Add training monkeypatches
Browse files
modeling_qwen3_shared_moe_monkeypatch_liger_cce.py
ADDED
@@ -0,0 +1,280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2025 Charles O. Goddard, The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
#
|
16 |
+
# The following monkeypatches were applied by Doctor Shotgun:
|
17 |
+
#
|
18 |
+
# Liger Kernel (https://github.com/linkedin/Liger-Kernel):
|
19 |
+
# 1. Liger RMSNorm
|
20 |
+
# 2. Liger RoPE
|
21 |
+
# 3. Liger SwiGLUMLP
|
22 |
+
#
|
23 |
+
# Cut Cross-Entropy (https://github.com/apple/ml-cross-entropy):
|
24 |
+
# 1. Cut Cross-Entropy
|
25 |
+
"""PyTorch Qwen3 model with shared expert support."""
|
26 |
+
|
27 |
+
from typing import List, Optional, Union
|
28 |
+
|
29 |
+
import torch
|
30 |
+
from torch import nn
|
31 |
+
import torch.nn.functional as F
|
32 |
+
|
33 |
+
# CCE Patch #
|
34 |
+
from cut_cross_entropy.linear_cross_entropy import LCE_IMPL_DEFAULT
|
35 |
+
from cut_cross_entropy.transformers.utils import (
|
36 |
+
PatchOptions,
|
37 |
+
apply_lce,
|
38 |
+
)
|
39 |
+
_PATCH_OPTS = PatchOptions(
|
40 |
+
impl=LCE_IMPL_DEFAULT,
|
41 |
+
reduction="mean",
|
42 |
+
filter_eps="auto",
|
43 |
+
accum_e_fp32=False,
|
44 |
+
accum_c_fp32=False,
|
45 |
+
filter_e_grad=True,
|
46 |
+
filter_c_grad=True,
|
47 |
+
train_only=False,
|
48 |
+
)
|
49 |
+
# CCE Patch #
|
50 |
+
|
51 |
+
# Liger Patch #
|
52 |
+
from liger_kernel.transformers.rms_norm import LigerRMSNorm
|
53 |
+
from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
|
54 |
+
from liger_kernel.transformers.rope import liger_rotary_pos_emb
|
55 |
+
|
56 |
+
import transformers.models.qwen3_moe.modeling_qwen3_moe
|
57 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
|
58 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeMLP = LigerQwen3MoeSwiGLUMLP
|
59 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
|
60 |
+
# Liger Patch #
|
61 |
+
|
62 |
+
from transformers.modeling_outputs import (
|
63 |
+
MoeCausalLMOutputWithPast,
|
64 |
+
MoeModelOutputWithPast,
|
65 |
+
)
|
66 |
+
from transformers.activations import ACT2FN
|
67 |
+
from transformers.utils import logging
|
68 |
+
from transformers.models.mixtral.modeling_mixtral import (
|
69 |
+
load_balancing_loss_func,
|
70 |
+
)
|
71 |
+
from transformers.models.qwen3_moe.modeling_qwen3_moe import (
|
72 |
+
Qwen3MoeMLP,
|
73 |
+
Qwen3MoeRMSNorm,
|
74 |
+
Qwen3MoeAttention,
|
75 |
+
Qwen3MoeDecoderLayer,
|
76 |
+
Qwen3MoeModel,
|
77 |
+
Qwen3MoeForCausalLM,
|
78 |
+
)
|
79 |
+
from .configuration_qwen3_shared_moe import Qwen3SharedMoeConfig
|
80 |
+
|
81 |
+
import scattermoe
|
82 |
+
|
83 |
+
|
84 |
+
logger = logging.get_logger(__name__)
|
85 |
+
|
86 |
+
|
87 |
+
class Qwen3SharedMoeSparseMoeBlock(nn.Module):
|
88 |
+
def __init__(self, config: Qwen3SharedMoeConfig):
|
89 |
+
super().__init__()
|
90 |
+
self.config = config
|
91 |
+
self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
|
92 |
+
if config.shared_expert_intermediate_size is not None:
|
93 |
+
self.shared_expert = Qwen3MoeMLP(
|
94 |
+
config, intermediate_size=config.shared_expert_intermediate_size
|
95 |
+
)
|
96 |
+
else:
|
97 |
+
self.shared_expert = None
|
98 |
+
self.moe_mlp = scattermoe.mlp.GLUMLP(
|
99 |
+
input_size=self.config.hidden_size,
|
100 |
+
hidden_size=self.config.moe_intermediate_size,
|
101 |
+
num_experts=self.config.num_experts,
|
102 |
+
top_k=self.config.num_experts_per_tok,
|
103 |
+
activation=ACT2FN[config.hidden_act],
|
104 |
+
)
|
105 |
+
|
106 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
107 |
+
# handling of gate/router logits copied from Qwen3MoeSparseMoeBlock
|
108 |
+
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
109 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
110 |
+
# router_logits: (batch * sequence_length, n_experts)
|
111 |
+
router_logits = self.gate(hidden_states)
|
112 |
+
|
113 |
+
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
114 |
+
routing_weights, selected_experts = torch.topk(
|
115 |
+
routing_weights, self.config.num_experts_per_tok, dim=-1
|
116 |
+
)
|
117 |
+
if self.config.norm_topk_prob: # only diff with mixtral sparse moe block!
|
118 |
+
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
119 |
+
# we cast back to the input dtype
|
120 |
+
routing_weights = routing_weights.to(hidden_states.dtype)
|
121 |
+
|
122 |
+
# modified here to use scattermoe + shared_expert
|
123 |
+
hs_0 = self.moe_mlp(hidden_states, routing_weights, selected_experts)
|
124 |
+
|
125 |
+
if self.shared_expert is not None:
|
126 |
+
shared_res = self.shared_expert(hidden_states)
|
127 |
+
res = hs_0 + shared_res
|
128 |
+
else:
|
129 |
+
res = hs_0
|
130 |
+
res = res.reshape(batch_size, sequence_length, hidden_dim)
|
131 |
+
return res, router_logits
|
132 |
+
|
133 |
+
|
134 |
+
class Qwen3SharedMoeDecoderLayer(Qwen3MoeDecoderLayer, nn.Module):
|
135 |
+
def __init__(self, config: Qwen3SharedMoeConfig, layer_idx: int):
|
136 |
+
super().__init__(config, layer_idx)
|
137 |
+
self.hidden_size = config.hidden_size
|
138 |
+
|
139 |
+
self.self_attn = Qwen3MoeAttention(config, layer_idx)
|
140 |
+
|
141 |
+
if (layer_idx not in config.mlp_only_layers) and (
|
142 |
+
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
|
143 |
+
):
|
144 |
+
self.mlp = Qwen3SharedMoeSparseMoeBlock(config)
|
145 |
+
else:
|
146 |
+
self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size)
|
147 |
+
|
148 |
+
self.input_layernorm = Qwen3MoeRMSNorm(
|
149 |
+
config.hidden_size, eps=config.rms_norm_eps
|
150 |
+
)
|
151 |
+
self.post_attention_layernorm = Qwen3MoeRMSNorm(
|
152 |
+
config.hidden_size, eps=config.rms_norm_eps
|
153 |
+
)
|
154 |
+
|
155 |
+
|
156 |
+
class Qwen3SharedMoeModel(Qwen3MoeModel):
|
157 |
+
config_class = Qwen3SharedMoeConfig
|
158 |
+
|
159 |
+
def __init__(self, config: Qwen3SharedMoeConfig):
|
160 |
+
super().__init__(config)
|
161 |
+
self.layers = nn.ModuleList(
|
162 |
+
[
|
163 |
+
Qwen3SharedMoeDecoderLayer(config, layer_idx)
|
164 |
+
for layer_idx in range(config.num_hidden_layers)
|
165 |
+
]
|
166 |
+
)
|
167 |
+
|
168 |
+
|
169 |
+
class Qwen3SharedMoeForCausalLM(Qwen3MoeForCausalLM):
|
170 |
+
config_class = Qwen3SharedMoeConfig
|
171 |
+
|
172 |
+
def __init__(self, config):
|
173 |
+
super().__init__(config)
|
174 |
+
self.model = Qwen3SharedMoeModel(config)
|
175 |
+
self.num_experts = config.num_experts
|
176 |
+
|
177 |
+
# CCE Patch #
|
178 |
+
def forward(
|
179 |
+
self,
|
180 |
+
input_ids: Optional[torch.LongTensor] = None,
|
181 |
+
attention_mask: Optional[torch.Tensor] = None,
|
182 |
+
position_ids: Optional[torch.LongTensor] = None,
|
183 |
+
past_key_values: Optional[list[torch.FloatTensor]] = None,
|
184 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
185 |
+
labels: Optional[torch.LongTensor] = None,
|
186 |
+
use_cache: Optional[bool] = None,
|
187 |
+
output_attentions: Optional[bool] = None,
|
188 |
+
output_hidden_states: Optional[bool] = None,
|
189 |
+
output_router_logits: Optional[bool] = None,
|
190 |
+
cache_position: Optional[torch.LongTensor] = None,
|
191 |
+
logits_to_keep: Union[int, torch.Tensor] = 0,
|
192 |
+
**kwargs,
|
193 |
+
) -> MoeCausalLMOutputWithPast:
|
194 |
+
|
195 |
+
output_attentions = (
|
196 |
+
output_attentions
|
197 |
+
if output_attentions is not None
|
198 |
+
else self.config.output_attentions
|
199 |
+
)
|
200 |
+
output_router_logits = (
|
201 |
+
output_router_logits
|
202 |
+
if output_router_logits is not None
|
203 |
+
else self.config.output_router_logits
|
204 |
+
)
|
205 |
+
|
206 |
+
output_hidden_states = (
|
207 |
+
output_hidden_states
|
208 |
+
if output_hidden_states is not None
|
209 |
+
else self.config.output_hidden_states
|
210 |
+
)
|
211 |
+
|
212 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
213 |
+
outputs: MoeModelOutputWithPast = self.model(
|
214 |
+
input_ids=input_ids,
|
215 |
+
attention_mask=attention_mask,
|
216 |
+
position_ids=position_ids,
|
217 |
+
past_key_values=past_key_values,
|
218 |
+
inputs_embeds=inputs_embeds,
|
219 |
+
use_cache=use_cache,
|
220 |
+
output_attentions=output_attentions,
|
221 |
+
output_hidden_states=output_hidden_states,
|
222 |
+
output_router_logits=output_router_logits,
|
223 |
+
cache_position=cache_position,
|
224 |
+
**kwargs,
|
225 |
+
)
|
226 |
+
|
227 |
+
hidden_states = outputs.last_hidden_state
|
228 |
+
|
229 |
+
if hidden_states is None:
|
230 |
+
raise ValueError("hidden_states is None")
|
231 |
+
|
232 |
+
loss = None
|
233 |
+
logits = None
|
234 |
+
|
235 |
+
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
236 |
+
slice_indices = (
|
237 |
+
slice(-logits_to_keep, None)
|
238 |
+
if isinstance(logits_to_keep, int)
|
239 |
+
else logits_to_keep
|
240 |
+
)
|
241 |
+
|
242 |
+
if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
|
243 |
+
assert labels is not None
|
244 |
+
loss = apply_lce(
|
245 |
+
hidden_states[:, slice_indices, :],
|
246 |
+
self.lm_head.weight,
|
247 |
+
labels,
|
248 |
+
_PATCH_OPTS,
|
249 |
+
**kwargs,
|
250 |
+
)
|
251 |
+
else:
|
252 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
253 |
+
|
254 |
+
if labels is not None:
|
255 |
+
loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
|
256 |
+
|
257 |
+
aux_loss = None
|
258 |
+
if output_router_logits:
|
259 |
+
aux_loss = load_balancing_loss_func(
|
260 |
+
outputs.router_logits,
|
261 |
+
self.num_experts,
|
262 |
+
self.num_experts_per_tok,
|
263 |
+
attention_mask,
|
264 |
+
)
|
265 |
+
if labels is not None:
|
266 |
+
loss += self.router_aux_loss_coef * aux_loss.to(
|
267 |
+
loss.device
|
268 |
+
) # make sure to reside in the same device
|
269 |
+
|
270 |
+
return MoeCausalLMOutputWithPast(
|
271 |
+
loss=loss,
|
272 |
+
aux_loss=aux_loss,
|
273 |
+
logits=logits,
|
274 |
+
past_key_values=outputs.past_key_values,
|
275 |
+
hidden_states=outputs.hidden_states,
|
276 |
+
attentions=outputs.attentions,
|
277 |
+
router_logits=outputs.router_logits,
|
278 |
+
)
|
279 |
+
# CCE Patch #
|
280 |
+
|
modeling_qwen3_shared_moe_monkeypatch_liger_flce.py
ADDED
@@ -0,0 +1,279 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2025 Charles O. Goddard, The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
#
|
16 |
+
# The following monkeypatches were applied by Doctor Shotgun:
|
17 |
+
#
|
18 |
+
# Liger Kernel (https://github.com/linkedin/Liger-Kernel):
|
19 |
+
# 1. Liger RMSNorm
|
20 |
+
# 2. Liger RoPE
|
21 |
+
# 3. Liger SwiGLUMLP
|
22 |
+
# 4. Liger Fused Linear Cross-Entropy
|
23 |
+
"""PyTorch Qwen3 model with shared expert support."""
|
24 |
+
|
25 |
+
from typing import List, Optional, Union
|
26 |
+
|
27 |
+
import torch
|
28 |
+
from torch import nn
|
29 |
+
import torch.nn.functional as F
|
30 |
+
|
31 |
+
# Liger Patch #
|
32 |
+
from liger_kernel.transformers.rms_norm import LigerRMSNorm
|
33 |
+
from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
|
34 |
+
from liger_kernel.transformers.rope import liger_rotary_pos_emb
|
35 |
+
from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
|
36 |
+
|
37 |
+
import transformers.models.qwen3_moe.modeling_qwen3_moe
|
38 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
|
39 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeMLP = LigerQwen3MoeSwiGLUMLP
|
40 |
+
transformers.models.qwen3_moe.modeling_qwen3_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
|
41 |
+
# Liger Patch #
|
42 |
+
|
43 |
+
from transformers.modeling_outputs import (
|
44 |
+
MoeCausalLMOutputWithPast,
|
45 |
+
MoeModelOutputWithPast,
|
46 |
+
)
|
47 |
+
from transformers.activations import ACT2FN
|
48 |
+
from transformers.utils import logging
|
49 |
+
from transformers.models.mixtral.modeling_mixtral import (
|
50 |
+
load_balancing_loss_func,
|
51 |
+
)
|
52 |
+
from transformers.models.qwen3_moe.modeling_qwen3_moe import (
|
53 |
+
Qwen3MoeMLP,
|
54 |
+
Qwen3MoeRMSNorm,
|
55 |
+
Qwen3MoeAttention,
|
56 |
+
Qwen3MoeDecoderLayer,
|
57 |
+
Qwen3MoeModel,
|
58 |
+
Qwen3MoeForCausalLM,
|
59 |
+
)
|
60 |
+
from .configuration_qwen3_shared_moe import Qwen3SharedMoeConfig
|
61 |
+
|
62 |
+
import scattermoe
|
63 |
+
|
64 |
+
|
65 |
+
logger = logging.get_logger(__name__)
|
66 |
+
|
67 |
+
|
68 |
+
class Qwen3SharedMoeSparseMoeBlock(nn.Module):
|
69 |
+
def __init__(self, config: Qwen3SharedMoeConfig):
|
70 |
+
super().__init__()
|
71 |
+
self.config = config
|
72 |
+
self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
|
73 |
+
if config.shared_expert_intermediate_size is not None:
|
74 |
+
self.shared_expert = Qwen3MoeMLP(
|
75 |
+
config, intermediate_size=config.shared_expert_intermediate_size
|
76 |
+
)
|
77 |
+
else:
|
78 |
+
self.shared_expert = None
|
79 |
+
self.moe_mlp = scattermoe.mlp.GLUMLP(
|
80 |
+
input_size=self.config.hidden_size,
|
81 |
+
hidden_size=self.config.moe_intermediate_size,
|
82 |
+
num_experts=self.config.num_experts,
|
83 |
+
top_k=self.config.num_experts_per_tok,
|
84 |
+
activation=ACT2FN[config.hidden_act],
|
85 |
+
)
|
86 |
+
|
87 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
88 |
+
# handling of gate/router logits copied from Qwen3MoeSparseMoeBlock
|
89 |
+
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
90 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
91 |
+
# router_logits: (batch * sequence_length, n_experts)
|
92 |
+
router_logits = self.gate(hidden_states)
|
93 |
+
|
94 |
+
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
95 |
+
routing_weights, selected_experts = torch.topk(
|
96 |
+
routing_weights, self.config.num_experts_per_tok, dim=-1
|
97 |
+
)
|
98 |
+
if self.config.norm_topk_prob: # only diff with mixtral sparse moe block!
|
99 |
+
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
100 |
+
# we cast back to the input dtype
|
101 |
+
routing_weights = routing_weights.to(hidden_states.dtype)
|
102 |
+
|
103 |
+
# modified here to use scattermoe + shared_expert
|
104 |
+
hs_0 = self.moe_mlp(hidden_states, routing_weights, selected_experts)
|
105 |
+
|
106 |
+
if self.shared_expert is not None:
|
107 |
+
shared_res = self.shared_expert(hidden_states)
|
108 |
+
res = hs_0 + shared_res
|
109 |
+
else:
|
110 |
+
res = hs_0
|
111 |
+
res = res.reshape(batch_size, sequence_length, hidden_dim)
|
112 |
+
return res, router_logits
|
113 |
+
|
114 |
+
|
115 |
+
class Qwen3SharedMoeDecoderLayer(Qwen3MoeDecoderLayer, nn.Module):
|
116 |
+
def __init__(self, config: Qwen3SharedMoeConfig, layer_idx: int):
|
117 |
+
super().__init__(config, layer_idx)
|
118 |
+
self.hidden_size = config.hidden_size
|
119 |
+
|
120 |
+
self.self_attn = Qwen3MoeAttention(config, layer_idx)
|
121 |
+
|
122 |
+
if (layer_idx not in config.mlp_only_layers) and (
|
123 |
+
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
|
124 |
+
):
|
125 |
+
self.mlp = Qwen3SharedMoeSparseMoeBlock(config)
|
126 |
+
else:
|
127 |
+
self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size)
|
128 |
+
|
129 |
+
self.input_layernorm = Qwen3MoeRMSNorm(
|
130 |
+
config.hidden_size, eps=config.rms_norm_eps
|
131 |
+
)
|
132 |
+
self.post_attention_layernorm = Qwen3MoeRMSNorm(
|
133 |
+
config.hidden_size, eps=config.rms_norm_eps
|
134 |
+
)
|
135 |
+
|
136 |
+
|
137 |
+
class Qwen3SharedMoeModel(Qwen3MoeModel):
|
138 |
+
config_class = Qwen3SharedMoeConfig
|
139 |
+
|
140 |
+
def __init__(self, config: Qwen3SharedMoeConfig):
|
141 |
+
super().__init__(config)
|
142 |
+
self.layers = nn.ModuleList(
|
143 |
+
[
|
144 |
+
Qwen3SharedMoeDecoderLayer(config, layer_idx)
|
145 |
+
for layer_idx in range(config.num_hidden_layers)
|
146 |
+
]
|
147 |
+
)
|
148 |
+
|
149 |
+
|
150 |
+
class Qwen3SharedMoeForCausalLM(Qwen3MoeForCausalLM):
|
151 |
+
config_class = Qwen3SharedMoeConfig
|
152 |
+
|
153 |
+
def __init__(self, config):
|
154 |
+
super().__init__(config)
|
155 |
+
self.model = Qwen3SharedMoeModel(config)
|
156 |
+
self.num_experts = config.num_experts
|
157 |
+
|
158 |
+
# Liger Patch #
|
159 |
+
def forward(
|
160 |
+
self,
|
161 |
+
input_ids: Optional[torch.LongTensor] = None,
|
162 |
+
attention_mask: Optional[torch.Tensor] = None,
|
163 |
+
position_ids: Optional[torch.LongTensor] = None,
|
164 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
165 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
166 |
+
labels: Optional[torch.LongTensor] = None,
|
167 |
+
use_cache: Optional[bool] = None,
|
168 |
+
output_attentions: Optional[bool] = None,
|
169 |
+
output_hidden_states: Optional[bool] = None,
|
170 |
+
output_router_logits: Optional[bool] = None,
|
171 |
+
cache_position: Optional[torch.LongTensor] = None,
|
172 |
+
logits_to_keep: Union[int, torch.Tensor] = 0,
|
173 |
+
skip_logits: Optional[bool] = None,
|
174 |
+
**kwargs,
|
175 |
+
) -> MoeCausalLMOutputWithPast:
|
176 |
+
r"""
|
177 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
178 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
179 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
180 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
181 |
+
|
182 |
+
logits_to_keep (`int` or `torch.Tensor`, *optional*):
|
183 |
+
If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
|
184 |
+
`input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
|
185 |
+
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
|
186 |
+
If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
|
187 |
+
This is useful when using packed tensor format (single dimension for batch and sequence length).
|
188 |
+
|
189 |
+
Returns:
|
190 |
+
|
191 |
+
Example:
|
192 |
+
|
193 |
+
```python
|
194 |
+
>>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
|
195 |
+
|
196 |
+
>>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
|
197 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
|
198 |
+
|
199 |
+
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
200 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
201 |
+
|
202 |
+
>>> # Generate
|
203 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
204 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
205 |
+
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
|
206 |
+
```"""
|
207 |
+
|
208 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
209 |
+
output_router_logits = (
|
210 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
211 |
+
)
|
212 |
+
|
213 |
+
output_hidden_states = (
|
214 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
215 |
+
)
|
216 |
+
|
217 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
218 |
+
outputs: MoeModelOutputWithPast = self.model(
|
219 |
+
input_ids=input_ids,
|
220 |
+
attention_mask=attention_mask,
|
221 |
+
position_ids=position_ids,
|
222 |
+
past_key_values=past_key_values,
|
223 |
+
inputs_embeds=inputs_embeds,
|
224 |
+
use_cache=use_cache,
|
225 |
+
output_attentions=output_attentions,
|
226 |
+
output_hidden_states=output_hidden_states,
|
227 |
+
output_router_logits=output_router_logits,
|
228 |
+
cache_position=cache_position,
|
229 |
+
**kwargs,
|
230 |
+
)
|
231 |
+
|
232 |
+
hidden_states = outputs.last_hidden_state
|
233 |
+
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
234 |
+
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
235 |
+
kept_hidden_states = hidden_states[:, slice_indices, :]
|
236 |
+
|
237 |
+
shift_labels = kwargs.pop("shift_labels", None)
|
238 |
+
logits = None
|
239 |
+
loss = None
|
240 |
+
|
241 |
+
if skip_logits is None:
|
242 |
+
skip_logits = self.training and (labels is not None or shift_labels is not None)
|
243 |
+
|
244 |
+
if skip_logits:
|
245 |
+
loss = LigerForCausalLMLoss(
|
246 |
+
hidden_states=kept_hidden_states,
|
247 |
+
lm_head_weight=self.lm_head.weight,
|
248 |
+
labels=labels,
|
249 |
+
shift_labels=shift_labels,
|
250 |
+
hidden_size=self.config.hidden_size,
|
251 |
+
**kwargs,
|
252 |
+
)
|
253 |
+
else: # if in inference model materialize logits
|
254 |
+
logits = self.lm_head(kept_hidden_states)
|
255 |
+
if labels is not None:
|
256 |
+
loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
|
257 |
+
|
258 |
+
aux_loss = None
|
259 |
+
if output_router_logits:
|
260 |
+
aux_loss = load_balancing_loss_func(
|
261 |
+
outputs.router_logits,
|
262 |
+
self.num_experts,
|
263 |
+
self.num_experts_per_tok,
|
264 |
+
attention_mask,
|
265 |
+
)
|
266 |
+
if labels is not None:
|
267 |
+
loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
|
268 |
+
|
269 |
+
return MoeCausalLMOutputWithPast(
|
270 |
+
loss=loss,
|
271 |
+
aux_loss=aux_loss,
|
272 |
+
logits=logits,
|
273 |
+
past_key_values=outputs.past_key_values,
|
274 |
+
hidden_states=outputs.hidden_states,
|
275 |
+
attentions=outputs.attentions,
|
276 |
+
router_logits=outputs.router_logits,
|
277 |
+
)
|
278 |
+
# Liger Patch #
|
279 |
+
|