A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 14:45 UTC
Add GLM-4-0414 support (#16338)
Signed-off-by: lvfei.lv <lvfei.lv@alibaba-inc.com> Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com> Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk> Signed-off-by: yihong0618 <zouzou0208@gmail.com> Signed-off-by: Lu Fang <fanglu@fb.com> Signed-off-by: Ajay Vohra <ajayvohr@amazon.com> Signed-off-by: NickLucche <nlucches@redhat.com> Signed-off-by: Guillaume Calmettes <gcalmettes@scaleway.com> Co-authored-by: Accelerator1996 <lvfei.lv@alibaba-inc.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk> Co-authored-by: Michael Goin <michael@neuralmagic.com> Co-authored-by: yihong <zouzou0208@gmail.com> Co-authored-by: Lucia Fang <116399278+luccafong@users.noreply.github.com> Co-authored-by: ajayvohra2005 <ajayvohr@amazon.com> Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com> Co-authored-by: Guillaume Calmettes <gcalmettes@scaleway.com>
Yuxuan Zhang committed
Apr 10, 2025 at 09:19 UTC
1e44ffc3ff5be0f7bd4c4e7efa888a80d2681743
4 files changed
+324
docs/source/models/supported_models.md
+5
@@ -303,6 +303,11 @@ See [this page](#generative-models) for more information on how to use generativ
303
* `THUDM/glm-4-9b-chat-hf`, etc.
304
* ✅︎
305
* ✅︎
306
+- * `Glm4ForCausalLM`
307
+ * GLM-4-0414
308
+ * `THUDM/GLM-4-32B-Chat-0414`, etc.
309
+ * ✅︎
310
+ * ✅︎
311
- * `GPT2LMHeadModel`
312
* GPT-2
313
* `gpt2`, `gpt2-xl`, etc.
tests/models/registry.py
+5
@@ -146,6 +146,11 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
146
"Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it",
147
min_transformers_version="4.50"),
148
"GlmForCausalLM": _HfExamplesInfo("THUDM/glm-4-9b-chat-hf"),
149
+ "Glm4ForCausalLM": _HfExamplesInfo(
150
+ "THUDM/GLM-4-32B-Chat-0414",
151
+ is_available_online=False,
152
+ min_transformers_version="4.52.dev0"
153
+ ),
154
"GPT2LMHeadModel": _HfExamplesInfo("gpt2"),
155
"GPTBigCodeForCausalLM": _HfExamplesInfo("bigcode/starcoder"),
156
"GPTJForCausalLM": _HfExamplesInfo("EleutherAI/gpt-j-6b"),
vllm/model_executor/models/glm4.py
new
+313
@@ -0,0 +1,313 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+
3
+# Copyright 2025 The Zhipu AI team.
4
+# Copyright 2023 The vLLM team.
5
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
6
+#
7
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
8
+# and OPT implementations in this library. It has been modified from its
9
+# original forms to accommodate minor architectural differences compared
10
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
11
+#
12
+# Licensed under the Apache License, Version 2.0 (the "License");
13
+# you may not use this file except in compliance with the License.
14
+# You may obtain a copy of the License at
15
+#
16
+# http://www.apache.org/licenses/LICENSE-2.0
17
+#
18
+# Unless required by applicable law or agreed to in writing, software
19
+# distributed under the License is distributed on an "AS IS" BASIS,
20
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
21
+# See the License for the specific language governing permissions and
22
+# limitations under the License.
23
+"""Inference-only GLM-4-0414 model compatible with HuggingFace weights."""
24
+from typing import Iterable, Optional, Set, Tuple, Union
25
+
26
+import torch
27
+from torch import nn
28
+from transformers import Glm4Config
29
+
30
+from vllm.attention import Attention, AttentionType
31
+from vllm.compilation.decorators import support_torch_compile
32
+from vllm.config import CacheConfig, VllmConfig
33
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
34
+from vllm.model_executor.layers.layernorm import RMSNorm
35
+from vllm.model_executor.layers.linear import (QKVParallelLinear,
36
+ RowParallelLinear)
37
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
38
+from vllm.model_executor.layers.quantization import QuantizationConfig
39
+from vllm.model_executor.layers.rotary_embedding import get_rope
40
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
41
+from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
42
+from vllm.model_executor.sampling_metadata import SamplingMetadata
43
+from vllm.sequence import IntermediateTensors
44
+
45
+from .interfaces import SupportsLoRA, SupportsPP
46
+from .llama import LlamaMLP as Glm4MLP
47
+from .llama import LlamaModel
48
+from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
49
+
50
+
51
+class Glm4Attention(nn.Module):
52
+
53
+ def __init__(self,
54
+ config: Glm4Config,
55
+ hidden_size: int,
56
+ num_heads: int,
57
+ num_kv_heads: int,
58
+ max_position: int = 4096 * 32,
59
+ head_dim: Optional[int] = None,
60
+ qkv_bias: bool = False,
61
+ rope_theta: float = 10000,
62
+ cache_config: Optional[CacheConfig] = None,
63
+ quant_config: Optional[QuantizationConfig] = None,
64
+ rope_scaling: Optional[Tuple] = None,
65
+ prefix: str = "",
66
+ attn_type: str = AttentionType.DECODER) -> None:
67
+ super().__init__()
68
+ self.hidden_size = hidden_size
69
+ tp_size = get_tensor_model_parallel_world_size()
70
+ self.total_num_heads = num_heads
71
+ assert self.total_num_heads % tp_size == 0
72
+ self.num_heads = self.total_num_heads // tp_size
73
+ self.total_num_kv_heads = num_kv_heads
74
+ if self.total_num_kv_heads >= tp_size:
75
+ # Number of KV heads is greater than TP size, so we partition
76
+ # the KV heads across multiple tensor parallel GPUs.
77
+ assert self.total_num_kv_heads % tp_size == 0
78
+ else:
79
+ # Number of KV heads is less than TP size, so we replicate
80
+ # the KV heads across multiple tensor parallel GPUs.
81
+ assert tp_size % self.total_num_kv_heads == 0
82
+ partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
83
+ self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
84
+ self.head_dim = head_dim or hidden_size // self.total_num_heads
85
+ self.rotary_dim = int(partial_rotary_factor * self.head_dim)
86
+ self.q_size = self.num_heads * self.head_dim
87
+ self.kv_size = self.num_kv_heads * self.head_dim
88
+ self.scaling = self.head_dim**-0.5
89
+ self.rope_theta = rope_theta
90
+ self.qkv_proj = QKVParallelLinear(
91
+ hidden_size,
92
+ self.head_dim,
93
+ self.total_num_heads,
94
+ self.total_num_kv_heads,
95
+ bias=qkv_bias,
96
+ quant_config=quant_config,
97
+ prefix=f"{prefix}.qkv_proj",
98
+ )
99
+ self.o_proj = RowParallelLinear(
100
+ self.total_num_heads * self.head_dim,
101
+ hidden_size,
102
+ bias=False,
103
+ quant_config=quant_config,
104
+ prefix=f"{prefix}.o_proj",
105
+ )
106
+ self.rotary_emb = get_rope(
107
+ self.head_dim,
108
+ rotary_dim=self.rotary_dim,
109
+ max_position=max_position,
110
+ base=self.rope_theta,
111
+ rope_scaling=rope_scaling,
112
+ partial_rotary_factor=partial_rotary_factor,
113
+ )
114
+ self.attn = Attention(self.num_heads,
115
+ self.head_dim,
116
+ self.scaling,
117
+ num_kv_heads=self.num_kv_heads,
118
+ cache_config=cache_config,
119
+ quant_config=quant_config,
120
+ prefix=f"{prefix}.attn",
121
+ attn_type=attn_type)
122
+
123
+ def forward(
124
+ self,
125
+ positions: torch.Tensor,
126
+ hidden_states: torch.Tensor,
127
+ ) -> torch.Tensor:
128
+ qkv, _ = self.qkv_proj(hidden_states)
129
+ q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
130
+ q, k = self.rotary_emb(positions, q, k)
131
+ attn_output = self.attn(q, k, v)
132
+ output, _ = self.o_proj(attn_output)
133
+ return output
134
+
135
+
136
+class Glm4DecoderLayer(nn.Module):
137
+
138
+ def __init__(
139
+ self,
140
+ config: Glm4Config,
141
+ cache_config: Optional[CacheConfig] = None,
142
+ quant_config: Optional[QuantizationConfig] = None,
143
+ prefix: str = "",
144
+ ) -> None:
145
+ super().__init__()
146
+ self.hidden_size = config.hidden_size
147
+ rope_theta = getattr(config, "rope_theta", 1000000)
148
+ rope_scaling = getattr(config, "rope_scaling", None)
149
+
150
+ self.self_attn = Glm4Attention(
151
+ config=config,
152
+ hidden_size=self.hidden_size,
153
+ num_heads=config.num_attention_heads,
154
+ max_position=config.max_position_embeddings,
155
+ num_kv_heads=config.num_key_value_heads,
156
+ rope_theta=rope_theta,
157
+ qkv_bias=getattr(config, 'attention_bias', False),
158
+ head_dim=getattr(config, 'head_dim', None),
159
+ cache_config=cache_config,
160
+ quant_config=quant_config,
161
+ rope_scaling=rope_scaling,
162
+ prefix=f"{prefix}.self_attn",
163
+ attn_type=AttentionType.DECODER,
164
+ )
165
+ self.mlp = Glm4MLP(
166
+ hidden_size=self.hidden_size,
167
+ intermediate_size=config.intermediate_size,
168
+ hidden_act=config.hidden_act,
169
+ quant_config=quant_config,
170
+ prefix=f"{prefix}.mlp",
171
+ )
172
+ self.input_layernorm = RMSNorm(config.hidden_size,
173
+ eps=config.rms_norm_eps)
174
+ self.post_attention_layernorm = RMSNorm(config.hidden_size,
175
+ eps=config.rms_norm_eps)
176
+ self.post_self_attn_layernorm = RMSNorm(config.hidden_size,
177
+ eps=config.rms_norm_eps)
178
+ self.post_mlp_layernorm = RMSNorm(config.hidden_size,
179
+ eps=config.rms_norm_eps)
180
+
181
+ def forward(
182
+ self,
183
+ positions: torch.Tensor,
184
+ hidden_states: torch.Tensor,
185
+ residual: Optional[torch.Tensor],
186
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
187
+ # Self Attention
188
+ if residual is None:
189
+ residual = hidden_states
190
+ hidden_states = self.input_layernorm(hidden_states)
191
+ else:
192
+ hidden_states, residual = self.input_layernorm(
193
+ hidden_states, residual)
194
+ hidden_states = self.self_attn(
195
+ positions=positions,
196
+ hidden_states=hidden_states,
197
+ )
198
+
199
+ hidden_states = self.post_self_attn_layernorm(hidden_states)
200
+ hidden_states = residual + hidden_states
201
+
202
+ # Fully Connected
203
+ hidden_states = self.post_attention_layernorm(hidden_states, residual)
204
+ hidden_states = self.mlp(hidden_states)
205
+ hidden_states = self.post_mlp_layernorm(hidden_states)
206
+ hidden_states = residual + hidden_states
207
+
208
+ return hidden_states, residual
209
+
210
+
211
+ALL_DECODER_LAYER_TYPES = {
212
+ "attention": Glm4DecoderLayer,
213
+}
214
+
215
+
216
+@support_torch_compile(
217
+ dynamic_arg_dims={
218
+ "input_ids": 0,
219
+ "positions": -1,
220
+ "intermediate_tensors": 0,
221
+ "inputs_embeds": 0,
222
+ })
223
+class Glm4Model(LlamaModel):
224
+
225
+ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
226
+ super().__init__(vllm_config=vllm_config,
227
+ prefix=prefix,
228
+ layer_type=Glm4DecoderLayer)
229
+
230
+
231
+class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
232
+ packed_modules_mapping = {
233
+ "qkv_proj": [
234
+ "q_proj",
235
+ "k_proj",
236
+ "v_proj",
237
+ ],
238
+ "gate_up_proj": [
239
+ "gate_proj",
240
+ "up_proj",
241
+ ],
242
+ }
243
+
244
+ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
245
+ super().__init__()
246
+ config = vllm_config.model_config.hf_config
247
+ quant_config = vllm_config.quant_config
248
+ lora_config = vllm_config.lora_config
249
+
250
+ self.config = config
251
+ self.lora_config = lora_config
252
+
253
+ self.quant_config = quant_config
254
+ self.model = Glm4Model(vllm_config=vllm_config,
255
+ prefix=maybe_prefix(prefix, "model"))
256
+
257
+ if get_pp_group().is_last_rank:
258
+ if config.tie_word_embeddings:
259
+ self.lm_head = self.model.embed_tokens
260
+ else:
261
+ self.lm_head = ParallelLMHead(config.vocab_size,
262
+ config.hidden_size,
263
+ quant_config=quant_config,
264
+ prefix=maybe_prefix(
265
+ prefix, "lm_head"))
266
+ else:
267
+ self.lm_head = PPMissingLayer()
268
+
269
+ self.logits_processor = LogitsProcessor(config.vocab_size)
270
+ self.sampler = get_sampler()
271
+
272
+ self.make_empty_intermediate_tensors = (
273
+ self.model.make_empty_intermediate_tensors)
274
+
275
+ def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
276
+ return self.model.get_input_embeddings(input_ids)
277
+
278
+ def forward(
279
+ self,
280
+ input_ids: torch.Tensor,
281
+ positions: torch.Tensor,
282
+ intermediate_tensors: Optional[IntermediateTensors] = None,
283
+ inputs_embeds: Optional[torch.Tensor] = None,
284
+ ) -> Union[torch.Tensor, IntermediateTensors]:
285
+ hidden_states = self.model(input_ids, positions, intermediate_tensors,
286
+ inputs_embeds)
287
+ return hidden_states
288
+
289
+ def compute_logits(
290
+ self,
291
+ hidden_states: torch.Tensor,
292
+ sampling_metadata: SamplingMetadata,
293
+ ) -> Optional[torch.Tensor]:
294
+ logits = self.logits_processor(self.lm_head, hidden_states,
295
+ sampling_metadata)
296
+ return logits
297
+
298
+ def sample(
299
+ self,
300
+ logits: torch.Tensor,
301
+ sampling_metadata: SamplingMetadata,
302
+ ) -> Optional[SamplerOutput]:
303
+ next_tokens = self.sampler(logits, sampling_metadata)
304
+ return next_tokens
305
+
306
+ def load_weights(self, weights: Iterable[Tuple[str,
307
+ torch.Tensor]]) -> Set[str]:
308
+ loader = AutoWeightsLoader(
309
+ self,
310
+ skip_prefixes=(["lm_head."]
311
+ if self.config.tie_word_embeddings else None),
312
+ )
313
+ return loader.load_weights(weights)
vllm/model_executor/models/registry.py
+1
@@ -58,6 +58,7 @@ _TEXT_GENERATION_MODELS = {
58
"Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
59
"Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
60
"GlmForCausalLM": ("glm", "GlmForCausalLM"),
61
+ "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
62
"GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
63
"GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
64
"GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),