From 9afb904b158dce9870c987480423bba6f343ca4c Mon Sep 17 00:00:00 2001 From: Cyril Vallez Date: Tue, 4 Feb 2025 11:18:43 +0100 Subject: [PATCH] Refactor (and fix) gpt_neox (#35610) * start a nice modular * Update modular_gpt_neox.py * Update modular_gpt_neox.py * Update modular_gpt_neox.py * Update modular_gpt_neox.py * update * Update modular_gpt_neox.py * convert * fix attribute * fix attrs * oups * fix * fix * fix * fix * fix * fix order to pass test (see with accelerate team) * trigger CIs * modular * update * up * Update test_modeling_gpt_neox.py * Update test_modeling_gpt_neox.py * trigger CIs * correctly pass arg * simplify * remove key warning * update tp -> it's compatible since the view is before * trigger CIs --- .../integrations/flex_attention.py | 3 + .../models/gpt_neox/configuration_gpt_neox.py | 6 + .../models/gpt_neox/modeling_gpt_neox.py | 905 ++++++------------ .../models/gpt_neox/modular_gpt_neox.py | 837 ++++++++++++++++ .../models/gpt_neox/test_modeling_gpt_neox.py | 15 +- utils/check_config_attributes.py | 1 + 6 files changed, 1159 insertions(+), 608 deletions(-) create mode 100644 src/transformers/models/gpt_neox/modular_gpt_neox.py diff --git a/src/transformers/integrations/flex_attention.py b/src/transformers/integrations/flex_attention.py index 53bd80bf04..5181b2c1a0 100644 --- a/src/transformers/integrations/flex_attention.py +++ b/src/transformers/integrations/flex_attention.py @@ -17,6 +17,7 @@ def flex_attention_forward( attention_mask: Optional[torch.Tensor], scaling: Optional[float] = None, softcap: Optional[float] = None, + head_mask: Optional[torch.Tensor] = None, **kwargs, ) -> Tuple[torch.Tensor, torch.Tensor]: causal_mask = attention_mask @@ -28,6 +29,8 @@ def flex_attention_forward( score = softcap * torch.tanh(score / softcap) if causal_mask is not None: score = score + causal_mask[b][0][q_idx][kv_idx] + if head_mask is not None: + score = score + head_mask[b][h][0][0] return score attn_output, attention_weights = flex_attention( diff --git a/src/transformers/models/gpt_neox/configuration_gpt_neox.py b/src/transformers/models/gpt_neox/configuration_gpt_neox.py index 6d73403ed1..e570662c10 100644 --- a/src/transformers/models/gpt_neox/configuration_gpt_neox.py +++ b/src/transformers/models/gpt_neox/configuration_gpt_neox.py @@ -131,6 +131,12 @@ class GPTNeoXConfig(PretrainedConfig): model_type = "gpt_neox" keys_to_ignore_at_inference = ["past_key_values"] + base_model_tp_plan = { + "layers.*.attention.query_key_value": "colwise", + "layers.*.attention.dense": "rowwise", + "layers.*.mlp.dense_h_to_4h": "colwise", + "layers.*.mlp.dense_4h_to_h": "rowwise", + } def __init__( self, diff --git a/src/transformers/models/gpt_neox/modeling_gpt_neox.py b/src/transformers/models/gpt_neox/modeling_gpt_neox.py index e4814ce4e7..876f9049ad 100755 --- a/src/transformers/models/gpt_neox/modeling_gpt_neox.py +++ b/src/transformers/models/gpt_neox/modeling_gpt_neox.py @@ -1,36 +1,19 @@ -# coding=utf-8 -# Copyright 2022 EleutherAI The HuggingFace Inc. team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -"""PyTorch GPTNeoX model.""" - -from typing import Optional, Tuple, Union +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# This file was automatically generated from src/transformers/models/gpt_neox/modular_gpt_neox.py. +# Do NOT edit this file manually as any edits will be overwritten by the generation of +# the file from the modular. If any change should be done, please apply the change to the +# modular_gpt_neox.py file directly. One of our CI enforces this. +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +from typing import Callable, Optional, Tuple, Union import torch -import torch.utils.checkpoint from torch import nn -from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, StaticCache -from ...file_utils import ( - add_code_sample_docstrings, - add_start_docstrings, - add_start_docstrings_to_model_forward, - replace_return_docstrings, -) from ...generation import GenerationMixin from ...modeling_attn_mask_utils import AttentionMaskConverter +from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -39,99 +22,109 @@ from ...modeling_outputs import ( TokenClassifierOutput, ) from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS -from ...modeling_utils import PreTrainedModel +from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel +from ...processing_utils import Unpack from ...utils import ( - is_flash_attn_2_available, - is_flash_attn_greater_or_equal_2_10, - is_torch_flex_attn_available, + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, logging, + replace_return_docstrings, ) from .configuration_gpt_neox import GPTNeoXConfig -if is_flash_attn_2_available(): - from ...modeling_flash_attention_utils import _flash_attention_forward - -if is_torch_flex_attn_available(): - from torch.nn.attention.flex_attention import flex_attention - logger = logging.get_logger(__name__) + _CHECKPOINT_FOR_DOC = "trl-internal-testing/tiny-random-GPTNeoXForCausalLM" -_REAL_CHECKPOINT_FOR_DOC = "EleutherAI/gpt-neox-20b" _CONFIG_FOR_DOC = "GPTNeoXConfig" -class GPTNeoXPreTrainedModel(PreTrainedModel): - """ - An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained - models. - """ +class GPTNeoXMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size) + self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size) + self.act = ACT2FN[config.hidden_act] - config_class = GPTNeoXConfig - base_model_prefix = "gpt_neox" - supports_gradient_checkpointing = True - _no_split_modules = ["GPTNeoXLayer"] - _skip_keys_device_placement = "past_key_values" - _supports_flash_attn_2 = True - _supports_cache_class = True - _supports_quantized_cache = True - _supports_static_cache = True - _supports_sdpa = True - _supports_flex_attn = True + def forward(self, hidden_states): + hidden_states = self.dense_h_to_4h(hidden_states) + hidden_states = self.act(hidden_states) + hidden_states = self.dense_4h_to_h(hidden_states) + return hidden_states - def _init_weights(self, module): - """Initialize the weights""" - if isinstance(module, nn.Linear): - 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) + +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + + # Keep half or full tensor for later concatenation + rotary_dim = cos.shape[-1] + q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] + k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] + + # Apply rotary embeddings on the first half or full tensor + q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin) + k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin) + + # Concatenate back to full shape + q_embed = torch.cat([q_embed, q_pass], dim=-1) + k_embed = torch.cat([k_embed, k_pass], dim=-1) + return q_embed, k_embed def eager_attention_forward( - query, key, value, attention_mask, head_mask, norm_factor, attention_dropout, training, **_kwargs + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: torch.Tensor, + scaling: float, + dropout: float = 0.0, + head_mask: Optional[torch.Tensor] = None, + **kwargs, ): - # q, k, v: [bs, num_attention_heads, seq_len, attn_head_size] - batch_size, num_attention_heads, query_length, attn_head_size = query.size() - key_length = key.size(-2) - - query = query.view(batch_size * num_attention_heads, query_length, attn_head_size) - key = key.view(batch_size * num_attention_heads, key_length, attn_head_size) - attn_scores = torch.zeros( - batch_size * num_attention_heads, - query_length, - key_length, - dtype=query.dtype, - device=key.device, - ) - attn_scores = torch.baddbmm( - attn_scores, - query, - key.transpose(1, 2), - beta=1.0, - alpha=norm_factor, - ) - attn_scores = attn_scores.view(batch_size, num_attention_heads, query_length, key_length) + attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling if attention_mask is not None: # no matter the length, we just slice it causal_mask = attention_mask[:, :, :, : key.shape[-2]] - attn_scores = attn_scores + causal_mask + attn_weights = attn_weights + causal_mask - attn_weights = nn.functional.softmax(attn_scores, dim=-1) - attn_weights = attn_weights.to(value.dtype) + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) # Mask heads if we want to if head_mask is not None: attn_weights = attn_weights * head_mask - attn_weights = nn.functional.dropout(attn_weights, p=attention_dropout, training=training) + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value) # Reshape outputs @@ -140,192 +133,49 @@ def eager_attention_forward( return attn_output, attn_weights -def flash_attention_forward( - query, - key, - value, - attention_mask, - norm_factor, - attention_dropout, - training, - position_ids=None, - target_dtype=None, - **_kwargs, -): - query_length = query.shape[-2] - - # GPT-neo-X casts query and key in fp32 to apply rotary embedding in full precision - query = query.to(value.dtype) - key = key.to(value.dtype) - - # Permute to get the expected shape for Flash Attention - query = query.transpose(1, 2) - key = key.transpose(1, 2) - value = value.transpose(1, 2) - - attention_dropout = attention_dropout if training else 0.0 - flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() - - # Compute attention - attn_output = _flash_attention_forward( - query, - key, - value, - attention_mask, - query_length, - dropout=attention_dropout, - position_ids=position_ids, - softmax_scale=norm_factor, - is_causal=True, - use_top_left_mask=flash_attn_uses_top_left_mask, - target_dtype=target_dtype, - ) - - return attn_output, None - - -def sdpa_attention_forward(query, key, value, attention_mask, attention_dropout, training, **_kwargs): - q_len = query.shape[-2] - - causal_mask = attention_mask - if attention_mask is not None: - causal_mask = causal_mask[:, :, :, : key.shape[-2]] - - # GPT-neo-X casts query and key in fp32 to apply rotary embedding in full precision - query = query.to(value.dtype) - key = key.to(value.dtype) - - # Avoid torch==2.1.2 specific bug for the memory-efficient backend in SDPA - query = query.contiguous() - key = key.contiguous() - value = value.contiguous() - - # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment - # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. - is_causal = True if causal_mask is None and q_len > 1 else False - - attn_output = torch.nn.functional.scaled_dot_product_attention( - query=query, - key=key, - value=value, - attn_mask=causal_mask, - dropout_p=attention_dropout if training else 0.0, - is_causal=is_causal, - ) - - # Reshape outputs - attn_output = attn_output.transpose(1, 2).contiguous() - - return attn_output, None - - -def flex_attention_forward(query, key, value, attention_mask, head_mask, norm_factor, **_kwargs): - causal_mask = attention_mask - if causal_mask is not None: - causal_mask = causal_mask[:, :, :, : key.shape[-2]] - - def causal_mod(score, b, h, q_idx, kv_idx): - if causal_mask is not None: - score += causal_mask[b][0][q_idx][kv_idx] - if head_mask is not None: - score += head_mask[b][h][0][0] - return score - - attn_output, attn_weights = flex_attention( - query, - key, - value, - score_mod=causal_mod, - enable_gqa=True, - scale=norm_factor, - # Last time checked on PyTorch == 2.5.1: Flex Attention always computes the lse regardless. - # For simplification, we thus always return it as no additional computations are introduced. - return_lse=True, - ) - - # lse is returned in float32 - attn_weights = attn_weights.to(value.dtype) - - # Reshape outputs - attn_output = attn_output.transpose(1, 2).contiguous() - - return attn_output, attn_weights - - -GPTNEOX_ATTENTION_FUNCTION = { - "eager": eager_attention_forward, - "flash_attention_2": flash_attention_forward, - "sdpa": sdpa_attention_forward, - "flex_attention": flex_attention_forward, -} - - class GPTNeoXAttention(nn.Module): def __init__(self, config, layer_idx=None): super().__init__() self.config = config - self.num_attention_heads = config.num_attention_heads - self.hidden_size = config.hidden_size - if self.hidden_size % self.num_attention_heads != 0: - raise ValueError( - "The hidden size is not divisble by the number of attention heads! Make sure to update them" - ) - self.head_size = self.hidden_size // self.num_attention_heads + self.head_size = config.hidden_size // config.num_attention_heads + self.attention_dropout = config.attention_dropout self.rotary_ndims = int(self.head_size * config.rotary_pct) - self.rope_theta = config.rotary_emb_base - self._init_bias(config.max_position_embeddings) - - self.register_buffer("masked_bias", torch.tensor(-1e9), persistent=False) - self.rotary_emb = GPTNeoXRotaryEmbedding(config=self.config) - - if layer_idx is None: - logger.warning_once( - f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " - "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " - "when creating this class." - ) - self.norm_factor = self.head_size**-0.5 - self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=config.attention_bias) - self.dense = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias) - self.attention_dropout = nn.Dropout(config.attention_dropout) + self.scaling = self.head_size**-0.5 self.is_causal = True self.layer_idx = layer_idx - def _init_bias(self, max_positions, device=None): - self.register_buffer( - "bias", - torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view( - 1, 1, max_positions, max_positions - ), - persistent=False, - ) - if device is not None: - self.bias = self.bias.to(device) + self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=config.attention_bias) + self.dense = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias) def forward( self, hidden_states: torch.FloatTensor, attention_mask: torch.FloatTensor, - position_ids: torch.LongTensor, head_mask: Optional[torch.FloatTensor] = None, layer_past: Optional[Cache] = None, - use_cache: Optional[bool] = False, output_attentions: Optional[bool] = False, - padding_mask: Optional[torch.Tensor] = None, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], ): - bsz, seq_len, _ = hidden_states.shape + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, 3 * self.head_size) - # Apply attention-specific projections and rope - query, key, value, present = self._attn_projections_and_rope( - hidden_states=hidden_states, - position_ids=position_ids, - layer_past=layer_past, - use_cache=use_cache, - cache_position=cache_position, - position_embeddings=position_embeddings, - ) + qkv = self.query_key_value(hidden_states).view(hidden_shape).transpose(1, 2) + query_states, key_states, value_states = qkv.chunk(3, dim=-1) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + # Cache QKV values + if layer_past is not None: + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_ndims, + "cache_position": cache_position, + } + key_states, value_states = layer_past.update(key_states, value_states, self.layer_idx, cache_kwargs) # Checking for fallbacks in case an unsupported feature is requested attention_type = self.config._attn_implementation @@ -339,135 +189,97 @@ class GPTNeoXAttention(nn.Module): ) attention_type = "eager" - elif ( - self.training - and self.config.attention_dropout > 0 - and self.config._attn_implementation == "flex_attention" - ): + elif self.training and self.attention_dropout > 0 and self.config._attn_implementation == "flex_attention": logger.warning_once( f"Setting `attention_type` to `eager` because `dropout` is not supported in `{attention_type}`." ) attention_type = "eager" + attention_interface: Callable = eager_attention_forward + attention_interface = ( + ALL_ATTENTION_FUNCTIONS[attention_type] if attention_type != "eager" else attention_interface + ) + # Compute attention - attn_output, attn_weights = GPTNEOX_ATTENTION_FUNCTION[attention_type]( - query, - key, - value, - attention_mask=attention_mask, - position_ids=position_ids, + attn_output, attn_weights = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask, + scaling=self.scaling, + dropout=0.0 if not self.training else self.attention_dropout, head_mask=head_mask, - norm_factor=self.norm_factor, - attention_dropout=self.config.attention_dropout, - training=self.training, - # Flash Attention 2 specific PEFT check - target_dtype=self._fa_peft_dtype_check(value), + **kwargs, ) # Reshape outputs and final projection - attn_output = attn_output.contiguous() - attn_output = attn_output.view(bsz, seq_len, -1) + attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.dense(attn_output) - outputs = (attn_output, present) + return attn_output, attn_weights + + +class GPTNeoXLayer(nn.Module): + def __init__(self, config, layer_idx): + super().__init__() + self.use_parallel_residual = config.use_parallel_residual + self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.post_attention_dropout = nn.Dropout(config.hidden_dropout) + self.post_mlp_dropout = nn.Dropout(config.hidden_dropout) + self.attention = GPTNeoXAttention(config, layer_idx) + self.mlp = GPTNeoXMLP(config) + + def forward( + self, + hidden_states: Optional[torch.FloatTensor], + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = False, + layer_past: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ): + attn_output, attn_weights = self.attention( + self.input_layernorm(hidden_states), + attention_mask=attention_mask, + position_ids=position_ids, + layer_past=layer_past, + head_mask=head_mask, + use_cache=use_cache, + output_attentions=output_attentions, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + attn_output = self.post_attention_dropout(attn_output) + + if self.use_parallel_residual: + # pseudocode: + # x = x + attn(ln1(x)) + mlp(ln2(x)) + mlp_output = self.mlp(self.post_attention_layernorm(hidden_states)) + mlp_output = self.post_mlp_dropout(mlp_output) + hidden_states = mlp_output + attn_output + hidden_states + else: + # pseudocode: + # x = x + attn(ln1(x)) + # x = x + mlp(ln2(x)) + attn_output = attn_output + hidden_states + mlp_output = self.mlp(self.post_attention_layernorm(attn_output)) + mlp_output = self.post_mlp_dropout(mlp_output) + hidden_states = mlp_output + attn_output + + outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) return outputs - @classmethod - def _split_heads(cls, tensor, num_attention_heads, attn_head_size): - """ - Splits hidden dim into attn_head_size and num_attention_heads - """ - # tensor: [bs, seq_len, hidden_size] - new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size) - # -> [bs, seq_len, num_attention_heads, attn_head_size] - tensor = tensor.view(new_shape) - # -> [bs, num_attention_heads, seq_len, attn_head_size] - tensor = tensor.permute(0, 2, 1, 3) - return tensor - @classmethod - def _merge_heads(cls, tensor, num_attention_heads, attn_head_size): - """ - Merges attn_head_size dim and num_attn_heads dim into hidden dim - """ - # tensor [bs, num_attention_heads, seq_len, attn_head_size] - tensor = tensor.permute(0, 2, 1, 3).contiguous() - # -> [bs, seq_len, num_attention_heads, attn_head_size] - tensor = tensor.view(tensor.size(0), tensor.size(1), num_attention_heads * attn_head_size) - # -> [bs, seq_len, hidden_size] - return tensor - - def _attn_projections_and_rope( - self, - hidden_states: torch.FloatTensor, - position_ids: torch.LongTensor, - layer_past: Optional[Tuple[torch.Tensor]] = None, - use_cache: Optional[bool] = False, - cache_position: Optional[torch.LongTensor] = None, - position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC - ): - # Compute QKV - # Attention heads [batch, seq_len, hidden_size] - # --> [batch, seq_len, (np * 3 * head_size)] - qkv = self.query_key_value(hidden_states) - - # [batch, seq_len, (num_heads * 3 * head_size)] - # --> [batch, seq_len, num_heads, 3 * head_size] - new_qkv_shape = qkv.size()[:-1] + (self.num_attention_heads, 3 * self.head_size) - qkv = qkv.view(*new_qkv_shape) - - # [batch, seq_len, num_attention_heads, 3 * head_size] --> 3 [batch, num_attention_heads, seq_len, head_size] - query = qkv[..., : self.head_size].permute(0, 2, 1, 3) - key = qkv[..., self.head_size : 2 * self.head_size].permute(0, 2, 1, 3) - value = qkv[..., 2 * self.head_size :].permute(0, 2, 1, 3) - - # Compute rotary embeddings on rotary_ndims - query_rot = query[..., : self.rotary_ndims] - query_pass = query[..., self.rotary_ndims :] - key_rot = key[..., : self.rotary_ndims] - key_pass = key[..., self.rotary_ndims :] - - cos, sin = position_embeddings - query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin) - query = torch.cat((query, query_pass), dim=-1) - key = torch.cat((key, key_pass), dim=-1) - - # Cache QKV values - if layer_past is not None: - cache_kwargs = { - "sin": sin, - "cos": cos, - "partial_rotation_size": self.rotary_ndims, - "cache_position": cache_position, - } - key, value = layer_past.update(key, value, self.layer_idx, cache_kwargs) - - return query, key, value, layer_past - - def _fa_peft_dtype_check(self, value): - """ - PEFT can silently cast the dtype to float32 - this method returns the target dtype to which - FA should convert back to (if necessary). For now, we can not move this to the forward pass - itself due to the dependency on checking on some part of its own weights (last case). - """ - target_dtype = None - if self.config._attn_implementation == "flash_attention_2": - input_dtype = value.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.query_key_value.weight.dtype - return target_dtype - - -# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->GPTNeoX class GPTNeoXRotaryEmbedding(nn.Module): def __init__(self, config: GPTNeoXConfig, device=None): super().__init__() @@ -529,158 +341,95 @@ class GPTNeoXRotaryEmbedding(nn.Module): return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) -def rotate_half(x): - """Rotates half the hidden dims of the input.""" - x1 = x[..., : x.shape[-1] // 2] - x2 = x[..., x.shape[-1] // 2 :] - return torch.cat((-x2, x1), dim=-1) - - -# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb -def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): - """Applies Rotary Position Embedding to the query and key tensors. - - Args: - q (`torch.Tensor`): The query tensor. - k (`torch.Tensor`): The key tensor. - cos (`torch.Tensor`): The cosine part of the rotary embedding. - sin (`torch.Tensor`): The sine part of the rotary embedding. - position_ids (`torch.Tensor`, *optional*): - Deprecated and unused. - unsqueeze_dim (`int`, *optional*, defaults to 1): - The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and - sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note - that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and - k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes - cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have - the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. - Returns: - `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. - """ - cos = cos.unsqueeze(unsqueeze_dim) - sin = sin.unsqueeze(unsqueeze_dim) - q_embed = (q * cos) + (rotate_half(q) * sin) - k_embed = (k * cos) + (rotate_half(k) * sin) - return q_embed, k_embed - - -class GPTNeoXMLP(nn.Module): - def __init__(self, config): - super().__init__() - self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size) - self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size) - self.act = ACT2FN[config.hidden_act] - - def forward(self, hidden_states): - hidden_states = self.dense_h_to_4h(hidden_states) - hidden_states = self.act(hidden_states) - hidden_states = self.dense_4h_to_h(hidden_states) - return hidden_states - - -class GPTNeoXLayer(nn.Module): - def __init__(self, config, layer_idx): - super().__init__() - self.use_parallel_residual = config.use_parallel_residual - self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) - self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) - self.post_attention_dropout = nn.Dropout(config.hidden_dropout) - self.post_mlp_dropout = nn.Dropout(config.hidden_dropout) - self.attention = GPTNeoXAttention(config, layer_idx) - self.mlp = GPTNeoXMLP(config) - - def forward( - self, - hidden_states: Optional[torch.FloatTensor], - attention_mask: Optional[torch.FloatTensor] = None, - position_ids: Optional[torch.LongTensor] = None, - head_mask: Optional[torch.FloatTensor] = None, - use_cache: Optional[bool] = False, - layer_past: Optional[Cache] = None, - output_attentions: Optional[bool] = False, - cache_position: Optional[torch.LongTensor] = None, - position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC - ): - attention_layer_outputs = self.attention( - self.input_layernorm(hidden_states), - attention_mask=attention_mask, - position_ids=position_ids, - layer_past=layer_past, - head_mask=head_mask, - use_cache=use_cache, - output_attentions=output_attentions, - cache_position=cache_position, - position_embeddings=position_embeddings, - ) - attn_output = attention_layer_outputs[0] # output_attn: attn_output, present, (attn_weights) - attn_output = self.post_attention_dropout(attn_output) - outputs = attention_layer_outputs[1:] - - if self.use_parallel_residual: - # pseudocode: - # x = x + attn(ln1(x)) + mlp(ln2(x)) - mlp_output = self.mlp(self.post_attention_layernorm(hidden_states)) - mlp_output = self.post_mlp_dropout(mlp_output) - hidden_states = mlp_output + attn_output + hidden_states - else: - # pseudocode: - # x = x + attn(ln1(x)) - # x = x + mlp(ln2(x)) - attn_output = attn_output + hidden_states - mlp_output = self.mlp(self.post_attention_layernorm(attn_output)) - mlp_output = self.post_mlp_dropout(mlp_output) - hidden_states = mlp_output + attn_output - - if use_cache: - outputs = (hidden_states,) + outputs # hidden_states, present, (attn_weights) - else: - outputs = (hidden_states,) + outputs[1:] # hidden_states, (attn_weights) - - return outputs - - GPT_NEOX_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. + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. Parameters: - config ([`~GPTNeoXConfig`]): 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. + config ([`GPTNeoXConfig`]): + 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. """ + +@add_start_docstrings( + "The bare GPTNeoX Model outputting raw hidden-states without any specific head on top.", + GPT_NEOX_START_DOCSTRING, +) +class GPTNeoXPreTrainedModel(PreTrainedModel): + config_class = GPTNeoXConfig + base_model_prefix = "gpt_neox" + supports_gradient_checkpointing = True + _no_split_modules = ["GPTNeoXLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_flex_attn = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = True + _supports_attention_backend = True + _keys_to_ignore_on_load_unexpected = [r"attention.bias", r"attention.masked_bias"] + + def _init_weights(self, module): + """Initialize the weights""" + if isinstance(module, nn.Linear): + 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) + + +_REAL_CHECKPOINT_FOR_DOC = "EleutherAI/gpt-neox-20b" + + GPT_NEOX_INPUTS_DOCSTRING = r""" Args: - input_ids (`torch.LongTensor` of shape `({0})`): - Indices of input sequence tokens in the vocabulary. + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. 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*): + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *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) - position_ids (`torch.LongTensor` of shape `({0})`, *optional*): + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *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_heads,)` or `(num_layers, num_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_size)`, *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. past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` @@ -699,6 +448,13 @@ GPT_NEOX_INPUTS_DOCSTRING = r""" If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *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. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. @@ -706,7 +462,7 @@ GPT_NEOX_INPUTS_DOCSTRING = r""" 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 [`~file_utils.ModelOutput`] instead of a plain tuple. + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer @@ -715,10 +471,17 @@ GPT_NEOX_INPUTS_DOCSTRING = r""" @add_start_docstrings( - "The bare GPTNeoX Model transformer outputting raw hidden-states without any specific head on top.", + "The bare GPTNeoX Model outputting raw hidden-states without any specific head on top.", GPT_NEOX_START_DOCSTRING, ) class GPTNeoXModel(GPTNeoXPreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GPTNeoXDecoderLayer`] + + Args: + config: GPTNeoXConfig + """ + def __init__(self, config): super().__init__(config) self.config = config @@ -728,9 +491,6 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): self.layers = nn.ModuleList([GPTNeoXLayer(config, i) for i in range(config.num_hidden_layers)]) self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.rotary_emb = GPTNeoXRotaryEmbedding(config=config) - - self._attn_implementation = config._attn_implementation - self.gradient_checkpointing = False # Initialize weights and apply final processing @@ -756,18 +516,14 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, - past_key_values: Optional[Union[Cache, Tuple[Tuple[torch.FloatTensor]]]] = None, + past_key_values: Optional[Cache] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], ) -> Union[Tuple, BaseModelOutputWithPast]: - r""" - use_cache (`bool`, *optional*): - If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see - `past_key_values`). - """ 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 @@ -788,24 +544,14 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): if inputs_embeds is None: inputs_embeds = self.embed_in(input_ids) - # kept for BC (non `Cache` `past_key_values` inputs) - return_legacy_cache = False - if use_cache and not isinstance(past_key_values, Cache): - return_legacy_cache = True - if past_key_values is None: - past_key_values = DynamicCache() - else: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - logger.warning_once( - "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and " - "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class " - "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)" - ) + if use_cache and past_key_values is None: + past_key_values = DynamicCache() - seq_length = inputs_embeds.shape[1] if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 - cache_position = torch.arange(past_seen_tokens, past_seen_tokens + seq_length, device=inputs_embeds.device) + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) if position_ids is None: position_ids = cache_position.unsqueeze(0) @@ -831,12 +577,9 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): # create position embeddings to be shared across the decoder layers position_embeddings = self.rotary_emb(hidden_states, position_ids) - next_decoder_cache = None all_attentions = () if output_attentions else None all_hidden_states = () if output_hidden_states else None - for i, layer in enumerate( - self.layers, - ): + for i, layer in enumerate(self.layers): if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) @@ -848,7 +591,7 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): position_ids, head_mask[i], use_cache, - None, + past_key_values, output_attentions, cache_position, position_embeddings, @@ -864,33 +607,26 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): output_attentions=output_attentions, cache_position=cache_position, position_embeddings=position_embeddings, + **flash_attn_kwargs, ) hidden_states = outputs[0] - if use_cache is True: - next_decoder_cache = outputs[1] + if output_attentions: - all_attentions = all_attentions + (outputs[2 if use_cache else 1],) + all_attentions = all_attentions + (outputs[1],) hidden_states = self.final_layer_norm(hidden_states) # Add last hidden state if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) - next_cache = next_decoder_cache if use_cache else None - if return_legacy_cache: - next_cache = next_cache.to_legacy_cache() - - if not return_dict: - return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_attentions] if v is not None) - - return BaseModelOutputWithPast( + output = BaseModelOutputWithPast( last_hidden_state=hidden_states, - past_key_values=next_cache, + past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_attentions, ) + return output if return_dict else output.to_tuple() - # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask def _update_causal_mask( self, attention_mask: torch.Tensor, @@ -957,7 +693,6 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): return causal_mask @staticmethod - # Copied from transformers.models.llama.modeling_llama.LlamaModel._prepare_4d_causal_attention_mask_with_cache_position def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, @@ -1014,11 +749,15 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel): return causal_mask +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + @add_start_docstrings( """GPTNeoX Model with a `language modeling` head on top for CLM fine-tuning.""", GPT_NEOX_START_DOCSTRING ) class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin): _tied_weights_keys = ["embed_out.weight"] + _tp_plan = {"embed_out": "colwise_rep"} def __init__(self, config): super().__init__(config) @@ -1051,6 +790,8 @@ class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin): output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], ) -> Union[Tuple, CausalLMOutputWithPast]: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): @@ -1093,42 +834,30 @@ class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin): output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, + **kwargs, ) hidden_states = outputs[0] - lm_logits = self.embed_out(hidden_states) + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.embed_out(hidden_states[:, slice_indices, :]) - lm_loss = None + loss = None if labels is not None: - # move labels to correct device to enable model parallelism - labels = labels.to(lm_logits.device) - # we are doing next-token prediction; shift prediction scores and input ids by one - shift_logits = lm_logits[:, :-1, :].contiguous() - labels = labels[:, 1:].contiguous() - loss_fct = CrossEntropyLoss() - lm_loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)) + loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) if not return_dict: - output = (lm_logits,) + outputs[1:] - return ((lm_loss,) + output) if lm_loss is not None else output + output = (logits,) + outputs[1:] + return ((loss,) + output) if loss is not None else output return CausalLMOutputWithPast( - loss=lm_loss, - logits=lm_logits, + loss=loss, + logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) - def _reorder_cache(self, past_key_values, beam_idx): - reordered_past = () - for layer_past in past_key_values: - reordered_past += ( - tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past[:2]) - + layer_past[2:], - ) - return reordered_past - @add_start_docstrings( """ @@ -1224,27 +953,8 @@ class GPTNeoXForSequenceClassification(GPTNeoXPreTrainedModel): loss = None if labels is not None: - labels = labels.to(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" + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) - 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,) + outputs[1:] return ((loss,) + output) if loss is not None else output @@ -1319,9 +1029,7 @@ class GPTNeoXForTokenClassification(GPTNeoXPreTrainedModel): loss = None if labels is not None: - labels = labels.to(logits.device) - loss_fct = CrossEntropyLoss() - loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) + loss = self.loss_function(logits, labels, self.config) if not return_dict: output = (logits,) + outputs[2:] @@ -1403,29 +1111,16 @@ class GPTNeoXForQuestionAnswering(GPTNeoXPreTrainedModel): start_logits = start_logits.squeeze(-1).contiguous() end_logits = end_logits.squeeze(-1).contiguous() - total_loss = None + 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 + loss = self.loss_function(start_logits, end_logits, start_positions, end_positions) if not return_dict: output = (start_logits, end_logits) + outputs[2:] - return ((total_loss,) + output) if total_loss is not None else output + return ((loss,) + output) if loss is not None else output return QuestionAnsweringModelOutput( - loss=total_loss, + loss=loss, start_logits=start_logits, end_logits=end_logits, hidden_states=outputs.hidden_states, diff --git a/src/transformers/models/gpt_neox/modular_gpt_neox.py b/src/transformers/models/gpt_neox/modular_gpt_neox.py new file mode 100644 index 0000000000..95bbdaa776 --- /dev/null +++ b/src/transformers/models/gpt_neox/modular_gpt_neox.py @@ -0,0 +1,837 @@ +from typing import Callable, Optional, Tuple, Union + +import torch +import torch.utils.checkpoint +from torch import nn + +from ...activations import ACT2FN +from ...cache_utils import Cache, DynamicCache +from ...generation import GenerationMixin +from ...modeling_flash_attention_utils import FlashAttentionKwargs +from ...modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + QuestionAnsweringModelOutput, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from ...modeling_utils import ALL_ATTENTION_FUNCTIONS +from ...processing_utils import Unpack +from ...utils import ( + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + logging, + replace_return_docstrings, +) +from ..llama.modeling_llama import ( + LlamaModel, + LlamaPreTrainedModel, + LlamaRotaryEmbedding, + rotate_half, +) + + +logger = logging.get_logger(__name__) + + +_CHECKPOINT_FOR_DOC = "trl-internal-testing/tiny-random-GPTNeoXForCausalLM" +_REAL_CHECKPOINT_FOR_DOC = "EleutherAI/gpt-neox-20b" +_CONFIG_FOR_DOC = "GPTNeoXConfig" + + +class GPTNeoXMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size) + self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size) + self.act = ACT2FN[config.hidden_act] + + def forward(self, hidden_states): + hidden_states = self.dense_h_to_4h(hidden_states) + hidden_states = self.act(hidden_states) + hidden_states = self.dense_4h_to_h(hidden_states) + return hidden_states + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + + # Keep half or full tensor for later concatenation + rotary_dim = cos.shape[-1] + q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] + k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] + + # Apply rotary embeddings on the first half or full tensor + q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin) + k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin) + + # Concatenate back to full shape + q_embed = torch.cat([q_embed, q_pass], dim=-1) + k_embed = torch.cat([k_embed, k_pass], dim=-1) + return q_embed, k_embed + + +def eager_attention_forward( + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: torch.Tensor, + scaling: float, + dropout: float = 0.0, + head_mask: Optional[torch.Tensor] = None, + **kwargs, +): + attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling + + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key.shape[-2]] + attn_weights = attn_weights + causal_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) + + # Mask heads if we want to + if head_mask is not None: + attn_weights = attn_weights * head_mask + + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) + attn_output = torch.matmul(attn_weights, value) + + # Reshape outputs + attn_output = attn_output.transpose(1, 2).contiguous() + + return attn_output, attn_weights + + +class GPTNeoXAttention(nn.Module): + def __init__(self, config, layer_idx=None): + super().__init__() + self.config = config + self.head_size = config.hidden_size // config.num_attention_heads + self.attention_dropout = config.attention_dropout + self.rotary_ndims = int(self.head_size * config.rotary_pct) + self.scaling = self.head_size**-0.5 + self.is_causal = True + self.layer_idx = layer_idx + + self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=config.attention_bias) + self.dense = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias) + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: torch.FloatTensor, + head_mask: Optional[torch.FloatTensor] = None, + layer_past: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ): + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, 3 * self.head_size) + + qkv = self.query_key_value(hidden_states).view(hidden_shape).transpose(1, 2) + query_states, key_states, value_states = qkv.chunk(3, dim=-1) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + # Cache QKV values + if layer_past is not None: + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_ndims, + "cache_position": cache_position, + } + key_states, value_states = layer_past.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # Checking for fallbacks in case an unsupported feature is requested + attention_type = self.config._attn_implementation + if (output_attentions or head_mask is not None) and self.config._attn_implementation in [ + "sdpa", + "flash_attention_2", + ]: + logger.warning_once( + f"Setting `attention_type` to `eager` because `{attention_type}` does not support" + f" `output_attentions=True` or `head_mask`." + ) + attention_type = "eager" + + elif self.training and self.attention_dropout > 0 and self.config._attn_implementation == "flex_attention": + logger.warning_once( + f"Setting `attention_type` to `eager` because `dropout` is not supported in `{attention_type}`." + ) + attention_type = "eager" + + attention_interface: Callable = eager_attention_forward + attention_interface = ( + ALL_ATTENTION_FUNCTIONS[attention_type] if attention_type != "eager" else attention_interface + ) + + # Compute attention + attn_output, attn_weights = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask, + scaling=self.scaling, + dropout=0.0 if not self.training else self.attention_dropout, + head_mask=head_mask, + **kwargs, + ) + + # Reshape outputs and final projection + attn_output = attn_output.reshape(*input_shape, -1).contiguous() + attn_output = self.dense(attn_output) + + return attn_output, attn_weights + + +class GPTNeoXLayer(nn.Module): + def __init__(self, config, layer_idx): + super().__init__() + self.use_parallel_residual = config.use_parallel_residual + self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.post_attention_dropout = nn.Dropout(config.hidden_dropout) + self.post_mlp_dropout = nn.Dropout(config.hidden_dropout) + self.attention = GPTNeoXAttention(config, layer_idx) + self.mlp = GPTNeoXMLP(config) + + def forward( + self, + hidden_states: Optional[torch.FloatTensor], + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = False, + layer_past: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ): + attn_output, attn_weights = self.attention( + self.input_layernorm(hidden_states), + attention_mask=attention_mask, + position_ids=position_ids, + layer_past=layer_past, + head_mask=head_mask, + use_cache=use_cache, + output_attentions=output_attentions, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + attn_output = self.post_attention_dropout(attn_output) + + if self.use_parallel_residual: + # pseudocode: + # x = x + attn(ln1(x)) + mlp(ln2(x)) + mlp_output = self.mlp(self.post_attention_layernorm(hidden_states)) + mlp_output = self.post_mlp_dropout(mlp_output) + hidden_states = mlp_output + attn_output + hidden_states + else: + # pseudocode: + # x = x + attn(ln1(x)) + # x = x + mlp(ln2(x)) + attn_output = attn_output + hidden_states + mlp_output = self.mlp(self.post_attention_layernorm(attn_output)) + mlp_output = self.post_mlp_dropout(mlp_output) + hidden_states = mlp_output + attn_output + + outputs = (hidden_states,) + if output_attentions: + outputs += (attn_weights,) + + return outputs + + +class GPTNeoXRotaryEmbedding(LlamaRotaryEmbedding): + pass + + +class GPTNeoXPreTrainedModel(LlamaPreTrainedModel): + base_model_prefix = "gpt_neox" + _no_split_modules = ["GPTNeoXLayer"] + _keys_to_ignore_on_load_unexpected = [r"attention.bias", r"attention.masked_bias"] + + def _init_weights(self, module): + """Initialize the weights""" + if isinstance(module, nn.Linear): + 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) + + +GPT_NEOX_START_DOCSTRING = None # Will be picked up by modular +GPT_NEOX_INPUTS_DOCSTRING = None # Will be picked up by modular + + +class GPTNeoXModel(LlamaModel, nn.Module): + def __init__(self, config): + nn.Module.__init__(config) + self.config = config + + self.embed_in = nn.Embedding(config.vocab_size, config.hidden_size) + self.emb_dropout = nn.Dropout(config.hidden_dropout) + self.layers = nn.ModuleList([GPTNeoXLayer(config, i) for i in range(config.num_hidden_layers)]) + self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.rotary_emb = GPTNeoXRotaryEmbedding(config=config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_in + + def set_input_embeddings(self, value): + self.embed_in = value + + @add_start_docstrings_to_model_forward(GPT_NEOX_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + real_checkpoint=_REAL_CHECKPOINT_FOR_DOC, + output_type=BaseModelOutputWithPast, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + past_key_values: Optional[Cache] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> 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 + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + use_cache = use_cache if use_cache is not None else self.config.use_cache + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + 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 + + if inputs_embeds is None: + inputs_embeds = self.embed_in(input_ids) + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape bsz x n_heads x N x N + # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] + # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length] + converted_head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers) + # Flex Attention converts it to a separate mask + if head_mask is not None: + converted_head_mask = ~converted_head_mask.bool() * torch.finfo(inputs_embeds.dtype).min + converted_head_mask = converted_head_mask.to(dtype=self.dtype, device=self.device) + head_mask = converted_head_mask + + hidden_states = self.emb_dropout(inputs_embeds) + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + all_attentions = () if output_attentions else None + all_hidden_states = () if output_hidden_states else None + for i, layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if self.gradient_checkpointing and self.training: + outputs = self._gradient_checkpointing_func( + layer.__call__, + hidden_states, + causal_mask, + position_ids, + head_mask[i], + use_cache, + past_key_values, + output_attentions, + cache_position, + position_embeddings, + ) + else: + outputs = layer( + hidden_states, + attention_mask=causal_mask, + position_ids=position_ids, + head_mask=head_mask[i], + layer_past=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + hidden_states = outputs[0] + + if output_attentions: + all_attentions = all_attentions + (outputs[1],) + + hidden_states = self.final_layer_norm(hidden_states) + # Add last hidden state + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + output = BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values, + hidden_states=all_hidden_states, + attentions=all_attentions, + ) + return output if return_dict else output.to_tuple() + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +@add_start_docstrings( + """GPTNeoX Model with a `language modeling` head on top for CLM fine-tuning.""", GPT_NEOX_START_DOCSTRING +) +class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin): + _tied_weights_keys = ["embed_out.weight"] + _tp_plan = {"embed_out": "colwise_rep"} + + def __init__(self, config): + super().__init__(config) + + self.gpt_neox = GPTNeoXModel(config) + self.embed_out = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_output_embeddings(self): + return self.embed_out + + def set_output_embeddings(self, new_embeddings): + self.embed_out = new_embeddings + + @add_start_docstrings_to_model_forward(GPT_NEOX_INPUTS_DOCSTRING.format("batch_size, sequence_length")) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + past_key_values: Optional[Union[Cache, Tuple[Tuple[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, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in + `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are + ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, GPTNeoXForCausalLM, GPTNeoXConfig + >>> import torch + + >>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b") + >>> config = GPTNeoXConfig.from_pretrained("EleutherAI/gpt-neox-20b") + >>> config.is_decoder = True + >>> model = GPTNeoXForCausalLM.from_pretrained("EleutherAI/gpt-neox-20b", config=config) + + >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt") + >>> outputs = model(**inputs) + + >>> prediction_logits = outputs.logits + ```""" + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + outputs = self.gpt_neox( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs[0] + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.embed_out(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) + + if not return_dict: + output = (logits,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The GPTNeoX Model transformer with a sequence classification head on top (linear layer). + + [`GPTNeoXForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-1) 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). + """, + GPT_NEOX_START_DOCSTRING, +) +class GPTNeoXForSequenceClassification(GPTNeoXPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.gpt_neox = GPTNeoXModel(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(GPT_NEOX_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=SequenceClassifierOutputWithPast, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + past_key_values: Optional[Union[Cache, Tuple[Tuple[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[torch.Tensor], 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 + + outputs = self.gpt_neox( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = outputs[0] + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size, sequence_length = input_ids.shape[:2] + else: + batch_size, sequence_length = inputs_embeds.shape[:2] + + 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_once( + 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: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + if not return_dict: + output = (pooled_logits,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +class GPTNeoXForTokenClassification(GPTNeoXPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + + self.gpt_neox = GPTNeoXModel(config) + self.dropout = nn.Dropout(config.classifier_dropout) + self.classifier = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(GPT_NEOX_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint="LarsJonasson/pythia-410m-deduped-sft-swedish", + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + expected_loss=0.25, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, 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, TokenClassifierOutput]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *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 + + outputs = self.gpt_neox( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + 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 = outputs[0] + hidden_states = self.dropout(hidden_states) + logits = self.classifier(hidden_states) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + if not return_dict: + output = (logits,) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The GPT-NeoX Model transformer with a span classification head on top for extractive question-answering tasks like + SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + GPT_NEOX_START_DOCSTRING, +) +class GPTNeoXForQuestionAnswering(GPTNeoXPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.gpt_neox = GPTNeoXModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, 2) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(GPT_NEOX_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.gpt_neox( + input_ids, + attention_mask=attention_mask, + 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() + + loss = None + if start_positions is not None and end_positions is not None: + loss = self.loss_function(start_logits, end_logits, start_positions, end_positions) + + if not return_dict: + output = (start_logits, end_logits) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return QuestionAnsweringModelOutput( + loss=loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +__all__ = [ + "GPTNeoXForCausalLM", + "GPTNeoXForQuestionAnswering", + "GPTNeoXForSequenceClassification", + "GPTNeoXForTokenClassification", + "GPTNeoXLayer", + "GPTNeoXModel", + "GPTNeoXPreTrainedModel", +] diff --git a/tests/models/gpt_neox/test_modeling_gpt_neox.py b/tests/models/gpt_neox/test_modeling_gpt_neox.py index 97403cb8e5..1580526b6f 100644 --- a/tests/models/gpt_neox/test_modeling_gpt_neox.py +++ b/tests/models/gpt_neox/test_modeling_gpt_neox.py @@ -18,7 +18,7 @@ import unittest from parameterized import parameterized -from transformers import AutoTokenizer, GPTNeoXConfig, is_torch_available, set_seed +from transformers import AutoTokenizer, DynamicCache, GPTNeoXConfig, is_torch_available, set_seed from transformers.testing_utils import require_torch, slow, torch_device from ...generation.test_utils import GenerationTesterMixin @@ -232,13 +232,22 @@ class GPTNeoXModelTester: cache_inputs = {"input_ids": input_ids[:, :cached_len], "attention_mask": input_mask[:, :cached_len]} non_cache_inputs = {"input_ids": input_ids[:, cached_len:], "attention_mask": input_mask} + def copy_cache(cache: DynamicCache): + """Deep copy a DynamicCache to reuse the same one multiple times.""" + new_cache = cache + for i in range(len(cache)): + new_cache.key_cache[i] = cache.key_cache[i].clone() + new_cache.value_cache[i] = cache.value_cache[i].clone() + # Cached forward once with the attention mask provided and the other time without it (which should assume full attention) + # We need to run both on a copy of the cache, otherwise it is modified in-place cache_outputs = model(**cache_inputs) + cache = cache_outputs.past_key_values full_outputs_with_attention_mask = model( - **non_cache_inputs, past_key_values=cache_outputs.past_key_values + **non_cache_inputs, past_key_values=copy_cache(cache) ).last_hidden_state full_outputs_without_attention_mask = model( - non_cache_inputs["input_ids"], past_key_values=cache_outputs.past_key_values + non_cache_inputs["input_ids"], past_key_values=copy_cache(cache) ).last_hidden_state self.parent.assertTrue( diff --git a/utils/check_config_attributes.py b/utils/check_config_attributes.py index 6262b1902a..5b5fd52e6f 100644 --- a/utils/check_config_attributes.py +++ b/utils/check_config_attributes.py @@ -201,6 +201,7 @@ SPECIAL_CASES_TO_ALLOW = { "giou_cost", "giou_loss_coefficient", ], + "GPTNeoXConfig": ["rotary_emb_base"], }