add files
Browse files- added_tokens.json +7 -0
- config.json +36 -0
- configuration_scidfm.py +163 -0
- generation_config.json +6 -0
- model.safetensors.index.json +816 -0
- modeling_scidfm.py +1591 -0
- special_tokens_map.json +36 -0
- tokenizer.model +3 -0
- tokenizer_config.json +88 -0
added_tokens.json
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"<|assistant|>": 32195,
|
| 3 |
+
"<|eod|>": 32192,
|
| 4 |
+
"<|system|>": 32193,
|
| 5 |
+
"<|user|>": 32194,
|
| 6 |
+
"[PAD]": 32191
|
| 7 |
+
}
|
config.json
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "OpenDFM/SciDFM-MoE-A5.6B-v1.0",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"SciDFMForCausalLM"
|
| 5 |
+
],
|
| 6 |
+
"auto_map": {
|
| 7 |
+
"AutoConfig": "configuration_scidfm.SciDFMConfig",
|
| 8 |
+
"AutoModel": "modeling_scidfm.SciDFMModel",
|
| 9 |
+
"AutoModelForCausalLM": "modeling_scidfm.SciDFMForCausalLM"
|
| 10 |
+
},
|
| 11 |
+
"attention_dropout": 0.0,
|
| 12 |
+
"bos_token_id": 1,
|
| 13 |
+
"eos_token_id": 2,
|
| 14 |
+
"hidden_act": "silu",
|
| 15 |
+
"hidden_size": 3200,
|
| 16 |
+
"initializer_range": 0.02,
|
| 17 |
+
"intermediate_size": 8640,
|
| 18 |
+
"max_position_embeddings": 8192,
|
| 19 |
+
"model_type": "SciDFM",
|
| 20 |
+
"num_attention_heads": 32,
|
| 21 |
+
"num_experts_per_tok": 2,
|
| 22 |
+
"num_hidden_layers": 26,
|
| 23 |
+
"num_key_value_heads": 32,
|
| 24 |
+
"num_local_experts": 8,
|
| 25 |
+
"output_router_logits": false,
|
| 26 |
+
"rms_norm_eps": 1e-06,
|
| 27 |
+
"rope_theta": 10000,
|
| 28 |
+
"router_aux_loss_coef": 0.02,
|
| 29 |
+
"router_jitter_noise": 0.0,
|
| 30 |
+
"sliding_window": null,
|
| 31 |
+
"tie_word_embeddings": false,
|
| 32 |
+
"torch_dtype": "bfloat16",
|
| 33 |
+
"transformers_version": "4.40.1",
|
| 34 |
+
"use_cache": true,
|
| 35 |
+
"vocab_size": 32196
|
| 36 |
+
}
|
configuration_scidfm.py
ADDED
|
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 SciDFM 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 |
+
""" SciDFM model configuration"""
|
| 16 |
+
|
| 17 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 18 |
+
from transformers.utils import logging
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
logger = logging.get_logger(__name__)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
class SciDFMConfig(PretrainedConfig):
|
| 25 |
+
r"""
|
| 26 |
+
This is the configuration class to store the configuration of a [`SciDFMModel`]. It is used to instantiate an
|
| 27 |
+
SciDFM model according to the specified arguments, defining the model architecture.
|
| 28 |
+
|
| 29 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 30 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 31 |
+
|
| 32 |
+
Args:
|
| 33 |
+
vocab_size (`int`, *optional*, defaults to 32191):
|
| 34 |
+
Vocabulary size of the SciDFM model. Defines the number of different tokens that can be represented by the
|
| 35 |
+
`inputs_ids` passed when calling [`SciDFMModel`]
|
| 36 |
+
hidden_size (`int`, *optional*, defaults to 3200):
|
| 37 |
+
Dimension of the hidden representations.
|
| 38 |
+
intermediate_size (`int`, *optional*, defaults to 8640):
|
| 39 |
+
Dimension of the MLP representations.
|
| 40 |
+
num_hidden_layers (`int`, *optional*, defaults to 26):
|
| 41 |
+
Number of hidden layers in the Transformer encoder.
|
| 42 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 43 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 44 |
+
num_key_value_heads (`int`, *optional*, defaults to 32):
|
| 45 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
| 46 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
| 47 |
+
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
| 48 |
+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
| 49 |
+
by meanpooling all the original heads within that group. For more details checkout [this
|
| 50 |
+
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
|
| 51 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
| 52 |
+
The non-linear activation function (function or string) in the decoder.
|
| 53 |
+
max_position_embeddings (`int`, *optional*, defaults to 8192):
|
| 54 |
+
The maximum sequence length that this model might ever be used with. SciDFM allows sequence of up to 8192 tokens.
|
| 55 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 56 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 57 |
+
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
| 58 |
+
The epsilon used by the rms normalization layers.
|
| 59 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 60 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 61 |
+
relevant if `config.is_decoder=True`.
|
| 62 |
+
pad_token_id (`int`, *optional*):
|
| 63 |
+
The id of the padding token.
|
| 64 |
+
bos_token_id (`int`, *optional*, defaults to 1):
|
| 65 |
+
The id of the "beginning-of-sequence" token.
|
| 66 |
+
eos_token_id (`int`, *optional*, defaults to 2):
|
| 67 |
+
The id of the "end-of-sequence" token.
|
| 68 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 69 |
+
Whether the model's input and output word embeddings should be tied.
|
| 70 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 71 |
+
The base period of the RoPE embeddings.
|
| 72 |
+
sliding_window (`int`, *optional*):
|
| 73 |
+
Sliding window attention window size. If not specified, will default to False.
|
| 74 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 75 |
+
The dropout ratio for the attention probabilities.
|
| 76 |
+
num_experts_per_tok (`int`, *optional*, defaults to 2):
|
| 77 |
+
The number of experts to root per-token, can be also interpreted as the `top-p` routing
|
| 78 |
+
parameter
|
| 79 |
+
num_local_experts (`int`, *optional*, defaults to 8):
|
| 80 |
+
Number of experts per Sparse MLP layer.
|
| 81 |
+
output_router_logits (`bool`, *optional*, defaults to `False`):
|
| 82 |
+
Whether or not the router logits should be returned by the model. Enabeling this will also
|
| 83 |
+
allow the model to output the auxiliary loss. See [here]() for more details
|
| 84 |
+
router_aux_loss_coef (`float`, *optional*, defaults to 0.02):
|
| 85 |
+
The aux loss factor for the total loss.
|
| 86 |
+
router_jitter_noise (`float`, *optional*, defaults to 0.0):
|
| 87 |
+
Amount of noise to add to the router.
|
| 88 |
+
|
| 89 |
+
```python
|
| 90 |
+
>>> from transformers import SciDFMModel, SciDFMConfig
|
| 91 |
+
|
| 92 |
+
>>> # Initializing a SciDFM configuration
|
| 93 |
+
>>> configuration = SciDFMConfig()
|
| 94 |
+
|
| 95 |
+
>>> # Initializing a model from the SciDFM configuration
|
| 96 |
+
>>> model = SciDFMModel(configuration)
|
| 97 |
+
|
| 98 |
+
>>> # Accessing the model configuration
|
| 99 |
+
>>> configuration = model.config
|
| 100 |
+
```"""
|
| 101 |
+
|
| 102 |
+
model_type = "SciDFM"
|
| 103 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 104 |
+
|
| 105 |
+
def __init__(
|
| 106 |
+
self,
|
| 107 |
+
vocab_size=32191,
|
| 108 |
+
hidden_size=3200,
|
| 109 |
+
intermediate_size=8640,
|
| 110 |
+
num_hidden_layers=26,
|
| 111 |
+
num_attention_heads=32,
|
| 112 |
+
num_key_value_heads=32,
|
| 113 |
+
hidden_act="silu",
|
| 114 |
+
max_position_embeddings=8192,
|
| 115 |
+
initializer_range=0.02,
|
| 116 |
+
rms_norm_eps=1e-6,
|
| 117 |
+
use_cache=True,
|
| 118 |
+
pad_token_id=None,
|
| 119 |
+
bos_token_id=1,
|
| 120 |
+
eos_token_id=2,
|
| 121 |
+
tie_word_embeddings=False,
|
| 122 |
+
rope_theta=10000.0,
|
| 123 |
+
sliding_window=None,
|
| 124 |
+
attention_dropout=0.0,
|
| 125 |
+
num_experts_per_tok=2,
|
| 126 |
+
num_local_experts=8,
|
| 127 |
+
output_router_logits=False,
|
| 128 |
+
router_aux_loss_coef=0.02,
|
| 129 |
+
router_jitter_noise=0.0,
|
| 130 |
+
**kwargs,
|
| 131 |
+
):
|
| 132 |
+
self.vocab_size = vocab_size
|
| 133 |
+
self.max_position_embeddings = max_position_embeddings
|
| 134 |
+
self.hidden_size = hidden_size
|
| 135 |
+
self.intermediate_size = intermediate_size
|
| 136 |
+
self.num_hidden_layers = num_hidden_layers
|
| 137 |
+
self.num_attention_heads = num_attention_heads
|
| 138 |
+
self.sliding_window = sliding_window
|
| 139 |
+
|
| 140 |
+
# for backward compatibility
|
| 141 |
+
if num_key_value_heads is None:
|
| 142 |
+
num_key_value_heads = num_attention_heads
|
| 143 |
+
|
| 144 |
+
self.num_key_value_heads = num_key_value_heads
|
| 145 |
+
self.hidden_act = hidden_act
|
| 146 |
+
self.initializer_range = initializer_range
|
| 147 |
+
self.rms_norm_eps = rms_norm_eps
|
| 148 |
+
self.use_cache = use_cache
|
| 149 |
+
self.rope_theta = rope_theta
|
| 150 |
+
self.attention_dropout = attention_dropout
|
| 151 |
+
|
| 152 |
+
self.num_experts_per_tok = num_experts_per_tok
|
| 153 |
+
self.num_local_experts = num_local_experts
|
| 154 |
+
self.output_router_logits = output_router_logits
|
| 155 |
+
self.router_aux_loss_coef = router_aux_loss_coef
|
| 156 |
+
self.router_jitter_noise = router_jitter_noise
|
| 157 |
+
super().__init__(
|
| 158 |
+
pad_token_id=pad_token_id,
|
| 159 |
+
bos_token_id=bos_token_id,
|
| 160 |
+
eos_token_id=eos_token_id,
|
| 161 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 162 |
+
**kwargs,
|
| 163 |
+
)
|
generation_config.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 1,
|
| 4 |
+
"eos_token_id": 2,
|
| 5 |
+
"transformers_version": "4.40.1"
|
| 6 |
+
}
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,816 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 37048403200
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"lm_head.weight": "model-00008-of-00008.safetensors",
|
| 7 |
+
"model.embed_tokens.weight": "model-00001-of-00008.safetensors",
|
| 8 |
+
"model.layers.0.block_sparse_moe.experts.0.w1.weight": "model-00001-of-00008.safetensors",
|
| 9 |
+
"model.layers.0.block_sparse_moe.experts.0.w2.weight": "model-00001-of-00008.safetensors",
|
| 10 |
+
"model.layers.0.block_sparse_moe.experts.0.w3.weight": "model-00001-of-00008.safetensors",
|
| 11 |
+
"model.layers.0.block_sparse_moe.experts.1.w1.weight": "model-00001-of-00008.safetensors",
|
| 12 |
+
"model.layers.0.block_sparse_moe.experts.1.w2.weight": "model-00001-of-00008.safetensors",
|
| 13 |
+
"model.layers.0.block_sparse_moe.experts.1.w3.weight": "model-00001-of-00008.safetensors",
|
| 14 |
+
"model.layers.0.block_sparse_moe.experts.2.w1.weight": "model-00001-of-00008.safetensors",
|
| 15 |
+
"model.layers.0.block_sparse_moe.experts.2.w2.weight": "model-00001-of-00008.safetensors",
|
| 16 |
+
"model.layers.0.block_sparse_moe.experts.2.w3.weight": "model-00001-of-00008.safetensors",
|
| 17 |
+
"model.layers.0.block_sparse_moe.experts.3.w1.weight": "model-00001-of-00008.safetensors",
|
| 18 |
+
"model.layers.0.block_sparse_moe.experts.3.w2.weight": "model-00001-of-00008.safetensors",
|
| 19 |
+
"model.layers.0.block_sparse_moe.experts.3.w3.weight": "model-00001-of-00008.safetensors",
|
| 20 |
+
"model.layers.0.block_sparse_moe.experts.4.w1.weight": "model-00001-of-00008.safetensors",
|
| 21 |
+
"model.layers.0.block_sparse_moe.experts.4.w2.weight": "model-00001-of-00008.safetensors",
|
| 22 |
+
"model.layers.0.block_sparse_moe.experts.4.w3.weight": "model-00001-of-00008.safetensors",
|
| 23 |
+
"model.layers.0.block_sparse_moe.experts.5.w1.weight": "model-00001-of-00008.safetensors",
|
| 24 |
+
"model.layers.0.block_sparse_moe.experts.5.w2.weight": "model-00001-of-00008.safetensors",
|
| 25 |
+
"model.layers.0.block_sparse_moe.experts.5.w3.weight": "model-00001-of-00008.safetensors",
|
| 26 |
+
"model.layers.0.block_sparse_moe.experts.6.w1.weight": "model-00001-of-00008.safetensors",
|
| 27 |
+
"model.layers.0.block_sparse_moe.experts.6.w2.weight": "model-00001-of-00008.safetensors",
|
| 28 |
+
"model.layers.0.block_sparse_moe.experts.6.w3.weight": "model-00001-of-00008.safetensors",
|
| 29 |
+
"model.layers.0.block_sparse_moe.experts.7.w1.weight": "model-00001-of-00008.safetensors",
|
| 30 |
+
"model.layers.0.block_sparse_moe.experts.7.w2.weight": "model-00001-of-00008.safetensors",
|
| 31 |
+
"model.layers.0.block_sparse_moe.experts.7.w3.weight": "model-00001-of-00008.safetensors",
|
| 32 |
+
"model.layers.0.block_sparse_moe.gate.weight": "model-00001-of-00008.safetensors",
|
| 33 |
+
"model.layers.0.input_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 34 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 35 |
+
"model.layers.0.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
|
| 36 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
|
| 37 |
+
"model.layers.0.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
|
| 38 |
+
"model.layers.0.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
|
| 39 |
+
"model.layers.1.block_sparse_moe.experts.0.w1.weight": "model-00001-of-00008.safetensors",
|
| 40 |
+
"model.layers.1.block_sparse_moe.experts.0.w2.weight": "model-00001-of-00008.safetensors",
|
| 41 |
+
"model.layers.1.block_sparse_moe.experts.0.w3.weight": "model-00001-of-00008.safetensors",
|
| 42 |
+
"model.layers.1.block_sparse_moe.experts.1.w1.weight": "model-00001-of-00008.safetensors",
|
| 43 |
+
"model.layers.1.block_sparse_moe.experts.1.w2.weight": "model-00001-of-00008.safetensors",
|
| 44 |
+
"model.layers.1.block_sparse_moe.experts.1.w3.weight": "model-00001-of-00008.safetensors",
|
| 45 |
+
"model.layers.1.block_sparse_moe.experts.2.w1.weight": "model-00001-of-00008.safetensors",
|
| 46 |
+
"model.layers.1.block_sparse_moe.experts.2.w2.weight": "model-00001-of-00008.safetensors",
|
| 47 |
+
"model.layers.1.block_sparse_moe.experts.2.w3.weight": "model-00001-of-00008.safetensors",
|
| 48 |
+
"model.layers.1.block_sparse_moe.experts.3.w1.weight": "model-00001-of-00008.safetensors",
|
| 49 |
+
"model.layers.1.block_sparse_moe.experts.3.w2.weight": "model-00001-of-00008.safetensors",
|
| 50 |
+
"model.layers.1.block_sparse_moe.experts.3.w3.weight": "model-00001-of-00008.safetensors",
|
| 51 |
+
"model.layers.1.block_sparse_moe.experts.4.w1.weight": "model-00001-of-00008.safetensors",
|
| 52 |
+
"model.layers.1.block_sparse_moe.experts.4.w2.weight": "model-00001-of-00008.safetensors",
|
| 53 |
+
"model.layers.1.block_sparse_moe.experts.4.w3.weight": "model-00001-of-00008.safetensors",
|
| 54 |
+
"model.layers.1.block_sparse_moe.experts.5.w1.weight": "model-00001-of-00008.safetensors",
|
| 55 |
+
"model.layers.1.block_sparse_moe.experts.5.w2.weight": "model-00001-of-00008.safetensors",
|
| 56 |
+
"model.layers.1.block_sparse_moe.experts.5.w3.weight": "model-00001-of-00008.safetensors",
|
| 57 |
+
"model.layers.1.block_sparse_moe.experts.6.w1.weight": "model-00001-of-00008.safetensors",
|
| 58 |
+
"model.layers.1.block_sparse_moe.experts.6.w2.weight": "model-00001-of-00008.safetensors",
|
| 59 |
+
"model.layers.1.block_sparse_moe.experts.6.w3.weight": "model-00001-of-00008.safetensors",
|
| 60 |
+
"model.layers.1.block_sparse_moe.experts.7.w1.weight": "model-00001-of-00008.safetensors",
|
| 61 |
+
"model.layers.1.block_sparse_moe.experts.7.w2.weight": "model-00001-of-00008.safetensors",
|
| 62 |
+
"model.layers.1.block_sparse_moe.experts.7.w3.weight": "model-00001-of-00008.safetensors",
|
| 63 |
+
"model.layers.1.block_sparse_moe.gate.weight": "model-00001-of-00008.safetensors",
|
| 64 |
+
"model.layers.1.input_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 65 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 66 |
+
"model.layers.1.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
|
| 67 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
|
| 68 |
+
"model.layers.1.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
|
| 69 |
+
"model.layers.1.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
|
| 70 |
+
"model.layers.10.block_sparse_moe.experts.0.w1.weight": "model-00003-of-00008.safetensors",
|
| 71 |
+
"model.layers.10.block_sparse_moe.experts.0.w2.weight": "model-00003-of-00008.safetensors",
|
| 72 |
+
"model.layers.10.block_sparse_moe.experts.0.w3.weight": "model-00003-of-00008.safetensors",
|
| 73 |
+
"model.layers.10.block_sparse_moe.experts.1.w1.weight": "model-00003-of-00008.safetensors",
|
| 74 |
+
"model.layers.10.block_sparse_moe.experts.1.w2.weight": "model-00003-of-00008.safetensors",
|
| 75 |
+
"model.layers.10.block_sparse_moe.experts.1.w3.weight": "model-00003-of-00008.safetensors",
|
| 76 |
+
"model.layers.10.block_sparse_moe.experts.2.w1.weight": "model-00003-of-00008.safetensors",
|
| 77 |
+
"model.layers.10.block_sparse_moe.experts.2.w2.weight": "model-00003-of-00008.safetensors",
|
| 78 |
+
"model.layers.10.block_sparse_moe.experts.2.w3.weight": "model-00003-of-00008.safetensors",
|
| 79 |
+
"model.layers.10.block_sparse_moe.experts.3.w1.weight": "model-00004-of-00008.safetensors",
|
| 80 |
+
"model.layers.10.block_sparse_moe.experts.3.w2.weight": "model-00004-of-00008.safetensors",
|
| 81 |
+
"model.layers.10.block_sparse_moe.experts.3.w3.weight": "model-00004-of-00008.safetensors",
|
| 82 |
+
"model.layers.10.block_sparse_moe.experts.4.w1.weight": "model-00004-of-00008.safetensors",
|
| 83 |
+
"model.layers.10.block_sparse_moe.experts.4.w2.weight": "model-00004-of-00008.safetensors",
|
| 84 |
+
"model.layers.10.block_sparse_moe.experts.4.w3.weight": "model-00004-of-00008.safetensors",
|
| 85 |
+
"model.layers.10.block_sparse_moe.experts.5.w1.weight": "model-00004-of-00008.safetensors",
|
| 86 |
+
"model.layers.10.block_sparse_moe.experts.5.w2.weight": "model-00004-of-00008.safetensors",
|
| 87 |
+
"model.layers.10.block_sparse_moe.experts.5.w3.weight": "model-00004-of-00008.safetensors",
|
| 88 |
+
"model.layers.10.block_sparse_moe.experts.6.w1.weight": "model-00004-of-00008.safetensors",
|
| 89 |
+
"model.layers.10.block_sparse_moe.experts.6.w2.weight": "model-00004-of-00008.safetensors",
|
| 90 |
+
"model.layers.10.block_sparse_moe.experts.6.w3.weight": "model-00004-of-00008.safetensors",
|
| 91 |
+
"model.layers.10.block_sparse_moe.experts.7.w1.weight": "model-00004-of-00008.safetensors",
|
| 92 |
+
"model.layers.10.block_sparse_moe.experts.7.w2.weight": "model-00004-of-00008.safetensors",
|
| 93 |
+
"model.layers.10.block_sparse_moe.experts.7.w3.weight": "model-00004-of-00008.safetensors",
|
| 94 |
+
"model.layers.10.block_sparse_moe.gate.weight": "model-00003-of-00008.safetensors",
|
| 95 |
+
"model.layers.10.input_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 96 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 97 |
+
"model.layers.10.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
|
| 98 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
|
| 99 |
+
"model.layers.10.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
|
| 100 |
+
"model.layers.10.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
|
| 101 |
+
"model.layers.11.block_sparse_moe.experts.0.w1.weight": "model-00004-of-00008.safetensors",
|
| 102 |
+
"model.layers.11.block_sparse_moe.experts.0.w2.weight": "model-00004-of-00008.safetensors",
|
| 103 |
+
"model.layers.11.block_sparse_moe.experts.0.w3.weight": "model-00004-of-00008.safetensors",
|
| 104 |
+
"model.layers.11.block_sparse_moe.experts.1.w1.weight": "model-00004-of-00008.safetensors",
|
| 105 |
+
"model.layers.11.block_sparse_moe.experts.1.w2.weight": "model-00004-of-00008.safetensors",
|
| 106 |
+
"model.layers.11.block_sparse_moe.experts.1.w3.weight": "model-00004-of-00008.safetensors",
|
| 107 |
+
"model.layers.11.block_sparse_moe.experts.2.w1.weight": "model-00004-of-00008.safetensors",
|
| 108 |
+
"model.layers.11.block_sparse_moe.experts.2.w2.weight": "model-00004-of-00008.safetensors",
|
| 109 |
+
"model.layers.11.block_sparse_moe.experts.2.w3.weight": "model-00004-of-00008.safetensors",
|
| 110 |
+
"model.layers.11.block_sparse_moe.experts.3.w1.weight": "model-00004-of-00008.safetensors",
|
| 111 |
+
"model.layers.11.block_sparse_moe.experts.3.w2.weight": "model-00004-of-00008.safetensors",
|
| 112 |
+
"model.layers.11.block_sparse_moe.experts.3.w3.weight": "model-00004-of-00008.safetensors",
|
| 113 |
+
"model.layers.11.block_sparse_moe.experts.4.w1.weight": "model-00004-of-00008.safetensors",
|
| 114 |
+
"model.layers.11.block_sparse_moe.experts.4.w2.weight": "model-00004-of-00008.safetensors",
|
| 115 |
+
"model.layers.11.block_sparse_moe.experts.4.w3.weight": "model-00004-of-00008.safetensors",
|
| 116 |
+
"model.layers.11.block_sparse_moe.experts.5.w1.weight": "model-00004-of-00008.safetensors",
|
| 117 |
+
"model.layers.11.block_sparse_moe.experts.5.w2.weight": "model-00004-of-00008.safetensors",
|
| 118 |
+
"model.layers.11.block_sparse_moe.experts.5.w3.weight": "model-00004-of-00008.safetensors",
|
| 119 |
+
"model.layers.11.block_sparse_moe.experts.6.w1.weight": "model-00004-of-00008.safetensors",
|
| 120 |
+
"model.layers.11.block_sparse_moe.experts.6.w2.weight": "model-00004-of-00008.safetensors",
|
| 121 |
+
"model.layers.11.block_sparse_moe.experts.6.w3.weight": "model-00004-of-00008.safetensors",
|
| 122 |
+
"model.layers.11.block_sparse_moe.experts.7.w1.weight": "model-00004-of-00008.safetensors",
|
| 123 |
+
"model.layers.11.block_sparse_moe.experts.7.w2.weight": "model-00004-of-00008.safetensors",
|
| 124 |
+
"model.layers.11.block_sparse_moe.experts.7.w3.weight": "model-00004-of-00008.safetensors",
|
| 125 |
+
"model.layers.11.block_sparse_moe.gate.weight": "model-00004-of-00008.safetensors",
|
| 126 |
+
"model.layers.11.input_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 127 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 128 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
|
| 129 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
|
| 130 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
|
| 131 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
|
| 132 |
+
"model.layers.12.block_sparse_moe.experts.0.w1.weight": "model-00004-of-00008.safetensors",
|
| 133 |
+
"model.layers.12.block_sparse_moe.experts.0.w2.weight": "model-00004-of-00008.safetensors",
|
| 134 |
+
"model.layers.12.block_sparse_moe.experts.0.w3.weight": "model-00004-of-00008.safetensors",
|
| 135 |
+
"model.layers.12.block_sparse_moe.experts.1.w1.weight": "model-00004-of-00008.safetensors",
|
| 136 |
+
"model.layers.12.block_sparse_moe.experts.1.w2.weight": "model-00004-of-00008.safetensors",
|
| 137 |
+
"model.layers.12.block_sparse_moe.experts.1.w3.weight": "model-00004-of-00008.safetensors",
|
| 138 |
+
"model.layers.12.block_sparse_moe.experts.2.w1.weight": "model-00004-of-00008.safetensors",
|
| 139 |
+
"model.layers.12.block_sparse_moe.experts.2.w2.weight": "model-00004-of-00008.safetensors",
|
| 140 |
+
"model.layers.12.block_sparse_moe.experts.2.w3.weight": "model-00004-of-00008.safetensors",
|
| 141 |
+
"model.layers.12.block_sparse_moe.experts.3.w1.weight": "model-00004-of-00008.safetensors",
|
| 142 |
+
"model.layers.12.block_sparse_moe.experts.3.w2.weight": "model-00004-of-00008.safetensors",
|
| 143 |
+
"model.layers.12.block_sparse_moe.experts.3.w3.weight": "model-00004-of-00008.safetensors",
|
| 144 |
+
"model.layers.12.block_sparse_moe.experts.4.w1.weight": "model-00004-of-00008.safetensors",
|
| 145 |
+
"model.layers.12.block_sparse_moe.experts.4.w2.weight": "model-00004-of-00008.safetensors",
|
| 146 |
+
"model.layers.12.block_sparse_moe.experts.4.w3.weight": "model-00004-of-00008.safetensors",
|
| 147 |
+
"model.layers.12.block_sparse_moe.experts.5.w1.weight": "model-00004-of-00008.safetensors",
|
| 148 |
+
"model.layers.12.block_sparse_moe.experts.5.w2.weight": "model-00004-of-00008.safetensors",
|
| 149 |
+
"model.layers.12.block_sparse_moe.experts.5.w3.weight": "model-00004-of-00008.safetensors",
|
| 150 |
+
"model.layers.12.block_sparse_moe.experts.6.w1.weight": "model-00004-of-00008.safetensors",
|
| 151 |
+
"model.layers.12.block_sparse_moe.experts.6.w2.weight": "model-00004-of-00008.safetensors",
|
| 152 |
+
"model.layers.12.block_sparse_moe.experts.6.w3.weight": "model-00004-of-00008.safetensors",
|
| 153 |
+
"model.layers.12.block_sparse_moe.experts.7.w1.weight": "model-00004-of-00008.safetensors",
|
| 154 |
+
"model.layers.12.block_sparse_moe.experts.7.w2.weight": "model-00004-of-00008.safetensors",
|
| 155 |
+
"model.layers.12.block_sparse_moe.experts.7.w3.weight": "model-00004-of-00008.safetensors",
|
| 156 |
+
"model.layers.12.block_sparse_moe.gate.weight": "model-00004-of-00008.safetensors",
|
| 157 |
+
"model.layers.12.input_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 158 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
|
| 159 |
+
"model.layers.12.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
|
| 160 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
|
| 161 |
+
"model.layers.12.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
|
| 162 |
+
"model.layers.12.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
|
| 163 |
+
"model.layers.13.block_sparse_moe.experts.0.w1.weight": "model-00004-of-00008.safetensors",
|
| 164 |
+
"model.layers.13.block_sparse_moe.experts.0.w2.weight": "model-00004-of-00008.safetensors",
|
| 165 |
+
"model.layers.13.block_sparse_moe.experts.0.w3.weight": "model-00004-of-00008.safetensors",
|
| 166 |
+
"model.layers.13.block_sparse_moe.experts.1.w1.weight": "model-00004-of-00008.safetensors",
|
| 167 |
+
"model.layers.13.block_sparse_moe.experts.1.w2.weight": "model-00004-of-00008.safetensors",
|
| 168 |
+
"model.layers.13.block_sparse_moe.experts.1.w3.weight": "model-00004-of-00008.safetensors",
|
| 169 |
+
"model.layers.13.block_sparse_moe.experts.2.w1.weight": "model-00004-of-00008.safetensors",
|
| 170 |
+
"model.layers.13.block_sparse_moe.experts.2.w2.weight": "model-00004-of-00008.safetensors",
|
| 171 |
+
"model.layers.13.block_sparse_moe.experts.2.w3.weight": "model-00004-of-00008.safetensors",
|
| 172 |
+
"model.layers.13.block_sparse_moe.experts.3.w1.weight": "model-00004-of-00008.safetensors",
|
| 173 |
+
"model.layers.13.block_sparse_moe.experts.3.w2.weight": "model-00004-of-00008.safetensors",
|
| 174 |
+
"model.layers.13.block_sparse_moe.experts.3.w3.weight": "model-00004-of-00008.safetensors",
|
| 175 |
+
"model.layers.13.block_sparse_moe.experts.4.w1.weight": "model-00004-of-00008.safetensors",
|
| 176 |
+
"model.layers.13.block_sparse_moe.experts.4.w2.weight": "model-00004-of-00008.safetensors",
|
| 177 |
+
"model.layers.13.block_sparse_moe.experts.4.w3.weight": "model-00004-of-00008.safetensors",
|
| 178 |
+
"model.layers.13.block_sparse_moe.experts.5.w1.weight": "model-00004-of-00008.safetensors",
|
| 179 |
+
"model.layers.13.block_sparse_moe.experts.5.w2.weight": "model-00004-of-00008.safetensors",
|
| 180 |
+
"model.layers.13.block_sparse_moe.experts.5.w3.weight": "model-00004-of-00008.safetensors",
|
| 181 |
+
"model.layers.13.block_sparse_moe.experts.6.w1.weight": "model-00004-of-00008.safetensors",
|
| 182 |
+
"model.layers.13.block_sparse_moe.experts.6.w2.weight": "model-00004-of-00008.safetensors",
|
| 183 |
+
"model.layers.13.block_sparse_moe.experts.6.w3.weight": "model-00004-of-00008.safetensors",
|
| 184 |
+
"model.layers.13.block_sparse_moe.experts.7.w1.weight": "model-00004-of-00008.safetensors",
|
| 185 |
+
"model.layers.13.block_sparse_moe.experts.7.w2.weight": "model-00005-of-00008.safetensors",
|
| 186 |
+
"model.layers.13.block_sparse_moe.experts.7.w3.weight": "model-00005-of-00008.safetensors",
|
| 187 |
+
"model.layers.13.block_sparse_moe.gate.weight": "model-00004-of-00008.safetensors",
|
| 188 |
+
"model.layers.13.input_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 189 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 190 |
+
"model.layers.13.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
|
| 191 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
|
| 192 |
+
"model.layers.13.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
|
| 193 |
+
"model.layers.13.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
|
| 194 |
+
"model.layers.14.block_sparse_moe.experts.0.w1.weight": "model-00005-of-00008.safetensors",
|
| 195 |
+
"model.layers.14.block_sparse_moe.experts.0.w2.weight": "model-00005-of-00008.safetensors",
|
| 196 |
+
"model.layers.14.block_sparse_moe.experts.0.w3.weight": "model-00005-of-00008.safetensors",
|
| 197 |
+
"model.layers.14.block_sparse_moe.experts.1.w1.weight": "model-00005-of-00008.safetensors",
|
| 198 |
+
"model.layers.14.block_sparse_moe.experts.1.w2.weight": "model-00005-of-00008.safetensors",
|
| 199 |
+
"model.layers.14.block_sparse_moe.experts.1.w3.weight": "model-00005-of-00008.safetensors",
|
| 200 |
+
"model.layers.14.block_sparse_moe.experts.2.w1.weight": "model-00005-of-00008.safetensors",
|
| 201 |
+
"model.layers.14.block_sparse_moe.experts.2.w2.weight": "model-00005-of-00008.safetensors",
|
| 202 |
+
"model.layers.14.block_sparse_moe.experts.2.w3.weight": "model-00005-of-00008.safetensors",
|
| 203 |
+
"model.layers.14.block_sparse_moe.experts.3.w1.weight": "model-00005-of-00008.safetensors",
|
| 204 |
+
"model.layers.14.block_sparse_moe.experts.3.w2.weight": "model-00005-of-00008.safetensors",
|
| 205 |
+
"model.layers.14.block_sparse_moe.experts.3.w3.weight": "model-00005-of-00008.safetensors",
|
| 206 |
+
"model.layers.14.block_sparse_moe.experts.4.w1.weight": "model-00005-of-00008.safetensors",
|
| 207 |
+
"model.layers.14.block_sparse_moe.experts.4.w2.weight": "model-00005-of-00008.safetensors",
|
| 208 |
+
"model.layers.14.block_sparse_moe.experts.4.w3.weight": "model-00005-of-00008.safetensors",
|
| 209 |
+
"model.layers.14.block_sparse_moe.experts.5.w1.weight": "model-00005-of-00008.safetensors",
|
| 210 |
+
"model.layers.14.block_sparse_moe.experts.5.w2.weight": "model-00005-of-00008.safetensors",
|
| 211 |
+
"model.layers.14.block_sparse_moe.experts.5.w3.weight": "model-00005-of-00008.safetensors",
|
| 212 |
+
"model.layers.14.block_sparse_moe.experts.6.w1.weight": "model-00005-of-00008.safetensors",
|
| 213 |
+
"model.layers.14.block_sparse_moe.experts.6.w2.weight": "model-00005-of-00008.safetensors",
|
| 214 |
+
"model.layers.14.block_sparse_moe.experts.6.w3.weight": "model-00005-of-00008.safetensors",
|
| 215 |
+
"model.layers.14.block_sparse_moe.experts.7.w1.weight": "model-00005-of-00008.safetensors",
|
| 216 |
+
"model.layers.14.block_sparse_moe.experts.7.w2.weight": "model-00005-of-00008.safetensors",
|
| 217 |
+
"model.layers.14.block_sparse_moe.experts.7.w3.weight": "model-00005-of-00008.safetensors",
|
| 218 |
+
"model.layers.14.block_sparse_moe.gate.weight": "model-00005-of-00008.safetensors",
|
| 219 |
+
"model.layers.14.input_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 220 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 221 |
+
"model.layers.14.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
|
| 222 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
|
| 223 |
+
"model.layers.14.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
|
| 224 |
+
"model.layers.14.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
|
| 225 |
+
"model.layers.15.block_sparse_moe.experts.0.w1.weight": "model-00005-of-00008.safetensors",
|
| 226 |
+
"model.layers.15.block_sparse_moe.experts.0.w2.weight": "model-00005-of-00008.safetensors",
|
| 227 |
+
"model.layers.15.block_sparse_moe.experts.0.w3.weight": "model-00005-of-00008.safetensors",
|
| 228 |
+
"model.layers.15.block_sparse_moe.experts.1.w1.weight": "model-00005-of-00008.safetensors",
|
| 229 |
+
"model.layers.15.block_sparse_moe.experts.1.w2.weight": "model-00005-of-00008.safetensors",
|
| 230 |
+
"model.layers.15.block_sparse_moe.experts.1.w3.weight": "model-00005-of-00008.safetensors",
|
| 231 |
+
"model.layers.15.block_sparse_moe.experts.2.w1.weight": "model-00005-of-00008.safetensors",
|
| 232 |
+
"model.layers.15.block_sparse_moe.experts.2.w2.weight": "model-00005-of-00008.safetensors",
|
| 233 |
+
"model.layers.15.block_sparse_moe.experts.2.w3.weight": "model-00005-of-00008.safetensors",
|
| 234 |
+
"model.layers.15.block_sparse_moe.experts.3.w1.weight": "model-00005-of-00008.safetensors",
|
| 235 |
+
"model.layers.15.block_sparse_moe.experts.3.w2.weight": "model-00005-of-00008.safetensors",
|
| 236 |
+
"model.layers.15.block_sparse_moe.experts.3.w3.weight": "model-00005-of-00008.safetensors",
|
| 237 |
+
"model.layers.15.block_sparse_moe.experts.4.w1.weight": "model-00005-of-00008.safetensors",
|
| 238 |
+
"model.layers.15.block_sparse_moe.experts.4.w2.weight": "model-00005-of-00008.safetensors",
|
| 239 |
+
"model.layers.15.block_sparse_moe.experts.4.w3.weight": "model-00005-of-00008.safetensors",
|
| 240 |
+
"model.layers.15.block_sparse_moe.experts.5.w1.weight": "model-00005-of-00008.safetensors",
|
| 241 |
+
"model.layers.15.block_sparse_moe.experts.5.w2.weight": "model-00005-of-00008.safetensors",
|
| 242 |
+
"model.layers.15.block_sparse_moe.experts.5.w3.weight": "model-00005-of-00008.safetensors",
|
| 243 |
+
"model.layers.15.block_sparse_moe.experts.6.w1.weight": "model-00005-of-00008.safetensors",
|
| 244 |
+
"model.layers.15.block_sparse_moe.experts.6.w2.weight": "model-00005-of-00008.safetensors",
|
| 245 |
+
"model.layers.15.block_sparse_moe.experts.6.w3.weight": "model-00005-of-00008.safetensors",
|
| 246 |
+
"model.layers.15.block_sparse_moe.experts.7.w1.weight": "model-00005-of-00008.safetensors",
|
| 247 |
+
"model.layers.15.block_sparse_moe.experts.7.w2.weight": "model-00005-of-00008.safetensors",
|
| 248 |
+
"model.layers.15.block_sparse_moe.experts.7.w3.weight": "model-00005-of-00008.safetensors",
|
| 249 |
+
"model.layers.15.block_sparse_moe.gate.weight": "model-00005-of-00008.safetensors",
|
| 250 |
+
"model.layers.15.input_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 251 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 252 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
|
| 253 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
|
| 254 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
|
| 255 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
|
| 256 |
+
"model.layers.16.block_sparse_moe.experts.0.w1.weight": "model-00005-of-00008.safetensors",
|
| 257 |
+
"model.layers.16.block_sparse_moe.experts.0.w2.weight": "model-00005-of-00008.safetensors",
|
| 258 |
+
"model.layers.16.block_sparse_moe.experts.0.w3.weight": "model-00005-of-00008.safetensors",
|
| 259 |
+
"model.layers.16.block_sparse_moe.experts.1.w1.weight": "model-00005-of-00008.safetensors",
|
| 260 |
+
"model.layers.16.block_sparse_moe.experts.1.w2.weight": "model-00005-of-00008.safetensors",
|
| 261 |
+
"model.layers.16.block_sparse_moe.experts.1.w3.weight": "model-00005-of-00008.safetensors",
|
| 262 |
+
"model.layers.16.block_sparse_moe.experts.2.w1.weight": "model-00005-of-00008.safetensors",
|
| 263 |
+
"model.layers.16.block_sparse_moe.experts.2.w2.weight": "model-00005-of-00008.safetensors",
|
| 264 |
+
"model.layers.16.block_sparse_moe.experts.2.w3.weight": "model-00005-of-00008.safetensors",
|
| 265 |
+
"model.layers.16.block_sparse_moe.experts.3.w1.weight": "model-00005-of-00008.safetensors",
|
| 266 |
+
"model.layers.16.block_sparse_moe.experts.3.w2.weight": "model-00005-of-00008.safetensors",
|
| 267 |
+
"model.layers.16.block_sparse_moe.experts.3.w3.weight": "model-00005-of-00008.safetensors",
|
| 268 |
+
"model.layers.16.block_sparse_moe.experts.4.w1.weight": "model-00005-of-00008.safetensors",
|
| 269 |
+
"model.layers.16.block_sparse_moe.experts.4.w2.weight": "model-00005-of-00008.safetensors",
|
| 270 |
+
"model.layers.16.block_sparse_moe.experts.4.w3.weight": "model-00005-of-00008.safetensors",
|
| 271 |
+
"model.layers.16.block_sparse_moe.experts.5.w1.weight": "model-00005-of-00008.safetensors",
|
| 272 |
+
"model.layers.16.block_sparse_moe.experts.5.w2.weight": "model-00005-of-00008.safetensors",
|
| 273 |
+
"model.layers.16.block_sparse_moe.experts.5.w3.weight": "model-00005-of-00008.safetensors",
|
| 274 |
+
"model.layers.16.block_sparse_moe.experts.6.w1.weight": "model-00005-of-00008.safetensors",
|
| 275 |
+
"model.layers.16.block_sparse_moe.experts.6.w2.weight": "model-00005-of-00008.safetensors",
|
| 276 |
+
"model.layers.16.block_sparse_moe.experts.6.w3.weight": "model-00005-of-00008.safetensors",
|
| 277 |
+
"model.layers.16.block_sparse_moe.experts.7.w1.weight": "model-00005-of-00008.safetensors",
|
| 278 |
+
"model.layers.16.block_sparse_moe.experts.7.w2.weight": "model-00005-of-00008.safetensors",
|
| 279 |
+
"model.layers.16.block_sparse_moe.experts.7.w3.weight": "model-00005-of-00008.safetensors",
|
| 280 |
+
"model.layers.16.block_sparse_moe.gate.weight": "model-00005-of-00008.safetensors",
|
| 281 |
+
"model.layers.16.input_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 282 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
|
| 283 |
+
"model.layers.16.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
|
| 284 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
|
| 285 |
+
"model.layers.16.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
|
| 286 |
+
"model.layers.16.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
|
| 287 |
+
"model.layers.17.block_sparse_moe.experts.0.w1.weight": "model-00005-of-00008.safetensors",
|
| 288 |
+
"model.layers.17.block_sparse_moe.experts.0.w2.weight": "model-00005-of-00008.safetensors",
|
| 289 |
+
"model.layers.17.block_sparse_moe.experts.0.w3.weight": "model-00005-of-00008.safetensors",
|
| 290 |
+
"model.layers.17.block_sparse_moe.experts.1.w1.weight": "model-00005-of-00008.safetensors",
|
| 291 |
+
"model.layers.17.block_sparse_moe.experts.1.w2.weight": "model-00005-of-00008.safetensors",
|
| 292 |
+
"model.layers.17.block_sparse_moe.experts.1.w3.weight": "model-00005-of-00008.safetensors",
|
| 293 |
+
"model.layers.17.block_sparse_moe.experts.2.w1.weight": "model-00005-of-00008.safetensors",
|
| 294 |
+
"model.layers.17.block_sparse_moe.experts.2.w2.weight": "model-00005-of-00008.safetensors",
|
| 295 |
+
"model.layers.17.block_sparse_moe.experts.2.w3.weight": "model-00005-of-00008.safetensors",
|
| 296 |
+
"model.layers.17.block_sparse_moe.experts.3.w1.weight": "model-00005-of-00008.safetensors",
|
| 297 |
+
"model.layers.17.block_sparse_moe.experts.3.w2.weight": "model-00006-of-00008.safetensors",
|
| 298 |
+
"model.layers.17.block_sparse_moe.experts.3.w3.weight": "model-00006-of-00008.safetensors",
|
| 299 |
+
"model.layers.17.block_sparse_moe.experts.4.w1.weight": "model-00006-of-00008.safetensors",
|
| 300 |
+
"model.layers.17.block_sparse_moe.experts.4.w2.weight": "model-00006-of-00008.safetensors",
|
| 301 |
+
"model.layers.17.block_sparse_moe.experts.4.w3.weight": "model-00006-of-00008.safetensors",
|
| 302 |
+
"model.layers.17.block_sparse_moe.experts.5.w1.weight": "model-00006-of-00008.safetensors",
|
| 303 |
+
"model.layers.17.block_sparse_moe.experts.5.w2.weight": "model-00006-of-00008.safetensors",
|
| 304 |
+
"model.layers.17.block_sparse_moe.experts.5.w3.weight": "model-00006-of-00008.safetensors",
|
| 305 |
+
"model.layers.17.block_sparse_moe.experts.6.w1.weight": "model-00006-of-00008.safetensors",
|
| 306 |
+
"model.layers.17.block_sparse_moe.experts.6.w2.weight": "model-00006-of-00008.safetensors",
|
| 307 |
+
"model.layers.17.block_sparse_moe.experts.6.w3.weight": "model-00006-of-00008.safetensors",
|
| 308 |
+
"model.layers.17.block_sparse_moe.experts.7.w1.weight": "model-00006-of-00008.safetensors",
|
| 309 |
+
"model.layers.17.block_sparse_moe.experts.7.w2.weight": "model-00006-of-00008.safetensors",
|
| 310 |
+
"model.layers.17.block_sparse_moe.experts.7.w3.weight": "model-00006-of-00008.safetensors",
|
| 311 |
+
"model.layers.17.block_sparse_moe.gate.weight": "model-00005-of-00008.safetensors",
|
| 312 |
+
"model.layers.17.input_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 313 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 314 |
+
"model.layers.17.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
|
| 315 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
|
| 316 |
+
"model.layers.17.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
|
| 317 |
+
"model.layers.17.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
|
| 318 |
+
"model.layers.18.block_sparse_moe.experts.0.w1.weight": "model-00006-of-00008.safetensors",
|
| 319 |
+
"model.layers.18.block_sparse_moe.experts.0.w2.weight": "model-00006-of-00008.safetensors",
|
| 320 |
+
"model.layers.18.block_sparse_moe.experts.0.w3.weight": "model-00006-of-00008.safetensors",
|
| 321 |
+
"model.layers.18.block_sparse_moe.experts.1.w1.weight": "model-00006-of-00008.safetensors",
|
| 322 |
+
"model.layers.18.block_sparse_moe.experts.1.w2.weight": "model-00006-of-00008.safetensors",
|
| 323 |
+
"model.layers.18.block_sparse_moe.experts.1.w3.weight": "model-00006-of-00008.safetensors",
|
| 324 |
+
"model.layers.18.block_sparse_moe.experts.2.w1.weight": "model-00006-of-00008.safetensors",
|
| 325 |
+
"model.layers.18.block_sparse_moe.experts.2.w2.weight": "model-00006-of-00008.safetensors",
|
| 326 |
+
"model.layers.18.block_sparse_moe.experts.2.w3.weight": "model-00006-of-00008.safetensors",
|
| 327 |
+
"model.layers.18.block_sparse_moe.experts.3.w1.weight": "model-00006-of-00008.safetensors",
|
| 328 |
+
"model.layers.18.block_sparse_moe.experts.3.w2.weight": "model-00006-of-00008.safetensors",
|
| 329 |
+
"model.layers.18.block_sparse_moe.experts.3.w3.weight": "model-00006-of-00008.safetensors",
|
| 330 |
+
"model.layers.18.block_sparse_moe.experts.4.w1.weight": "model-00006-of-00008.safetensors",
|
| 331 |
+
"model.layers.18.block_sparse_moe.experts.4.w2.weight": "model-00006-of-00008.safetensors",
|
| 332 |
+
"model.layers.18.block_sparse_moe.experts.4.w3.weight": "model-00006-of-00008.safetensors",
|
| 333 |
+
"model.layers.18.block_sparse_moe.experts.5.w1.weight": "model-00006-of-00008.safetensors",
|
| 334 |
+
"model.layers.18.block_sparse_moe.experts.5.w2.weight": "model-00006-of-00008.safetensors",
|
| 335 |
+
"model.layers.18.block_sparse_moe.experts.5.w3.weight": "model-00006-of-00008.safetensors",
|
| 336 |
+
"model.layers.18.block_sparse_moe.experts.6.w1.weight": "model-00006-of-00008.safetensors",
|
| 337 |
+
"model.layers.18.block_sparse_moe.experts.6.w2.weight": "model-00006-of-00008.safetensors",
|
| 338 |
+
"model.layers.18.block_sparse_moe.experts.6.w3.weight": "model-00006-of-00008.safetensors",
|
| 339 |
+
"model.layers.18.block_sparse_moe.experts.7.w1.weight": "model-00006-of-00008.safetensors",
|
| 340 |
+
"model.layers.18.block_sparse_moe.experts.7.w2.weight": "model-00006-of-00008.safetensors",
|
| 341 |
+
"model.layers.18.block_sparse_moe.experts.7.w3.weight": "model-00006-of-00008.safetensors",
|
| 342 |
+
"model.layers.18.block_sparse_moe.gate.weight": "model-00006-of-00008.safetensors",
|
| 343 |
+
"model.layers.18.input_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 344 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 345 |
+
"model.layers.18.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
|
| 346 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
|
| 347 |
+
"model.layers.18.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
|
| 348 |
+
"model.layers.18.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
|
| 349 |
+
"model.layers.19.block_sparse_moe.experts.0.w1.weight": "model-00006-of-00008.safetensors",
|
| 350 |
+
"model.layers.19.block_sparse_moe.experts.0.w2.weight": "model-00006-of-00008.safetensors",
|
| 351 |
+
"model.layers.19.block_sparse_moe.experts.0.w3.weight": "model-00006-of-00008.safetensors",
|
| 352 |
+
"model.layers.19.block_sparse_moe.experts.1.w1.weight": "model-00006-of-00008.safetensors",
|
| 353 |
+
"model.layers.19.block_sparse_moe.experts.1.w2.weight": "model-00006-of-00008.safetensors",
|
| 354 |
+
"model.layers.19.block_sparse_moe.experts.1.w3.weight": "model-00006-of-00008.safetensors",
|
| 355 |
+
"model.layers.19.block_sparse_moe.experts.2.w1.weight": "model-00006-of-00008.safetensors",
|
| 356 |
+
"model.layers.19.block_sparse_moe.experts.2.w2.weight": "model-00006-of-00008.safetensors",
|
| 357 |
+
"model.layers.19.block_sparse_moe.experts.2.w3.weight": "model-00006-of-00008.safetensors",
|
| 358 |
+
"model.layers.19.block_sparse_moe.experts.3.w1.weight": "model-00006-of-00008.safetensors",
|
| 359 |
+
"model.layers.19.block_sparse_moe.experts.3.w2.weight": "model-00006-of-00008.safetensors",
|
| 360 |
+
"model.layers.19.block_sparse_moe.experts.3.w3.weight": "model-00006-of-00008.safetensors",
|
| 361 |
+
"model.layers.19.block_sparse_moe.experts.4.w1.weight": "model-00006-of-00008.safetensors",
|
| 362 |
+
"model.layers.19.block_sparse_moe.experts.4.w2.weight": "model-00006-of-00008.safetensors",
|
| 363 |
+
"model.layers.19.block_sparse_moe.experts.4.w3.weight": "model-00006-of-00008.safetensors",
|
| 364 |
+
"model.layers.19.block_sparse_moe.experts.5.w1.weight": "model-00006-of-00008.safetensors",
|
| 365 |
+
"model.layers.19.block_sparse_moe.experts.5.w2.weight": "model-00006-of-00008.safetensors",
|
| 366 |
+
"model.layers.19.block_sparse_moe.experts.5.w3.weight": "model-00006-of-00008.safetensors",
|
| 367 |
+
"model.layers.19.block_sparse_moe.experts.6.w1.weight": "model-00006-of-00008.safetensors",
|
| 368 |
+
"model.layers.19.block_sparse_moe.experts.6.w2.weight": "model-00006-of-00008.safetensors",
|
| 369 |
+
"model.layers.19.block_sparse_moe.experts.6.w3.weight": "model-00006-of-00008.safetensors",
|
| 370 |
+
"model.layers.19.block_sparse_moe.experts.7.w1.weight": "model-00006-of-00008.safetensors",
|
| 371 |
+
"model.layers.19.block_sparse_moe.experts.7.w2.weight": "model-00006-of-00008.safetensors",
|
| 372 |
+
"model.layers.19.block_sparse_moe.experts.7.w3.weight": "model-00006-of-00008.safetensors",
|
| 373 |
+
"model.layers.19.block_sparse_moe.gate.weight": "model-00006-of-00008.safetensors",
|
| 374 |
+
"model.layers.19.input_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 375 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
|
| 376 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
|
| 377 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
|
| 378 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
|
| 379 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
|
| 380 |
+
"model.layers.2.block_sparse_moe.experts.0.w1.weight": "model-00001-of-00008.safetensors",
|
| 381 |
+
"model.layers.2.block_sparse_moe.experts.0.w2.weight": "model-00001-of-00008.safetensors",
|
| 382 |
+
"model.layers.2.block_sparse_moe.experts.0.w3.weight": "model-00001-of-00008.safetensors",
|
| 383 |
+
"model.layers.2.block_sparse_moe.experts.1.w1.weight": "model-00001-of-00008.safetensors",
|
| 384 |
+
"model.layers.2.block_sparse_moe.experts.1.w2.weight": "model-00001-of-00008.safetensors",
|
| 385 |
+
"model.layers.2.block_sparse_moe.experts.1.w3.weight": "model-00001-of-00008.safetensors",
|
| 386 |
+
"model.layers.2.block_sparse_moe.experts.2.w1.weight": "model-00001-of-00008.safetensors",
|
| 387 |
+
"model.layers.2.block_sparse_moe.experts.2.w2.weight": "model-00001-of-00008.safetensors",
|
| 388 |
+
"model.layers.2.block_sparse_moe.experts.2.w3.weight": "model-00001-of-00008.safetensors",
|
| 389 |
+
"model.layers.2.block_sparse_moe.experts.3.w1.weight": "model-00001-of-00008.safetensors",
|
| 390 |
+
"model.layers.2.block_sparse_moe.experts.3.w2.weight": "model-00001-of-00008.safetensors",
|
| 391 |
+
"model.layers.2.block_sparse_moe.experts.3.w3.weight": "model-00001-of-00008.safetensors",
|
| 392 |
+
"model.layers.2.block_sparse_moe.experts.4.w1.weight": "model-00001-of-00008.safetensors",
|
| 393 |
+
"model.layers.2.block_sparse_moe.experts.4.w2.weight": "model-00001-of-00008.safetensors",
|
| 394 |
+
"model.layers.2.block_sparse_moe.experts.4.w3.weight": "model-00001-of-00008.safetensors",
|
| 395 |
+
"model.layers.2.block_sparse_moe.experts.5.w1.weight": "model-00001-of-00008.safetensors",
|
| 396 |
+
"model.layers.2.block_sparse_moe.experts.5.w2.weight": "model-00001-of-00008.safetensors",
|
| 397 |
+
"model.layers.2.block_sparse_moe.experts.5.w3.weight": "model-00001-of-00008.safetensors",
|
| 398 |
+
"model.layers.2.block_sparse_moe.experts.6.w1.weight": "model-00001-of-00008.safetensors",
|
| 399 |
+
"model.layers.2.block_sparse_moe.experts.6.w2.weight": "model-00001-of-00008.safetensors",
|
| 400 |
+
"model.layers.2.block_sparse_moe.experts.6.w3.weight": "model-00001-of-00008.safetensors",
|
| 401 |
+
"model.layers.2.block_sparse_moe.experts.7.w1.weight": "model-00001-of-00008.safetensors",
|
| 402 |
+
"model.layers.2.block_sparse_moe.experts.7.w2.weight": "model-00001-of-00008.safetensors",
|
| 403 |
+
"model.layers.2.block_sparse_moe.experts.7.w3.weight": "model-00001-of-00008.safetensors",
|
| 404 |
+
"model.layers.2.block_sparse_moe.gate.weight": "model-00001-of-00008.safetensors",
|
| 405 |
+
"model.layers.2.input_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 406 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00001-of-00008.safetensors",
|
| 407 |
+
"model.layers.2.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
|
| 408 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
|
| 409 |
+
"model.layers.2.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
|
| 410 |
+
"model.layers.2.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
|
| 411 |
+
"model.layers.20.block_sparse_moe.experts.0.w1.weight": "model-00006-of-00008.safetensors",
|
| 412 |
+
"model.layers.20.block_sparse_moe.experts.0.w2.weight": "model-00006-of-00008.safetensors",
|
| 413 |
+
"model.layers.20.block_sparse_moe.experts.0.w3.weight": "model-00006-of-00008.safetensors",
|
| 414 |
+
"model.layers.20.block_sparse_moe.experts.1.w1.weight": "model-00006-of-00008.safetensors",
|
| 415 |
+
"model.layers.20.block_sparse_moe.experts.1.w2.weight": "model-00006-of-00008.safetensors",
|
| 416 |
+
"model.layers.20.block_sparse_moe.experts.1.w3.weight": "model-00006-of-00008.safetensors",
|
| 417 |
+
"model.layers.20.block_sparse_moe.experts.2.w1.weight": "model-00006-of-00008.safetensors",
|
| 418 |
+
"model.layers.20.block_sparse_moe.experts.2.w2.weight": "model-00006-of-00008.safetensors",
|
| 419 |
+
"model.layers.20.block_sparse_moe.experts.2.w3.weight": "model-00006-of-00008.safetensors",
|
| 420 |
+
"model.layers.20.block_sparse_moe.experts.3.w1.weight": "model-00006-of-00008.safetensors",
|
| 421 |
+
"model.layers.20.block_sparse_moe.experts.3.w2.weight": "model-00006-of-00008.safetensors",
|
| 422 |
+
"model.layers.20.block_sparse_moe.experts.3.w3.weight": "model-00006-of-00008.safetensors",
|
| 423 |
+
"model.layers.20.block_sparse_moe.experts.4.w1.weight": "model-00006-of-00008.safetensors",
|
| 424 |
+
"model.layers.20.block_sparse_moe.experts.4.w2.weight": "model-00006-of-00008.safetensors",
|
| 425 |
+
"model.layers.20.block_sparse_moe.experts.4.w3.weight": "model-00006-of-00008.safetensors",
|
| 426 |
+
"model.layers.20.block_sparse_moe.experts.5.w1.weight": "model-00006-of-00008.safetensors",
|
| 427 |
+
"model.layers.20.block_sparse_moe.experts.5.w2.weight": "model-00006-of-00008.safetensors",
|
| 428 |
+
"model.layers.20.block_sparse_moe.experts.5.w3.weight": "model-00006-of-00008.safetensors",
|
| 429 |
+
"model.layers.20.block_sparse_moe.experts.6.w1.weight": "model-00006-of-00008.safetensors",
|
| 430 |
+
"model.layers.20.block_sparse_moe.experts.6.w2.weight": "model-00006-of-00008.safetensors",
|
| 431 |
+
"model.layers.20.block_sparse_moe.experts.6.w3.weight": "model-00006-of-00008.safetensors",
|
| 432 |
+
"model.layers.20.block_sparse_moe.experts.7.w1.weight": "model-00006-of-00008.safetensors",
|
| 433 |
+
"model.layers.20.block_sparse_moe.experts.7.w2.weight": "model-00006-of-00008.safetensors",
|
| 434 |
+
"model.layers.20.block_sparse_moe.experts.7.w3.weight": "model-00007-of-00008.safetensors",
|
| 435 |
+
"model.layers.20.block_sparse_moe.gate.weight": "model-00006-of-00008.safetensors",
|
| 436 |
+
"model.layers.20.input_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 437 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 438 |
+
"model.layers.20.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
|
| 439 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
|
| 440 |
+
"model.layers.20.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
|
| 441 |
+
"model.layers.20.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
|
| 442 |
+
"model.layers.21.block_sparse_moe.experts.0.w1.weight": "model-00007-of-00008.safetensors",
|
| 443 |
+
"model.layers.21.block_sparse_moe.experts.0.w2.weight": "model-00007-of-00008.safetensors",
|
| 444 |
+
"model.layers.21.block_sparse_moe.experts.0.w3.weight": "model-00007-of-00008.safetensors",
|
| 445 |
+
"model.layers.21.block_sparse_moe.experts.1.w1.weight": "model-00007-of-00008.safetensors",
|
| 446 |
+
"model.layers.21.block_sparse_moe.experts.1.w2.weight": "model-00007-of-00008.safetensors",
|
| 447 |
+
"model.layers.21.block_sparse_moe.experts.1.w3.weight": "model-00007-of-00008.safetensors",
|
| 448 |
+
"model.layers.21.block_sparse_moe.experts.2.w1.weight": "model-00007-of-00008.safetensors",
|
| 449 |
+
"model.layers.21.block_sparse_moe.experts.2.w2.weight": "model-00007-of-00008.safetensors",
|
| 450 |
+
"model.layers.21.block_sparse_moe.experts.2.w3.weight": "model-00007-of-00008.safetensors",
|
| 451 |
+
"model.layers.21.block_sparse_moe.experts.3.w1.weight": "model-00007-of-00008.safetensors",
|
| 452 |
+
"model.layers.21.block_sparse_moe.experts.3.w2.weight": "model-00007-of-00008.safetensors",
|
| 453 |
+
"model.layers.21.block_sparse_moe.experts.3.w3.weight": "model-00007-of-00008.safetensors",
|
| 454 |
+
"model.layers.21.block_sparse_moe.experts.4.w1.weight": "model-00007-of-00008.safetensors",
|
| 455 |
+
"model.layers.21.block_sparse_moe.experts.4.w2.weight": "model-00007-of-00008.safetensors",
|
| 456 |
+
"model.layers.21.block_sparse_moe.experts.4.w3.weight": "model-00007-of-00008.safetensors",
|
| 457 |
+
"model.layers.21.block_sparse_moe.experts.5.w1.weight": "model-00007-of-00008.safetensors",
|
| 458 |
+
"model.layers.21.block_sparse_moe.experts.5.w2.weight": "model-00007-of-00008.safetensors",
|
| 459 |
+
"model.layers.21.block_sparse_moe.experts.5.w3.weight": "model-00007-of-00008.safetensors",
|
| 460 |
+
"model.layers.21.block_sparse_moe.experts.6.w1.weight": "model-00007-of-00008.safetensors",
|
| 461 |
+
"model.layers.21.block_sparse_moe.experts.6.w2.weight": "model-00007-of-00008.safetensors",
|
| 462 |
+
"model.layers.21.block_sparse_moe.experts.6.w3.weight": "model-00007-of-00008.safetensors",
|
| 463 |
+
"model.layers.21.block_sparse_moe.experts.7.w1.weight": "model-00007-of-00008.safetensors",
|
| 464 |
+
"model.layers.21.block_sparse_moe.experts.7.w2.weight": "model-00007-of-00008.safetensors",
|
| 465 |
+
"model.layers.21.block_sparse_moe.experts.7.w3.weight": "model-00007-of-00008.safetensors",
|
| 466 |
+
"model.layers.21.block_sparse_moe.gate.weight": "model-00007-of-00008.safetensors",
|
| 467 |
+
"model.layers.21.input_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 468 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 469 |
+
"model.layers.21.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
|
| 470 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
|
| 471 |
+
"model.layers.21.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
|
| 472 |
+
"model.layers.21.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
|
| 473 |
+
"model.layers.22.block_sparse_moe.experts.0.w1.weight": "model-00007-of-00008.safetensors",
|
| 474 |
+
"model.layers.22.block_sparse_moe.experts.0.w2.weight": "model-00007-of-00008.safetensors",
|
| 475 |
+
"model.layers.22.block_sparse_moe.experts.0.w3.weight": "model-00007-of-00008.safetensors",
|
| 476 |
+
"model.layers.22.block_sparse_moe.experts.1.w1.weight": "model-00007-of-00008.safetensors",
|
| 477 |
+
"model.layers.22.block_sparse_moe.experts.1.w2.weight": "model-00007-of-00008.safetensors",
|
| 478 |
+
"model.layers.22.block_sparse_moe.experts.1.w3.weight": "model-00007-of-00008.safetensors",
|
| 479 |
+
"model.layers.22.block_sparse_moe.experts.2.w1.weight": "model-00007-of-00008.safetensors",
|
| 480 |
+
"model.layers.22.block_sparse_moe.experts.2.w2.weight": "model-00007-of-00008.safetensors",
|
| 481 |
+
"model.layers.22.block_sparse_moe.experts.2.w3.weight": "model-00007-of-00008.safetensors",
|
| 482 |
+
"model.layers.22.block_sparse_moe.experts.3.w1.weight": "model-00007-of-00008.safetensors",
|
| 483 |
+
"model.layers.22.block_sparse_moe.experts.3.w2.weight": "model-00007-of-00008.safetensors",
|
| 484 |
+
"model.layers.22.block_sparse_moe.experts.3.w3.weight": "model-00007-of-00008.safetensors",
|
| 485 |
+
"model.layers.22.block_sparse_moe.experts.4.w1.weight": "model-00007-of-00008.safetensors",
|
| 486 |
+
"model.layers.22.block_sparse_moe.experts.4.w2.weight": "model-00007-of-00008.safetensors",
|
| 487 |
+
"model.layers.22.block_sparse_moe.experts.4.w3.weight": "model-00007-of-00008.safetensors",
|
| 488 |
+
"model.layers.22.block_sparse_moe.experts.5.w1.weight": "model-00007-of-00008.safetensors",
|
| 489 |
+
"model.layers.22.block_sparse_moe.experts.5.w2.weight": "model-00007-of-00008.safetensors",
|
| 490 |
+
"model.layers.22.block_sparse_moe.experts.5.w3.weight": "model-00007-of-00008.safetensors",
|
| 491 |
+
"model.layers.22.block_sparse_moe.experts.6.w1.weight": "model-00007-of-00008.safetensors",
|
| 492 |
+
"model.layers.22.block_sparse_moe.experts.6.w2.weight": "model-00007-of-00008.safetensors",
|
| 493 |
+
"model.layers.22.block_sparse_moe.experts.6.w3.weight": "model-00007-of-00008.safetensors",
|
| 494 |
+
"model.layers.22.block_sparse_moe.experts.7.w1.weight": "model-00007-of-00008.safetensors",
|
| 495 |
+
"model.layers.22.block_sparse_moe.experts.7.w2.weight": "model-00007-of-00008.safetensors",
|
| 496 |
+
"model.layers.22.block_sparse_moe.experts.7.w3.weight": "model-00007-of-00008.safetensors",
|
| 497 |
+
"model.layers.22.block_sparse_moe.gate.weight": "model-00007-of-00008.safetensors",
|
| 498 |
+
"model.layers.22.input_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 499 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 500 |
+
"model.layers.22.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
|
| 501 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
|
| 502 |
+
"model.layers.22.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
|
| 503 |
+
"model.layers.22.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
|
| 504 |
+
"model.layers.23.block_sparse_moe.experts.0.w1.weight": "model-00007-of-00008.safetensors",
|
| 505 |
+
"model.layers.23.block_sparse_moe.experts.0.w2.weight": "model-00007-of-00008.safetensors",
|
| 506 |
+
"model.layers.23.block_sparse_moe.experts.0.w3.weight": "model-00007-of-00008.safetensors",
|
| 507 |
+
"model.layers.23.block_sparse_moe.experts.1.w1.weight": "model-00007-of-00008.safetensors",
|
| 508 |
+
"model.layers.23.block_sparse_moe.experts.1.w2.weight": "model-00007-of-00008.safetensors",
|
| 509 |
+
"model.layers.23.block_sparse_moe.experts.1.w3.weight": "model-00007-of-00008.safetensors",
|
| 510 |
+
"model.layers.23.block_sparse_moe.experts.2.w1.weight": "model-00007-of-00008.safetensors",
|
| 511 |
+
"model.layers.23.block_sparse_moe.experts.2.w2.weight": "model-00007-of-00008.safetensors",
|
| 512 |
+
"model.layers.23.block_sparse_moe.experts.2.w3.weight": "model-00007-of-00008.safetensors",
|
| 513 |
+
"model.layers.23.block_sparse_moe.experts.3.w1.weight": "model-00007-of-00008.safetensors",
|
| 514 |
+
"model.layers.23.block_sparse_moe.experts.3.w2.weight": "model-00007-of-00008.safetensors",
|
| 515 |
+
"model.layers.23.block_sparse_moe.experts.3.w3.weight": "model-00007-of-00008.safetensors",
|
| 516 |
+
"model.layers.23.block_sparse_moe.experts.4.w1.weight": "model-00007-of-00008.safetensors",
|
| 517 |
+
"model.layers.23.block_sparse_moe.experts.4.w2.weight": "model-00007-of-00008.safetensors",
|
| 518 |
+
"model.layers.23.block_sparse_moe.experts.4.w3.weight": "model-00007-of-00008.safetensors",
|
| 519 |
+
"model.layers.23.block_sparse_moe.experts.5.w1.weight": "model-00007-of-00008.safetensors",
|
| 520 |
+
"model.layers.23.block_sparse_moe.experts.5.w2.weight": "model-00007-of-00008.safetensors",
|
| 521 |
+
"model.layers.23.block_sparse_moe.experts.5.w3.weight": "model-00007-of-00008.safetensors",
|
| 522 |
+
"model.layers.23.block_sparse_moe.experts.6.w1.weight": "model-00007-of-00008.safetensors",
|
| 523 |
+
"model.layers.23.block_sparse_moe.experts.6.w2.weight": "model-00007-of-00008.safetensors",
|
| 524 |
+
"model.layers.23.block_sparse_moe.experts.6.w3.weight": "model-00007-of-00008.safetensors",
|
| 525 |
+
"model.layers.23.block_sparse_moe.experts.7.w1.weight": "model-00007-of-00008.safetensors",
|
| 526 |
+
"model.layers.23.block_sparse_moe.experts.7.w2.weight": "model-00007-of-00008.safetensors",
|
| 527 |
+
"model.layers.23.block_sparse_moe.experts.7.w3.weight": "model-00007-of-00008.safetensors",
|
| 528 |
+
"model.layers.23.block_sparse_moe.gate.weight": "model-00007-of-00008.safetensors",
|
| 529 |
+
"model.layers.23.input_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 530 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
|
| 531 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
|
| 532 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
|
| 533 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
|
| 534 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
|
| 535 |
+
"model.layers.24.block_sparse_moe.experts.0.w1.weight": "model-00007-of-00008.safetensors",
|
| 536 |
+
"model.layers.24.block_sparse_moe.experts.0.w2.weight": "model-00007-of-00008.safetensors",
|
| 537 |
+
"model.layers.24.block_sparse_moe.experts.0.w3.weight": "model-00007-of-00008.safetensors",
|
| 538 |
+
"model.layers.24.block_sparse_moe.experts.1.w1.weight": "model-00007-of-00008.safetensors",
|
| 539 |
+
"model.layers.24.block_sparse_moe.experts.1.w2.weight": "model-00007-of-00008.safetensors",
|
| 540 |
+
"model.layers.24.block_sparse_moe.experts.1.w3.weight": "model-00007-of-00008.safetensors",
|
| 541 |
+
"model.layers.24.block_sparse_moe.experts.2.w1.weight": "model-00007-of-00008.safetensors",
|
| 542 |
+
"model.layers.24.block_sparse_moe.experts.2.w2.weight": "model-00007-of-00008.safetensors",
|
| 543 |
+
"model.layers.24.block_sparse_moe.experts.2.w3.weight": "model-00007-of-00008.safetensors",
|
| 544 |
+
"model.layers.24.block_sparse_moe.experts.3.w1.weight": "model-00007-of-00008.safetensors",
|
| 545 |
+
"model.layers.24.block_sparse_moe.experts.3.w2.weight": "model-00007-of-00008.safetensors",
|
| 546 |
+
"model.layers.24.block_sparse_moe.experts.3.w3.weight": "model-00008-of-00008.safetensors",
|
| 547 |
+
"model.layers.24.block_sparse_moe.experts.4.w1.weight": "model-00008-of-00008.safetensors",
|
| 548 |
+
"model.layers.24.block_sparse_moe.experts.4.w2.weight": "model-00008-of-00008.safetensors",
|
| 549 |
+
"model.layers.24.block_sparse_moe.experts.4.w3.weight": "model-00008-of-00008.safetensors",
|
| 550 |
+
"model.layers.24.block_sparse_moe.experts.5.w1.weight": "model-00008-of-00008.safetensors",
|
| 551 |
+
"model.layers.24.block_sparse_moe.experts.5.w2.weight": "model-00008-of-00008.safetensors",
|
| 552 |
+
"model.layers.24.block_sparse_moe.experts.5.w3.weight": "model-00008-of-00008.safetensors",
|
| 553 |
+
"model.layers.24.block_sparse_moe.experts.6.w1.weight": "model-00008-of-00008.safetensors",
|
| 554 |
+
"model.layers.24.block_sparse_moe.experts.6.w2.weight": "model-00008-of-00008.safetensors",
|
| 555 |
+
"model.layers.24.block_sparse_moe.experts.6.w3.weight": "model-00008-of-00008.safetensors",
|
| 556 |
+
"model.layers.24.block_sparse_moe.experts.7.w1.weight": "model-00008-of-00008.safetensors",
|
| 557 |
+
"model.layers.24.block_sparse_moe.experts.7.w2.weight": "model-00008-of-00008.safetensors",
|
| 558 |
+
"model.layers.24.block_sparse_moe.experts.7.w3.weight": "model-00008-of-00008.safetensors",
|
| 559 |
+
"model.layers.24.block_sparse_moe.gate.weight": "model-00007-of-00008.safetensors",
|
| 560 |
+
"model.layers.24.input_layernorm.weight": "model-00008-of-00008.safetensors",
|
| 561 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00008-of-00008.safetensors",
|
| 562 |
+
"model.layers.24.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
|
| 563 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
|
| 564 |
+
"model.layers.24.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
|
| 565 |
+
"model.layers.24.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
|
| 566 |
+
"model.layers.25.block_sparse_moe.experts.0.w1.weight": "model-00008-of-00008.safetensors",
|
| 567 |
+
"model.layers.25.block_sparse_moe.experts.0.w2.weight": "model-00008-of-00008.safetensors",
|
| 568 |
+
"model.layers.25.block_sparse_moe.experts.0.w3.weight": "model-00008-of-00008.safetensors",
|
| 569 |
+
"model.layers.25.block_sparse_moe.experts.1.w1.weight": "model-00008-of-00008.safetensors",
|
| 570 |
+
"model.layers.25.block_sparse_moe.experts.1.w2.weight": "model-00008-of-00008.safetensors",
|
| 571 |
+
"model.layers.25.block_sparse_moe.experts.1.w3.weight": "model-00008-of-00008.safetensors",
|
| 572 |
+
"model.layers.25.block_sparse_moe.experts.2.w1.weight": "model-00008-of-00008.safetensors",
|
| 573 |
+
"model.layers.25.block_sparse_moe.experts.2.w2.weight": "model-00008-of-00008.safetensors",
|
| 574 |
+
"model.layers.25.block_sparse_moe.experts.2.w3.weight": "model-00008-of-00008.safetensors",
|
| 575 |
+
"model.layers.25.block_sparse_moe.experts.3.w1.weight": "model-00008-of-00008.safetensors",
|
| 576 |
+
"model.layers.25.block_sparse_moe.experts.3.w2.weight": "model-00008-of-00008.safetensors",
|
| 577 |
+
"model.layers.25.block_sparse_moe.experts.3.w3.weight": "model-00008-of-00008.safetensors",
|
| 578 |
+
"model.layers.25.block_sparse_moe.experts.4.w1.weight": "model-00008-of-00008.safetensors",
|
| 579 |
+
"model.layers.25.block_sparse_moe.experts.4.w2.weight": "model-00008-of-00008.safetensors",
|
| 580 |
+
"model.layers.25.block_sparse_moe.experts.4.w3.weight": "model-00008-of-00008.safetensors",
|
| 581 |
+
"model.layers.25.block_sparse_moe.experts.5.w1.weight": "model-00008-of-00008.safetensors",
|
| 582 |
+
"model.layers.25.block_sparse_moe.experts.5.w2.weight": "model-00008-of-00008.safetensors",
|
| 583 |
+
"model.layers.25.block_sparse_moe.experts.5.w3.weight": "model-00008-of-00008.safetensors",
|
| 584 |
+
"model.layers.25.block_sparse_moe.experts.6.w1.weight": "model-00008-of-00008.safetensors",
|
| 585 |
+
"model.layers.25.block_sparse_moe.experts.6.w2.weight": "model-00008-of-00008.safetensors",
|
| 586 |
+
"model.layers.25.block_sparse_moe.experts.6.w3.weight": "model-00008-of-00008.safetensors",
|
| 587 |
+
"model.layers.25.block_sparse_moe.experts.7.w1.weight": "model-00008-of-00008.safetensors",
|
| 588 |
+
"model.layers.25.block_sparse_moe.experts.7.w2.weight": "model-00008-of-00008.safetensors",
|
| 589 |
+
"model.layers.25.block_sparse_moe.experts.7.w3.weight": "model-00008-of-00008.safetensors",
|
| 590 |
+
"model.layers.25.block_sparse_moe.gate.weight": "model-00008-of-00008.safetensors",
|
| 591 |
+
"model.layers.25.input_layernorm.weight": "model-00008-of-00008.safetensors",
|
| 592 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00008-of-00008.safetensors",
|
| 593 |
+
"model.layers.25.self_attn.k_proj.weight": "model-00008-of-00008.safetensors",
|
| 594 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00008-of-00008.safetensors",
|
| 595 |
+
"model.layers.25.self_attn.q_proj.weight": "model-00008-of-00008.safetensors",
|
| 596 |
+
"model.layers.25.self_attn.v_proj.weight": "model-00008-of-00008.safetensors",
|
| 597 |
+
"model.layers.3.block_sparse_moe.experts.0.w1.weight": "model-00001-of-00008.safetensors",
|
| 598 |
+
"model.layers.3.block_sparse_moe.experts.0.w2.weight": "model-00001-of-00008.safetensors",
|
| 599 |
+
"model.layers.3.block_sparse_moe.experts.0.w3.weight": "model-00001-of-00008.safetensors",
|
| 600 |
+
"model.layers.3.block_sparse_moe.experts.1.w1.weight": "model-00001-of-00008.safetensors",
|
| 601 |
+
"model.layers.3.block_sparse_moe.experts.1.w2.weight": "model-00001-of-00008.safetensors",
|
| 602 |
+
"model.layers.3.block_sparse_moe.experts.1.w3.weight": "model-00001-of-00008.safetensors",
|
| 603 |
+
"model.layers.3.block_sparse_moe.experts.2.w1.weight": "model-00001-of-00008.safetensors",
|
| 604 |
+
"model.layers.3.block_sparse_moe.experts.2.w2.weight": "model-00001-of-00008.safetensors",
|
| 605 |
+
"model.layers.3.block_sparse_moe.experts.2.w3.weight": "model-00002-of-00008.safetensors",
|
| 606 |
+
"model.layers.3.block_sparse_moe.experts.3.w1.weight": "model-00002-of-00008.safetensors",
|
| 607 |
+
"model.layers.3.block_sparse_moe.experts.3.w2.weight": "model-00002-of-00008.safetensors",
|
| 608 |
+
"model.layers.3.block_sparse_moe.experts.3.w3.weight": "model-00002-of-00008.safetensors",
|
| 609 |
+
"model.layers.3.block_sparse_moe.experts.4.w1.weight": "model-00002-of-00008.safetensors",
|
| 610 |
+
"model.layers.3.block_sparse_moe.experts.4.w2.weight": "model-00002-of-00008.safetensors",
|
| 611 |
+
"model.layers.3.block_sparse_moe.experts.4.w3.weight": "model-00002-of-00008.safetensors",
|
| 612 |
+
"model.layers.3.block_sparse_moe.experts.5.w1.weight": "model-00002-of-00008.safetensors",
|
| 613 |
+
"model.layers.3.block_sparse_moe.experts.5.w2.weight": "model-00002-of-00008.safetensors",
|
| 614 |
+
"model.layers.3.block_sparse_moe.experts.5.w3.weight": "model-00002-of-00008.safetensors",
|
| 615 |
+
"model.layers.3.block_sparse_moe.experts.6.w1.weight": "model-00002-of-00008.safetensors",
|
| 616 |
+
"model.layers.3.block_sparse_moe.experts.6.w2.weight": "model-00002-of-00008.safetensors",
|
| 617 |
+
"model.layers.3.block_sparse_moe.experts.6.w3.weight": "model-00002-of-00008.safetensors",
|
| 618 |
+
"model.layers.3.block_sparse_moe.experts.7.w1.weight": "model-00002-of-00008.safetensors",
|
| 619 |
+
"model.layers.3.block_sparse_moe.experts.7.w2.weight": "model-00002-of-00008.safetensors",
|
| 620 |
+
"model.layers.3.block_sparse_moe.experts.7.w3.weight": "model-00002-of-00008.safetensors",
|
| 621 |
+
"model.layers.3.block_sparse_moe.gate.weight": "model-00001-of-00008.safetensors",
|
| 622 |
+
"model.layers.3.input_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 623 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 624 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
|
| 625 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
|
| 626 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
|
| 627 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
|
| 628 |
+
"model.layers.4.block_sparse_moe.experts.0.w1.weight": "model-00002-of-00008.safetensors",
|
| 629 |
+
"model.layers.4.block_sparse_moe.experts.0.w2.weight": "model-00002-of-00008.safetensors",
|
| 630 |
+
"model.layers.4.block_sparse_moe.experts.0.w3.weight": "model-00002-of-00008.safetensors",
|
| 631 |
+
"model.layers.4.block_sparse_moe.experts.1.w1.weight": "model-00002-of-00008.safetensors",
|
| 632 |
+
"model.layers.4.block_sparse_moe.experts.1.w2.weight": "model-00002-of-00008.safetensors",
|
| 633 |
+
"model.layers.4.block_sparse_moe.experts.1.w3.weight": "model-00002-of-00008.safetensors",
|
| 634 |
+
"model.layers.4.block_sparse_moe.experts.2.w1.weight": "model-00002-of-00008.safetensors",
|
| 635 |
+
"model.layers.4.block_sparse_moe.experts.2.w2.weight": "model-00002-of-00008.safetensors",
|
| 636 |
+
"model.layers.4.block_sparse_moe.experts.2.w3.weight": "model-00002-of-00008.safetensors",
|
| 637 |
+
"model.layers.4.block_sparse_moe.experts.3.w1.weight": "model-00002-of-00008.safetensors",
|
| 638 |
+
"model.layers.4.block_sparse_moe.experts.3.w2.weight": "model-00002-of-00008.safetensors",
|
| 639 |
+
"model.layers.4.block_sparse_moe.experts.3.w3.weight": "model-00002-of-00008.safetensors",
|
| 640 |
+
"model.layers.4.block_sparse_moe.experts.4.w1.weight": "model-00002-of-00008.safetensors",
|
| 641 |
+
"model.layers.4.block_sparse_moe.experts.4.w2.weight": "model-00002-of-00008.safetensors",
|
| 642 |
+
"model.layers.4.block_sparse_moe.experts.4.w3.weight": "model-00002-of-00008.safetensors",
|
| 643 |
+
"model.layers.4.block_sparse_moe.experts.5.w1.weight": "model-00002-of-00008.safetensors",
|
| 644 |
+
"model.layers.4.block_sparse_moe.experts.5.w2.weight": "model-00002-of-00008.safetensors",
|
| 645 |
+
"model.layers.4.block_sparse_moe.experts.5.w3.weight": "model-00002-of-00008.safetensors",
|
| 646 |
+
"model.layers.4.block_sparse_moe.experts.6.w1.weight": "model-00002-of-00008.safetensors",
|
| 647 |
+
"model.layers.4.block_sparse_moe.experts.6.w2.weight": "model-00002-of-00008.safetensors",
|
| 648 |
+
"model.layers.4.block_sparse_moe.experts.6.w3.weight": "model-00002-of-00008.safetensors",
|
| 649 |
+
"model.layers.4.block_sparse_moe.experts.7.w1.weight": "model-00002-of-00008.safetensors",
|
| 650 |
+
"model.layers.4.block_sparse_moe.experts.7.w2.weight": "model-00002-of-00008.safetensors",
|
| 651 |
+
"model.layers.4.block_sparse_moe.experts.7.w3.weight": "model-00002-of-00008.safetensors",
|
| 652 |
+
"model.layers.4.block_sparse_moe.gate.weight": "model-00002-of-00008.safetensors",
|
| 653 |
+
"model.layers.4.input_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 654 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 655 |
+
"model.layers.4.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
|
| 656 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
|
| 657 |
+
"model.layers.4.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
|
| 658 |
+
"model.layers.4.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
|
| 659 |
+
"model.layers.5.block_sparse_moe.experts.0.w1.weight": "model-00002-of-00008.safetensors",
|
| 660 |
+
"model.layers.5.block_sparse_moe.experts.0.w2.weight": "model-00002-of-00008.safetensors",
|
| 661 |
+
"model.layers.5.block_sparse_moe.experts.0.w3.weight": "model-00002-of-00008.safetensors",
|
| 662 |
+
"model.layers.5.block_sparse_moe.experts.1.w1.weight": "model-00002-of-00008.safetensors",
|
| 663 |
+
"model.layers.5.block_sparse_moe.experts.1.w2.weight": "model-00002-of-00008.safetensors",
|
| 664 |
+
"model.layers.5.block_sparse_moe.experts.1.w3.weight": "model-00002-of-00008.safetensors",
|
| 665 |
+
"model.layers.5.block_sparse_moe.experts.2.w1.weight": "model-00002-of-00008.safetensors",
|
| 666 |
+
"model.layers.5.block_sparse_moe.experts.2.w2.weight": "model-00002-of-00008.safetensors",
|
| 667 |
+
"model.layers.5.block_sparse_moe.experts.2.w3.weight": "model-00002-of-00008.safetensors",
|
| 668 |
+
"model.layers.5.block_sparse_moe.experts.3.w1.weight": "model-00002-of-00008.safetensors",
|
| 669 |
+
"model.layers.5.block_sparse_moe.experts.3.w2.weight": "model-00002-of-00008.safetensors",
|
| 670 |
+
"model.layers.5.block_sparse_moe.experts.3.w3.weight": "model-00002-of-00008.safetensors",
|
| 671 |
+
"model.layers.5.block_sparse_moe.experts.4.w1.weight": "model-00002-of-00008.safetensors",
|
| 672 |
+
"model.layers.5.block_sparse_moe.experts.4.w2.weight": "model-00002-of-00008.safetensors",
|
| 673 |
+
"model.layers.5.block_sparse_moe.experts.4.w3.weight": "model-00002-of-00008.safetensors",
|
| 674 |
+
"model.layers.5.block_sparse_moe.experts.5.w1.weight": "model-00002-of-00008.safetensors",
|
| 675 |
+
"model.layers.5.block_sparse_moe.experts.5.w2.weight": "model-00002-of-00008.safetensors",
|
| 676 |
+
"model.layers.5.block_sparse_moe.experts.5.w3.weight": "model-00002-of-00008.safetensors",
|
| 677 |
+
"model.layers.5.block_sparse_moe.experts.6.w1.weight": "model-00002-of-00008.safetensors",
|
| 678 |
+
"model.layers.5.block_sparse_moe.experts.6.w2.weight": "model-00002-of-00008.safetensors",
|
| 679 |
+
"model.layers.5.block_sparse_moe.experts.6.w3.weight": "model-00002-of-00008.safetensors",
|
| 680 |
+
"model.layers.5.block_sparse_moe.experts.7.w1.weight": "model-00002-of-00008.safetensors",
|
| 681 |
+
"model.layers.5.block_sparse_moe.experts.7.w2.weight": "model-00002-of-00008.safetensors",
|
| 682 |
+
"model.layers.5.block_sparse_moe.experts.7.w3.weight": "model-00002-of-00008.safetensors",
|
| 683 |
+
"model.layers.5.block_sparse_moe.gate.weight": "model-00002-of-00008.safetensors",
|
| 684 |
+
"model.layers.5.input_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 685 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
|
| 686 |
+
"model.layers.5.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
|
| 687 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
|
| 688 |
+
"model.layers.5.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
|
| 689 |
+
"model.layers.5.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
|
| 690 |
+
"model.layers.6.block_sparse_moe.experts.0.w1.weight": "model-00002-of-00008.safetensors",
|
| 691 |
+
"model.layers.6.block_sparse_moe.experts.0.w2.weight": "model-00002-of-00008.safetensors",
|
| 692 |
+
"model.layers.6.block_sparse_moe.experts.0.w3.weight": "model-00002-of-00008.safetensors",
|
| 693 |
+
"model.layers.6.block_sparse_moe.experts.1.w1.weight": "model-00002-of-00008.safetensors",
|
| 694 |
+
"model.layers.6.block_sparse_moe.experts.1.w2.weight": "model-00002-of-00008.safetensors",
|
| 695 |
+
"model.layers.6.block_sparse_moe.experts.1.w3.weight": "model-00002-of-00008.safetensors",
|
| 696 |
+
"model.layers.6.block_sparse_moe.experts.2.w1.weight": "model-00002-of-00008.safetensors",
|
| 697 |
+
"model.layers.6.block_sparse_moe.experts.2.w2.weight": "model-00002-of-00008.safetensors",
|
| 698 |
+
"model.layers.6.block_sparse_moe.experts.2.w3.weight": "model-00002-of-00008.safetensors",
|
| 699 |
+
"model.layers.6.block_sparse_moe.experts.3.w1.weight": "model-00002-of-00008.safetensors",
|
| 700 |
+
"model.layers.6.block_sparse_moe.experts.3.w2.weight": "model-00002-of-00008.safetensors",
|
| 701 |
+
"model.layers.6.block_sparse_moe.experts.3.w3.weight": "model-00002-of-00008.safetensors",
|
| 702 |
+
"model.layers.6.block_sparse_moe.experts.4.w1.weight": "model-00002-of-00008.safetensors",
|
| 703 |
+
"model.layers.6.block_sparse_moe.experts.4.w2.weight": "model-00002-of-00008.safetensors",
|
| 704 |
+
"model.layers.6.block_sparse_moe.experts.4.w3.weight": "model-00002-of-00008.safetensors",
|
| 705 |
+
"model.layers.6.block_sparse_moe.experts.5.w1.weight": "model-00002-of-00008.safetensors",
|
| 706 |
+
"model.layers.6.block_sparse_moe.experts.5.w2.weight": "model-00002-of-00008.safetensors",
|
| 707 |
+
"model.layers.6.block_sparse_moe.experts.5.w3.weight": "model-00002-of-00008.safetensors",
|
| 708 |
+
"model.layers.6.block_sparse_moe.experts.6.w1.weight": "model-00002-of-00008.safetensors",
|
| 709 |
+
"model.layers.6.block_sparse_moe.experts.6.w2.weight": "model-00002-of-00008.safetensors",
|
| 710 |
+
"model.layers.6.block_sparse_moe.experts.6.w3.weight": "model-00002-of-00008.safetensors",
|
| 711 |
+
"model.layers.6.block_sparse_moe.experts.7.w1.weight": "model-00003-of-00008.safetensors",
|
| 712 |
+
"model.layers.6.block_sparse_moe.experts.7.w2.weight": "model-00003-of-00008.safetensors",
|
| 713 |
+
"model.layers.6.block_sparse_moe.experts.7.w3.weight": "model-00003-of-00008.safetensors",
|
| 714 |
+
"model.layers.6.block_sparse_moe.gate.weight": "model-00002-of-00008.safetensors",
|
| 715 |
+
"model.layers.6.input_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 716 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 717 |
+
"model.layers.6.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
|
| 718 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
|
| 719 |
+
"model.layers.6.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
|
| 720 |
+
"model.layers.6.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
|
| 721 |
+
"model.layers.7.block_sparse_moe.experts.0.w1.weight": "model-00003-of-00008.safetensors",
|
| 722 |
+
"model.layers.7.block_sparse_moe.experts.0.w2.weight": "model-00003-of-00008.safetensors",
|
| 723 |
+
"model.layers.7.block_sparse_moe.experts.0.w3.weight": "model-00003-of-00008.safetensors",
|
| 724 |
+
"model.layers.7.block_sparse_moe.experts.1.w1.weight": "model-00003-of-00008.safetensors",
|
| 725 |
+
"model.layers.7.block_sparse_moe.experts.1.w2.weight": "model-00003-of-00008.safetensors",
|
| 726 |
+
"model.layers.7.block_sparse_moe.experts.1.w3.weight": "model-00003-of-00008.safetensors",
|
| 727 |
+
"model.layers.7.block_sparse_moe.experts.2.w1.weight": "model-00003-of-00008.safetensors",
|
| 728 |
+
"model.layers.7.block_sparse_moe.experts.2.w2.weight": "model-00003-of-00008.safetensors",
|
| 729 |
+
"model.layers.7.block_sparse_moe.experts.2.w3.weight": "model-00003-of-00008.safetensors",
|
| 730 |
+
"model.layers.7.block_sparse_moe.experts.3.w1.weight": "model-00003-of-00008.safetensors",
|
| 731 |
+
"model.layers.7.block_sparse_moe.experts.3.w2.weight": "model-00003-of-00008.safetensors",
|
| 732 |
+
"model.layers.7.block_sparse_moe.experts.3.w3.weight": "model-00003-of-00008.safetensors",
|
| 733 |
+
"model.layers.7.block_sparse_moe.experts.4.w1.weight": "model-00003-of-00008.safetensors",
|
| 734 |
+
"model.layers.7.block_sparse_moe.experts.4.w2.weight": "model-00003-of-00008.safetensors",
|
| 735 |
+
"model.layers.7.block_sparse_moe.experts.4.w3.weight": "model-00003-of-00008.safetensors",
|
| 736 |
+
"model.layers.7.block_sparse_moe.experts.5.w1.weight": "model-00003-of-00008.safetensors",
|
| 737 |
+
"model.layers.7.block_sparse_moe.experts.5.w2.weight": "model-00003-of-00008.safetensors",
|
| 738 |
+
"model.layers.7.block_sparse_moe.experts.5.w3.weight": "model-00003-of-00008.safetensors",
|
| 739 |
+
"model.layers.7.block_sparse_moe.experts.6.w1.weight": "model-00003-of-00008.safetensors",
|
| 740 |
+
"model.layers.7.block_sparse_moe.experts.6.w2.weight": "model-00003-of-00008.safetensors",
|
| 741 |
+
"model.layers.7.block_sparse_moe.experts.6.w3.weight": "model-00003-of-00008.safetensors",
|
| 742 |
+
"model.layers.7.block_sparse_moe.experts.7.w1.weight": "model-00003-of-00008.safetensors",
|
| 743 |
+
"model.layers.7.block_sparse_moe.experts.7.w2.weight": "model-00003-of-00008.safetensors",
|
| 744 |
+
"model.layers.7.block_sparse_moe.experts.7.w3.weight": "model-00003-of-00008.safetensors",
|
| 745 |
+
"model.layers.7.block_sparse_moe.gate.weight": "model-00003-of-00008.safetensors",
|
| 746 |
+
"model.layers.7.input_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 747 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 748 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
|
| 749 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
|
| 750 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
|
| 751 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
|
| 752 |
+
"model.layers.8.block_sparse_moe.experts.0.w1.weight": "model-00003-of-00008.safetensors",
|
| 753 |
+
"model.layers.8.block_sparse_moe.experts.0.w2.weight": "model-00003-of-00008.safetensors",
|
| 754 |
+
"model.layers.8.block_sparse_moe.experts.0.w3.weight": "model-00003-of-00008.safetensors",
|
| 755 |
+
"model.layers.8.block_sparse_moe.experts.1.w1.weight": "model-00003-of-00008.safetensors",
|
| 756 |
+
"model.layers.8.block_sparse_moe.experts.1.w2.weight": "model-00003-of-00008.safetensors",
|
| 757 |
+
"model.layers.8.block_sparse_moe.experts.1.w3.weight": "model-00003-of-00008.safetensors",
|
| 758 |
+
"model.layers.8.block_sparse_moe.experts.2.w1.weight": "model-00003-of-00008.safetensors",
|
| 759 |
+
"model.layers.8.block_sparse_moe.experts.2.w2.weight": "model-00003-of-00008.safetensors",
|
| 760 |
+
"model.layers.8.block_sparse_moe.experts.2.w3.weight": "model-00003-of-00008.safetensors",
|
| 761 |
+
"model.layers.8.block_sparse_moe.experts.3.w1.weight": "model-00003-of-00008.safetensors",
|
| 762 |
+
"model.layers.8.block_sparse_moe.experts.3.w2.weight": "model-00003-of-00008.safetensors",
|
| 763 |
+
"model.layers.8.block_sparse_moe.experts.3.w3.weight": "model-00003-of-00008.safetensors",
|
| 764 |
+
"model.layers.8.block_sparse_moe.experts.4.w1.weight": "model-00003-of-00008.safetensors",
|
| 765 |
+
"model.layers.8.block_sparse_moe.experts.4.w2.weight": "model-00003-of-00008.safetensors",
|
| 766 |
+
"model.layers.8.block_sparse_moe.experts.4.w3.weight": "model-00003-of-00008.safetensors",
|
| 767 |
+
"model.layers.8.block_sparse_moe.experts.5.w1.weight": "model-00003-of-00008.safetensors",
|
| 768 |
+
"model.layers.8.block_sparse_moe.experts.5.w2.weight": "model-00003-of-00008.safetensors",
|
| 769 |
+
"model.layers.8.block_sparse_moe.experts.5.w3.weight": "model-00003-of-00008.safetensors",
|
| 770 |
+
"model.layers.8.block_sparse_moe.experts.6.w1.weight": "model-00003-of-00008.safetensors",
|
| 771 |
+
"model.layers.8.block_sparse_moe.experts.6.w2.weight": "model-00003-of-00008.safetensors",
|
| 772 |
+
"model.layers.8.block_sparse_moe.experts.6.w3.weight": "model-00003-of-00008.safetensors",
|
| 773 |
+
"model.layers.8.block_sparse_moe.experts.7.w1.weight": "model-00003-of-00008.safetensors",
|
| 774 |
+
"model.layers.8.block_sparse_moe.experts.7.w2.weight": "model-00003-of-00008.safetensors",
|
| 775 |
+
"model.layers.8.block_sparse_moe.experts.7.w3.weight": "model-00003-of-00008.safetensors",
|
| 776 |
+
"model.layers.8.block_sparse_moe.gate.weight": "model-00003-of-00008.safetensors",
|
| 777 |
+
"model.layers.8.input_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 778 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 779 |
+
"model.layers.8.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
|
| 780 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
|
| 781 |
+
"model.layers.8.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
|
| 782 |
+
"model.layers.8.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
|
| 783 |
+
"model.layers.9.block_sparse_moe.experts.0.w1.weight": "model-00003-of-00008.safetensors",
|
| 784 |
+
"model.layers.9.block_sparse_moe.experts.0.w2.weight": "model-00003-of-00008.safetensors",
|
| 785 |
+
"model.layers.9.block_sparse_moe.experts.0.w3.weight": "model-00003-of-00008.safetensors",
|
| 786 |
+
"model.layers.9.block_sparse_moe.experts.1.w1.weight": "model-00003-of-00008.safetensors",
|
| 787 |
+
"model.layers.9.block_sparse_moe.experts.1.w2.weight": "model-00003-of-00008.safetensors",
|
| 788 |
+
"model.layers.9.block_sparse_moe.experts.1.w3.weight": "model-00003-of-00008.safetensors",
|
| 789 |
+
"model.layers.9.block_sparse_moe.experts.2.w1.weight": "model-00003-of-00008.safetensors",
|
| 790 |
+
"model.layers.9.block_sparse_moe.experts.2.w2.weight": "model-00003-of-00008.safetensors",
|
| 791 |
+
"model.layers.9.block_sparse_moe.experts.2.w3.weight": "model-00003-of-00008.safetensors",
|
| 792 |
+
"model.layers.9.block_sparse_moe.experts.3.w1.weight": "model-00003-of-00008.safetensors",
|
| 793 |
+
"model.layers.9.block_sparse_moe.experts.3.w2.weight": "model-00003-of-00008.safetensors",
|
| 794 |
+
"model.layers.9.block_sparse_moe.experts.3.w3.weight": "model-00003-of-00008.safetensors",
|
| 795 |
+
"model.layers.9.block_sparse_moe.experts.4.w1.weight": "model-00003-of-00008.safetensors",
|
| 796 |
+
"model.layers.9.block_sparse_moe.experts.4.w2.weight": "model-00003-of-00008.safetensors",
|
| 797 |
+
"model.layers.9.block_sparse_moe.experts.4.w3.weight": "model-00003-of-00008.safetensors",
|
| 798 |
+
"model.layers.9.block_sparse_moe.experts.5.w1.weight": "model-00003-of-00008.safetensors",
|
| 799 |
+
"model.layers.9.block_sparse_moe.experts.5.w2.weight": "model-00003-of-00008.safetensors",
|
| 800 |
+
"model.layers.9.block_sparse_moe.experts.5.w3.weight": "model-00003-of-00008.safetensors",
|
| 801 |
+
"model.layers.9.block_sparse_moe.experts.6.w1.weight": "model-00003-of-00008.safetensors",
|
| 802 |
+
"model.layers.9.block_sparse_moe.experts.6.w2.weight": "model-00003-of-00008.safetensors",
|
| 803 |
+
"model.layers.9.block_sparse_moe.experts.6.w3.weight": "model-00003-of-00008.safetensors",
|
| 804 |
+
"model.layers.9.block_sparse_moe.experts.7.w1.weight": "model-00003-of-00008.safetensors",
|
| 805 |
+
"model.layers.9.block_sparse_moe.experts.7.w2.weight": "model-00003-of-00008.safetensors",
|
| 806 |
+
"model.layers.9.block_sparse_moe.experts.7.w3.weight": "model-00003-of-00008.safetensors",
|
| 807 |
+
"model.layers.9.block_sparse_moe.gate.weight": "model-00003-of-00008.safetensors",
|
| 808 |
+
"model.layers.9.input_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 809 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
|
| 810 |
+
"model.layers.9.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
|
| 811 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
|
| 812 |
+
"model.layers.9.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
|
| 813 |
+
"model.layers.9.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
|
| 814 |
+
"model.norm.weight": "model-00008-of-00008.safetensors"
|
| 815 |
+
}
|
| 816 |
+
}
|
modeling_scidfm.py
ADDED
|
@@ -0,0 +1,1591 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 SciDFM 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 |
+
""" PyTorch SciDFM model."""
|
| 21 |
+
import inspect
|
| 22 |
+
import math
|
| 23 |
+
import warnings
|
| 24 |
+
from typing import List, Optional, Tuple, Union
|
| 25 |
+
|
| 26 |
+
import torch
|
| 27 |
+
import torch.nn.functional as F
|
| 28 |
+
import torch.utils.checkpoint
|
| 29 |
+
from torch import nn
|
| 30 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 31 |
+
|
| 32 |
+
from transformers.activations import ACT2FN
|
| 33 |
+
from transformers.cache_utils import Cache, DynamicCache
|
| 34 |
+
from transformers.modeling_attn_mask_utils import (
|
| 35 |
+
_prepare_4d_causal_attention_mask,
|
| 36 |
+
_prepare_4d_causal_attention_mask_for_sdpa,
|
| 37 |
+
)
|
| 38 |
+
from transformers.modeling_outputs import (
|
| 39 |
+
MoeCausalLMOutputWithPast,
|
| 40 |
+
MoeModelOutputWithPast,
|
| 41 |
+
SequenceClassifierOutputWithPast,
|
| 42 |
+
)
|
| 43 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 44 |
+
from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
|
| 45 |
+
from transformers.utils import (
|
| 46 |
+
add_start_docstrings,
|
| 47 |
+
add_start_docstrings_to_model_forward,
|
| 48 |
+
is_flash_attn_2_available,
|
| 49 |
+
is_flash_attn_greater_or_equal_2_10,
|
| 50 |
+
logging,
|
| 51 |
+
replace_return_docstrings,
|
| 52 |
+
)
|
| 53 |
+
from transformers.utils.import_utils import is_torch_fx_available
|
| 54 |
+
from .configuration_scidfm import SciDFMConfig
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
if is_flash_attn_2_available():
|
| 58 |
+
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
| 59 |
+
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
| 60 |
+
|
| 61 |
+
_flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
|
| 62 |
+
|
| 63 |
+
# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
|
| 64 |
+
# It means that the function will not be traced through and simply appear as a node in the graph.
|
| 65 |
+
if is_torch_fx_available():
|
| 66 |
+
if not is_torch_greater_or_equal_than_1_13:
|
| 67 |
+
import torch.fx
|
| 68 |
+
|
| 69 |
+
_prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
logger = logging.get_logger(__name__)
|
| 73 |
+
|
| 74 |
+
_CONFIG_FOR_DOC = "SciDFMConfig"
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def load_balancing_loss_func(
|
| 78 |
+
gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
|
| 79 |
+
) -> float:
|
| 80 |
+
r"""
|
| 81 |
+
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
|
| 82 |
+
|
| 83 |
+
See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
|
| 84 |
+
function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
|
| 85 |
+
experts is too unbalanced.
|
| 86 |
+
|
| 87 |
+
Args:
|
| 88 |
+
gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
|
| 89 |
+
Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
|
| 90 |
+
shape [batch_size X sequence_length, num_experts].
|
| 91 |
+
attention_mask (`torch.Tensor`, None):
|
| 92 |
+
The attention_mask used in forward function
|
| 93 |
+
shape [batch_size X sequence_length] if not None.
|
| 94 |
+
num_experts (`int`, *optional*):
|
| 95 |
+
Number of experts
|
| 96 |
+
|
| 97 |
+
Returns:
|
| 98 |
+
The auxiliary loss.
|
| 99 |
+
"""
|
| 100 |
+
if gate_logits is None or not isinstance(gate_logits, tuple):
|
| 101 |
+
return 0
|
| 102 |
+
|
| 103 |
+
if isinstance(gate_logits, tuple):
|
| 104 |
+
compute_device = gate_logits[0].device
|
| 105 |
+
concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
|
| 106 |
+
|
| 107 |
+
routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
|
| 108 |
+
|
| 109 |
+
_, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
|
| 110 |
+
|
| 111 |
+
expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
|
| 112 |
+
|
| 113 |
+
if attention_mask is None:
|
| 114 |
+
# Compute the percentage of tokens routed to each experts
|
| 115 |
+
tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
|
| 116 |
+
|
| 117 |
+
# Compute the average probability of routing to these experts
|
| 118 |
+
router_prob_per_expert = torch.mean(routing_weights, dim=0)
|
| 119 |
+
else:
|
| 120 |
+
batch_size, sequence_length = attention_mask.shape
|
| 121 |
+
num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
|
| 122 |
+
|
| 123 |
+
# Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
|
| 124 |
+
expert_attention_mask = (
|
| 125 |
+
attention_mask[None, :, :, None, None]
|
| 126 |
+
.expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
|
| 127 |
+
.reshape(-1, top_k, num_experts)
|
| 128 |
+
.to(compute_device)
|
| 129 |
+
)
|
| 130 |
+
|
| 131 |
+
# Compute the percentage of tokens routed to each experts
|
| 132 |
+
tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
|
| 133 |
+
expert_attention_mask, dim=0
|
| 134 |
+
)
|
| 135 |
+
|
| 136 |
+
# Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
|
| 137 |
+
router_per_expert_attention_mask = (
|
| 138 |
+
attention_mask[None, :, :, None]
|
| 139 |
+
.expand((num_hidden_layers, batch_size, sequence_length, num_experts))
|
| 140 |
+
.reshape(-1, num_experts)
|
| 141 |
+
.to(compute_device)
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
# Compute the average probability of routing to these experts
|
| 145 |
+
router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
|
| 146 |
+
router_per_expert_attention_mask, dim=0
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
|
| 150 |
+
return overall_loss * num_experts
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
# Copied from transformers.models.llama.modeling_llama._get_unpad_data
|
| 154 |
+
def _get_unpad_data(attention_mask):
|
| 155 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
| 156 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
| 157 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
| 158 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
|
| 159 |
+
return (
|
| 160 |
+
indices,
|
| 161 |
+
cu_seqlens,
|
| 162 |
+
max_seqlen_in_batch,
|
| 163 |
+
)
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->SciDFM
|
| 167 |
+
class SciDFMRMSNorm(nn.Module):
|
| 168 |
+
def __init__(self, hidden_size, eps=1e-6):
|
| 169 |
+
"""
|
| 170 |
+
SciDFMRMSNorm is equivalent to T5LayerNorm
|
| 171 |
+
"""
|
| 172 |
+
super().__init__()
|
| 173 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 174 |
+
self.variance_epsilon = eps
|
| 175 |
+
|
| 176 |
+
def forward(self, hidden_states):
|
| 177 |
+
input_dtype = hidden_states.dtype
|
| 178 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 179 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 180 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 181 |
+
return self.weight * hidden_states.to(input_dtype)
|
| 182 |
+
|
| 183 |
+
|
| 184 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->SciDFM
|
| 185 |
+
class SciDFMRotaryEmbedding(nn.Module):
|
| 186 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 187 |
+
super().__init__()
|
| 188 |
+
|
| 189 |
+
self.dim = dim
|
| 190 |
+
self.max_position_embeddings = max_position_embeddings
|
| 191 |
+
self.base = base
|
| 192 |
+
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
|
| 193 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 194 |
+
|
| 195 |
+
# Build here to make `torch.jit.trace` work.
|
| 196 |
+
self._set_cos_sin_cache(
|
| 197 |
+
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
|
| 198 |
+
)
|
| 199 |
+
|
| 200 |
+
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
| 201 |
+
self.max_seq_len_cached = seq_len
|
| 202 |
+
t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
|
| 203 |
+
|
| 204 |
+
freqs = torch.outer(t, self.inv_freq)
|
| 205 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 206 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 207 |
+
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
|
| 208 |
+
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
|
| 209 |
+
|
| 210 |
+
def forward(self, x, seq_len=None):
|
| 211 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
| 212 |
+
if seq_len > self.max_seq_len_cached:
|
| 213 |
+
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
| 214 |
+
|
| 215 |
+
return (
|
| 216 |
+
self.cos_cached[:seq_len].to(dtype=x.dtype),
|
| 217 |
+
self.sin_cached[:seq_len].to(dtype=x.dtype),
|
| 218 |
+
)
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
# Copied from transformers.models.llama.modeling_llama.rotate_half
|
| 222 |
+
def rotate_half(x):
|
| 223 |
+
"""Rotates half the hidden dims of the input."""
|
| 224 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 225 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 226 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 227 |
+
|
| 228 |
+
|
| 229 |
+
# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
|
| 230 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
|
| 231 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
| 232 |
+
|
| 233 |
+
Args:
|
| 234 |
+
q (`torch.Tensor`): The query tensor.
|
| 235 |
+
k (`torch.Tensor`): The key tensor.
|
| 236 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
| 237 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
| 238 |
+
position_ids (`torch.Tensor`):
|
| 239 |
+
The position indices of the tokens corresponding to the query and key tensors. For example, this can be
|
| 240 |
+
used to pass offsetted position ids when working with a KV-cache.
|
| 241 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
| 242 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
| 243 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
| 244 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
| 245 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
| 246 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
| 247 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
| 248 |
+
Returns:
|
| 249 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
| 250 |
+
"""
|
| 251 |
+
cos = cos[position_ids].unsqueeze(unsqueeze_dim)
|
| 252 |
+
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
|
| 253 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 254 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 255 |
+
return q_embed, k_embed
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
# Copied from transformers.models.llama.modeling_llama.repeat_kv
|
| 259 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 260 |
+
"""
|
| 261 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 262 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 263 |
+
"""
|
| 264 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 265 |
+
if n_rep == 1:
|
| 266 |
+
return hidden_states
|
| 267 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 268 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 269 |
+
|
| 270 |
+
|
| 271 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->SciDFM
|
| 272 |
+
class SciDFMAttention(nn.Module):
|
| 273 |
+
"""
|
| 274 |
+
Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
|
| 275 |
+
and "Generating Long Sequences with Sparse Transformers".
|
| 276 |
+
"""
|
| 277 |
+
|
| 278 |
+
def __init__(self, config: SciDFMConfig, layer_idx: Optional[int] = None):
|
| 279 |
+
super().__init__()
|
| 280 |
+
self.config = config
|
| 281 |
+
self.layer_idx = layer_idx
|
| 282 |
+
if layer_idx is None:
|
| 283 |
+
logger.warning_once(
|
| 284 |
+
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
| 285 |
+
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
| 286 |
+
"when creating this class."
|
| 287 |
+
)
|
| 288 |
+
|
| 289 |
+
self.hidden_size = config.hidden_size
|
| 290 |
+
self.num_heads = config.num_attention_heads
|
| 291 |
+
self.head_dim = self.hidden_size // self.num_heads
|
| 292 |
+
self.num_key_value_heads = config.num_key_value_heads
|
| 293 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
| 294 |
+
self.max_position_embeddings = config.max_position_embeddings
|
| 295 |
+
self.rope_theta = config.rope_theta
|
| 296 |
+
self.is_causal = True
|
| 297 |
+
self.attention_dropout = config.attention_dropout
|
| 298 |
+
|
| 299 |
+
if (self.head_dim * self.num_heads) != self.hidden_size:
|
| 300 |
+
raise ValueError(
|
| 301 |
+
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
| 302 |
+
f" and `num_heads`: {self.num_heads})."
|
| 303 |
+
)
|
| 304 |
+
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
| 305 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
| 306 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
| 307 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
| 308 |
+
|
| 309 |
+
self.rotary_emb = SciDFMRotaryEmbedding(
|
| 310 |
+
self.head_dim,
|
| 311 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 312 |
+
base=self.rope_theta,
|
| 313 |
+
)
|
| 314 |
+
|
| 315 |
+
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
| 316 |
+
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
| 317 |
+
|
| 318 |
+
def forward(
|
| 319 |
+
self,
|
| 320 |
+
hidden_states: torch.Tensor,
|
| 321 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 322 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 323 |
+
past_key_value: Optional[Cache] = None,
|
| 324 |
+
output_attentions: bool = False,
|
| 325 |
+
use_cache: bool = False,
|
| 326 |
+
**kwargs,
|
| 327 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 328 |
+
if "padding_mask" in kwargs:
|
| 329 |
+
warnings.warn(
|
| 330 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 331 |
+
)
|
| 332 |
+
bsz, q_len, _ = hidden_states.size()
|
| 333 |
+
|
| 334 |
+
query_states = self.q_proj(hidden_states)
|
| 335 |
+
key_states = self.k_proj(hidden_states)
|
| 336 |
+
value_states = self.v_proj(hidden_states)
|
| 337 |
+
|
| 338 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 339 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 340 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 341 |
+
|
| 342 |
+
kv_seq_len = key_states.shape[-2]
|
| 343 |
+
if past_key_value is not None:
|
| 344 |
+
if self.layer_idx is None:
|
| 345 |
+
raise ValueError(
|
| 346 |
+
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 347 |
+
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 348 |
+
"with a layer index."
|
| 349 |
+
)
|
| 350 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 351 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
| 352 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 353 |
+
|
| 354 |
+
if past_key_value is not None:
|
| 355 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 356 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 357 |
+
|
| 358 |
+
# repeat k/v heads if n_kv_heads < n_heads
|
| 359 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 360 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 361 |
+
|
| 362 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
| 363 |
+
|
| 364 |
+
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
| 365 |
+
raise ValueError(
|
| 366 |
+
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
|
| 367 |
+
f" {attn_weights.size()}"
|
| 368 |
+
)
|
| 369 |
+
|
| 370 |
+
if attention_mask is not None:
|
| 371 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
| 372 |
+
raise ValueError(
|
| 373 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 374 |
+
)
|
| 375 |
+
|
| 376 |
+
attn_weights = attn_weights + attention_mask
|
| 377 |
+
|
| 378 |
+
# upcast attention to fp32
|
| 379 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
| 380 |
+
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
| 381 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 382 |
+
|
| 383 |
+
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
| 384 |
+
raise ValueError(
|
| 385 |
+
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
| 386 |
+
f" {attn_output.size()}"
|
| 387 |
+
)
|
| 388 |
+
|
| 389 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 390 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
| 391 |
+
|
| 392 |
+
attn_output = self.o_proj(attn_output)
|
| 393 |
+
|
| 394 |
+
if not output_attentions:
|
| 395 |
+
attn_weights = None
|
| 396 |
+
|
| 397 |
+
return attn_output, attn_weights, past_key_value
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->SciDFM
|
| 401 |
+
class SciDFMFlashAttention2(SciDFMAttention):
|
| 402 |
+
"""
|
| 403 |
+
SciDFM flash attention module. This module inherits from `SciDFMAttention` as the weights of the module stays
|
| 404 |
+
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
| 405 |
+
flash attention and deal with padding tokens in case the input contains any of them.
|
| 406 |
+
"""
|
| 407 |
+
|
| 408 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
|
| 409 |
+
def __init__(self, *args, **kwargs):
|
| 410 |
+
super().__init__(*args, **kwargs)
|
| 411 |
+
|
| 412 |
+
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
| 413 |
+
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
| 414 |
+
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
| 415 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
| 416 |
+
|
| 417 |
+
def forward(
|
| 418 |
+
self,
|
| 419 |
+
hidden_states: torch.Tensor,
|
| 420 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 421 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 422 |
+
past_key_value: Optional[Cache] = None,
|
| 423 |
+
output_attentions: bool = False,
|
| 424 |
+
use_cache: bool = False,
|
| 425 |
+
**kwargs,
|
| 426 |
+
):
|
| 427 |
+
if "padding_mask" in kwargs:
|
| 428 |
+
warnings.warn(
|
| 429 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# overwrite attention_mask with padding_mask
|
| 433 |
+
attention_mask = kwargs.pop("padding_mask")
|
| 434 |
+
bsz, q_len, _ = hidden_states.size()
|
| 435 |
+
|
| 436 |
+
query_states = self.q_proj(hidden_states)
|
| 437 |
+
key_states = self.k_proj(hidden_states)
|
| 438 |
+
value_states = self.v_proj(hidden_states)
|
| 439 |
+
|
| 440 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 441 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 442 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 443 |
+
|
| 444 |
+
kv_seq_len = key_states.shape[-2]
|
| 445 |
+
if past_key_value is not None:
|
| 446 |
+
if self.layer_idx is None:
|
| 447 |
+
raise ValueError(
|
| 448 |
+
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 449 |
+
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 450 |
+
"with a layer index."
|
| 451 |
+
)
|
| 452 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 453 |
+
|
| 454 |
+
# Because the input can be padded, the absolute sequence length depends on the max position id.
|
| 455 |
+
rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
|
| 456 |
+
cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
|
| 457 |
+
|
| 458 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 459 |
+
|
| 460 |
+
use_sliding_windows = (
|
| 461 |
+
_flash_supports_window_size
|
| 462 |
+
and getattr(self.config, "sliding_window", None) is not None
|
| 463 |
+
and kv_seq_len > self.config.sliding_window
|
| 464 |
+
)
|
| 465 |
+
|
| 466 |
+
# if not _flash_supports_window_size:
|
| 467 |
+
# logger.warning_once(
|
| 468 |
+
# "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
|
| 469 |
+
# " make sure to upgrade flash-attn library."
|
| 470 |
+
# )
|
| 471 |
+
|
| 472 |
+
if past_key_value is not None:
|
| 473 |
+
# Activate slicing cache only if the config has a value `sliding_windows` attribute
|
| 474 |
+
cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
|
| 475 |
+
if (
|
| 476 |
+
getattr(self.config, "sliding_window", None) is not None
|
| 477 |
+
and kv_seq_len > self.config.sliding_window
|
| 478 |
+
and cache_has_contents
|
| 479 |
+
):
|
| 480 |
+
slicing_tokens = 1 - self.config.sliding_window
|
| 481 |
+
|
| 482 |
+
past_key = past_key_value[self.layer_idx][0]
|
| 483 |
+
past_value = past_key_value[self.layer_idx][1]
|
| 484 |
+
|
| 485 |
+
past_key = past_key[:, :, slicing_tokens:, :].contiguous()
|
| 486 |
+
past_value = past_value[:, :, slicing_tokens:, :].contiguous()
|
| 487 |
+
|
| 488 |
+
if past_key.shape[-2] != self.config.sliding_window - 1:
|
| 489 |
+
raise ValueError(
|
| 490 |
+
f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
|
| 491 |
+
f" {past_key.shape}"
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
if attention_mask is not None:
|
| 495 |
+
attention_mask = attention_mask[:, slicing_tokens:]
|
| 496 |
+
attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
|
| 497 |
+
|
| 498 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 499 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 500 |
+
|
| 501 |
+
# repeat k/v heads if n_kv_heads < n_heads
|
| 502 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 503 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 504 |
+
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
| 505 |
+
|
| 506 |
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
| 507 |
+
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
| 508 |
+
# cast them back in float16 just to be sure everything works as expected.
|
| 509 |
+
input_dtype = query_states.dtype
|
| 510 |
+
if input_dtype == torch.float32:
|
| 511 |
+
if torch.is_autocast_enabled():
|
| 512 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
| 513 |
+
# Handle the case where the model is quantized
|
| 514 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
| 515 |
+
target_dtype = self.config._pre_quantization_dtype
|
| 516 |
+
else:
|
| 517 |
+
target_dtype = self.q_proj.weight.dtype
|
| 518 |
+
|
| 519 |
+
logger.warning_once(
|
| 520 |
+
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
| 521 |
+
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
| 522 |
+
f" {target_dtype}."
|
| 523 |
+
)
|
| 524 |
+
|
| 525 |
+
query_states = query_states.to(target_dtype)
|
| 526 |
+
key_states = key_states.to(target_dtype)
|
| 527 |
+
value_states = value_states.to(target_dtype)
|
| 528 |
+
|
| 529 |
+
# Reashape to the expected shape for Flash Attention
|
| 530 |
+
query_states = query_states.transpose(1, 2)
|
| 531 |
+
key_states = key_states.transpose(1, 2)
|
| 532 |
+
value_states = value_states.transpose(1, 2)
|
| 533 |
+
|
| 534 |
+
attn_output = self._flash_attention_forward(
|
| 535 |
+
query_states,
|
| 536 |
+
key_states,
|
| 537 |
+
value_states,
|
| 538 |
+
attention_mask,
|
| 539 |
+
q_len,
|
| 540 |
+
dropout=dropout_rate,
|
| 541 |
+
use_sliding_windows=use_sliding_windows,
|
| 542 |
+
)
|
| 543 |
+
|
| 544 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
| 545 |
+
attn_output = self.o_proj(attn_output)
|
| 546 |
+
|
| 547 |
+
if not output_attentions:
|
| 548 |
+
attn_weights = None
|
| 549 |
+
|
| 550 |
+
return attn_output, attn_weights, past_key_value
|
| 551 |
+
|
| 552 |
+
def _flash_attention_forward(
|
| 553 |
+
self,
|
| 554 |
+
query_states,
|
| 555 |
+
key_states,
|
| 556 |
+
value_states,
|
| 557 |
+
attention_mask,
|
| 558 |
+
query_length,
|
| 559 |
+
dropout=0.0,
|
| 560 |
+
softmax_scale=None,
|
| 561 |
+
use_sliding_windows=False,
|
| 562 |
+
):
|
| 563 |
+
"""
|
| 564 |
+
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
|
| 565 |
+
first unpad the input, then computes the attention scores and pad the final attention scores.
|
| 566 |
+
|
| 567 |
+
Args:
|
| 568 |
+
query_states (`torch.Tensor`):
|
| 569 |
+
Input query states to be passed to Flash Attention API
|
| 570 |
+
key_states (`torch.Tensor`):
|
| 571 |
+
Input key states to be passed to Flash Attention API
|
| 572 |
+
value_states (`torch.Tensor`):
|
| 573 |
+
Input value states to be passed to Flash Attention API
|
| 574 |
+
attention_mask (`torch.Tensor`):
|
| 575 |
+
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
| 576 |
+
position of padding tokens and 1 for the position of non-padding tokens.
|
| 577 |
+
dropout (`float`):
|
| 578 |
+
Attention dropout
|
| 579 |
+
softmax_scale (`float`, *optional*):
|
| 580 |
+
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
| 581 |
+
use_sliding_windows (`bool`, *optional*):
|
| 582 |
+
Whether to activate sliding window attention.
|
| 583 |
+
"""
|
| 584 |
+
if not self._flash_attn_uses_top_left_mask:
|
| 585 |
+
causal = self.is_causal
|
| 586 |
+
else:
|
| 587 |
+
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
|
| 588 |
+
causal = self.is_causal and query_length != 1
|
| 589 |
+
|
| 590 |
+
# Contains at least one padding token in the sequence
|
| 591 |
+
if attention_mask is not None:
|
| 592 |
+
batch_size = query_states.shape[0]
|
| 593 |
+
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
| 594 |
+
query_states, key_states, value_states, attention_mask, query_length
|
| 595 |
+
)
|
| 596 |
+
|
| 597 |
+
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
| 598 |
+
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
| 599 |
+
|
| 600 |
+
if not use_sliding_windows:
|
| 601 |
+
attn_output_unpad = flash_attn_varlen_func(
|
| 602 |
+
query_states,
|
| 603 |
+
key_states,
|
| 604 |
+
value_states,
|
| 605 |
+
cu_seqlens_q=cu_seqlens_q,
|
| 606 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 607 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
| 608 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
| 609 |
+
dropout_p=dropout,
|
| 610 |
+
softmax_scale=softmax_scale,
|
| 611 |
+
causal=causal,
|
| 612 |
+
)
|
| 613 |
+
else:
|
| 614 |
+
attn_output_unpad = flash_attn_varlen_func(
|
| 615 |
+
query_states,
|
| 616 |
+
key_states,
|
| 617 |
+
value_states,
|
| 618 |
+
cu_seqlens_q=cu_seqlens_q,
|
| 619 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 620 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
| 621 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
| 622 |
+
dropout_p=dropout,
|
| 623 |
+
softmax_scale=softmax_scale,
|
| 624 |
+
causal=causal,
|
| 625 |
+
window_size=(self.config.sliding_window, self.config.sliding_window),
|
| 626 |
+
)
|
| 627 |
+
|
| 628 |
+
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
| 629 |
+
else:
|
| 630 |
+
if not use_sliding_windows:
|
| 631 |
+
attn_output = flash_attn_func(
|
| 632 |
+
query_states,
|
| 633 |
+
key_states,
|
| 634 |
+
value_states,
|
| 635 |
+
dropout,
|
| 636 |
+
softmax_scale=softmax_scale,
|
| 637 |
+
causal=causal,
|
| 638 |
+
)
|
| 639 |
+
else:
|
| 640 |
+
attn_output = flash_attn_func(
|
| 641 |
+
query_states,
|
| 642 |
+
key_states,
|
| 643 |
+
value_states,
|
| 644 |
+
dropout,
|
| 645 |
+
softmax_scale=softmax_scale,
|
| 646 |
+
causal=causal,
|
| 647 |
+
window_size=(self.config.sliding_window, self.config.sliding_window),
|
| 648 |
+
)
|
| 649 |
+
|
| 650 |
+
return attn_output
|
| 651 |
+
|
| 652 |
+
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
| 653 |
+
batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
|
| 654 |
+
|
| 655 |
+
# On the first iteration we need to properly re-create the padding mask
|
| 656 |
+
# by slicing it on the proper place
|
| 657 |
+
if kv_seq_len != attention_mask.shape[-1]:
|
| 658 |
+
attention_mask_num_tokens = attention_mask.shape[-1]
|
| 659 |
+
attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
|
| 660 |
+
|
| 661 |
+
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
| 662 |
+
|
| 663 |
+
key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
|
| 664 |
+
value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
|
| 665 |
+
|
| 666 |
+
if query_length == kv_seq_len:
|
| 667 |
+
query_layer = index_first_axis(
|
| 668 |
+
query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
|
| 669 |
+
)
|
| 670 |
+
cu_seqlens_q = cu_seqlens_k
|
| 671 |
+
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
| 672 |
+
indices_q = indices_k
|
| 673 |
+
elif query_length == 1:
|
| 674 |
+
max_seqlen_in_batch_q = 1
|
| 675 |
+
cu_seqlens_q = torch.arange(
|
| 676 |
+
batch_size + 1, dtype=torch.int32, device=query_layer.device
|
| 677 |
+
) # There is a memcpy here, that is very bad.
|
| 678 |
+
indices_q = cu_seqlens_q[:-1]
|
| 679 |
+
query_layer = query_layer.squeeze(1)
|
| 680 |
+
else:
|
| 681 |
+
# The -q_len: slice assumes left padding.
|
| 682 |
+
attention_mask = attention_mask[:, -query_length:]
|
| 683 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
|
| 684 |
+
|
| 685 |
+
return (
|
| 686 |
+
query_layer,
|
| 687 |
+
key_layer,
|
| 688 |
+
value_layer,
|
| 689 |
+
indices_q,
|
| 690 |
+
(cu_seqlens_q, cu_seqlens_k),
|
| 691 |
+
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
| 692 |
+
)
|
| 693 |
+
|
| 694 |
+
|
| 695 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->SciDFM
|
| 696 |
+
class SciDFMSdpaAttention(SciDFMAttention):
|
| 697 |
+
"""
|
| 698 |
+
SciDFM attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
| 699 |
+
`SciDFMAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
| 700 |
+
SDPA API.
|
| 701 |
+
"""
|
| 702 |
+
|
| 703 |
+
# Adapted from SciDFMAttention.forward
|
| 704 |
+
def forward(
|
| 705 |
+
self,
|
| 706 |
+
hidden_states: torch.Tensor,
|
| 707 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 708 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 709 |
+
past_key_value: Optional[Cache] = None,
|
| 710 |
+
output_attentions: bool = False,
|
| 711 |
+
use_cache: bool = False,
|
| 712 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 713 |
+
if output_attentions:
|
| 714 |
+
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
| 715 |
+
logger.warning_once(
|
| 716 |
+
"SciDFMModel is using SciDFMSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
| 717 |
+
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 718 |
+
)
|
| 719 |
+
return super().forward(
|
| 720 |
+
hidden_states=hidden_states,
|
| 721 |
+
attention_mask=attention_mask,
|
| 722 |
+
position_ids=position_ids,
|
| 723 |
+
past_key_value=past_key_value,
|
| 724 |
+
output_attentions=output_attentions,
|
| 725 |
+
use_cache=use_cache,
|
| 726 |
+
)
|
| 727 |
+
|
| 728 |
+
bsz, q_len, _ = hidden_states.size()
|
| 729 |
+
|
| 730 |
+
query_states = self.q_proj(hidden_states)
|
| 731 |
+
key_states = self.k_proj(hidden_states)
|
| 732 |
+
value_states = self.v_proj(hidden_states)
|
| 733 |
+
|
| 734 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 735 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 736 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 737 |
+
|
| 738 |
+
kv_seq_len = key_states.shape[-2]
|
| 739 |
+
if past_key_value is not None:
|
| 740 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 741 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
| 742 |
+
|
| 743 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 744 |
+
|
| 745 |
+
if past_key_value is not None:
|
| 746 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 747 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 748 |
+
|
| 749 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 750 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 751 |
+
|
| 752 |
+
if attention_mask is not None:
|
| 753 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
| 754 |
+
raise ValueError(
|
| 755 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 756 |
+
)
|
| 757 |
+
|
| 758 |
+
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 759 |
+
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
| 760 |
+
if query_states.device.type == "cuda" and attention_mask is not None:
|
| 761 |
+
query_states = query_states.contiguous()
|
| 762 |
+
key_states = key_states.contiguous()
|
| 763 |
+
value_states = value_states.contiguous()
|
| 764 |
+
|
| 765 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 766 |
+
query_states,
|
| 767 |
+
key_states,
|
| 768 |
+
value_states,
|
| 769 |
+
attn_mask=attention_mask,
|
| 770 |
+
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 771 |
+
# The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
|
| 772 |
+
is_causal=self.is_causal and attention_mask is None and q_len > 1,
|
| 773 |
+
)
|
| 774 |
+
|
| 775 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 776 |
+
attn_output = attn_output.view(bsz, q_len, self.hidden_size)
|
| 777 |
+
|
| 778 |
+
attn_output = self.o_proj(attn_output)
|
| 779 |
+
|
| 780 |
+
return attn_output, None, past_key_value
|
| 781 |
+
|
| 782 |
+
|
| 783 |
+
SciDFM_ATTENTION_CLASSES = {
|
| 784 |
+
"eager": SciDFMAttention,
|
| 785 |
+
"flash_attention_2": SciDFMFlashAttention2,
|
| 786 |
+
"sdpa": SciDFMSdpaAttention,
|
| 787 |
+
}
|
| 788 |
+
|
| 789 |
+
|
| 790 |
+
class SciDFMBlockSparseTop2MLP(nn.Module):
|
| 791 |
+
def __init__(self, config: SciDFMConfig):
|
| 792 |
+
super().__init__()
|
| 793 |
+
self.ffn_dim = config.intermediate_size
|
| 794 |
+
self.hidden_dim = config.hidden_size
|
| 795 |
+
|
| 796 |
+
self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
| 797 |
+
self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
|
| 798 |
+
self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
| 799 |
+
|
| 800 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
| 801 |
+
|
| 802 |
+
def forward(self, hidden_states):
|
| 803 |
+
current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
|
| 804 |
+
current_hidden_states = self.w2(current_hidden_states)
|
| 805 |
+
return current_hidden_states
|
| 806 |
+
|
| 807 |
+
|
| 808 |
+
class SciDFMBLockSparseTop2MLP(SciDFMBlockSparseTop2MLP):
|
| 809 |
+
def __init__(self, *args, **kwargs):
|
| 810 |
+
logger.warning_once(
|
| 811 |
+
"SciDFMBLockSparseTop2MLP is deprecated by SciDFMBlockSparseTop2MLP and will be removed in v4.40."
|
| 812 |
+
)
|
| 813 |
+
super().__init__(*args, **kwargs)
|
| 814 |
+
|
| 815 |
+
|
| 816 |
+
class SciDFMSparseMoeBlock(nn.Module):
|
| 817 |
+
"""
|
| 818 |
+
This implementation is
|
| 819 |
+
strictly equivalent to standard MoE with full capacity (no
|
| 820 |
+
dropped tokens). It's faster since it formulates MoE operations
|
| 821 |
+
in terms of block-sparse operations to accomodate imbalanced
|
| 822 |
+
assignments of tokens to experts, whereas standard MoE either
|
| 823 |
+
(1) drop tokens at the cost of reduced performance or (2) set
|
| 824 |
+
capacity factor to number of experts and thus waste computation
|
| 825 |
+
and memory on padding.
|
| 826 |
+
"""
|
| 827 |
+
|
| 828 |
+
def __init__(self, config):
|
| 829 |
+
super().__init__()
|
| 830 |
+
self.hidden_dim = config.hidden_size
|
| 831 |
+
self.ffn_dim = config.intermediate_size
|
| 832 |
+
self.num_experts = config.num_local_experts
|
| 833 |
+
self.top_k = config.num_experts_per_tok
|
| 834 |
+
|
| 835 |
+
# gating
|
| 836 |
+
self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
| 837 |
+
|
| 838 |
+
self.experts = nn.ModuleList([SciDFMBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
|
| 839 |
+
|
| 840 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 841 |
+
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
| 842 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
| 843 |
+
# router_logits: (batch * sequence_length, n_experts)
|
| 844 |
+
router_logits = self.gate(hidden_states)
|
| 845 |
+
|
| 846 |
+
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
| 847 |
+
topk_weight, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
|
| 848 |
+
topk_weight /= topk_weight.sum(dim=-1, keepdim=True)
|
| 849 |
+
# we cast back to the input dtype
|
| 850 |
+
topk_weight = topk_weight.to(hidden_states.dtype)
|
| 851 |
+
|
| 852 |
+
hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
|
| 853 |
+
y = torch.empty_like(hidden_states)
|
| 854 |
+
flat_topk_idx = topk_idx.view(-1)
|
| 855 |
+
for i in range(self.num_experts):
|
| 856 |
+
expert = self.experts[i]
|
| 857 |
+
y[flat_topk_idx == i] = expert(hidden_states[flat_topk_idx == i])
|
| 858 |
+
y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
|
| 859 |
+
final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
|
| 860 |
+
|
| 861 |
+
return final_hidden_states, router_logits
|
| 862 |
+
|
| 863 |
+
|
| 864 |
+
class SciDFMDecoderLayer(nn.Module):
|
| 865 |
+
def __init__(self, config: SciDFMConfig, layer_idx: int):
|
| 866 |
+
super().__init__()
|
| 867 |
+
self.hidden_size = config.hidden_size
|
| 868 |
+
|
| 869 |
+
self.self_attn = SciDFM_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
|
| 870 |
+
|
| 871 |
+
self.block_sparse_moe = SciDFMSparseMoeBlock(config)
|
| 872 |
+
self.input_layernorm = SciDFMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 873 |
+
self.post_attention_layernorm = SciDFMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 874 |
+
self.layer_idx = layer_idx
|
| 875 |
+
|
| 876 |
+
def forward(
|
| 877 |
+
self,
|
| 878 |
+
hidden_states: torch.Tensor,
|
| 879 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 880 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 881 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
| 882 |
+
output_attentions: Optional[bool] = False,
|
| 883 |
+
output_router_logits: Optional[bool] = False,
|
| 884 |
+
use_cache: Optional[bool] = False,
|
| 885 |
+
**kwargs,
|
| 886 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 887 |
+
if "padding_mask" in kwargs:
|
| 888 |
+
warnings.warn(
|
| 889 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 890 |
+
)
|
| 891 |
+
"""
|
| 892 |
+
Args:
|
| 893 |
+
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 894 |
+
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 895 |
+
`(batch, sequence_length)` where padding elements are indicated by 0.
|
| 896 |
+
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
| 897 |
+
output_attentions (`bool`, *optional*):
|
| 898 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 899 |
+
returned tensors for more detail.
|
| 900 |
+
output_router_logits (`bool`, *optional*):
|
| 901 |
+
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
|
| 902 |
+
should not be returned during inference.
|
| 903 |
+
use_cache (`bool`, *optional*):
|
| 904 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 905 |
+
(see `past_key_values`).
|
| 906 |
+
"""
|
| 907 |
+
|
| 908 |
+
residual = hidden_states
|
| 909 |
+
|
| 910 |
+
hidden_states = self.input_layernorm(hidden_states)
|
| 911 |
+
|
| 912 |
+
# Self Attention
|
| 913 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
| 914 |
+
hidden_states=hidden_states,
|
| 915 |
+
attention_mask=attention_mask,
|
| 916 |
+
position_ids=position_ids,
|
| 917 |
+
past_key_value=past_key_value,
|
| 918 |
+
output_attentions=output_attentions,
|
| 919 |
+
use_cache=use_cache,
|
| 920 |
+
)
|
| 921 |
+
hidden_states = residual + hidden_states
|
| 922 |
+
|
| 923 |
+
# Fully Connected
|
| 924 |
+
residual = hidden_states
|
| 925 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 926 |
+
hidden_states, router_logits = self.block_sparse_moe(hidden_states)
|
| 927 |
+
hidden_states = residual + hidden_states
|
| 928 |
+
|
| 929 |
+
outputs = (hidden_states,)
|
| 930 |
+
|
| 931 |
+
if output_attentions:
|
| 932 |
+
outputs += (self_attn_weights,)
|
| 933 |
+
|
| 934 |
+
if use_cache:
|
| 935 |
+
outputs += (present_key_value,)
|
| 936 |
+
|
| 937 |
+
if output_router_logits:
|
| 938 |
+
outputs += (router_logits,)
|
| 939 |
+
|
| 940 |
+
return outputs
|
| 941 |
+
|
| 942 |
+
|
| 943 |
+
SciDFM_START_DOCSTRING = r"""
|
| 944 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 945 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 946 |
+
etc.)
|
| 947 |
+
|
| 948 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 949 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 950 |
+
and behavior.
|
| 951 |
+
|
| 952 |
+
Parameters:
|
| 953 |
+
config ([`SciDFMConfig`]):
|
| 954 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
| 955 |
+
load the weights associated with the model, only the configuration. Check out the
|
| 956 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 957 |
+
"""
|
| 958 |
+
|
| 959 |
+
|
| 960 |
+
@add_start_docstrings(
|
| 961 |
+
"The bare SciDFM Model outputting raw hidden-states without any specific head on top.",
|
| 962 |
+
SciDFM_START_DOCSTRING,
|
| 963 |
+
)
|
| 964 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->SciDFM
|
| 965 |
+
class SciDFMPreTrainedModel(PreTrainedModel):
|
| 966 |
+
config_class = SciDFMConfig
|
| 967 |
+
base_model_prefix = "model"
|
| 968 |
+
supports_gradient_checkpointing = True
|
| 969 |
+
_no_split_modules = ["SciDFMDecoderLayer"]
|
| 970 |
+
_skip_keys_device_placement = "past_key_values"
|
| 971 |
+
_supports_flash_attn_2 = True
|
| 972 |
+
_supports_sdpa = True
|
| 973 |
+
_supports_cache_class = True
|
| 974 |
+
|
| 975 |
+
def _init_weights(self, module):
|
| 976 |
+
std = self.config.initializer_range
|
| 977 |
+
if isinstance(module, nn.Linear):
|
| 978 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 979 |
+
if module.bias is not None:
|
| 980 |
+
module.bias.data.zero_()
|
| 981 |
+
elif isinstance(module, nn.Embedding):
|
| 982 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 983 |
+
if module.padding_idx is not None:
|
| 984 |
+
module.weight.data[module.padding_idx].zero_()
|
| 985 |
+
|
| 986 |
+
|
| 987 |
+
SciDFM_INPUTS_DOCSTRING = r"""
|
| 988 |
+
Args:
|
| 989 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 990 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 991 |
+
it.
|
| 992 |
+
|
| 993 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 994 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 995 |
+
|
| 996 |
+
[What are input IDs?](../glossary#input-ids)
|
| 997 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 998 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 999 |
+
|
| 1000 |
+
- 1 for tokens that are **not masked**,
|
| 1001 |
+
- 0 for tokens that are **masked**.
|
| 1002 |
+
|
| 1003 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 1004 |
+
|
| 1005 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 1006 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 1007 |
+
|
| 1008 |
+
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
| 1009 |
+
`past_key_values`).
|
| 1010 |
+
|
| 1011 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
| 1012 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
| 1013 |
+
information on the default strategy.
|
| 1014 |
+
|
| 1015 |
+
- 1 indicates the head is **not masked**,
|
| 1016 |
+
- 0 indicates the head is **masked**.
|
| 1017 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1018 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 1019 |
+
config.n_positions - 1]`.
|
| 1020 |
+
|
| 1021 |
+
[What are position IDs?](../glossary#position-ids)
|
| 1022 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 1023 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
| 1024 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
| 1025 |
+
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
| 1026 |
+
|
| 1027 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
| 1028 |
+
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
| 1029 |
+
|
| 1030 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
| 1031 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
| 1032 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
| 1033 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
| 1034 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
| 1035 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
| 1036 |
+
model's internal embedding lookup matrix.
|
| 1037 |
+
use_cache (`bool`, *optional*):
|
| 1038 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
| 1039 |
+
`past_key_values`).
|
| 1040 |
+
output_attentions (`bool`, *optional*):
|
| 1041 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 1042 |
+
tensors for more detail.
|
| 1043 |
+
output_hidden_states (`bool`, *optional*):
|
| 1044 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 1045 |
+
more detail.
|
| 1046 |
+
output_router_logits (`bool`, *optional*):
|
| 1047 |
+
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
|
| 1048 |
+
should not be returned during inference.
|
| 1049 |
+
return_dict (`bool`, *optional*):
|
| 1050 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 1051 |
+
"""
|
| 1052 |
+
|
| 1053 |
+
|
| 1054 |
+
@add_start_docstrings(
|
| 1055 |
+
"The bare SciDFM Model outputting raw hidden-states without any specific head on top.",
|
| 1056 |
+
SciDFM_START_DOCSTRING,
|
| 1057 |
+
)
|
| 1058 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->SciDFM,Mistral->SciDFM
|
| 1059 |
+
class SciDFMModel(SciDFMPreTrainedModel):
|
| 1060 |
+
"""
|
| 1061 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`SciDFMDecoderLayer`]
|
| 1062 |
+
|
| 1063 |
+
Args:
|
| 1064 |
+
config: SciDFMConfig
|
| 1065 |
+
"""
|
| 1066 |
+
|
| 1067 |
+
def __init__(self, config: SciDFMConfig):
|
| 1068 |
+
super().__init__(config)
|
| 1069 |
+
self.padding_idx = config.pad_token_id
|
| 1070 |
+
self.vocab_size = config.vocab_size
|
| 1071 |
+
|
| 1072 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 1073 |
+
self.layers = nn.ModuleList(
|
| 1074 |
+
[SciDFMDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 1075 |
+
)
|
| 1076 |
+
self._attn_implementation = config._attn_implementation
|
| 1077 |
+
self.norm = SciDFMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1078 |
+
|
| 1079 |
+
self.gradient_checkpointing = False
|
| 1080 |
+
# Initialize weights and apply final processing
|
| 1081 |
+
self.post_init()
|
| 1082 |
+
|
| 1083 |
+
def get_input_embeddings(self):
|
| 1084 |
+
return self.embed_tokens
|
| 1085 |
+
|
| 1086 |
+
def set_input_embeddings(self, value):
|
| 1087 |
+
self.embed_tokens = value
|
| 1088 |
+
|
| 1089 |
+
# Ignore copy
|
| 1090 |
+
@add_start_docstrings_to_model_forward(SciDFM_INPUTS_DOCSTRING)
|
| 1091 |
+
def forward(
|
| 1092 |
+
self,
|
| 1093 |
+
input_ids: torch.LongTensor = None,
|
| 1094 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1095 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1096 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1097 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1098 |
+
use_cache: Optional[bool] = None,
|
| 1099 |
+
output_attentions: Optional[bool] = None,
|
| 1100 |
+
output_hidden_states: Optional[bool] = None,
|
| 1101 |
+
output_router_logits: Optional[bool] = None,
|
| 1102 |
+
return_dict: Optional[bool] = None,
|
| 1103 |
+
) -> Union[Tuple, MoeModelOutputWithPast]:
|
| 1104 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1105 |
+
output_router_logits = (
|
| 1106 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
| 1107 |
+
)
|
| 1108 |
+
output_hidden_states = (
|
| 1109 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1110 |
+
)
|
| 1111 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 1112 |
+
|
| 1113 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1114 |
+
|
| 1115 |
+
# retrieve input_ids and inputs_embeds
|
| 1116 |
+
if input_ids is not None and inputs_embeds is not None:
|
| 1117 |
+
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
|
| 1118 |
+
elif input_ids is not None:
|
| 1119 |
+
batch_size, seq_length = input_ids.shape
|
| 1120 |
+
elif inputs_embeds is not None:
|
| 1121 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
| 1122 |
+
else:
|
| 1123 |
+
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
|
| 1124 |
+
|
| 1125 |
+
past_key_values_length = 0
|
| 1126 |
+
|
| 1127 |
+
if self.gradient_checkpointing and self.training:
|
| 1128 |
+
if use_cache:
|
| 1129 |
+
logger.warning_once(
|
| 1130 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
| 1131 |
+
)
|
| 1132 |
+
use_cache = False
|
| 1133 |
+
|
| 1134 |
+
if use_cache:
|
| 1135 |
+
use_legacy_cache = not isinstance(past_key_values, Cache)
|
| 1136 |
+
if use_legacy_cache:
|
| 1137 |
+
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
| 1138 |
+
past_key_values_length = past_key_values.get_usable_length(seq_length)
|
| 1139 |
+
|
| 1140 |
+
if position_ids is None:
|
| 1141 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
| 1142 |
+
position_ids = torch.arange(
|
| 1143 |
+
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
| 1144 |
+
)
|
| 1145 |
+
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
| 1146 |
+
else:
|
| 1147 |
+
position_ids = position_ids.view(-1, seq_length).long()
|
| 1148 |
+
|
| 1149 |
+
if inputs_embeds is None:
|
| 1150 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
| 1151 |
+
|
| 1152 |
+
if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
|
| 1153 |
+
is_padding_right = attention_mask[:, -1].sum().item() != batch_size
|
| 1154 |
+
if is_padding_right:
|
| 1155 |
+
raise ValueError(
|
| 1156 |
+
"You are attempting to perform batched generation with padding_side='right'"
|
| 1157 |
+
" this may lead to unexpected behaviour for Flash Attention version of SciDFM. Make sure to "
|
| 1158 |
+
" call `tokenizer.padding_side = 'left'` before tokenizing the input. "
|
| 1159 |
+
)
|
| 1160 |
+
|
| 1161 |
+
if self._attn_implementation == "flash_attention_2":
|
| 1162 |
+
# 2d mask is passed through the layers
|
| 1163 |
+
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
|
| 1164 |
+
elif self._attn_implementation == "sdpa" and not output_attentions:
|
| 1165 |
+
# output_attentions=True can not be supported when using SDPA, and we fall back on
|
| 1166 |
+
# the manual implementation that requires a 4D causal mask in all cases.
|
| 1167 |
+
attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
|
| 1168 |
+
attention_mask,
|
| 1169 |
+
(batch_size, seq_length),
|
| 1170 |
+
inputs_embeds,
|
| 1171 |
+
past_key_values_length,
|
| 1172 |
+
)
|
| 1173 |
+
else:
|
| 1174 |
+
# 4d mask is passed through the layers
|
| 1175 |
+
attention_mask = _prepare_4d_causal_attention_mask(
|
| 1176 |
+
attention_mask,
|
| 1177 |
+
(batch_size, seq_length),
|
| 1178 |
+
inputs_embeds,
|
| 1179 |
+
past_key_values_length,
|
| 1180 |
+
sliding_window=self.config.sliding_window,
|
| 1181 |
+
)
|
| 1182 |
+
|
| 1183 |
+
hidden_states = inputs_embeds
|
| 1184 |
+
|
| 1185 |
+
# decoder layers
|
| 1186 |
+
all_hidden_states = () if output_hidden_states else None
|
| 1187 |
+
all_self_attns = () if output_attentions else None
|
| 1188 |
+
all_router_logits = () if output_router_logits else None
|
| 1189 |
+
next_decoder_cache = None
|
| 1190 |
+
|
| 1191 |
+
for decoder_layer in self.layers:
|
| 1192 |
+
if output_hidden_states:
|
| 1193 |
+
all_hidden_states += (hidden_states,)
|
| 1194 |
+
|
| 1195 |
+
if self.gradient_checkpointing and self.training:
|
| 1196 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 1197 |
+
decoder_layer.__call__,
|
| 1198 |
+
hidden_states,
|
| 1199 |
+
attention_mask,
|
| 1200 |
+
position_ids,
|
| 1201 |
+
past_key_values,
|
| 1202 |
+
output_attentions,
|
| 1203 |
+
output_router_logits,
|
| 1204 |
+
use_cache,
|
| 1205 |
+
)
|
| 1206 |
+
else:
|
| 1207 |
+
layer_outputs = decoder_layer(
|
| 1208 |
+
hidden_states,
|
| 1209 |
+
attention_mask=attention_mask,
|
| 1210 |
+
position_ids=position_ids,
|
| 1211 |
+
past_key_value=past_key_values,
|
| 1212 |
+
output_attentions=output_attentions,
|
| 1213 |
+
output_router_logits=output_router_logits,
|
| 1214 |
+
use_cache=use_cache,
|
| 1215 |
+
)
|
| 1216 |
+
|
| 1217 |
+
hidden_states = layer_outputs[0]
|
| 1218 |
+
|
| 1219 |
+
if use_cache:
|
| 1220 |
+
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
| 1221 |
+
|
| 1222 |
+
if output_attentions:
|
| 1223 |
+
all_self_attns += (layer_outputs[1],)
|
| 1224 |
+
|
| 1225 |
+
if output_router_logits:
|
| 1226 |
+
all_router_logits += (layer_outputs[-1],)
|
| 1227 |
+
|
| 1228 |
+
hidden_states = self.norm(hidden_states)
|
| 1229 |
+
|
| 1230 |
+
# add hidden states from the last decoder layer
|
| 1231 |
+
if output_hidden_states:
|
| 1232 |
+
all_hidden_states += (hidden_states,)
|
| 1233 |
+
|
| 1234 |
+
next_cache = None
|
| 1235 |
+
if use_cache:
|
| 1236 |
+
next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
|
| 1237 |
+
|
| 1238 |
+
if not return_dict:
|
| 1239 |
+
return tuple(
|
| 1240 |
+
v
|
| 1241 |
+
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
|
| 1242 |
+
if v is not None
|
| 1243 |
+
)
|
| 1244 |
+
return MoeModelOutputWithPast(
|
| 1245 |
+
last_hidden_state=hidden_states,
|
| 1246 |
+
past_key_values=next_cache,
|
| 1247 |
+
hidden_states=all_hidden_states,
|
| 1248 |
+
attentions=all_self_attns,
|
| 1249 |
+
router_logits=all_router_logits,
|
| 1250 |
+
)
|
| 1251 |
+
|
| 1252 |
+
|
| 1253 |
+
class SciDFMForCausalLM(SciDFMPreTrainedModel):
|
| 1254 |
+
_tied_weights_keys = ["lm_head.weight"]
|
| 1255 |
+
|
| 1256 |
+
def __init__(self, config):
|
| 1257 |
+
super().__init__(config)
|
| 1258 |
+
self.model = SciDFMModel(config)
|
| 1259 |
+
self.vocab_size = config.vocab_size
|
| 1260 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 1261 |
+
self.router_aux_loss_coef = config.router_aux_loss_coef
|
| 1262 |
+
self.num_experts = config.num_local_experts
|
| 1263 |
+
self.num_experts_per_tok = config.num_experts_per_tok
|
| 1264 |
+
# Initialize weights and apply final processing
|
| 1265 |
+
self.post_init()
|
| 1266 |
+
|
| 1267 |
+
def get_input_embeddings(self):
|
| 1268 |
+
return self.model.embed_tokens
|
| 1269 |
+
|
| 1270 |
+
def set_input_embeddings(self, value):
|
| 1271 |
+
self.model.embed_tokens = value
|
| 1272 |
+
|
| 1273 |
+
def get_output_embeddings(self):
|
| 1274 |
+
return self.lm_head
|
| 1275 |
+
|
| 1276 |
+
def set_output_embeddings(self, new_embeddings):
|
| 1277 |
+
self.lm_head = new_embeddings
|
| 1278 |
+
|
| 1279 |
+
def set_decoder(self, decoder):
|
| 1280 |
+
self.model = decoder
|
| 1281 |
+
|
| 1282 |
+
def get_decoder(self):
|
| 1283 |
+
return self.model
|
| 1284 |
+
|
| 1285 |
+
@add_start_docstrings_to_model_forward(SciDFM_INPUTS_DOCSTRING)
|
| 1286 |
+
@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
| 1287 |
+
# Ignore copy
|
| 1288 |
+
def forward(
|
| 1289 |
+
self,
|
| 1290 |
+
input_ids: torch.LongTensor = None,
|
| 1291 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1292 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1293 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1294 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1295 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1296 |
+
use_cache: Optional[bool] = None,
|
| 1297 |
+
output_attentions: Optional[bool] = None,
|
| 1298 |
+
output_hidden_states: Optional[bool] = None,
|
| 1299 |
+
output_router_logits: Optional[bool] = None,
|
| 1300 |
+
return_dict: Optional[bool] = None,
|
| 1301 |
+
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
| 1302 |
+
r"""
|
| 1303 |
+
Args:
|
| 1304 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1305 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 1306 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 1307 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 1308 |
+
|
| 1309 |
+
Returns:
|
| 1310 |
+
|
| 1311 |
+
Example:
|
| 1312 |
+
|
| 1313 |
+
```python
|
| 1314 |
+
>>> from transformers import AutoTokenizer, SciDFMForCausalLM
|
| 1315 |
+
|
| 1316 |
+
>>> model = SciDFMForCausalLM.from_pretrained("OpenDFM/SciDFM-MoE-A5.6B-v1.0")
|
| 1317 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("OpenDFM/SciDFM-MoE-A5.6B-v1.0")
|
| 1318 |
+
|
| 1319 |
+
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
| 1320 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
| 1321 |
+
|
| 1322 |
+
>>> # Generate
|
| 1323 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
| 1324 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 1325 |
+
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
|
| 1326 |
+
```"""
|
| 1327 |
+
|
| 1328 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1329 |
+
output_router_logits = (
|
| 1330 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
| 1331 |
+
)
|
| 1332 |
+
|
| 1333 |
+
output_hidden_states = (
|
| 1334 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1335 |
+
)
|
| 1336 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1337 |
+
|
| 1338 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 1339 |
+
outputs = self.model(
|
| 1340 |
+
input_ids=input_ids,
|
| 1341 |
+
attention_mask=attention_mask,
|
| 1342 |
+
position_ids=position_ids,
|
| 1343 |
+
past_key_values=past_key_values,
|
| 1344 |
+
inputs_embeds=inputs_embeds,
|
| 1345 |
+
use_cache=use_cache,
|
| 1346 |
+
output_attentions=output_attentions,
|
| 1347 |
+
output_hidden_states=output_hidden_states,
|
| 1348 |
+
output_router_logits=output_router_logits,
|
| 1349 |
+
return_dict=return_dict,
|
| 1350 |
+
)
|
| 1351 |
+
|
| 1352 |
+
hidden_states = outputs[0]
|
| 1353 |
+
logits = self.lm_head(hidden_states)
|
| 1354 |
+
logits = logits.float()
|
| 1355 |
+
|
| 1356 |
+
loss = None
|
| 1357 |
+
if labels is not None:
|
| 1358 |
+
# Shift so that tokens < n predict n
|
| 1359 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 1360 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 1361 |
+
# Flatten the tokens
|
| 1362 |
+
loss_fct = CrossEntropyLoss()
|
| 1363 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
| 1364 |
+
shift_labels = shift_labels.view(-1)
|
| 1365 |
+
# Enable model parallelism
|
| 1366 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
| 1367 |
+
loss = loss_fct(shift_logits, shift_labels)
|
| 1368 |
+
|
| 1369 |
+
aux_loss = None
|
| 1370 |
+
if output_router_logits:
|
| 1371 |
+
aux_loss = load_balancing_loss_func(
|
| 1372 |
+
outputs.router_logits if return_dict else outputs[-1],
|
| 1373 |
+
self.num_experts,
|
| 1374 |
+
self.num_experts_per_tok,
|
| 1375 |
+
attention_mask,
|
| 1376 |
+
)
|
| 1377 |
+
if labels is not None:
|
| 1378 |
+
loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
|
| 1379 |
+
|
| 1380 |
+
if not return_dict:
|
| 1381 |
+
output = (logits,) + outputs[1:]
|
| 1382 |
+
if output_router_logits:
|
| 1383 |
+
output = (aux_loss,) + output
|
| 1384 |
+
return (loss,) + output if loss is not None else output
|
| 1385 |
+
|
| 1386 |
+
return MoeCausalLMOutputWithPast(
|
| 1387 |
+
loss=loss,
|
| 1388 |
+
aux_loss=aux_loss,
|
| 1389 |
+
logits=logits,
|
| 1390 |
+
past_key_values=outputs.past_key_values,
|
| 1391 |
+
hidden_states=outputs.hidden_states,
|
| 1392 |
+
attentions=outputs.attentions,
|
| 1393 |
+
router_logits=outputs.router_logits,
|
| 1394 |
+
)
|
| 1395 |
+
|
| 1396 |
+
def prepare_inputs_for_generation(
|
| 1397 |
+
self,
|
| 1398 |
+
input_ids,
|
| 1399 |
+
past_key_values=None,
|
| 1400 |
+
attention_mask=None,
|
| 1401 |
+
inputs_embeds=None,
|
| 1402 |
+
output_router_logits=False,
|
| 1403 |
+
**kwargs,
|
| 1404 |
+
):
|
| 1405 |
+
# Omit tokens covered by past_key_values
|
| 1406 |
+
if past_key_values is not None:
|
| 1407 |
+
if isinstance(past_key_values, Cache):
|
| 1408 |
+
cache_length = past_key_values.get_seq_length()
|
| 1409 |
+
past_length = past_key_values.seen_tokens
|
| 1410 |
+
max_cache_length = past_key_values.get_max_length()
|
| 1411 |
+
else:
|
| 1412 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
| 1413 |
+
max_cache_length = None
|
| 1414 |
+
|
| 1415 |
+
# Keep only the unprocessed tokens:
|
| 1416 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
| 1417 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
| 1418 |
+
# input)
|
| 1419 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
| 1420 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
| 1421 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
| 1422 |
+
# input_ids based on the past_length.
|
| 1423 |
+
elif past_length < input_ids.shape[1]:
|
| 1424 |
+
input_ids = input_ids[:, past_length:]
|
| 1425 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
| 1426 |
+
|
| 1427 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
| 1428 |
+
if (
|
| 1429 |
+
max_cache_length is not None
|
| 1430 |
+
and attention_mask is not None
|
| 1431 |
+
and cache_length + input_ids.shape[1] > max_cache_length
|
| 1432 |
+
):
|
| 1433 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
| 1434 |
+
|
| 1435 |
+
position_ids = kwargs.get("position_ids", None)
|
| 1436 |
+
if attention_mask is not None and position_ids is None:
|
| 1437 |
+
# create position_ids on the fly for batch generation
|
| 1438 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1439 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1440 |
+
if past_key_values:
|
| 1441 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 1442 |
+
|
| 1443 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 1444 |
+
if inputs_embeds is not None and past_key_values is None:
|
| 1445 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 1446 |
+
else:
|
| 1447 |
+
model_inputs = {"input_ids": input_ids}
|
| 1448 |
+
|
| 1449 |
+
model_inputs.update(
|
| 1450 |
+
{
|
| 1451 |
+
"position_ids": position_ids,
|
| 1452 |
+
"past_key_values": past_key_values,
|
| 1453 |
+
"use_cache": kwargs.get("use_cache"),
|
| 1454 |
+
"attention_mask": attention_mask,
|
| 1455 |
+
"output_router_logits": output_router_logits,
|
| 1456 |
+
}
|
| 1457 |
+
)
|
| 1458 |
+
return model_inputs
|
| 1459 |
+
|
| 1460 |
+
@staticmethod
|
| 1461 |
+
def _reorder_cache(past_key_values, beam_idx):
|
| 1462 |
+
reordered_past = ()
|
| 1463 |
+
for layer_past in past_key_values:
|
| 1464 |
+
reordered_past += (
|
| 1465 |
+
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
|
| 1466 |
+
)
|
| 1467 |
+
return reordered_past
|
| 1468 |
+
|
| 1469 |
+
|
| 1470 |
+
@add_start_docstrings(
|
| 1471 |
+
"""
|
| 1472 |
+
The SciDFM Model transformer with a sequence classification head on top (linear layer).
|
| 1473 |
+
|
| 1474 |
+
[`SciDFMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
| 1475 |
+
(e.g. GPT-2) do.
|
| 1476 |
+
|
| 1477 |
+
Since it does classification on the last token, it requires to know the position of the last token. If a
|
| 1478 |
+
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
| 1479 |
+
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
| 1480 |
+
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
| 1481 |
+
each row of the batch).
|
| 1482 |
+
""",
|
| 1483 |
+
SciDFM_START_DOCSTRING,
|
| 1484 |
+
)
|
| 1485 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->SciDFM, LLAMA->SciDFM
|
| 1486 |
+
class SciDFMForSequenceClassification(SciDFMPreTrainedModel):
|
| 1487 |
+
def __init__(self, config):
|
| 1488 |
+
super().__init__(config)
|
| 1489 |
+
self.num_labels = config.num_labels
|
| 1490 |
+
self.model = SciDFMModel(config)
|
| 1491 |
+
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
| 1492 |
+
|
| 1493 |
+
# Initialize weights and apply final processing
|
| 1494 |
+
self.post_init()
|
| 1495 |
+
|
| 1496 |
+
def get_input_embeddings(self):
|
| 1497 |
+
return self.model.embed_tokens
|
| 1498 |
+
|
| 1499 |
+
def set_input_embeddings(self, value):
|
| 1500 |
+
self.model.embed_tokens = value
|
| 1501 |
+
|
| 1502 |
+
@add_start_docstrings_to_model_forward(SciDFM_INPUTS_DOCSTRING)
|
| 1503 |
+
def forward(
|
| 1504 |
+
self,
|
| 1505 |
+
input_ids: torch.LongTensor = None,
|
| 1506 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1507 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1508 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1509 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1510 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1511 |
+
use_cache: Optional[bool] = None,
|
| 1512 |
+
output_attentions: Optional[bool] = None,
|
| 1513 |
+
output_hidden_states: Optional[bool] = None,
|
| 1514 |
+
return_dict: Optional[bool] = None,
|
| 1515 |
+
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
| 1516 |
+
r"""
|
| 1517 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1518 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1519 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1520 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1521 |
+
"""
|
| 1522 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1523 |
+
|
| 1524 |
+
transformer_outputs = self.model(
|
| 1525 |
+
input_ids,
|
| 1526 |
+
attention_mask=attention_mask,
|
| 1527 |
+
position_ids=position_ids,
|
| 1528 |
+
past_key_values=past_key_values,
|
| 1529 |
+
inputs_embeds=inputs_embeds,
|
| 1530 |
+
use_cache=use_cache,
|
| 1531 |
+
output_attentions=output_attentions,
|
| 1532 |
+
output_hidden_states=output_hidden_states,
|
| 1533 |
+
return_dict=return_dict,
|
| 1534 |
+
)
|
| 1535 |
+
hidden_states = transformer_outputs[0]
|
| 1536 |
+
logits = self.score(hidden_states)
|
| 1537 |
+
|
| 1538 |
+
if input_ids is not None:
|
| 1539 |
+
batch_size = input_ids.shape[0]
|
| 1540 |
+
else:
|
| 1541 |
+
batch_size = inputs_embeds.shape[0]
|
| 1542 |
+
|
| 1543 |
+
if self.config.pad_token_id is None and batch_size != 1:
|
| 1544 |
+
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
| 1545 |
+
if self.config.pad_token_id is None:
|
| 1546 |
+
sequence_lengths = -1
|
| 1547 |
+
else:
|
| 1548 |
+
if input_ids is not None:
|
| 1549 |
+
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
|
| 1550 |
+
sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
|
| 1551 |
+
sequence_lengths = sequence_lengths % input_ids.shape[-1]
|
| 1552 |
+
sequence_lengths = sequence_lengths.to(logits.device)
|
| 1553 |
+
else:
|
| 1554 |
+
sequence_lengths = -1
|
| 1555 |
+
|
| 1556 |
+
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
|
| 1557 |
+
|
| 1558 |
+
loss = None
|
| 1559 |
+
if labels is not None:
|
| 1560 |
+
labels = labels.to(logits.device)
|
| 1561 |
+
if self.config.problem_type is None:
|
| 1562 |
+
if self.num_labels == 1:
|
| 1563 |
+
self.config.problem_type = "regression"
|
| 1564 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 1565 |
+
self.config.problem_type = "single_label_classification"
|
| 1566 |
+
else:
|
| 1567 |
+
self.config.problem_type = "multi_label_classification"
|
| 1568 |
+
|
| 1569 |
+
if self.config.problem_type == "regression":
|
| 1570 |
+
loss_fct = MSELoss()
|
| 1571 |
+
if self.num_labels == 1:
|
| 1572 |
+
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
| 1573 |
+
else:
|
| 1574 |
+
loss = loss_fct(pooled_logits, labels)
|
| 1575 |
+
elif self.config.problem_type == "single_label_classification":
|
| 1576 |
+
loss_fct = CrossEntropyLoss()
|
| 1577 |
+
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
|
| 1578 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 1579 |
+
loss_fct = BCEWithLogitsLoss()
|
| 1580 |
+
loss = loss_fct(pooled_logits, labels)
|
| 1581 |
+
if not return_dict:
|
| 1582 |
+
output = (pooled_logits,) + transformer_outputs[1:]
|
| 1583 |
+
return ((loss,) + output) if loss is not None else output
|
| 1584 |
+
|
| 1585 |
+
return SequenceClassifierOutputWithPast(
|
| 1586 |
+
loss=loss,
|
| 1587 |
+
logits=pooled_logits,
|
| 1588 |
+
past_key_values=transformer_outputs.past_key_values,
|
| 1589 |
+
hidden_states=transformer_outputs.hidden_states,
|
| 1590 |
+
attentions=transformer_outputs.attentions,
|
| 1591 |
+
)
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|eod|>",
|
| 4 |
+
"<|system|>",
|
| 5 |
+
"<|user|>",
|
| 6 |
+
"<|assistant|>"
|
| 7 |
+
],
|
| 8 |
+
"bos_token": {
|
| 9 |
+
"content": "<s>",
|
| 10 |
+
"lstrip": false,
|
| 11 |
+
"normalized": false,
|
| 12 |
+
"rstrip": false,
|
| 13 |
+
"single_word": false
|
| 14 |
+
},
|
| 15 |
+
"eos_token": {
|
| 16 |
+
"content": "</s>",
|
| 17 |
+
"lstrip": false,
|
| 18 |
+
"normalized": false,
|
| 19 |
+
"rstrip": false,
|
| 20 |
+
"single_word": false
|
| 21 |
+
},
|
| 22 |
+
"pad_token": {
|
| 23 |
+
"content": "[PAD]",
|
| 24 |
+
"lstrip": false,
|
| 25 |
+
"normalized": false,
|
| 26 |
+
"rstrip": false,
|
| 27 |
+
"single_word": false
|
| 28 |
+
},
|
| 29 |
+
"unk_token": {
|
| 30 |
+
"content": "<unk>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false
|
| 35 |
+
}
|
| 36 |
+
}
|
tokenizer.model
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:becc16550321ae552b9d3375c45b9cc24a5a73a9ae64d64ebd604ec9e6d48cf9
|
| 3 |
+
size 522336
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": true,
|
| 3 |
+
"add_eos_token": false,
|
| 4 |
+
"add_prefix_space": true,
|
| 5 |
+
"added_tokens_decoder": {
|
| 6 |
+
"0": {
|
| 7 |
+
"content": "<unk>",
|
| 8 |
+
"lstrip": false,
|
| 9 |
+
"normalized": false,
|
| 10 |
+
"rstrip": false,
|
| 11 |
+
"single_word": false,
|
| 12 |
+
"special": true
|
| 13 |
+
},
|
| 14 |
+
"1": {
|
| 15 |
+
"content": "<s>",
|
| 16 |
+
"lstrip": false,
|
| 17 |
+
"normalized": false,
|
| 18 |
+
"rstrip": false,
|
| 19 |
+
"single_word": false,
|
| 20 |
+
"special": true
|
| 21 |
+
},
|
| 22 |
+
"2": {
|
| 23 |
+
"content": "</s>",
|
| 24 |
+
"lstrip": false,
|
| 25 |
+
"normalized": false,
|
| 26 |
+
"rstrip": false,
|
| 27 |
+
"single_word": false,
|
| 28 |
+
"special": true
|
| 29 |
+
},
|
| 30 |
+
"32191": {
|
| 31 |
+
"content": "[PAD]",
|
| 32 |
+
"lstrip": false,
|
| 33 |
+
"normalized": false,
|
| 34 |
+
"rstrip": false,
|
| 35 |
+
"single_word": false,
|
| 36 |
+
"special": true
|
| 37 |
+
},
|
| 38 |
+
"32192": {
|
| 39 |
+
"content": "<|eod|>",
|
| 40 |
+
"lstrip": false,
|
| 41 |
+
"normalized": false,
|
| 42 |
+
"rstrip": false,
|
| 43 |
+
"single_word": false,
|
| 44 |
+
"special": true
|
| 45 |
+
},
|
| 46 |
+
"32193": {
|
| 47 |
+
"content": "<|system|>",
|
| 48 |
+
"lstrip": false,
|
| 49 |
+
"normalized": false,
|
| 50 |
+
"rstrip": false,
|
| 51 |
+
"single_word": false,
|
| 52 |
+
"special": true
|
| 53 |
+
},
|
| 54 |
+
"32194": {
|
| 55 |
+
"content": "<|user|>",
|
| 56 |
+
"lstrip": false,
|
| 57 |
+
"normalized": false,
|
| 58 |
+
"rstrip": false,
|
| 59 |
+
"single_word": false,
|
| 60 |
+
"special": true
|
| 61 |
+
},
|
| 62 |
+
"32195": {
|
| 63 |
+
"content": "<|assistant|>",
|
| 64 |
+
"lstrip": false,
|
| 65 |
+
"normalized": false,
|
| 66 |
+
"rstrip": false,
|
| 67 |
+
"single_word": false,
|
| 68 |
+
"special": true
|
| 69 |
+
}
|
| 70 |
+
},
|
| 71 |
+
"additional_special_tokens": [
|
| 72 |
+
"<|eod|>",
|
| 73 |
+
"<|system|>",
|
| 74 |
+
"<|user|>",
|
| 75 |
+
"<|assistant|>"
|
| 76 |
+
],
|
| 77 |
+
"bos_token": "<s>",
|
| 78 |
+
"clean_up_tokenization_spaces": false,
|
| 79 |
+
"eos_token": "</s>",
|
| 80 |
+
"legacy": true,
|
| 81 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 82 |
+
"pad_token": "[PAD]",
|
| 83 |
+
"sp_model_kwargs": {},
|
| 84 |
+
"spaces_between_special_tokens": false,
|
| 85 |
+
"tokenizer_class": "LlamaTokenizer",
|
| 86 |
+
"unk_token": "<unk>",
|
| 87 |
+
"use_default_system_prompt": false
|
| 88 |
+
}
|