| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2024 HuggingFace Inc. team. All rights reserved. |
| 5 | # Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. |
| 6 | # |
| 7 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 8 | # you may not use this file except in compliance with the License. |
| 9 | # You may obtain a copy of the License at |
| 10 | # |
| 11 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 12 | # |
| 13 | # Unless required by applicable law or agreed to in writing, software |
| 14 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 15 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 16 | # See the License for the specific language governing permissions and |
| 17 | # limitations under the License. |
| 18 | """NemotronH model configuration.""" |
| 19 | |
| 20 | import regex as re |
| 21 | from transformers.configuration_utils import PreTrainedConfig |
| 22 | from transformers.utils import logging |
| 23 | |
| 24 | logger = logging.get_logger(__name__) |
| 25 | |
| 26 | |
| 27 | class NemotronHConfig(PreTrainedConfig): |
| 28 | r"""This is the configuration class to store the configuration of a |
| 29 | [`NemotronHModel`]. It is used to instantiate a NemotronH model according |
| 30 | to the specified arguments, defining the model architecture. Instantiating |
| 31 | a configuration with the defaults will yield a similar configuration to |
| 32 | that of the NemotronH-v0.1 model. |
| 33 | |
| 34 | Args: |
| 35 | vocab_size (`int`, *optional*, defaults to 131072): |
| 36 | Vocabulary size of the NemotronH model. Defines the number of |
| 37 | different tokens that can be represented by the `inputs_ids` |
| 38 | passed when calling [`NemotronHModel`] |
| 39 | tie_word_embeddings (`bool`, *optional*, defaults to `False`): |
| 40 | Whether the model's input and output word embeddings should be |
| 41 | tied. Note that this is only relevant if the model has an output |
| 42 | word embedding layer. |
| 43 | hidden_size (`int`, *optional*, defaults to 4096): |
| 44 | Dimension of the hidden representations. |
| 45 | intermediate_size (`int`, *optional*, defaults to 21504): |
| 46 | Dimension of the MLP representations. |
| 47 | num_hidden_layers (`int`, *optional*, defaults to 52): |
| 48 | Number of hidden layers in the Transformer encoder. |
| 49 | hybrid_override_pattern (`str`, *optional*, defaults to |
| 50 | `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`): |
| 51 | The pattern of the hybrid model. The pattern is a string of |
| 52 | characters where each character represents |
| 53 | M: Mamba2, *: Attention, -: MLP |
| 54 | mtp_hybrid_override_pattern (`str`, *optional*, defaults to `"*E"`): |
| 55 | The pattern of the MTP layers. |
| 56 | num_attention_heads (`int`, *optional*, defaults to 32): |
| 57 | Number of attention heads for each attention layer in the |
| 58 | Transformer encoder. |
| 59 | attention_head_dim (`int`, *optional*, defaults to 128): |
| 60 | Dimension of each attention head. |
| 61 | num_key_value_heads (`int`, *optional*, defaults to 8): |
| 62 | This is the number of key_value heads that should be used to |
| 63 | implement Grouped Query Attention. If |
| 64 | `num_key_value_heads=num_attention_heads`, the model will use |
| 65 | Multi Head Attention (MHA), if `num_key_value_heads=1` the model |
| 66 | will use Multi Query Attention (MQA) otherwise GQA is used. |
| 67 | mlp_hidden_act (`str`, *optional*, defaults to "relu2"): |
| 68 | The non-linear activation function in the MLP layers. |
| 69 | attention_bias (`bool`, *optional*, defaults to `False`): |
| 70 | Whether to use bias in attention layers. |
| 71 | mlp_bias (`bool`, *optional*, defaults to `False`): |
| 72 | Whether to use bias in MLP layers. |
| 73 | use_bias (`bool`, *optional*, defaults to `False`): |
| 74 | Whether to use bias in the model. |
| 75 | initializer_range (`float`, *optional*, defaults to 0.02): |
| 76 | The standard deviation of the truncated_normal_initializer for |
| 77 | initializing all weight matrices. |
| 78 | layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): |
| 79 | The epsilon used by the layer normalization layers. |
| 80 | residual_in_fp32 (`bool`, *optional*, defaults to `False`): |
| 81 | Whether or not residuals should be in `float32`. If set to `False` |
| 82 | residuals will keep the same `dtype` as the rest of the model. |
| 83 | use_cache (`bool`, *optional*, defaults to `True`): |
| 84 | Whether or not the model should return the last key/values |
| 85 | attentions (not used by all models). Only relevant if |
| 86 | `config.is_decoder=True`. |
| 87 | num_logits_to_keep (`int` or `None`, *optional*, defaults to 1): |
| 88 | Number of prompt logits to calculate during generation. If `None`, |
| 89 | all logits will be calculated. If an integer value, only last |
| 90 | `num_logits_to_keep` logits will be calculated. |
| 91 | pad_token_id (`int`, *optional*, defaults to 0): |
| 92 | The id of the padding token. |
| 93 | bos_token_id (`int`, *optional*, defaults to 1): |
| 94 | The id of the "beginning-of-sequence" token. |
| 95 | eos_token_id (`int`, *optional*, defaults to 2): |
| 96 | The id of the "end-of-sequence" token. |
| 97 | sliding_window (`int`, *optional*, defaults to None): |
| 98 | Sliding window attention window size. |
| 99 | max_position_embeddings (`int`, *optional*, defaults to 4096): |
| 100 | The maximum sequence length that this model might ever be used |
| 101 | with. |
| 102 | attention_dropout (`float`, *optional*, defaults to 0.0): |
| 103 | The dropout ratio for the attention probabilities. |
| 104 | hidden_dropout (`float`, *optional*, defaults to 0.0): |
| 105 | The dropout ratio for the hidden states. |
| 106 | use_mamba_kernels (`bool`, *optional*, defaults to `True`): |
| 107 | Flag indicating whether or not to use the fast mamba kernels. |
| 108 | These are available only if `mamba-ssm` and `causal-conv1d` |
| 109 | are installed, and the mamba modules are running on a CUDA device. |
| 110 | ssm_state_size (`int`, *optional*, defaults to 128): |
| 111 | The dimension of the mamba state space latents. |
| 112 | mamba_num_heads (`int`, *optional*, defaults to 128): |
| 113 | Number of heads in Mamba layers. |
| 114 | mamba_n_groups (`int`, *optional*, defaults to 8): |
| 115 | Number of groups in Mamba layers. |
| 116 | mamba_head_dim (`int`, *optional*, defaults to 64): |
| 117 | Dimension of each Mamba head. |
| 118 | mamba_d_conv (`int`, *optional*, defaults to 4): |
| 119 | The size of the mamba convolution kernel. |
| 120 | mamba_expand (`int`, *optional*, defaults to 2): |
| 121 | Expanding factor used to determine the mamba intermediate size. |
| 122 | mamba_hidden_act (`str`, *optional*, defaults to "silu"): |
| 123 | The non-linear activation function in the Mamba layers. |
| 124 | mamba_dt_min (`float`, *optional*, defaults to 0.001): |
| 125 | Minimum value for the time step in Mamba. |
| 126 | mamba_dt_max (`float`, *optional*, defaults to 0.1): |
| 127 | Maximum value for the time step in Mamba. |
| 128 | mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))): |
| 129 | Limits for the time step in Mamba. |
| 130 | mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4): |
| 131 | Floor value for time step initialization in Mamba. |
| 132 | mamba_conv_bias (`bool`, *optional*, defaults to `True`): |
| 133 | Whether to use bias in the convolution layer of the mamba mixer |
| 134 | block. |
| 135 | mamba_proj_bias (`bool`, *optional*, defaults to `False`): |
| 136 | Whether to use bias in the input and output projections of the |
| 137 | mamba mixer block. |
| 138 | mamba_chunk_size (`int`, *optional*, defaults to 256): |
| 139 | Size of chunks for Mamba processing. |
| 140 | rescale_prenorm_residual (`bool`, *optional*, defaults to `True`): |
| 141 | Whether to rescale the pre-normalization residual connections. |
| 142 | |
| 143 | """ |
| 144 | |
| 145 | model_type = "nemotron_h" |
| 146 | keys_to_ignore_at_inference = ["past_key_values"] |
| 147 | |
| 148 | def __init__( |
| 149 | self, |
| 150 | vocab_size=131072, |
| 151 | tie_word_embeddings=False, |
| 152 | hidden_size=4096, |
| 153 | intermediate_size=21504, |
| 154 | num_hidden_layers=52, |
| 155 | hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-", |
| 156 | mtp_hybrid_override_pattern="*E", |
| 157 | num_attention_heads=32, |
| 158 | head_dim=128, |
| 159 | num_key_value_heads=8, # nemo: num_query_groups |
| 160 | mlp_hidden_act="relu2", |
| 161 | attention_bias=False, |
| 162 | mlp_bias=False, |
| 163 | use_bias=False, |
| 164 | initializer_range=0.02, # nemo: init_method_std |
| 165 | layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon |
| 166 | residual_in_fp32=False, # Megatron Core default value |
| 167 | use_cache=True, |
| 168 | num_logits_to_keep=1, |
| 169 | pad_token_id=0, |
| 170 | bos_token_id=1, |
| 171 | eos_token_id=2, |
| 172 | sliding_window=None, |
| 173 | max_position_embeddings=4096, |
| 174 | attention_dropout=0.0, |
| 175 | hidden_dropout=0.0, # * ADDED |
| 176 | use_mamba_kernels=True, |
| 177 | ssm_state_size=128, # mamba_state_size |
| 178 | mamba_num_heads=128, |
| 179 | mamba_n_groups=8, # nemo: mamba_ssm_ngroups = num_heads |
| 180 | mamba_head_dim=64, |
| 181 | mamba_d_conv=4, |
| 182 | mamba_expand=2, |
| 183 | mamba_hidden_act="silu", |
| 184 | mamba_dt_min=0.001, |
| 185 | mamba_dt_max=0.1, |
| 186 | mamba_dt_limit=(0.0, float("inf")), |
| 187 | mamba_dt_init_floor=1e-4, |
| 188 | mamba_conv_bias=True, |
| 189 | mamba_proj_bias=False, |
| 190 | mamba_chunk_size=256, |
| 191 | rescale_prenorm_residual=True, |
| 192 | n_routed_experts=8, |
| 193 | n_shared_experts=1, |
| 194 | moe_intermediate_size=7688, |
| 195 | moe_shared_expert_intermediate_size=7688, |
| 196 | moe_latent_size=None, |
| 197 | num_experts_per_tok=2, |
| 198 | routed_scaling_factor=1.0, |
| 199 | n_group=1, |
| 200 | topk_group=1, |
| 201 | norm_topk_prob=True, |
| 202 | **kwargs, |
| 203 | ): |
| 204 | self.vocab_size = vocab_size |
| 205 | self.tie_word_embeddings = tie_word_embeddings |
| 206 | self.hidden_size = hidden_size |
| 207 | self.intermediate_size = intermediate_size |
| 208 | self.num_hidden_layers = num_hidden_layers |
| 209 | self.hybrid_override_pattern = hybrid_override_pattern |
| 210 | self.mtp_hybrid_override_pattern = mtp_hybrid_override_pattern |
| 211 | self.num_attention_heads = num_attention_heads |
| 212 | self.head_dim = head_dim |
| 213 | self.sliding_window = sliding_window |
| 214 | self.max_position_embeddings = max_position_embeddings |
| 215 | self.attention_dropout = attention_dropout |
| 216 | self.hidden_dropout = hidden_dropout |
| 217 | |
| 218 | # Validate hybrid_override_pattern |
| 219 | # M: Mamba2, *: Attention, -: MLP |
| 220 | assert len(self.hybrid_override_pattern) == self.num_hidden_layers, ( |
| 221 | "hybrid_override_pattern must have same length as num_hidden_layers" |
| 222 | ) |
| 223 | assert re.match(r"^[*-ME]+$", self.hybrid_override_pattern), ( |
| 224 | "hybrid_override_pattern must only contain characters 'M', '*', '-', or 'E'" |
| 225 | ) |
| 226 | # for backward compatibility |
| 227 | if num_key_value_heads is None: |
| 228 | num_key_value_heads = num_attention_heads |
| 229 | |
| 230 | self.num_key_value_heads = num_key_value_heads |
| 231 | self.mlp_hidden_act = mlp_hidden_act |
| 232 | self.attention_bias = attention_bias |
| 233 | self.mlp_bias = mlp_bias |
| 234 | self.use_bias = use_bias |
| 235 | self.initializer_range = initializer_range |
| 236 | self.layer_norm_epsilon = layer_norm_epsilon |
| 237 | self.residual_in_fp32 = residual_in_fp32 |
| 238 | |
| 239 | self.use_cache = use_cache |
| 240 | self.num_logits_to_keep = num_logits_to_keep |
| 241 | |
| 242 | self.use_mamba_kernels = use_mamba_kernels |
| 243 | self.n_groups = mamba_n_groups |
| 244 | self.mamba_head_dim = mamba_head_dim |
| 245 | self.ssm_state_size = ssm_state_size |
| 246 | self.mamba_num_heads = mamba_num_heads |
| 247 | self.conv_kernel = mamba_d_conv |
| 248 | self.expand = mamba_expand |
| 249 | self.mamba_hidden_act = mamba_hidden_act |
| 250 | self.time_step_min = mamba_dt_min |
| 251 | self.time_step_max = mamba_dt_max |
| 252 | self.time_step_limit = mamba_dt_limit |
| 253 | self.time_step_floor = mamba_dt_init_floor |
| 254 | self.use_conv_bias = mamba_conv_bias |
| 255 | self.mamba_proj_bias = mamba_proj_bias |
| 256 | self.chunk_size = mamba_chunk_size |
| 257 | self.rescale_prenorm_residual = rescale_prenorm_residual |
| 258 | self.n_routed_experts = n_routed_experts |
| 259 | self.n_shared_experts = n_shared_experts |
| 260 | self.moe_intermediate_size = moe_intermediate_size |
| 261 | self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size # noqa: E501 |
| 262 | self.moe_latent_size = moe_latent_size |
| 263 | self.num_experts_per_tok = num_experts_per_tok |
| 264 | self.routed_scaling_factor = routed_scaling_factor |
| 265 | self.n_group = n_group |
| 266 | self.topk_group = topk_group |
| 267 | self.norm_topk_prob = norm_topk_prob |
| 268 | |
| 269 | super().__init__( |
| 270 | pad_token_id=pad_token_id, |
| 271 | bos_token_id=bos_token_id, |
| 272 | eos_token_id=eos_token_id, |
| 273 | tie_word_embeddings=tie_word_embeddings, |
| 274 | **kwargs, |
| 275 | ) |
| 276 | |
| 277 | @property |
| 278 | def layers_block_type(self): |
| 279 | return [ |
| 280 | "mamba" |
| 281 | if self.hybrid_override_pattern[i] == "M" |
| 282 | else "attention" |
| 283 | if self.hybrid_override_pattern[i] == "*" |
| 284 | else "mlp" |
| 285 | if self.hybrid_override_pattern[i] == "-" |
| 286 | else "moe" |
| 287 | for i in range(self.num_hidden_layers) |
| 288 | ] |