Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2024 Microsoft 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 at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# 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 and14# limitations under the License.15 16"""PyTorch Phi-3 model."""17 18from typing import Callable, Optional19 20import torch21from torch import nn22 23from ...activations import ACT2FN24from ...cache_utils import Cache25from ...generation import GenerationMixin26from ...modeling_flash_attention_utils import FlashAttentionKwargs27from ...modeling_utils import ALL_ATTENTION_FUNCTIONS28from ...processing_utils import Unpack29from ...utils import logging30from ...utils.deprecation import deprecate_kwarg31from ..mistral.modeling_mistral import (32    MistralDecoderLayer,33    MistralForCausalLM,34    MistralForSequenceClassification,35    MistralForTokenClassification,36    MistralPreTrainedModel,37    eager_attention_forward,38    rotate_half,39)40from .configuration_phi3 import Phi3Config41 42 43logger = logging.get_logger(__name__)44 45_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"46_CONFIG_FOR_DOC = "Phi3Config"47 48 49class Phi3MLP(nn.Module):50    def __init__(self, config):51        super().__init__()52 53        self.config = config54        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)55        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)56        self.activation_fn = ACT2FN[config.hidden_act]57 58    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:59        up_states = self.gate_up_proj(hidden_states)60 61        gate, up_states = up_states.chunk(2, dim=-1)62        up_states = up_states * self.activation_fn(gate)63 64        return self.down_proj(up_states)65 66 67def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):68    """Applies Rotary Position Embedding to the query and key tensors.69 70    Args:71        q (`torch.Tensor`): The query tensor.72        k (`torch.Tensor`): The key tensor.73        cos (`torch.Tensor`): The cosine part of the rotary embedding.74        sin (`torch.Tensor`): The sine part of the rotary embedding.75        position_ids (`torch.Tensor`, *optional*):76            Deprecated and unused.77        unsqueeze_dim (`int`, *optional*, defaults to 1):78            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and79            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note80            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and81            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes82            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have83            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.84    Returns:85        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.86    """87    cos = cos.unsqueeze(unsqueeze_dim)88    sin = sin.unsqueeze(unsqueeze_dim)89 90    rotary_dim = cos.shape[-1]91    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]92    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]93 94    q_embed = torch.cat([(q_rot * cos) + (rotate_half(q_rot) * sin), q_pass], dim=-1)95    k_embed = torch.cat([(k_rot * cos) + (rotate_half(k_rot) * sin), k_pass], dim=-1)96    return q_embed, k_embed97 98 99class Phi3Attention(nn.Module):100    """Multi-headed attention from 'Attention Is All You Need' paper"""101 102    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):103        super().__init__()104        self.config = config105        self.layer_idx = layer_idx106        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)107        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads108        self.num_key_value_heads = config.num_key_value_heads109        self.scaling = self.head_dim**-0.5110        self.attention_dropout = config.attention_dropout111        self.is_causal = True112 113        op_size = config.num_attention_heads * self.head_dim + 2 * (config.num_key_value_heads * self.head_dim)114        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)115        self.qkv_proj = nn.Linear(config.hidden_size, op_size, bias=False)116 117    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")118    def forward(119        self,120        hidden_states: torch.Tensor,121        position_embeddings: tuple[torch.Tensor, torch.Tensor],122        attention_mask: Optional[torch.Tensor],123        past_key_values: Optional[Cache] = None,124        cache_position: Optional[torch.LongTensor] = None,125        **kwargs: Unpack[FlashAttentionKwargs],126    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:127        input_shape = hidden_states.shape[:-1]128        hidden_shape = (*input_shape, -1, self.head_dim)129 130        qkv = self.qkv_proj(hidden_states)131        query_pos = self.config.num_attention_heads * self.head_dim132        query_states = qkv[..., :query_pos]133        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]134        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]135 136        query_states = query_states.view(hidden_shape).transpose(1, 2)137        key_states = key_states.view(hidden_shape).transpose(1, 2)138        value_states = value_states.view(hidden_shape).transpose(1, 2)139 140        cos, sin = position_embeddings141        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)142 143        if past_key_values is not None:144            # sin and cos are specific to RoPE models; cache_position needed for the static cache145            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}146            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)147 148        attention_interface: Callable = eager_attention_forward149        if self.config._attn_implementation != "eager":150            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]151 152        attn_output, attn_weights = attention_interface(153            self,154            query_states,155            key_states,156            value_states,157            attention_mask,158            dropout=0.0 if not self.training else self.attention_dropout,159            scaling=self.scaling,160            sliding_window=getattr(self.config, "sliding_window", None),161            **kwargs,162        )163 164        attn_output = attn_output.reshape(*input_shape, -1).contiguous()165        attn_output = self.o_proj(attn_output)166        return attn_output, attn_weights167 168 169class Phi3DecoderLayer(MistralDecoderLayer):170    def __init__(self, config: Phi3Config, layer_idx: int):171        super().__init__(config, layer_idx)172        self.config = config173        self.self_attn = Phi3Attention(config=config, layer_idx=layer_idx)174        self.mlp = Phi3MLP(config)175        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)176        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)177 178    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")179    def forward(180        self,181        hidden_states: torch.Tensor,182        attention_mask: Optional[torch.Tensor] = None,183        position_ids: Optional[torch.LongTensor] = None,184        past_key_values: Optional[Cache] = None,185        use_cache: Optional[bool] = False,186        cache_position: Optional[torch.LongTensor] = None,187        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC188        **kwargs: Unpack[FlashAttentionKwargs],189    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:190        residual = hidden_states191        hidden_states = self.input_layernorm(hidden_states)192 193        hidden_states, self_attn_weights = self.self_attn(194            hidden_states=hidden_states,195            attention_mask=attention_mask,196            position_ids=position_ids,197            past_key_values=past_key_values,198            use_cache=use_cache,199            cache_position=cache_position,200            position_embeddings=position_embeddings,201            **kwargs,202        )203        hidden_states = residual + self.resid_attn_dropout(hidden_states)  # main diff with Llama204 205        residual = hidden_states206        hidden_states = self.post_attention_layernorm(hidden_states)207        hidden_states = self.mlp(hidden_states)208        hidden_states = residual + self.resid_mlp_dropout(hidden_states)  # main diff with Llama209        return hidden_states210 211 212class Phi3PreTrainedModel(MistralPreTrainedModel):213    _version = "0.0.5"214 215 216class Phi3ForCausalLM(MistralForCausalLM):217    def prepare_inputs_for_generation(218        self,219        input_ids,220        past_key_values=None,221        attention_mask=None,222        inputs_embeds=None,223        cache_position=None,224        position_ids=None,225        use_cache=True,226        logits_to_keep=None,227        **kwargs,228    ):229        # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the230        # process231 232        # When the first time input length reached long and short factor switching point, enforce re-compute cache233        # It will cause downside of slower at this single token position, however, better than current failure.234        if (235            past_key_values236            and self.config.rope_scaling237            and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1238        ):239            past_length = cache_position[0]240            if past_length <= self.config.original_max_position_embeddings:241                past_key_values = None242 243        model_inputs = GenerationMixin.prepare_inputs_for_generation(244            self,245            input_ids=input_ids,246            past_key_values=past_key_values,247            attention_mask=attention_mask,248            inputs_embeds=inputs_embeds,249            cache_position=cache_position,250            position_ids=position_ids,251            use_cache=use_cache,252            logits_to_keep=logits_to_keep,253            **kwargs,254        )255        return model_inputs256 257 258class Phi3ForSequenceClassification(MistralForSequenceClassification):259    pass260 261 262class Phi3ForTokenClassification(MistralForTokenClassification):263    pass264 265 266__all__ = [267    "Phi3PreTrainedModel",268    "Phi3Model",  # noqa: F822269    "Phi3ForCausalLM",270    "Phi3ForSequenceClassification",271    "Phi3ForTokenClassification",272]273