gate_proj(x)
act_out = self.act_fn(gate_up_out)
@@ -204,7 +204,7 @@ class Qwen3DecoderLayer(nn.Layer):
# Fully Connected
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
hidden_states = self.mlp(hidden_states)
hidden_states = self.mlp(hidden_states, forward_meta)
return hidden_states, residual
@@ -416,7 +416,7 @@ class Qwen3MoeForCausalLM(ModelForCasualLM):
return logits
def empty_input_forward(self):
def empty_input_forward(self, forward_meta):
"""
empty_input_forward
"""
@@ -428,7 +428,7 @@ class Qwen3MoeForCausalLM(ModelForCasualLM):
self.fd_config.model_config.moe_layer_start_index,
self.fd_config.model_config.num_hidden_layers,
):
self.model.layers[i].mlp.experts(fake_hidden_states, self.model.layers[i].mlp.gate)
self.model.layers[i].mlp.experts(fake_hidden_states, self.model.layers[i].mlp.gate, forward_meta)
def forward(
self,