Aluode/PerceptionLabPortable

1#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨2#           This file was automatically generated from src/transformers/models/olmo2/modular_olmo2.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_olmo2.py file directly. One of our CI enforces this.6#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨7from typing import Callable, Optional, Union8 9import torch10import torch.nn as nn11 12from transformers.utils.generic import TransformersKwargs13 14from ...activations import ACT2FN15from ...cache_utils import Cache, DynamicCache16from ...generation import GenerationMixin17from ...integrations import use_kernel_forward_from_hub18from ...masking_utils import create_causal_mask19from ...modeling_layers import GradientCheckpointingLayer20from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast21from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update22from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel23from ...processing_utils import Unpack24from ...utils import auto_docstring, can_return_tuple25from ...utils.deprecation import deprecate_kwarg26from ...utils.generic import check_model_inputs27from .configuration_olmo2 import Olmo2Config28 29 30@use_kernel_forward_from_hub("RMSNorm")31class Olmo2RMSNorm(nn.Module):32    def __init__(self, hidden_size, eps=1e-6):33        """34        Olmo2RMSNorm is equivalent to T5LayerNorm35        """36        super().__init__()37        self.weight = nn.Parameter(torch.ones(hidden_size))38        self.variance_epsilon = eps39 40    def forward(self, hidden_states):41        input_dtype = hidden_states.dtype42        hidden_states = hidden_states.to(torch.float32)43        variance = hidden_states.pow(2).mean(-1, keepdim=True)44        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)45        return (self.weight * hidden_states).to(input_dtype)46 47    def extra_repr(self):48        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"49 50 51def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:52    """53    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,54    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)55    """56    batch, num_key_value_heads, slen, head_dim = hidden_states.shape57    if n_rep == 1:58        return hidden_states59    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)60    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)61 62 63def eager_attention_forward(64    module: nn.Module,65    query: torch.Tensor,66    key: torch.Tensor,67    value: torch.Tensor,68    attention_mask: Optional[torch.Tensor],69    scaling: float,70    dropout: float = 0.0,71    **kwargs: Unpack[TransformersKwargs],72):73    key_states = repeat_kv(key, module.num_key_value_groups)74    value_states = repeat_kv(value, module.num_key_value_groups)75 76    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling77    if attention_mask is not None:78        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]79        attn_weights = attn_weights + causal_mask80 81    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)82    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)83    attn_output = torch.matmul(attn_weights, value_states)84    attn_output = attn_output.transpose(1, 2).contiguous()85 86    return attn_output, attn_weights87 88 89def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):90    """Applies Rotary Position Embedding to the query and key tensors.91 92    Args:93        q (`torch.Tensor`): The query tensor.94        k (`torch.Tensor`): The key tensor.95        cos (`torch.Tensor`): The cosine part of the rotary embedding.96        sin (`torch.Tensor`): The sine part of the rotary embedding.97        position_ids (`torch.Tensor`, *optional*):98            Deprecated and unused.99        unsqueeze_dim (`int`, *optional*, defaults to 1):100            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and101            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note102            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and103            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes104            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have105            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.106    Returns:107        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.108    """109    q_type, k_type = q.dtype, k.dtype110    cos = cos.unsqueeze(unsqueeze_dim)111    sin = sin.unsqueeze(unsqueeze_dim)112    q_embed = (q * cos) + (rotate_half(q) * sin)113    k_embed = (k * cos) + (rotate_half(k) * sin)114    return q_embed.to(q_type), k_embed.to(k_type)115 116 117def rotate_half(x):118    """Rotates half the hidden dims of the input."""119    x1 = x[..., : x.shape[-1] // 2]120    x2 = x[..., x.shape[-1] // 2 :]121    return torch.cat((-x2, x1), dim=-1)122 123 124class Olmo2Attention(nn.Module):125    """Multi-headed attention from 'Attention Is All You Need' paper"""126 127    def __init__(self, config: Olmo2Config, layer_idx: Optional[int] = None):128        super().__init__()129        self.config = config130        self.layer_idx = layer_idx131        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)132        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads133        self.scaling = self.head_dim**-0.5134        self.attention_dropout = config.attention_dropout135        self.is_causal = True136 137        self.q_proj = nn.Linear(138            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias139        )140        self.k_proj = nn.Linear(141            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias142        )143        self.v_proj = nn.Linear(144            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias145        )146        self.o_proj = nn.Linear(147            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias148        )149        self.q_norm = Olmo2RMSNorm(config.num_attention_heads * self.head_dim, config.rms_norm_eps)150        self.k_norm = Olmo2RMSNorm(config.num_key_value_heads * self.head_dim, config.rms_norm_eps)151 152    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")153    def forward(154        self,155        hidden_states: torch.Tensor,156        position_embeddings: tuple[torch.Tensor, torch.Tensor],157        attention_mask: Optional[torch.Tensor],158        past_key_values: Optional[Cache] = None,159        cache_position: Optional[torch.LongTensor] = None,160        **kwargs: Unpack[TransformersKwargs],161    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:162        input_shape = hidden_states.shape[:-1]163        hidden_shape = (*input_shape, -1, self.head_dim)164 165        query_states = self.q_norm(self.q_proj(hidden_states))166        key_states = self.k_norm(self.k_proj(hidden_states))167        value_states = self.v_proj(hidden_states)168 169        query_states = query_states.view(hidden_shape).transpose(1, 2)170        key_states = key_states.view(hidden_shape).transpose(1, 2)171        value_states = value_states.view(hidden_shape).transpose(1, 2)172 173        cos, sin = position_embeddings174        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)175 176        if past_key_values is not None:177            # sin and cos are specific to RoPE models; cache_position needed for the static cache178            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}179            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)180 181        attention_interface: Callable = eager_attention_forward182        if self.config._attn_implementation != "eager":183            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]184 185        attn_output, attn_weights = attention_interface(186            self,187            query_states,188            key_states,189            value_states,190            attention_mask,191            dropout=0.0 if not self.training else self.attention_dropout,192            scaling=self.scaling,193            **kwargs,194        )195 196        attn_output = attn_output.reshape(*input_shape, -1).contiguous()197        attn_output = self.o_proj(attn_output)198        return attn_output, attn_weights199 200 201class Olmo2MLP(nn.Module):202    def __init__(self, config):203        super().__init__()204        self.config = config205        self.hidden_size = config.hidden_size206        self.intermediate_size = config.intermediate_size207        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)208        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)209        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)210        self.act_fn = ACT2FN[config.hidden_act]211 212    def forward(self, x):213        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))214        return down_proj215 216 217class Olmo2DecoderLayer(GradientCheckpointingLayer):218    def __init__(self, config: Olmo2Config, layer_idx: int):219        super().__init__()220        self.hidden_size = config.hidden_size221        self.self_attn = Olmo2Attention(config=config, layer_idx=layer_idx)222 223        self.mlp = Olmo2MLP(config)224        self.post_attention_layernorm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)225        self.post_feedforward_layernorm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)226 227    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")228    def forward(229        self,230        hidden_states: torch.Tensor,231        attention_mask: Optional[torch.Tensor] = None,232        position_ids: Optional[torch.LongTensor] = None,233        past_key_values: Optional[Cache] = None,234        use_cache: Optional[bool] = False,235        cache_position: Optional[torch.LongTensor] = None,236        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC237        **kwargs: Unpack[TransformersKwargs],238    ) -> torch.Tensor:239        residual = hidden_states240        hidden_states, _ = self.self_attn(241            hidden_states=hidden_states,242            attention_mask=attention_mask,243            position_ids=position_ids,244            past_key_values=past_key_values,245            use_cache=use_cache,246            cache_position=cache_position,247            position_embeddings=position_embeddings,248            **kwargs,249        )250        hidden_states = self.post_attention_layernorm(hidden_states)251        hidden_states = residual + hidden_states252 253        # Fully Connected254        residual = hidden_states255        hidden_states = self.mlp(hidden_states)256        hidden_states = self.post_feedforward_layernorm(hidden_states)257        hidden_states = residual + hidden_states258        return hidden_states259 260 261class Olmo2RotaryEmbedding(nn.Module):262    inv_freq: torch.Tensor  # fix linting for `register_buffer`263 264    def __init__(self, config: Olmo2Config, device=None):265        super().__init__()266        # BC: "rope_type" was originally "type"267        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):268            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))269        else:270            self.rope_type = "default"271        self.max_seq_len_cached = config.max_position_embeddings272        self.original_max_seq_len = config.max_position_embeddings273 274        self.config = config275        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]276 277        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)278        self.register_buffer("inv_freq", inv_freq, persistent=False)279        self.original_inv_freq = self.inv_freq280 281    @torch.no_grad()282    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)283    def forward(self, x, position_ids):284        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)285        position_ids_expanded = position_ids[:, None, :].float()286 287        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"288        with torch.autocast(device_type=device_type, enabled=False):  # Force float32289            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)290            emb = torch.cat((freqs, freqs), dim=-1)291            cos = emb.cos() * self.attention_scaling292            sin = emb.sin() * self.attention_scaling293            return cos, sin294 295 296@auto_docstring297class Olmo2PreTrainedModel(PreTrainedModel):298    config: Olmo2Config299    base_model_prefix = "model"300    supports_gradient_checkpointing = True301    _no_split_modules = ["Olmo2DecoderLayer"]302    _skip_keys_device_placement = ["past_key_values"]303    _supports_flash_attn = True304    _supports_sdpa = True305    _supports_flex_attn = True306 307    _can_compile_fullgraph = True308    _supports_attention_backend = True309    _can_record_outputs = {310        "hidden_states": Olmo2DecoderLayer,311        "attentions": Olmo2Attention,312    }313 314 315@auto_docstring316class Olmo2Model(Olmo2PreTrainedModel):317    def __init__(self, config: Olmo2Config):318        super().__init__(config)319        self.padding_idx = config.pad_token_id320        self.vocab_size = config.vocab_size321 322        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)323        self.layers = nn.ModuleList(324            [Olmo2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]325        )326        self.norm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)327        self.rotary_emb = Olmo2RotaryEmbedding(config=config)328        self.gradient_checkpointing = False329 330        # Initialize weights and apply final processing331        self.post_init()332 333    @check_model_inputs()334    @auto_docstring335    def forward(336        self,337        input_ids: Optional[torch.LongTensor] = None,338        attention_mask: Optional[torch.Tensor] = None,339        position_ids: Optional[torch.LongTensor] = None,340        past_key_values: Optional[Cache] = None,341        inputs_embeds: Optional[torch.FloatTensor] = None,342        cache_position: Optional[torch.LongTensor] = None,343        use_cache: Optional[bool] = None,344        **kwargs: Unpack[TransformersKwargs],345    ) -> BaseModelOutputWithPast:346        if (input_ids is None) ^ (inputs_embeds is not None):347            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")348 349        if inputs_embeds is None:350            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)351 352        if use_cache and past_key_values is None:353            past_key_values = DynamicCache(config=self.config)354 355        if cache_position is None:356            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0357            cache_position: torch.Tensor = torch.arange(358                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device359            )360 361        if position_ids is None:362            position_ids = cache_position.unsqueeze(0)363 364        causal_mask = create_causal_mask(365            config=self.config,366            input_embeds=inputs_embeds,367            attention_mask=attention_mask,368            cache_position=cache_position,369            past_key_values=past_key_values,370            position_ids=position_ids,371        )372 373        hidden_states = inputs_embeds374        position_embeddings = self.rotary_emb(hidden_states, position_ids)375 376        for decoder_layer in self.layers[: self.config.num_hidden_layers]:377            hidden_states = decoder_layer(378                hidden_states,379                attention_mask=causal_mask,380                position_ids=position_ids,381                past_key_values=past_key_values,382                cache_position=cache_position,383                position_embeddings=position_embeddings,384                **kwargs,385            )386 387        hidden_states = self.norm(hidden_states)388        return BaseModelOutputWithPast(389            last_hidden_state=hidden_states,390            past_key_values=past_key_values,391        )392 393 394@auto_docstring395class Olmo2ForCausalLM(Olmo2PreTrainedModel, GenerationMixin):396    _tied_weights_keys = ["lm_head.weight"]397    _tp_plan = {"lm_head": "colwise_rep"}398    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}399 400    def __init__(self, config):401        super().__init__(config)402        self.model = Olmo2Model(config)403        self.vocab_size = config.vocab_size404        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)405 406        # Initialize weights and apply final processing407        self.post_init()408 409    @can_return_tuple410    @auto_docstring411    def forward(412        self,413        input_ids: Optional[torch.LongTensor] = None,414        attention_mask: Optional[torch.Tensor] = None,415        position_ids: Optional[torch.LongTensor] = None,416        past_key_values: Optional[Cache] = None,417        inputs_embeds: Optional[torch.FloatTensor] = None,418        labels: Optional[torch.LongTensor] = None,419        use_cache: Optional[bool] = None,420        cache_position: Optional[torch.LongTensor] = None,421        logits_to_keep: Union[int, torch.Tensor] = 0,422        **kwargs: Unpack[TransformersKwargs],423    ) -> CausalLMOutputWithPast:424        r"""425        Example:426 427        ```python428        >>> from transformers import AutoTokenizer, Olmo2ForCausalLM429 430        >>> model = Olmo2ForCausalLM.from_pretrained("meta-olmo2/Olmo2-2-7b-hf")431        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo2/Olmo2-2-7b-hf")432 433        >>> prompt = "Hey, are you conscious? Can you talk to me?"434        >>> inputs = tokenizer(prompt, return_tensors="pt")435 436        >>> # Generate437        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)438        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]439        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."440        ```"""441        outputs: BaseModelOutputWithPast = self.model(442            input_ids=input_ids,443            attention_mask=attention_mask,444            position_ids=position_ids,445            past_key_values=past_key_values,446            inputs_embeds=inputs_embeds,447            use_cache=use_cache,448            cache_position=cache_position,449            **kwargs,450        )451 452        hidden_states = outputs.last_hidden_state453        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss454        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep455        logits = self.lm_head(hidden_states[:, slice_indices, :])456 457        loss = None458        if labels is not None:459            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)460 461        return CausalLMOutputWithPast(462            loss=loss,463            logits=logits,464            past_key_values=outputs.past_key_values,465            hidden_states=outputs.hidden_states,466            attentions=outputs.attentions,467        )468 469 470__all__ = ["Olmo2ForCausalLM", "Olmo2Model", "Olmo2PreTrainedModel"]471