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 16, 05:45 UTC
[Model] Support telechat2 (#10311)
Signed-off-by: Isotr0py <2037008807@qq.com> Co-authored-by: xiangw2 <xiangw2@chinatelecom.cn> Co-authored-by: Isotr0py <2037008807@qq.com>
shunxing12345 committed
Nov 27, 2024 at 19:32 UTC
1209261e937f7cc5a933da48d625d17e6ee8eea9
8 files changed
+210
-3
docs/source/models/supported_models.rst
+5
@@ -309,6 +309,11 @@ Text Generation
309
- :code:`upstage/solar-pro-preview-instruct`, etc.
310
- ✅︎
311
- ✅︎
312
+ * - :code:`TeleChat2ForCausalLM`
313
+ - TeleChat2
314
+ - :code:`TeleAI/TeleChat2-3B`, :code:`TeleAI/TeleChat2-7B`, :code:`TeleAI/TeleChat2-35B`, etc.
315
+ - ✅︎
316
+ - ✅︎
317
* - :code:`XverseForCausalLM`
318
- XVERSE
319
- :code:`xverse/XVERSE-7B-Chat`, :code:`xverse/XVERSE-13B-Chat`, :code:`xverse/XVERSE-65B-Chat`, etc.
tests/models/registry.py
+2
@@ -115,6 +115,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
115
"StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
116
"Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
117
"SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct"),
118
+ "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
119
+ trust_remote_code=True),
120
"XverseForCausalLM": _HfExamplesInfo("xverse/XVERSE-7B-Chat",
121
is_available_online=False,
122
trust_remote_code=True),
vllm/model_executor/models/llama.py
+4
-2
@@ -501,8 +501,7 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
501
self.config = config
502
self.lora_config = lora_config
503
504
- self.model = LlamaModel(vllm_config=vllm_config,
505
- prefix=maybe_prefix(prefix, "model"))
504
+ self.model = self._init_model(vllm_config=vllm_config, prefix=prefix)
505
if get_pp_group().is_last_rank:
506
self.unpadded_vocab_size = config.vocab_size
507
if lora_config:
@@ -539,6 +538,9 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
538
normalize=False,
539
softmax=False)
540
541
+ def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
542
+ return LlamaModel(vllm_config=vllm_config, prefix=prefix)
543
+
544
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
545
return self.model.get_input_embeddings(input_ids)
546
vllm/model_executor/models/registry.py
+2
@@ -91,6 +91,7 @@ _TEXT_GENERATION_MODELS = {
91
"StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
92
"Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
93
"SolarForCausalLM": ("solar", "SolarForCausalLM"),
94
+ "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
95
"XverseForCausalLM": ("xverse", "XverseForCausalLM"),
96
# [Encoder-decoder]
97
"BartModel": ("bart", "BartForConditionalGeneration"),
@@ -118,6 +119,7 @@ _EMBEDDING_MODELS = {
119
"Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
120
"Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
121
"Qwen2ForSequenceClassification": ("qwen2_cls", "Qwen2ForSequenceClassification"), # noqa: E501
122
+ "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
123
# [Multimodal]
124
"LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"), # noqa: E501
125
"Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
vllm/model_executor/models/telechat2.py
new
+131
@@ -0,0 +1,131 @@
1
+# Copyright 2023 The vLLM team.
2
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+#
4
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+# and OPT implementations in this library. It has been modified from its
6
+# original forms to accommodate minor architectural differences compared
7
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+#
9
+# Licensed under the Apache License, Version 2.0 (the "License");
10
+# you may not use this file except in compliance with the License.
11
+# You may obtain a copy of the License at
12
+#
13
+# http://www.apache.org/licenses/LICENSE-2.0
14
+#
15
+# Unless required by applicable law or agreed to in writing, software
16
+# distributed under the License is distributed on an "AS IS" BASIS,
17
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+# See the License for the specific language governing permissions and
19
+# limitations under the License.
20
+from typing import Iterable, Set, Tuple
21
+
22
+import torch
23
+
24
+from vllm.config import VllmConfig
25
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
26
+from vllm.model_executor.models.llama import LlamaForCausalLM, LlamaModel
27
+
28
+from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
29
+ is_pp_missing_parameter)
30
+
31
+
32
+class TeleChat2Model(LlamaModel):
33
+
34
+ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
35
+ # 1. Initialize the LlamaModel with bias
36
+ vllm_config.model_config.hf_config.bias = True
37
+ vllm_config.model_config.hf_config.mlp_bias = True
38
+ super().__init__(vllm_config=vllm_config, prefix=prefix)
39
+ # 2. Remove the bias from the qkv_proj and gate_up_proj based on config
40
+ # Telechat2's gate_up_proj and qkv_proj don't have bias
41
+ # see: https://github.com/vllm-project/vllm/pull/10311#issuecomment-2490297566
42
+ for layer in self.layers:
43
+ if not isinstance(layer, PPMissingLayer):
44
+ layer.self_attn.qkv_proj.bias = None
45
+ layer.self_attn.qkv_proj.skip_bias_add = True
46
+ layer.mlp.gate_up_proj.bias = None
47
+ layer.mlp.gate_up_proj.skip_bias_add = True
48
+
49
+ def load_weights(self, weights: Iterable[Tuple[str,
50
+ torch.Tensor]]) -> Set[str]:
51
+ stacked_params_mapping = [
52
+ ('gate_up_proj', 'gate_proj', 0),
53
+ ('gate_up_proj', 'up_proj', 1),
54
+ ]
55
+ params_dict = dict(self.named_parameters())
56
+ loaded_params: Set[str] = set()
57
+ total_num_heads = self.config.n_head
58
+ head_dim = self.config.hidden_size // total_num_heads
59
+ for name, loaded_weight in weights:
60
+ if "self_attn.key_value" in name:
61
+ k_weight = []
62
+ v_weight = []
63
+ for i in range(total_num_heads):
64
+ start = i * head_dim * 2
65
+ k_weight.append(loaded_weight[start:start + head_dim, :])
66
+ v_weight.append(loaded_weight[start + head_dim:start +
67
+ 2 * head_dim:])
68
+ k_weight = torch.cat(k_weight, dim=0)
69
+ v_weight = torch.cat(v_weight, dim=0)
70
+ name = name.replace("key_value", "qkv_proj")
71
+ if is_pp_missing_parameter(name, self):
72
+ continue
73
+ param = params_dict[name]
74
+ weight_loader = param.weight_loader
75
+ weight_loader(param, k_weight, "k")
76
+ weight_loader(param, v_weight, "v")
77
+ elif "query" in name:
78
+ name = name.replace("query", "qkv_proj")
79
+ if is_pp_missing_parameter(name, self):
80
+ continue
81
+ param = params_dict[name]
82
+ weight_loader = param.weight_loader
83
+ weight_loader(param, loaded_weight, "q")
84
+ else:
85
+ for param_name, weight_name, shard_id in stacked_params_mapping:
86
+ if weight_name not in name:
87
+ continue
88
+ name = name.replace(weight_name, param_name)
89
+ if is_pp_missing_parameter(name, self):
90
+ continue
91
+ param = params_dict[name]
92
+ weight_loader = param.weight_loader
93
+ weight_loader(param, loaded_weight, shard_id)
94
+ break
95
+ else:
96
+ if is_pp_missing_parameter(name, self):
97
+ continue
98
+ param = params_dict[name]
99
+ weight_loader = getattr(param, "weight_loader",
100
+ default_weight_loader)
101
+ weight_loader(param, loaded_weight)
102
+ loaded_params.add(name)
103
+ return loaded_params
104
+
105
+
106
+class TeleChat2ForCausalLM(LlamaForCausalLM):
107
+
108
+ def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
109
+ return TeleChat2Model(vllm_config=vllm_config, prefix=prefix)
110
+
111
+ def load_weights(self, weights: Iterable[Tuple[str,
112
+ torch.Tensor]]) -> Set[str]:
113
+
114
+ hf_to_vllm_mapper = WeightsMapper(
115
+ orig_to_new_prefix={
116
+ "transformer.": "model.",
117
+ },
118
+ orig_to_new_substr={
119
+ ".h.": ".layers.",
120
+ ".self_attention.": ".self_attn.",
121
+ ".word_embeddings.": ".embed_tokens.",
122
+ ".dense.": ".o_proj.",
123
+ ".ln_f.": ".norm.",
124
+ },
125
+ )
126
+ loader = AutoWeightsLoader(
127
+ self,
128
+ skip_prefixes=(["lm_head."]
129
+ if self.config.tie_word_embeddings else None),
130
+ )
131
+ return loader.load_weights(weights, mapper=hf_to_vllm_mapper)
vllm/transformers_utils/config.py
+3
-1
@@ -29,7 +29,8 @@ from vllm.transformers_utils.configs import (ChatGLMConfig, DbrxConfig,
29
MLPSpeculatorConfig, MPTConfig,
30
NemotronConfig, NVLM_D_Config,
31
Olmo2Config, RWConfig,
32
- SolarConfig, UltravoxConfig)
32
+ SolarConfig, Telechat2Config,
33
+ UltravoxConfig)
34
# yapf: enable
35
from vllm.transformers_utils.utils import check_gguf_file
36
from vllm.utils import resolve_obj_by_qualname
@@ -64,6 +65,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
65
"NVLM_D": NVLM_D_Config,
66
"olmo2": Olmo2Config,
67
"solar": SolarConfig,
68
+ "telechat": Telechat2Config,
69
"ultravox": UltravoxConfig,
70
**_CONFIG_REGISTRY_OVERRIDE_HF
71
}
vllm/transformers_utils/configs/__init__.py
+2
@@ -17,6 +17,7 @@ from vllm.transformers_utils.configs.nemotron import NemotronConfig
17
from vllm.transformers_utils.configs.nvlm_d import NVLM_D_Config
18
from vllm.transformers_utils.configs.olmo2 import Olmo2Config
19
from vllm.transformers_utils.configs.solar import SolarConfig
20
+from vllm.transformers_utils.configs.telechat2 import Telechat2Config
21
from vllm.transformers_utils.configs.ultravox import UltravoxConfig
22
23
__all__ = [
@@ -36,5 +37,6 @@ __all__ = [
37
"NVLM_D_Config",
38
"Olmo2Config",
39
"SolarConfig",
40
+ "Telechat2Config",
41
"UltravoxConfig",
42
]
\ No newline at end of file
vllm/transformers_utils/configs/telechat2.py
new
+61
@@ -0,0 +1,61 @@
1
+# adapted from https://www.modelscope.cn/models/TeleAI/TeleChat2-3B/resolve/master/configuration_telechat2.py
2
+""" Telechat configuration compatible with LlamaConfig. """
3
+
4
+from transformers.configuration_utils import PretrainedConfig
5
+
6
+
7
+class Telechat2Config(PretrainedConfig):
8
+
9
+ model_type = "telechat"
10
+ keys_to_ignore_at_inference = ["past_key_values"]
11
+ attribute_map = {
12
+ "num_hidden_layers": "n_layer",
13
+ "num_attention_heads": "n_head",
14
+ "intermediate_size": "ffn_hidden_size",
15
+ "rms_norm_eps": "layer_norm_epsilon"
16
+ }
17
+
18
+ def __init__(
19
+ self,
20
+ vocab_size=160256,
21
+ hidden_size=4096,
22
+ n_layer=30,
23
+ n_head=32,
24
+ layer_norm_epsilon=1e-5,
25
+ initializer_range=0.02,
26
+ use_cache=True,
27
+ bos_token_id=1,
28
+ eos_token_id=2,
29
+ apply_residual_connection_post_layernorm=False,
30
+ hidden_dropout=0.0,
31
+ attention_dropout=0.0,
32
+ ffn_hidden_size=12288,
33
+ training_seqlen=8192,
34
+ logn=True,
35
+ embed_layernorm=False,
36
+ hidden_act="silu",
37
+ **kwargs,
38
+ ):
39
+ self.vocab_size = vocab_size
40
+ n_embed = kwargs.pop("n_embed", None)
41
+ self.hidden_size = hidden_size if n_embed is None else n_embed
42
+ self.n_layer = n_layer
43
+ self.n_head = n_head
44
+ self.layer_norm_epsilon = layer_norm_epsilon
45
+ self.initializer_range = initializer_range
46
+ self.use_cache = use_cache
47
+ self.apply_residual_connection_post_layernorm = (
48
+ apply_residual_connection_post_layernorm)
49
+ self.hidden_dropout = hidden_dropout
50
+ self.attention_dropout = attention_dropout
51
+ self.bos_token_id = bos_token_id
52
+ self.eos_token_id = eos_token_id
53
+ self.logn = logn
54
+ self.training_seqlen = training_seqlen
55
+ self.embed_layernorm = embed_layernorm
56
+ self.num_key_value_heads = kwargs.pop("num_key_value_heads", None)
57
+ self.ffn_hidden_size = ffn_hidden_size
58
+ self.hidden_act = hidden_act
59
+ super().__init__(bos_token_id=bos_token_id,
60
+ eos_token_id=eos_token_id,
61
+ **kwargs)