diff --git a/Model_Architecture_Discussions/gptj/configuration_gptj.py b/Model_Architecture_Discussions/gptj/configuration_gptj.py new file mode 100644 index 0000000..3b916a5 --- /dev/null +++ b/Model_Architecture_Discussions/gptj/configuration_gptj.py @@ -0,0 +1,202 @@ +"""GPT-J model configuration""" + +from collections import OrderedDict +from typing import Any, List, Mapping, Optional + +from transformers import PreTrainedTokenizer, TensorType, is_torch_available +from transformers.configuration_utils import PretrainedConfig +from transformers.onnx import OnnxConfigWithPast, PatchingSpec +from transformers.utils import logging + + +logger = logging.get_logger(__name__) + + +class GPTJConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`GPTJModel`]. It is used to instantiate a GPT-J + model according to the specified arguments, defining the model architecture. Instantiating a configuration with the + defaults will yield a similar configuration to that of the GPT-J + [EleutherAI/gpt-j-6B](https://huggingface.co/EleutherAI/gpt-j-6B) architecture. Configuration objects inherit from + [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] + for more information. + + Args: + vocab_size (`int`, *optional*, defaults to 50400): + Vocabulary size of the GPT-J model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`GPTJModel`]. + n_positions (`int`, *optional*, defaults to 2048): + The maximum sequence length that this model might ever be used with. Typically set this to something large + just in case (e.g., 512 or 1024 or 2048). + n_embd (`int`, *optional*, defaults to 4096): + Dimensionality of the embeddings and hidden states. + n_layer (`int`, *optional*, defaults to 28): + Number of hidden layers in the Transformer encoder. + n_head (`int`, *optional*, defaults to 16): + Number of attention heads for each attention layer in the Transformer encoder. + rotary_dim (`int`, *optional*, defaults to 64): + Number of dimensions in the embedding that Rotary Position Embedding is applied to. + n_inner (`int`, *optional*, defaults to None): + Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd + activation_function (`str`, *optional*, defaults to `"gelu_new"`): + Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`. + resid_pdrop (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + embd_pdrop (`int`, *optional*, defaults to 0.1): + The dropout ratio for the embeddings. + attn_pdrop (`float`, *optional*, defaults to 0.1): + The dropout ratio for the attention. + layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): + The epsilon to use in the layer normalization layers. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). + + Example: + + ```python + >>> from transformers import GPTJModel, GPTJConfig + + >>> # Initializing a GPT-J 6B configuration + >>> configuration = GPTJConfig() + + >>> # Initializing a model from the configuration + >>> model = GPTJModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "gptj" + attribute_map = { + "max_position_embeddings": "n_positions", + "hidden_size": "n_embd", + "num_attention_heads": "n_head", + "num_hidden_layers": "n_layer", + } + + def __init__( + self, + vocab_size=50400, + n_positions=2048, + n_embd=4096, + n_layer=28, + n_head=16, + rotary_dim=64, + n_inner=None, + activation_function="gelu_new", + resid_pdrop=0.0, + embd_pdrop=0.0, + attn_pdrop=0.0, + layer_norm_epsilon=1e-5, + initializer_range=0.02, + use_cache=True, + bos_token_id=50256, + eos_token_id=50256, + tie_word_embeddings=False, + **kwargs, + ): + self.vocab_size = vocab_size + self.n_positions = n_positions + self.n_embd = n_embd + self.n_layer = n_layer + self.n_head = n_head + self.n_inner = n_inner + self.rotary_dim = rotary_dim + self.activation_function = activation_function + self.resid_pdrop = resid_pdrop + self.embd_pdrop = embd_pdrop + self.attn_pdrop = attn_pdrop + self.layer_norm_epsilon = layer_norm_epsilon + self.initializer_range = initializer_range + self.use_cache = use_cache + + self.bos_token_id = bos_token_id + self.eos_token_id = eos_token_id + + super().__init__( + bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs + ) + + +# Copied from transformers.models.gpt2.configuration_gpt2.GPT2OnnxConfig +class GPTJOnnxConfig(OnnxConfigWithPast): + def __init__( + self, + config: PretrainedConfig, + task: str = "default", + patching_specs: List[PatchingSpec] = None, + use_past: bool = False, + ): + super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past) + if not getattr(self._config, "pad_token_id", None): + # TODO: how to do that better? + self._config.pad_token_id = 0 + + @property + def inputs(self) -> Mapping[str, Mapping[int, str]]: + common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}}) + if self.use_past: + self.fill_with_past_key_values_(common_inputs, direction="inputs") + common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"} + else: + common_inputs["attention_mask"] = {0: "batch", 1: "sequence"} + + return common_inputs + + @property + def num_layers(self) -> int: + return self._config.n_layer + + @property + def num_attention_heads(self) -> int: + return self._config.n_head + + def generate_dummy_inputs( + self, + tokenizer: PreTrainedTokenizer, + batch_size: int = -1, + seq_length: int = -1, + is_pair: bool = False, + framework: Optional[TensorType] = None, + ) -> Mapping[str, Any]: + common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs( + tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework + ) + + # We need to order the input in the way they appears in the forward() + ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]}) + + # Need to add the past_keys + if self.use_past: + if not is_torch_available(): + raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.") + else: + import torch + + batch, seqlen = common_inputs["input_ids"].shape + # Not using the same length for past_key_values + past_key_values_length = seqlen + 2 + past_shape = ( + batch, + self.num_attention_heads, + past_key_values_length, + self._config.hidden_size // self.num_attention_heads, + ) + ordered_inputs["past_key_values"] = [ + (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers) + ] + + ordered_inputs["attention_mask"] = common_inputs["attention_mask"] + if self.use_past: + mask_dtype = ordered_inputs["attention_mask"].dtype + ordered_inputs["attention_mask"] = torch.cat( + [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1 + ) + + return ordered_inputs + + @property + def default_onnx_opset(self) -> int: + return 13 diff --git a/Model_Architecture_Discussions/gptj/gptj.ipynb b/Model_Architecture_Discussions/gptj/gptj.ipynb new file mode 100644 index 0000000..f34afa3 --- /dev/null +++ b/Model_Architecture_Discussions/gptj/gptj.ipynb @@ -0,0 +1,272 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 1, + "id": "c5cff513-7872-4207-8877-c1873a58545a", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "env: HF_ENDPOINT=https://hf-mirror.com\n" + ] + } + ], + "source": [ + "%env HF_ENDPOINT=https://hf-mirror.com" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "id": "f4c95697-a706-4fd1-bdb5-27f5868d6839", + "metadata": {}, + "outputs": [], + "source": [ + "import os\n", + "\n", + "# 设置 HF_HOME 环境变量 设置下载路径\n", + "os.environ['HF_HOME'] = '/data1/ckw'" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "id": "25ca2fbc-b24a-4355-83cb-8e721616e9e7", + "metadata": {}, + "outputs": [ + { + "data": { + "application/vnd.jupyter.widget-view+json": { + "model_id": "58c6aa1afc0a46f98744a945dec66497", + "version_major": 2, + "version_minor": 0 + }, + "text/plain": [ + "pytorch_model.bin: 95%|#########4| 22.9G/24.2G [00:00 torch.Tensor: + inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim)) + sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float() + return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1) + + +@torch.fx.wrap +def get_embed_positions(embed_positions, position_ids): + return embed_positions.to(position_ids.device).repeat(position_ids.shape[0], 1, 1) + + +def rotate_every_two(x: torch.Tensor) -> torch.Tensor: + x1 = x[:, :, :, ::2] + x2 = x[:, :, :, 1::2] + x = torch.stack((-x2, x1), dim=-1) + return x.flatten(-2) # in einsum notation: rearrange(x, '... d j -> ... (d j)') + + +def apply_rotary_pos_emb(tensor: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor) -> torch.Tensor: + sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3) + cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3) + return (tensor * cos) + (rotate_every_two(tensor) * sin) + + +class GPTJAttention(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + max_positions = config.max_position_embeddings + self.register_buffer( + "bias", + torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view( + 1, 1, max_positions, max_positions + ), + persistent=False, + ) + self.register_buffer("masked_bias", torch.tensor(-1e9), persistent=False) + + self.attn_dropout = nn.Dropout(config.attn_pdrop) + self.resid_dropout = nn.Dropout(config.resid_pdrop) + + self.is_causal = True + + self.embed_dim = config.hidden_size + self.num_attention_heads = config.num_attention_heads + self.head_dim = self.embed_dim // self.num_attention_heads + if self.head_dim * self.num_attention_heads != self.embed_dim: + raise ValueError( + f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and" + f" `num_attention_heads`: {self.num_attention_heads})." + ) + self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()) + + self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False) + self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False) + self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False) + self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False) + self.rotary_dim = config.rotary_dim + pos_embd_dim = self.rotary_dim or self.embed_dim + self.embed_positions = create_sinusoidal_positions(max_positions, pos_embd_dim) + + def _split_heads(self, tensor, num_attention_heads, attn_head_size, rotary): + """ + Splits hidden dim into attn_head_size and num_attention_heads + """ + new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size) + tensor = tensor.view(new_shape) + if rotary: + return tensor + if len(tensor.shape) == 5: + return tensor.permute(0, 1, 3, 2, 4) # (batch, blocks, head, block_length, head_features) + elif len(tensor.shape) == 4: + return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features) + else: + raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}") + + def _merge_heads(self, tensor, num_attention_heads, attn_head_size): + """ + Merges attn_head_size dim and num_attn_heads dim into hidden dim + """ + if len(tensor.shape) == 5: + tensor = tensor.permute(0, 1, 3, 2, 4).contiguous() + elif len(tensor.shape) == 4: + tensor = tensor.permute(0, 2, 1, 3).contiguous() + else: + raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}") + new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,) + return tensor.view(new_shape) + + def _attn( + self, + query, + key, + value, + attention_mask=None, + head_mask=None, + ): + # compute causal mask from causal mask buffer + query_length, key_length = query.size(-2), key.size(-2) + causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length] + + # Keep the attention weights computation in fp32 to avoid overflow issues + query = query.to(torch.float32) + key = key.to(torch.float32) + + attn_weights = torch.matmul(query, key.transpose(-1, -2)) + + mask_value = torch.finfo(attn_weights.dtype).min + # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`. + # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device` + mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device) + attn_weights = torch.where(causal_mask, attn_weights, mask_value) + + attn_weights = attn_weights / self.scale_attn + + if attention_mask is not None: + # Apply the attention mask + attn_weights = attn_weights + attention_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1) + attn_weights = attn_weights.to(value.dtype) + attn_weights = self.attn_dropout(attn_weights) + + # Mask heads if we want to + if head_mask is not None: + attn_weights = attn_weights * head_mask + + attn_output = torch.matmul(attn_weights, value) + + return attn_output, attn_weights + + def _get_embed_positions(self, position_ids): + embed_positions = self.embed_positions + if embed_positions.device != position_ids.device: + embed_positions = embed_positions.to(position_ids.device) + self.embed_positions = embed_positions + return embed_positions.repeat(position_ids.shape[0], 1, 1) + + def forward( + self, + hidden_states: torch.FloatTensor, + layer_past: Optional[Tuple[torch.Tensor]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = False, + output_attentions: Optional[bool] = False, + ) -> Union[ + Tuple[torch.Tensor, Tuple[torch.Tensor]], + Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]], + ]: + query = self.q_proj(hidden_states) + key = self.k_proj(hidden_states) + value = self.v_proj(hidden_states) + + query = self._split_heads(query, self.num_attention_heads, self.head_dim, True) + key = self._split_heads(key, self.num_attention_heads, self.head_dim, True) + value = self._split_heads(value, self.num_attention_heads, self.head_dim, False) + + if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing(): + # The logic to conditionally copy to GPU could not be traced, so we do this + # every time in the torch.fx case + embed_positions = get_embed_positions(self.embed_positions, position_ids) + else: + embed_positions = self._get_embed_positions(position_ids) + + repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1]) + sincos = torch.gather(embed_positions, 1, repeated_position_ids) + sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1) + + if self.rotary_dim is not None: + k_rot = key[:, :, :, : self.rotary_dim] + k_pass = key[:, :, :, self.rotary_dim :] + + q_rot = query[:, :, :, : self.rotary_dim] + q_pass = query[:, :, :, self.rotary_dim :] + + k_rot = apply_rotary_pos_emb(k_rot, sin, cos) + q_rot = apply_rotary_pos_emb(q_rot, sin, cos) + + key = torch.cat([k_rot, k_pass], dim=-1) + query = torch.cat([q_rot, q_pass], dim=-1) + else: + key = apply_rotary_pos_emb(key, sin, cos) + query = apply_rotary_pos_emb(query, sin, cos) + + key = key.permute(0, 2, 1, 3) + query = query.permute(0, 2, 1, 3) + + if layer_past is not None: + past_key = layer_past[0] + past_value = layer_past[1] + key = torch.cat((past_key, key), dim=-2) + value = torch.cat((past_value, value), dim=-2) + + if use_cache is True: + # Note that this cast is quite ugly, but is not implemented before ROPE as the original codebase keeps the key in float32 all along the computation. + # Reference: https://github.com/kingoflolz/mesh-transformer-jax/blob/f8315e3003033b23f21d78361b288953064e0e76/mesh_transformer/layers.py#L128 + present = (key.to(hidden_states.dtype), value) + else: + present = None + + # compute self-attention: V x Softmax(QK^T) + attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask) + + attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim) + attn_output = self.out_proj(attn_output) + attn_output = self.resid_dropout(attn_output) + + outputs = (attn_output, present) + if output_attentions: + outputs += (attn_weights,) + + return outputs # a, present, (attentions) + + +class GPTJFlashAttention2(GPTJAttention): + """ + GPTJ flash attention module. This module inherits from `GPTJAttention` as the weights of the module stays + untouched. The only required change would be on the forward pass where it needs to correctly call the public API of + flash attention and deal with padding tokens in case the input contains any of them. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. + # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. + # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). + self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() + + def forward( + self, + hidden_states: torch.FloatTensor, + layer_past: Optional[Tuple[torch.Tensor]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = False, + output_attentions: Optional[bool] = False, + ) -> Union[ + Tuple[torch.Tensor, Tuple[torch.Tensor]], + Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]], + ]: + query = self.q_proj(hidden_states) + key = self.k_proj(hidden_states) + value = self.v_proj(hidden_states) + + query = self._split_heads(query, self.num_attention_heads, self.head_dim, True) + key = self._split_heads(key, self.num_attention_heads, self.head_dim, True) + value = self._split_heads(value, self.num_attention_heads, self.head_dim, False) + + if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing(): + # The logic to conditionally copy to GPU could not be traced, so we do this + # every time in the torch.fx case + embed_positions = get_embed_positions(self.embed_positions, position_ids) + else: + embed_positions = self._get_embed_positions(position_ids) + + repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1]) + sincos = torch.gather(embed_positions, 1, repeated_position_ids) + sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1) + + if self.rotary_dim is not None: + k_rot = key[:, :, :, : self.rotary_dim] + k_pass = key[:, :, :, self.rotary_dim :] + + q_rot = query[:, :, :, : self.rotary_dim] + q_pass = query[:, :, :, self.rotary_dim :] + + k_rot = apply_rotary_pos_emb(k_rot, sin, cos) + q_rot = apply_rotary_pos_emb(q_rot, sin, cos) + + key = torch.cat([k_rot, k_pass], dim=-1) + query = torch.cat([q_rot, q_pass], dim=-1) + else: + key = apply_rotary_pos_emb(key, sin, cos) + query = apply_rotary_pos_emb(query, sin, cos) + + # tanspose to have the desired shape + # before transpose: batch_size x seq_length x num_attention_heads x head_dim + # after transpose: batch_size x num_attention_heads x seq_length x head_dim + key = key.permute(0, 2, 1, 3) + query = query.permute(0, 2, 1, 3) + # value: batch_size x num_attention_heads x seq_length x head_dim + + if layer_past is not None: + past_key = layer_past[0] + past_value = layer_past[1] + key = torch.cat((past_key, key), dim=-2) + value = torch.cat((past_value, value), dim=-2) + + if use_cache is True: + # Note that this cast is quite ugly, but is not implemented before ROPE as the original codebase keeps the key in float32 all along the computation. + # Reference: https://github.com/kingoflolz/mesh-transformer-jax/blob/f8315e3003033b23f21d78361b288953064e0e76/mesh_transformer/layers.py#L128 + present = (key.to(hidden_states.dtype), value) + else: + present = None + + # The Flash attention requires the input to have the shape + # batch_size x seq_length x head_dim x hidden_dim + # therefore we need to keep the original shape for query and key, and reshape value + # to have the correct shape. + key = key.permute(0, 2, 1, 3).contiguous() + query = query.permute(0, 2, 1, 3).contiguous() + value = value.permute(0, 2, 1, 3).contiguous() + + # In PEFT, usually we cast the layer norms in float32 for training stability reasons + # therefore the input hidden states gets silently casted in float32. Hence, we need + # cast them back in the correct dtype just to be sure everything works as expected. + # This might slowdown training & inference so it is recommended to not cast the LayerNorms + # in fp32. (LlamaRMSNorm handles it correctly) + + input_dtype = query.dtype + if input_dtype == torch.float32: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + # Handle the case where the model is quantized + elif hasattr(self.config, "_pre_quantization_dtype"): + target_dtype = self.config._pre_quantization_dtype + else: + target_dtype = self.q_proj.weight.dtype + + logger.warning_once( + f"The input hidden states seems to be silently casted in float32, this might be related to" + f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" + f" {target_dtype}." + ) + + query = query.to(target_dtype) + key = key.to(target_dtype) + value = value.to(target_dtype) + + attention_dropout = self.config.attn_pdrop if self.training else 0.0 # attn_pdrop in gptj + + query_length = query.shape[1] + + # Compute attention + attn_weights = self._flash_attention_forward( + query, + key, + value, + attention_mask, + query_length, + dropout=attention_dropout, + ) + + # Reshape outputs + attn_output = attn_weights.reshape( + attn_weights.shape[0], attn_weights.shape[1], attn_weights.shape[2] * attn_weights.shape[3] + ) + attn_output = self.out_proj(attn_output) + attn_output = self.resid_dropout(attn_output) + + outputs = (attn_output, present) + if output_attentions: + outputs += (attn_weights,) + + return outputs + + # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward + def _flash_attention_forward( + self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None + ): + """ + Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token + first unpad the input, then computes the attention scores and pad the final attention scores. + + Args: + query_states (`torch.Tensor`): + Input query states to be passed to Flash Attention API + key_states (`torch.Tensor`): + Input key states to be passed to Flash Attention API + value_states (`torch.Tensor`): + Input value states to be passed to Flash Attention API + attention_mask (`torch.Tensor`): + The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the + position of padding tokens and 1 for the position of non-padding tokens. + dropout (`float`): + Attention dropout + softmax_scale (`float`, *optional*): + The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) + """ + if not self._flash_attn_uses_top_left_mask: + causal = self.is_causal + else: + # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__. + causal = self.is_causal and query_length != 1 + + # Contains at least one padding token in the sequence + if attention_mask is not None: + batch_size = query_states.shape[0] + query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( + query_states, key_states, value_states, attention_mask, query_length + ) + + cu_seqlens_q, cu_seqlens_k = cu_seq_lens + max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens + + attn_output_unpad = flash_attn_varlen_func( + query_states, + key_states, + value_states, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_k=cu_seqlens_k, + max_seqlen_q=max_seqlen_in_batch_q, + max_seqlen_k=max_seqlen_in_batch_k, + dropout_p=dropout, + softmax_scale=softmax_scale, + causal=causal, + ) + + attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) + else: + attn_output = flash_attn_func( + query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal + ) + + return attn_output + + # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input with num_heads->num_attention_heads + def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): + indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) + batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape + + key_layer = index_first_axis( + key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k + ) + value_layer = index_first_axis( + value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k + ) + if query_length == kv_seq_len: + query_layer = index_first_axis( + query_layer.reshape(batch_size * kv_seq_len, self.num_attention_heads, head_dim), indices_k + ) + cu_seqlens_q = cu_seqlens_k + max_seqlen_in_batch_q = max_seqlen_in_batch_k + indices_q = indices_k + elif query_length == 1: + max_seqlen_in_batch_q = 1 + cu_seqlens_q = torch.arange( + batch_size + 1, dtype=torch.int32, device=query_layer.device + ) # There is a memcpy here, that is very bad. + indices_q = cu_seqlens_q[:-1] + query_layer = query_layer.squeeze(1) + else: + # The -q_len: slice assumes left padding. + attention_mask = attention_mask[:, -query_length:] + query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) + + return ( + query_layer, + key_layer, + value_layer, + indices_q, + (cu_seqlens_q, cu_seqlens_k), + (max_seqlen_in_batch_q, max_seqlen_in_batch_k), + ) + + +GPTJ_ATTENTION_CLASSES = { + "eager": GPTJAttention, + "flash_attention_2": GPTJFlashAttention2, +} + + +class GPTJMLP(nn.Module): + def __init__(self, intermediate_size, config): # in MLP: intermediate_size= 4 * embed_dim + super().__init__() + embed_dim = config.n_embd + + self.fc_in = nn.Linear(embed_dim, intermediate_size) + self.fc_out = nn.Linear(intermediate_size, embed_dim) + + self.act = ACT2FN[config.activation_function] + self.dropout = nn.Dropout(config.resid_pdrop) + + def forward(self, hidden_states: Optional[torch.FloatTensor]) -> torch.FloatTensor: + hidden_states = self.fc_in(hidden_states) + hidden_states = self.act(hidden_states) + hidden_states = self.fc_out(hidden_states) + hidden_states = self.dropout(hidden_states) + return hidden_states + + +class GPTJBlock(nn.Module): + def __init__(self, config): + super().__init__() + inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd + self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon) + self.attn = GPTJ_ATTENTION_CLASSES[config._attn_implementation](config) + self.mlp = GPTJMLP(inner_dim, config) + + def forward( + self, + hidden_states: Optional[torch.FloatTensor], + layer_past: Optional[Tuple[torch.Tensor]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = False, + output_attentions: Optional[bool] = False, + ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]: + residual = hidden_states + hidden_states = self.ln_1(hidden_states) + attn_outputs = self.attn( + hidden_states=hidden_states, + layer_past=layer_past, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask, + use_cache=use_cache, + output_attentions=output_attentions, + ) + attn_output = attn_outputs[0] # output_attn: a, present, (attentions) + outputs = attn_outputs[1:] + + feed_forward_hidden_states = self.mlp(hidden_states) + hidden_states = attn_output + feed_forward_hidden_states + residual + + if use_cache: + outputs = (hidden_states,) + outputs + else: + outputs = (hidden_states,) + outputs[1:] + + return outputs # hidden_states, present, (attentions) + + +class GPTJPreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = GPTJConfig + base_model_prefix = "transformer" + is_parallelizable = True + supports_gradient_checkpointing = True + _no_split_modules = ["GPTJBlock"] + _skip_keys_device_placement = "past_key_values" + _supports_flash_attn_2 = True + + def __init__(self, *inputs, **kwargs): + super().__init__(*inputs, **kwargs) + + def _init_weights(self, module): + """Initialize the weights.""" + if isinstance(module, (nn.Linear,)): + # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization + # cf https://github.com/pytorch/pytorch/pull/5617 + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + elif isinstance(module, nn.LayerNorm): + module.bias.data.zero_() + module.weight.data.fill_(1.0) + + +GPTJ_START_DOCSTRING = r""" + This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use + it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and + behavior. + + Parameters: + config ([`GPTJConfig`]): Model configuration class with all the parameters of the model. + Initializing with a config file does not load the weights associated with the model, only the + configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + +GPTJ_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `({0})`): + Indices of input sequence tokens in the vocabulary. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*): + Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, + 1]`: + + - 0 corresponds to a *sentence A* token, + - 1 corresponds to a *sentence B* token. + + [What are token type IDs?](../glossary#token-type-ids) + position_ids (`torch.LongTensor` of shape `({0})`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. + + [What are position IDs?](../glossary#position-ids) + head_mask (`torch.FloatTensor` of shape `(num_attention_heads,)` or `(n_layer, num_attention_heads)`, *optional*): + Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_dim)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert *input_ids* indices into associated vectors than the + model's internal embedding lookup matrix. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + +PARALLELIZE_DOCSTRING = r""" + This is an experimental feature and is a subject to change at a moment's notice. Uses a device map to distribute + attention modules of the model across several devices. If no device map is given, it will evenly distribute blocks + across all devices. + + Args: + device_map (`Dict[int, list]`, optional, defaults to None): + A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always + automatically mapped to the first device (for esoteric reasons). That means that the first device should + have fewer attention modules mapped to it than other devices. For reference, the GPT-J models have the + following number of attention modules: + + - gpt-j-6B: 28 + + Example: + + ```python + # Here is an example of a device map on a machine with 4 GPUs using gpt-j-6B, which has a total of 28 attention modules: + model = GPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B") + device_map = { + 0: [0, 1, 2, 3, 4, 5, 6], + 1: [7, 8, 9, 10, 11, 12, 13], + 2: [14, 15, 16, 17, 18, 19, 20], + 3: [21, 22, 23, 24, 25, 26, 27], + } + model.parallelize(device_map) + ``` +""" + +DEPARALLELIZE_DOCSTRING = r""" + Moves the model to CPU from a model parallel state. + + Example: + + ```python + # On a 4 GPU machine with gpt-j-6B: + model = GPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B") + device_map = { + 0: [0, 1, 2, 3, 4, 5, 6], + 1: [7, 8, 9, 10, 11, 12, 13], + 2: [14, 15, 16, 17, 18, 19, 20], + 3: [21, 22, 23, 24, 25, 26, 27], + } + model.parallelize(device_map) # Splits the model across several devices + model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache() + ``` +""" + + +@add_start_docstrings( + "The bare GPT-J Model transformer outputting raw hidden-states without any specific head on top.", + GPTJ_START_DOCSTRING, +) +class GPTJModel(GPTJPreTrainedModel): + def __init__(self, config): + super().__init__(config) + + self.embed_dim = config.n_embd + self.vocab_size = config.vocab_size + self.wte = nn.Embedding(config.vocab_size, self.embed_dim) + self.drop = nn.Dropout(config.embd_pdrop) + self.h = nn.ModuleList([GPTJBlock(config) for _ in range(config.n_layer)]) + self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon) + + # Model parallel + self.model_parallel = False + self.device_map = None + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2" + + @add_start_docstrings(PARALLELIZE_DOCSTRING) + def parallelize(self, device_map=None): + warnings.warn( + "`GPTJModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your" + " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own" + " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1," + " ...}", + FutureWarning, + ) + # Check validity of device_map + self.device_map = ( + get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map + ) + assert_device_map(self.device_map, len(self.h)) + self.model_parallel = True + self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys())) + self.last_device = "cuda:" + str(max(self.device_map.keys())) + self.wte = self.wte.to(self.first_device) + # Load onto devices + for k, v in self.device_map.items(): + for block in v: + cuda_device = "cuda:" + str(k) + self.h[block] = self.h[block].to(cuda_device) + # ln_f to last + self.ln_f = self.ln_f.to(self.last_device) + + @add_start_docstrings(DEPARALLELIZE_DOCSTRING) + def deparallelize(self): + warnings.warn( + "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.", + FutureWarning, + ) + self.model_parallel = False + self.device_map = None + self.first_device = "cpu" + self.last_device = "cpu" + self.wte = self.wte.to("cpu") + for index in range(len(self.h)): + self.h[index] = self.h[index].to("cpu") + self.ln_f = self.ln_f.to("cpu") + torch.cuda.empty_cache() + + def get_input_embeddings(self): + return self.wte + + def set_input_embeddings(self, new_embeddings): + self.wte = new_embeddings + + @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=BaseModelOutputWithPast, + config_class=_CONFIG_FOR_DOC, + real_checkpoint=_REAL_CHECKPOINT_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + batch_size = input_ids.shape[0] + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + batch_size = inputs_embeds.shape[0] + else: + raise ValueError("You have to specify either input_ids or inputs_embeds") + + device = input_ids.device if input_ids is not None else inputs_embeds.device + + if token_type_ids is not None: + token_type_ids = token_type_ids.view(-1, input_shape[-1]) + + if past_key_values is None: + past_length = 0 + past_key_values = tuple([None] * len(self.h)) + else: + past_length = past_key_values[0][0].size(-2) + + if position_ids is None: + position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device) + position_ids = position_ids.unsqueeze(0) + + if not self._use_flash_attention_2: + # Attention mask. + if attention_mask is not None: + if batch_size <= 0: + raise ValueError("batch_size has to be defined and > 0") + attention_mask = attention_mask.view(batch_size, -1) + # We create a 3D attention mask from a 2D tensor mask. + # Sizes are [batch_size, 1, 1, to_seq_length] + # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] + # this attention mask is more simple than the triangular masking of causal attention + # used in OpenAI GPT, we just need to prepare the broadcast dimension here. + attention_mask = attention_mask[:, None, None, :] + + # Since attention_mask is 1.0 for positions we want to attend and 0.0 for + # masked positions, this operation will create a tensor which is 0.0 for + # positions we want to attend and the dtype's smallest value for masked positions. + # Since we are adding it to the raw scores before the softmax, this is + # effectively the same as removing these entirely. + attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility + attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min + + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape bsz x num_attention_heads x N x N + # head_mask has shape n_layer x batch x num_attention_heads x N x N + head_mask = self.get_head_mask(head_mask, self.config.n_layer) + + if inputs_embeds is None: + inputs_embeds = self.wte(input_ids) + + hidden_states = inputs_embeds + + if token_type_ids is not None: + token_type_embeds = self.wte(token_type_ids) + hidden_states = hidden_states + token_type_embeds + + hidden_states = self.drop(hidden_states) + + output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + presents = () if use_cache else None + all_self_attentions = () if output_attentions else None + all_hidden_states = () if output_hidden_states else None + for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): + # Model parallel + if self.model_parallel: + torch.cuda.set_device(hidden_states.device) + # Ensure layer_past is on same device as hidden_states (might not be correct) + if layer_past is not None: + layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past) + # Ensure that attention_mask is always on the same device as hidden_states + if attention_mask is not None: + attention_mask = attention_mask.to(hidden_states.device) + if isinstance(head_mask, torch.Tensor): + head_mask = head_mask.to(hidden_states.device) + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if self.gradient_checkpointing and self.training: + outputs = self._gradient_checkpointing_func( + block.__call__, + hidden_states, + None, + attention_mask, + position_ids, + head_mask[i], + use_cache, + output_attentions, + ) + else: + outputs = block( + hidden_states=hidden_states, + layer_past=layer_past, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask[i], + use_cache=use_cache, + output_attentions=output_attentions, + ) + + hidden_states = outputs[0] + if use_cache is True: + presents = presents + (outputs[1],) + + if output_attentions: + all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) + + # Model Parallel: If it's the last layer for that device, put things on the next device + if self.model_parallel: + for k, v in self.device_map.items(): + if i == v[-1] and "cuda:" + str(k) != self.last_device: + hidden_states = hidden_states.to("cuda:" + str(k + 1)) + + hidden_states = self.ln_f(hidden_states) + + hidden_states = hidden_states.view(output_shape) + # Add last hidden state + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if not return_dict: + return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=presents, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + ) + + +@add_start_docstrings( + """ + The GPT-J Model transformer with a language modeling head on top. + """, + GPTJ_START_DOCSTRING, +) +class GPTJForCausalLM(GPTJPreTrainedModel): + _tied_weights_keys = ["lm_head.weight"] + + def __init__(self, config): + super().__init__(config) + self.transformer = GPTJModel(config) + self.lm_head = nn.Linear(config.n_embd, config.vocab_size) + + # Model parallel + self.model_parallel = False + self.device_map = None + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings(PARALLELIZE_DOCSTRING) + def parallelize(self, device_map=None): + warnings.warn( + "`GPTJForCausalLM.parallelize` is deprecated and will be removed in v5 of Transformers, you should load" + " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own" + " `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':" + " 0, 'transformer.h.1': 1, ...}", + FutureWarning, + ) + self.device_map = ( + get_device_map(len(self.transformer.h), range(torch.cuda.device_count())) + if device_map is None + else device_map + ) + assert_device_map(self.device_map, len(self.transformer.h)) + self.transformer.parallelize(self.device_map) + self.lm_head = self.lm_head.to(self.transformer.first_device) + self.model_parallel = True + + @add_start_docstrings(DEPARALLELIZE_DOCSTRING) + def deparallelize(self): + warnings.warn( + "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.", + FutureWarning, + ) + self.transformer.deparallelize() + self.transformer = self.transformer.to("cpu") + self.lm_head = self.lm_head.to("cpu") + self.model_parallel = False + torch.cuda.empty_cache() + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs): + token_type_ids = kwargs.get("token_type_ids", None) + # Omit tokens covered by past_key_values + if past_key_values: + past_length = past_key_values[0][0].shape[2] + + # Some generation methods already pass only the last input ID + if input_ids.shape[1] > past_length: + remove_prefix_length = past_length + else: + # Default to old behavior: keep only final ID + remove_prefix_length = input_ids.shape[1] - 1 + + input_ids = input_ids[:, remove_prefix_length:] + if token_type_ids is not None: + token_type_ids = token_type_ids[:, -input_ids.shape[1] :] + + attention_mask = kwargs.get("attention_mask", None) + position_ids = kwargs.get("position_ids", None) + + if attention_mask is not None and position_ids is None: + # create position_ids on the fly for batch generation + position_ids = attention_mask.long().cumsum(-1) - 1 + position_ids.masked_fill_(attention_mask == 0, 1) + if past_key_values: + position_ids = position_ids[:, -input_ids.shape[1] :] + + # if `inputs_embeds` are passed, we only want to use them in the 1st generation step + if inputs_embeds is not None and past_key_values is None: + model_inputs = {"inputs_embeds": inputs_embeds} + else: + model_inputs = {"input_ids": input_ids} + + model_inputs.update( + { + "past_key_values": past_key_values, + "use_cache": kwargs.get("use_cache"), + "position_ids": position_ids, + "attention_mask": attention_mask, + "token_type_ids": token_type_ids, + } + ) + + return model_inputs + + @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=CausalLMOutputWithPast, + config_class=_CONFIG_FOR_DOC, + real_checkpoint=_REAL_CHECKPOINT_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set + `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` + are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + transformer_outputs = self.transformer( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + token_type_ids=token_type_ids, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = transformer_outputs[0] + + # Set device for model parallelism + if self.model_parallel: + torch.cuda.set_device(self.transformer.first_device) + hidden_states = hidden_states.to(self.lm_head.weight.device) + + # make sure sampling in fp16 works correctly and + # compute loss in fp32 to match with mesh-tf version + # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179 + lm_logits = self.lm_head(hidden_states).to(torch.float32) + + loss = None + if labels is not None: + # move labels to correct device to enable model parallelism + labels = labels.to(lm_logits.device) + # Shift so that tokens < n predict n + shift_logits = lm_logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss() + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + + loss = loss.to(hidden_states.dtype) + + if not return_dict: + output = (lm_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=lm_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + @staticmethod + def _reorder_cache( + past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor + ) -> Tuple[Tuple[torch.Tensor]]: + """ + This function is used to re-order the `past_key_values` cache if [`~PretrainedModel.beam_search`] or + [`~PretrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct + beam_idx at every generation step. + """ + return tuple( + tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past) + for layer_past in past_key_values + ) + + +@add_start_docstrings( + """ + The GPT-J Model transformer with a sequence classification head on top (linear layer). + + [`GPTJForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT, GPT-2, GPT-Neo) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """, + GPTJ_START_DOCSTRING, +) +class GPTJForSequenceClassification(GPTJPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.transformer = GPTJModel(config) + self.score = nn.Linear(config.n_embd, self.num_labels, bias=False) + + # Model parallel + self.model_parallel = False + self.device_map = None + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @add_code_sample_docstrings( + checkpoint="ydshieh/tiny-random-gptj-for-sequence-classification", + output_type=SequenceClassifierOutputWithPast, + config_class=_CONFIG_FOR_DOC, + real_checkpoint=_REAL_CHECKPOINT_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, SequenceClassifierOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + transformer_outputs = self.transformer( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + token_type_ids=token_type_ids, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = transformer_outputs[0] + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") + if self.config.pad_token_id is None: + sequence_lengths = -1 + else: + if input_ids is not None: + # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility + sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 + sequence_lengths = sequence_lengths % input_ids.shape[-1] + sequence_lengths = sequence_lengths.to(logits.device) + else: + sequence_lengths = -1 + logger.warning( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] + + loss = None + if labels is not None: + labels = labels.to(pooled_logits.device) + if self.config.problem_type is None: + if self.num_labels == 1: + self.config.problem_type = "regression" + elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): + self.config.problem_type = "single_label_classification" + else: + self.config.problem_type = "multi_label_classification" + + if self.config.problem_type == "regression": + loss_fct = MSELoss() + if self.num_labels == 1: + loss = loss_fct(pooled_logits.squeeze(), labels.squeeze()) + else: + loss = loss_fct(pooled_logits, labels) + elif self.config.problem_type == "single_label_classification": + loss_fct = CrossEntropyLoss() + loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1)) + elif self.config.problem_type == "multi_label_classification": + loss_fct = BCEWithLogitsLoss() + loss = loss_fct(pooled_logits, labels) + if not return_dict: + output = (pooled_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@add_start_docstrings( + """ + The GPT-J Model transformer with a span classification head on top for extractive question-answering tasks like + SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + GPTJ_START_DOCSTRING, +) +class GPTJForQuestionAnswering(GPTJPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.transformer = GPTJModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels) + + # Model parallel + self.model_parallel = False + self.device_map = None + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=QuestionAnsweringModelOutput, + config_class=_CONFIG_FOR_DOC, + real_checkpoint=_REAL_CHECKPOINT_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, QuestionAnsweringModelOutput]: + r""" + start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the start of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the end of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + outputs = self.transformer( + input_ids, + attention_mask=attention_mask, + token_type_ids=token_type_ids, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + sequence_output = outputs[0] + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + total_loss = None + if start_positions is not None and end_positions is not None: + # If we are on multi-GPU, split add a dimension + if len(start_positions.size()) > 1: + start_positions = start_positions.squeeze(-1).to(start_logits.device) + if len(end_positions.size()) > 1: + end_positions = end_positions.squeeze(-1).to(end_logits.device) + # sometimes the start/end positions are outside our model inputs, we ignore these terms + ignored_index = start_logits.size(1) + start_positions = start_positions.clamp(0, ignored_index) + end_positions = end_positions.clamp(0, ignored_index) + + loss_fct = CrossEntropyLoss(ignore_index=ignored_index) + start_loss = loss_fct(start_logits, start_positions) + end_loss = loss_fct(end_logits, end_positions) + total_loss = (start_loss + end_loss) / 2 + + if not return_dict: + output = (start_logits, end_logits) + outputs[2:] + return ((total_loss,) + output) if total_loss is not None else output + + return QuestionAnsweringModelOutput( + loss=total_loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + )