Skip to content

modules.qwen2_moe.configuration_qwen2_moe

Qwen2MoeConfig

Bases: EasyDeLPretrainedConfig

Source code in src/python/easydel/modules/qwen2_moe/configuration_qwen2_moe.py 
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
class Qwen2MoeConfig(EasyDeLPretrainedConfig):
    model_type: str = "qwen2_moe"

    def __init__(
            self,
            vocab_size=151936,
            hidden_size=2048,
            intermediate_size=5632,
            num_hidden_layers=24,
            num_attention_heads=16,
            num_key_value_heads=16,
            hidden_act="silu",
            max_position_embeddings=32768,
            initializer_range=0.02,
            rms_norm_eps=1e-6,
            use_cache=True,
            tie_word_embeddings=False,
            rope_theta=10000.0,
            use_sliding_window=False,
            sliding_window=4096,
            max_window_layers=28,
            attention_dropout=0.0,
            decoder_sparse_step=1,
            moe_intermediate_size=1408,
            shared_expert_intermediate_size=5632,
            num_experts_per_tok=4,
            num_experts=60,
            norm_topk_prob=False,
            output_router_logits=False,
            router_aux_loss_coef=0.001,
            gradient_checkpointing: str = "nothing_saveable",
            bits: Optional[int] = None,
            **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.max_window_layers = max_window_layers

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout

        # MoE arguments
        self.decoder_sparse_step = decoder_sparse_step
        self.moe_intermediate_size = moe_intermediate_size
        self.shared_expert_intermediate_size = shared_expert_intermediate_size
        self.num_experts_per_tok = num_experts_per_tok
        self.num_experts = num_experts
        self.norm_topk_prob = norm_topk_prob
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.gradient_checkpointing = gradient_checkpointing
        self.bits = bits
        super().__init__(
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )

    def get_partition_rules(self, fully_sharded_data_parallel: bool = True):
        """
        The get_partition_rules function is used to define the partitioning scheme for a model.
        It returns a list of tuples, where each tuple contains two elements:
            1) A regex string that matches the name of one or more parameters in the model.
            2) A PartitionScheme object that defines how those parameters should be partitioned across devices.

        :param fully_sharded_data_parallel: bool: Determine whether to partition the model fully or not
        :return: A list of tuples

        """
        return (

            ("model/embed_tokens/embedding", PartitionSpec("tp", ("fsdp", "sp"))),

            ("self_attn/(q_proj|k_proj|v_proj)/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            ("self_attn/o_proj/kernel", PartitionSpec("tp", ("sp", "fsdp"))),

            ("gate_proj/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            ("down_proj/kernel", PartitionSpec("tp", ("fsdp", "sp"))),
            ("up_proj/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            ("shared_expert_gate/kernel", PartitionSpec(("fsdp", "sp"))),
            ("gate/kernel", PartitionSpec(("fsdp", "sp"))),

            ("input_layernorm/kernel", PartitionSpec(None)),
            ("post_attention_layernorm/kernel", PartitionSpec(None)),

            ("model/norm/kernel", PartitionSpec(None)),
            ("lm_head/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            (".*", PartitionSpec(None)),
        ) if not fully_sharded_data_parallel else (

            ("model/embed_tokens/embedding", PartitionSpec("tp", ("fsdp", "sp"))),

            ("self_attn/(q_proj|k_proj|v_proj)/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            ("self_attn/o_proj/kernel", PartitionSpec("tp", ("sp", "fsdp"))),

            ("gate_proj/kernel", PartitionSpec(("fsdp", "sp"))),
            ("down_proj/kernel", PartitionSpec(("fsdp", "sp"))),
            ("up_proj/kernel", PartitionSpec(("fsdp", "sp"))),
            ("shared_expert_gate/kernel", PartitionSpec(("fsdp", "sp"))),
            ("gate/kernel", PartitionSpec(("fsdp", "sp"))),

            ("input_layernorm/kernel", PartitionSpec(None)),
            ("post_attention_layernorm/kernel", PartitionSpec(None)),

            ("model/norm/kernel", PartitionSpec(None)),
            ("lm_head/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
            (".*", PartitionSpec(("fsdp", "sp"))),
        )

    def add_jax_args(
            self,
            gradient_checkpointing: str = "nothing_saveable",
            bits: Optional[int] = None,
            **kwargs,
    ):
        """
        The add_jax_args function adds the following arguments to the Transformer class:

        :param self: Refer to the current object

        :param gradient_checkpointing: str: Control the amount of memory used by jax
        :param bits: Optional[int]: Determine the number of bits used in the quantization
        :return: The following:

        """
        self.gradient_checkpointing = gradient_checkpointing
        self.bits = bits

    @staticmethod
    def get_weight_decay_exclusions():
        return tuple()

    @staticmethod
    def rng_keys():
        return "params", "dropout"

add_jax_args(gradient_checkpointing='nothing_saveable', bits=None, **kwargs)

The add_jax_args function adds the following arguments to the Transformer class:

Parameters:

Name Type Description Default
self

Refer to the current object

 required
gradient_checkpointing str

str: Control the amount of memory used by jax

 'nothing_saveable'
bits Optional[int]

Optional[int]: Determine the number of bits used in the quantization

 None

Returns:

Type Description

The following:
Source code in src/python/easydel/modules/qwen2_moe/configuration_qwen2_moe.py 
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
def add_jax_args(
        self,
        gradient_checkpointing: str = "nothing_saveable",
        bits: Optional[int] = None,
        **kwargs,
):
    """
    The add_jax_args function adds the following arguments to the Transformer class:

    :param self: Refer to the current object

    :param gradient_checkpointing: str: Control the amount of memory used by jax
    :param bits: Optional[int]: Determine the number of bits used in the quantization
    :return: The following:

    """
    self.gradient_checkpointing = gradient_checkpointing
    self.bits = bits

get_partition_rules(fully_sharded_data_parallel=True)

The get_partition_rules function is used to define the partitioning scheme for a model. It returns a list of tuples, where each tuple contains two elements: 1) A regex string that matches the name of one or more parameters in the model. 2) A PartitionScheme object that defines how those parameters should be partitioned across devices.

Parameters:

Name Type Description Default
fully_sharded_data_parallel bool

bool: Determine whether to partition the model fully or not

 True

Returns:

Type Description

A list of tuples
Source code in src/python/easydel/modules/qwen2_moe/configuration_qwen2_moe.py 
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
def get_partition_rules(self, fully_sharded_data_parallel: bool = True):
    """
    The get_partition_rules function is used to define the partitioning scheme for a model.
    It returns a list of tuples, where each tuple contains two elements:
        1) A regex string that matches the name of one or more parameters in the model.
        2) A PartitionScheme object that defines how those parameters should be partitioned across devices.

    :param fully_sharded_data_parallel: bool: Determine whether to partition the model fully or not
    :return: A list of tuples

    """
    return (

        ("model/embed_tokens/embedding", PartitionSpec("tp", ("fsdp", "sp"))),

        ("self_attn/(q_proj|k_proj|v_proj)/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        ("self_attn/o_proj/kernel", PartitionSpec("tp", ("sp", "fsdp"))),

        ("gate_proj/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        ("down_proj/kernel", PartitionSpec("tp", ("fsdp", "sp"))),
        ("up_proj/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        ("shared_expert_gate/kernel", PartitionSpec(("fsdp", "sp"))),
        ("gate/kernel", PartitionSpec(("fsdp", "sp"))),

        ("input_layernorm/kernel", PartitionSpec(None)),
        ("post_attention_layernorm/kernel", PartitionSpec(None)),

        ("model/norm/kernel", PartitionSpec(None)),
        ("lm_head/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        (".*", PartitionSpec(None)),
    ) if not fully_sharded_data_parallel else (

        ("model/embed_tokens/embedding", PartitionSpec("tp", ("fsdp", "sp"))),

        ("self_attn/(q_proj|k_proj|v_proj)/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        ("self_attn/o_proj/kernel", PartitionSpec("tp", ("sp", "fsdp"))),

        ("gate_proj/kernel", PartitionSpec(("fsdp", "sp"))),
        ("down_proj/kernel", PartitionSpec(("fsdp", "sp"))),
        ("up_proj/kernel", PartitionSpec(("fsdp", "sp"))),
        ("shared_expert_gate/kernel", PartitionSpec(("fsdp", "sp"))),
        ("gate/kernel", PartitionSpec(("fsdp", "sp"))),

        ("input_layernorm/kernel", PartitionSpec(None)),
        ("post_attention_layernorm/kernel", PartitionSpec(None)),

        ("model/norm/kernel", PartitionSpec(None)),
        ("lm_head/kernel", PartitionSpec(("fsdp", "sp"), "tp")),
        (".*", PartitionSpec(("fsdp", "sp"))),
    )