FrontiersMind/Nandi-Mini-150M-Instruct

1#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨2#           This file was automatically generated from src/transformers/models/nandi/modular_nandi.py.3#               Do NOT edit this file manually as any edits will be overwritten by the generation of4#             the file from the modular. If any change should be done, please apply the change to the5#                          modular_nandi.py file directly. One of our CI enforces this.6#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨7# Copyright 2026 The HuggingFace Inc. team. All rights reserved.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 at12#13#     http://www.apache.org/licenses/LICENSE-2.014#15# Unless required by applicable law or agreed to in writing, software16# 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 and19# limitations under the License.20 21from collections.abc import Callable22 23import torch24import torch.nn as nn25 26from transformers.activations import ACT2FN27from transformers.cache_utils import Cache, DynamicCache, DynamicLayer28from transformers.generation import GenerationMixin29from transformers.integrations import use_kernel_forward_from_hub30from transformers.masking_utils import create_causal_mask31from transformers.modeling_layers import GradientCheckpointingLayer32from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast33from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update34from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel35from transformers.processing_utils import Unpack36from transformers.utils import TransformersKwargs, auto_docstring37from transformers.utils.deprecation import deprecate_kwarg38from transformers.utils.generic import can_return_tuple, merge_with_config_defaults39from transformers.utils.output_capturing import capture_outputs40from .configuration_nandi import NandiConfig41 42 43@use_kernel_forward_from_hub("RMSNorm")44class NandiRMSNorm(nn.Module):45    def __init__(self, hidden_size, eps=1e-6):46        super().__init__()47        self.weight = nn.Parameter(torch.ones(hidden_size))48        self.variance_epsilon = eps49 50    def forward(self, hidden_states):51        input_dtype = hidden_states.dtype52        hidden_states = hidden_states.to(torch.float32)53        variance = hidden_states.pow(2).mean(-1, keepdim=True)54        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)55        return self.weight * hidden_states.to(input_dtype)56 57    def extra_repr(self):58        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"59 60 61class NandiRotaryEmbedding(nn.Module):62    inv_freq: torch.Tensor63 64    def __init__(self, config: NandiConfig, device=None):65        super().__init__()66        self.max_seq_len_cached = config.max_position_embeddings67        self.original_max_seq_len = config.max_position_embeddings68 69        self.config = config70        self.rope_type = self.config.rope_parameters.get("rope_type", "default")71        rope_init_fn: Callable = self.compute_default_rope_parameters72        if self.rope_type != "default":73            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]74        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)75 76        self.register_buffer("inv_freq", inv_freq, persistent=False)77        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)78 79    @staticmethod80    def compute_default_rope_parameters(81        config: NandiConfig | None = None,82        device: torch.device | None = None,83        seq_len: int | None = None,84    ) -> tuple[torch.Tensor, float]:85        del seq_len86        base = config.rope_parameters["rope_theta"]87        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads88        attention_factor = 1.089        inv_freq = 1.0 / (90            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)91        )92        return inv_freq, attention_factor93 94    @torch.no_grad()95    @dynamic_rope_update96    def forward(self, x, position_ids):97        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)98        position_ids_expanded = position_ids[:, None, :].float()99 100        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"101        with torch.autocast(device_type=device_type, enabled=False):102            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)103            emb = torch.cat((freqs, freqs), dim=-1)104            cos = emb.cos() * self.attention_scaling105            sin = emb.sin() * self.attention_scaling106 107        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)108 109 110def rotate_half(x):111    """Rotates half the hidden dims of the input."""112    x1 = x[..., : x.shape[-1] // 2]113    x2 = x[..., x.shape[-1] // 2 :]114    return torch.cat((-x2, x1), dim=-1)115 116 117def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):118    del position_ids119    cos = cos.unsqueeze(unsqueeze_dim)120    sin = sin.unsqueeze(unsqueeze_dim)121    q_embed = (q * cos) + (rotate_half(q) * sin)122    k_embed = (k * cos) + (rotate_half(k) * sin)123    return q_embed, k_embed124 125 126def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:127    batch, num_key_value_heads, slen, head_dim = hidden_states.shape128    if n_rep == 1:129        return hidden_states130    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)131    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)132 133 134def eager_attention_forward(135    module: nn.Module,136    query: torch.Tensor,137    key: torch.Tensor,138    value: torch.Tensor,139    attention_mask: torch.Tensor | None,140    scaling: float,141    dropout: float = 0.0,142    **kwargs: Unpack[TransformersKwargs],143):144    del kwargs145    key_states = repeat_kv(key, module.num_key_value_groups)146    value_states = repeat_kv(value, module.num_key_value_groups)147 148    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling149    if attention_mask is not None:150        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]151        attn_weights = attn_weights + causal_mask152 153    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)154    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)155    attn_output = torch.matmul(attn_weights, value_states)156    attn_output = attn_output.transpose(1, 2).contiguous()157 158    return attn_output, attn_weights159 160 161class NandiAttention(nn.Module):162    def __init__(self, config: NandiConfig, layer_idx: int):163        super().__init__()164        self.config = config165        self.layer_idx = layer_idx166        self.head_dim = config.head_dim167        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads168        self.scaling = self.head_dim**-0.5169        self.attention_dropout = config.attention_dropout170        self.is_causal = True171 172        self.q_proj = nn.Linear(173            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias174        )175        self.k_proj = nn.Linear(176            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias177        )178        self.v_proj = nn.Linear(179            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias180        )181        self.o_proj = nn.Linear(182            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias183        )184 185    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")186    def forward(187        self,188        hidden_states: torch.Tensor,189        position_embeddings: tuple[torch.Tensor, torch.Tensor],190        attention_mask: torch.Tensor | None,191        past_key_values: Cache | None = None,192        **kwargs: Unpack[TransformersKwargs],193    ) -> tuple[torch.Tensor, torch.Tensor]:194        input_shape = hidden_states.shape[:-1]195        hidden_shape = (*input_shape, -1, self.head_dim)196 197        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)198        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)199        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)200 201        cos, sin = position_embeddings202        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)203 204        if past_key_values is not None:205            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)206 207        attention_interface: Callable = eager_attention_forward208        if self.config._attn_implementation != "eager":209            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]210 211        attn_output, attn_weights = attention_interface(212            self,213            query_states,214            key_states,215            value_states,216            attention_mask,217            dropout=0.0 if not self.training else self.attention_dropout,218            scaling=self.scaling,219            **kwargs,220        )221 222        attn_output = attn_output.reshape(*input_shape, -1).contiguous()223        attn_output = self.o_proj(attn_output)224        return attn_output, attn_weights225 226 227class NandiMLP(nn.Module):228    def __init__(self, config):229        super().__init__()230        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=config.mlp_bias)231        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=config.mlp_bias)232        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=config.mlp_bias)233        self.act_fn = ACT2FN[config.hidden_act]234 235    def forward(self, x):236        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))237 238 239class NandiDecoderLayer(GradientCheckpointingLayer):240    def __init__(self, config: NandiConfig, layer_idx: int):241        super().__init__()242        self.hidden_size = config.hidden_size243        self.self_attn = NandiAttention(config=config, layer_idx=layer_idx)244        self.mlp = NandiMLP(config)245        self.input_layernorm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)246        self.post_attention_layernorm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)247 248    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")249    def forward(250        self,251        hidden_states: torch.Tensor,252        attention_mask: torch.Tensor | None = None,253        position_ids: torch.LongTensor | None = None,254        past_key_values: Cache | None = None,255        use_cache: bool | None = False,256        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,257        **kwargs: Unpack[TransformersKwargs],258    ) -> torch.Tensor:259        residual = hidden_states260        hidden_states = self.input_layernorm(hidden_states)261 262        hidden_states, _ = self.self_attn(263            hidden_states=hidden_states,264            attention_mask=attention_mask,265            position_ids=position_ids,266            past_key_values=past_key_values,267            use_cache=use_cache,268            position_embeddings=position_embeddings,269            **kwargs,270        )271        hidden_states = residual + hidden_states272 273        residual = hidden_states274        hidden_states = self.post_attention_layernorm(hidden_states)275        hidden_states = self.mlp(hidden_states)276        hidden_states = residual + hidden_states277        return hidden_states278 279 280class _VirtualLayerCache:281    """Proxy that shifts cache layer indices by `offset` to give each repeat its own virtual slots."""282 283    def __init__(self, cache: Cache, offset: int):284        self._cache = cache285        self._offset = offset286 287    def __getattr__(self, name):288        return getattr(self._cache, name)289 290    def update(self, key_states, value_states, layer_idx, cache_kwargs=None):291        virtual_idx = layer_idx + self._offset292        # grow the backing cache if generate() pre-allocated fewer slots than needed293        while len(self._cache.layers) <= virtual_idx:294            self._cache.layers.append(DynamicLayer())295        return self._cache.update(key_states, value_states, virtual_idx, cache_kwargs)296 297    def get_seq_length(self, layer_idx: int = 0) -> int:298        return self._cache.get_seq_length(layer_idx + self._offset)299 300 301@auto_docstring302class NandiPreTrainedModel(PreTrainedModel):303    config: NandiConfig304    base_model_prefix = "model"305    supports_gradient_checkpointing = True306    _no_split_modules = ["NandiDecoderLayer"]307    _skip_keys_device_placement = ["past_key_values"]308    _supports_flash_attn = True309    _supports_sdpa = True310    _supports_flex_attn = True311    _can_compile_fullgraph = True312    _supports_attention_backend = True313    _can_record_outputs = {314        "hidden_states": NandiDecoderLayer,315        "attentions": NandiAttention,316    }317 318    def __init__(self, config: NandiConfig):319        super().__init__(config)320 321 322@auto_docstring323class NandiModel(NandiPreTrainedModel):324    def __init__(self, config: NandiConfig):325        super().__init__(config)326        self.padding_idx = config.pad_token_id327        self.vocab_size = config.vocab_size328        embedding_dim = config.embedding_rank if config.factorized_embedding else config.hidden_size329 330        self.embed_tokens = nn.Embedding(config.vocab_size, embedding_dim, self.padding_idx)331        self.embedding_proj = (332            nn.Linear(config.embedding_rank, config.hidden_size, bias=False) if config.factorized_embedding else None333        )334        self.layers = nn.ModuleList(335            [NandiDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]336        )337        self.norm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)338        self.rotary_emb = NandiRotaryEmbedding(config=config)339        self.gradient_checkpointing = False340 341        self.post_init()342 343    @merge_with_config_defaults344    @capture_outputs345    @auto_docstring346    def forward(347        self,348        input_ids: torch.LongTensor | None = None,349        attention_mask: torch.Tensor | None = None,350        position_ids: torch.LongTensor | None = None,351        past_key_values: Cache | None = None,352        inputs_embeds: torch.FloatTensor | None = None,353        use_cache: bool | None = None,354        **kwargs: Unpack[TransformersKwargs],355    ) -> BaseModelOutputWithPast:356        if (input_ids is None) ^ (inputs_embeds is not None):357            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")358 359        if inputs_embeds is None:360            inputs_embeds = self.embed_tokens(input_ids)361 362        if self.embedding_proj is not None:363            inputs_embeds = self.embedding_proj(inputs_embeds)364 365        repeats = self.config.layer_sharing_repeats if self.config.layer_sharing else 1366 367        if use_cache and past_key_values is None:368            # Use lazy DynamicCache (no config) so it grows to accommodate369            # num_hidden_layers * repeats virtual slots for layer-sharing.370            past_key_values = DynamicCache()371 372        if position_ids is None:373            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0374            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens375            position_ids = position_ids.unsqueeze(0)376 377        causal_mask = create_causal_mask(378            config=self.config,379            inputs_embeds=inputs_embeds,380            attention_mask=attention_mask,381            past_key_values=past_key_values,382            position_ids=position_ids,383        )384 385        hidden_states = inputs_embeds386        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)387 388        for decoder_layer in self.layers[: self.config.num_hidden_layers]:389            for repeat_idx in range(repeats):390                # Each repeat gets its own virtual cache slots offset by num_hidden_layers,391                # so repeat 0 uses slots 0..N-1 and repeat 1 uses slots N..2N-1, etc.392                repeat_cache = (393                    _VirtualLayerCache(past_key_values, repeat_idx * self.config.num_hidden_layers)394                    if (past_key_values is not None and repeat_idx > 0)395                    else past_key_values396                )397                hidden_states = decoder_layer(398                    hidden_states,399                    attention_mask=causal_mask,400                    position_embeddings=position_embeddings,401                    position_ids=position_ids,402                    past_key_values=repeat_cache,403                    use_cache=use_cache,404                    **kwargs,405                )406 407        hidden_states = self.norm(hidden_states)408        return BaseModelOutputWithPast(409            last_hidden_state=hidden_states,410            past_key_values=past_key_values,411        )412 413 414@auto_docstring415class NandiForCausalLM(NandiPreTrainedModel, GenerationMixin):416    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}417    _tp_plan = {"lm_head": "colwise_gather_output"}418    _pp_plan = {419        "lm_head_proj": (["hidden_states"], ["hidden_states"]),420        "lm_head": (["hidden_states"], ["logits"]),421    }422 423    def __init__(self, config):424        super().__init__(config)425        self.model = NandiModel(config)426        self.vocab_size = config.vocab_size427 428        lm_head_in_features = config.embedding_rank if config.factorized_embedding else config.hidden_size429        self.lm_head_proj = (430            nn.Linear(config.hidden_size, config.embedding_rank, bias=False) if config.factorized_embedding else None431        )432        self.lm_head = nn.Linear(lm_head_in_features, config.vocab_size, bias=False)433 434        self.post_init()435 436    @can_return_tuple437    @auto_docstring438    def forward(439        self,440        input_ids: torch.LongTensor | None = None,441        attention_mask: torch.Tensor | None = None,442        position_ids: torch.LongTensor | None = None,443        past_key_values: Cache | None = None,444        inputs_embeds: torch.FloatTensor | None = None,445        labels: torch.LongTensor | None = None,446        use_cache: bool | None = None,447        logits_to_keep: int | torch.Tensor = 0,448        **kwargs: Unpack[TransformersKwargs],449    ) -> CausalLMOutputWithPast:450        outputs: BaseModelOutputWithPast = self.model(451            input_ids=input_ids,452            attention_mask=attention_mask,453            position_ids=position_ids,454            past_key_values=past_key_values,455            inputs_embeds=inputs_embeds,456            use_cache=use_cache,457            **kwargs,458        )459 460        hidden_states = outputs.last_hidden_state461        if self.lm_head_proj is not None:462            hidden_states = self.lm_head_proj(hidden_states)463 464        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep465        logits = self.lm_head(hidden_states[:, slice_indices, :])466 467        loss = None468        if labels is not None:469            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)470 471        return CausalLMOutputWithPast(472            loss=loss,473            logits=logits,474            past_key_values=outputs.past_key_values,475            hidden_states=outputs.hidden_states,476            attentions=outputs.attentions,477        )478 479 480__all__ = ["NandiPreTrainedModel", "NandiModel", "NandiForCausalLM"]481