upd

config.json

@@ -0,0 +1,37 @@
+{
+  "_name_or_path": "bigcode/santacoder",
+  "activation_function": "gelu_fast",
+  "architectures": [
+    "GPT2LMHeadCustomModel"
+  ],
+  "attention_head_type": "multiquery",
+  "attn_pdrop": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_gpt2_mq.GPT2CustomConfig",
+    "AutoModelForCausalLM": "modeling_gpt2_mq.GPT2LMHeadCustomModel"
+  },
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_embd": 2048,
+  "n_head": 16,
+  "n_inner": 8192,
+  "n_layer": 24,
+  "n_positions": 2048,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.0.dev0",
+  "use_cache": false,
+  "vocab_size": 49280
+}

last-checkpoint/config.json

@@ -0,0 +1,37 @@
+{
+  "_name_or_path": "bigcode/santacoder",
+  "activation_function": "gelu_fast",
+  "architectures": [
+    "GPT2LMHeadCustomModel"
+  ],
+  "attention_head_type": "multiquery",
+  "attn_pdrop": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_gpt2_mq.GPT2CustomConfig",
+    "AutoModelForCausalLM": "modeling_gpt2_mq.GPT2LMHeadCustomModel"
+  },
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_embd": 2048,
+  "n_head": 16,
+  "n_inner": 8192,
+  "n_layer": 24,
+  "n_positions": 2048,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.0.dev0",
+  "use_cache": false,
+  "vocab_size": 49280
+}

last-checkpoint/configuration_gpt2_mq.py

@@ -0,0 +1,201 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and Hugging Face Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  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.
+""" Custom GPT-2 configuration"""
+from collections import OrderedDict
+from typing import Any, List, Mapping, Optional
+from enum import Enum
+
+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__)
+
+GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "gpt2": "https://huggingface.co/gpt2/resolve/main/config.json",
+    "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/config.json",
+    "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/config.json",
+    "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/config.json",
+    "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/config.json",
+}
+
+MULTI_HEAD = "multihead"
+MULTI_QUERY = "multiquery"
+
+
+class GPT2CustomConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a [`GPT2Model`] or a [`TFGPT2Model`]. It is used to
+    instantiate a GPT-2 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-2
+    [gpt2](https://huggingface.co/gpt2) 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 50257):
+            Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`].
+        n_positions (`int`, *optional*, defaults to 1024):
+            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 768):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        n_head (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        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"`):
+            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.
+        summary_type (`string`, *optional*, defaults to `"cls_index"`):
+            Argument used when doing sequence summary, used in the models [`GPT2DoubleHeadsModel`] and
+            [`TFGPT2DoubleHeadsModel`].
+
+            Has to be one of the following options:
+
+                - `"last"`: Take the last token hidden state (like XLNet).
+                - `"first"`: Take the first token hidden state (like BERT).
+                - `"mean"`: Take the mean of all tokens hidden states.
+                - `"cls_index"`: Supply a Tensor of classification token position (like GPT/GPT-2).
+                - `"attn"`: Not implemented now, use multi-head attention.
+        summary_use_proj (`bool`, *optional*, defaults to `True`):
+            Argument used when doing sequence summary, used in the models [`GPT2DoubleHeadsModel`] and
+            [`TFGPT2DoubleHeadsModel`].
+
+            Whether or not to add a projection after the vector extraction.
+        summary_activation (`str`, *optional*):
+            Argument used when doing sequence summary. Used in for the multiple choice head in
+            [`GPT2DoubleHeadsModel`].
+
+            Pass `"tanh"` for a tanh activation to the output, any other value will result in no activation.
+        summary_proj_to_labels (`bool`, *optional*, defaults to `True`):
+            Argument used when doing sequence summary, used in the models [`GPT2DoubleHeadsModel`] and
+            [`TFGPT2DoubleHeadsModel`].
+
+            Whether the projection outputs should have `config.num_labels` or `config.hidden_size` classes.
+        summary_first_dropout (`float`, *optional*, defaults to 0.1):
+            Argument used when doing sequence summary, used in the models [`GPT2DoubleHeadsModel`] and
+            [`TFGPT2DoubleHeadsModel`].
+
+            The dropout ratio to be used after the projection and activation.
+        scale_attn_weights (`bool`, *optional*, defaults to `True`):
+            Scale attention weights by dividing by sqrt(head_dim)..
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
+            Whether to additionally scale attention weights by `1 / layer_idx + 1`.
+        reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
+            Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
+            dot-product/softmax to float() when training with mixed precision.
+
+    Example:
+
+    ```python
+    >>> from transformers import GPT2Config, GPT2Model
+
+    >>> # Initializing a GPT2 configuration
+    >>> configuration = GPT2Config()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = GPT2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "gpt2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "n_embd",
+        "max_position_embeddings": "n_positions",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50257,
+        n_positions=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        summary_type="cls_index",
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        scale_attn_weights=True,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        scale_attn_by_inverse_layer_idx=False,
+        reorder_and_upcast_attn=False,
+        attention_head_type=MULTI_HEAD,
+        **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.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.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+        self.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
+        self.reorder_and_upcast_attn = reorder_and_upcast_attn
+        self.attention_head_type = attention_head_type
+        # assert attention_head_type in [AttentionType.MULTI_HEAD, AttentionType.MULTI_QUERY]
+        assert attention_head_type in [MULTI_HEAD, MULTI_QUERY]
+
+        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, **kwargs)

last-checkpoint/modeling_gpt2_mq.py

@@ -0,0 +1,346 @@
+"""PyTorch OpenAI GPT-2 model modified with MultiQuery attention"""
+
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.cuda.amp import autocast
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
+
+from transformers.utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2Block, GPT2PreTrainedModel, GPT2LMHeadModel
+from .configuration_gpt2_mq import GPT2CustomConfig, MULTI_QUERY, MULTI_HEAD
+
+
+
+class GPT2MQAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False, layer_idx=None):
+        super().__init__()
+        assert config.attention_head_type == MULTI_QUERY
+
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
+                1, 1, max_positions, max_positions
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e4))
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.split_size = self.embed_dim
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+
+        self.scale_attn_weights = config.scale_attn_weights
+        if is_cross_attention:
+            raise NotImplementedError("Cross-attention not implemented for MQA")
+        self.is_cross_attention = is_cross_attention
+
+        # Layer-wise attention scaling, reordering, and upcasting
+        self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
+        self.layer_idx = layer_idx
+        self.reorder_and_upcast_attn = config.reorder_and_upcast_attn
+
+        if self.is_cross_attention:
+            self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
+            self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
+        else:
+            # self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+            self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
+            # Keys and values are shared across heads
+            self.kv_attn = Conv1D(2 * self.head_dim, self.embed_dim)
+        self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
+        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
+
+        # Prune conv1d layers
+        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
+        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
+        # Update hyper params
+        self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
+        self.num_heads = self.num_heads - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
+        # query: (b, num_heads * sq, head_dim)
+        # key: (b, head_dim, sk)
+        # value: (b, sk, head_dim)
+        batch_size = query.size(0)
+        query_length = query.size(1) // self.num_heads
+        key_length = key.size(2)
+        # (b, num_heads * sq, head_dim) x (b, head_dim, sk) -> (b, num_heads * sq, sk)
+        attn_weights = torch.bmm(query, key)
+        # -> (b, num_heads, sq, sk)
+        attn_weights = attn_weights.view(batch_size, self.num_heads, query_length, key_length)
+
+        if self.scale_attn_weights:
+            attn_weights = attn_weights / torch.tensor(
+                value.size(-1) ** 0.5, dtype=attn_weights.dtype, device=attn_weights.device
+            )
+
+        # Layer-wise attention scaling
+        if self.scale_attn_by_inverse_layer_idx:
+            attn_weights = attn_weights / float(self.layer_idx + 1)
+
+        if not self.is_cross_attention:
+            # if only "normal" attention layer implements causal mask
+            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].to(torch.bool)
+            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)
+
+        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)
+
+        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
+        attn_weights = attn_weights.type(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
+
+        # (b, num_heads, sq, sk) -> (b, num_heads * sq, sk)
+        _attn_weights = attn_weights.view(batch_size, self.num_heads * query_length, key_length)
+        # (b, num_heads * sq, sk) x (b, sk, head_dim) -> (b, num_heads * sq, head_dim)
+        attn_output = torch.bmm(_attn_weights, value)
+        attn_output = attn_output.view(batch_size, self.num_heads, query_length, self.head_dim)
+
+        return attn_output, attn_weights
+
+    def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None):
+        # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
+        bsz, num_heads, q_seq_len, dk = query.size()
+        _, _, k_seq_len, _ = key.size()
+
+        # Preallocate attn_weights for `baddbmm`
+        attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)
+
+        # Compute Scale Factor
+        scale_factor = 1.0
+        if self.scale_attn_weights:
+            scale_factor /= float(value.size(-1)) ** 0.5
+
+        if self.scale_attn_by_inverse_layer_idx:
+            scale_factor /= float(self.layer_idx + 1)
+
+        # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
+        with autocast(enabled=False):
+            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
+            attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
+            attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
+
+        if not self.is_cross_attention:
+            # if only "normal" attention layer implements causal mask
+            query_length, key_length = query.size(-2), key.size(-2)
+            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()
+            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)
+
+        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)
+
+        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
+        if attn_weights.dtype != torch.float32:
+            raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
+        attn_weights = attn_weights.type(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 _split_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Splits hidden_size dim into attn_head_size and num_heads
+        """
+        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
+        tensor = tensor.view(new_shape)
+        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def _merge_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into hidden_size
+        """
+        tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
+        if encoder_hidden_states is not None:
+            raise NotImplementedError("Cross-attention not implemented for MQA")
+            if not hasattr(self, "q_attn"):
+                raise ValueError(
+                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
+                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
+                )
+
+            query = self.q_attn(hidden_states)
+            key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
+            attention_mask = encoder_attention_mask
+        else:
+            query = self.q_attn(hidden_states)
+            key, value = self.kv_attn(hidden_states).split(self.head_dim, dim=2)
+
+
+        batch_size, seq_length = query.shape[:2]
+        # (query_length, batch, num_heads, head_dim)
+        # (batch, num_heads * query_length, head_dim)\
+
+        # (batch, query_length, hidden_size) -> (batch, num_heads, query_length, head_dim)
+        query = query.view(batch_size, seq_length, self.num_heads, self.head_dim).permute([0, 2, 1, 3])
+        # -> (batch, num_heads * query_length, head_dim)
+        query = query.reshape(batch_size, self.num_heads * seq_length, self.head_dim)
+
+        # (batch, query_length, hidden_size) -> (batch, query_length * num_heads, head_dim)
+        # query = query.view(
+        #     batch_size, seq_length, self.num_heads, self.head_dim,
+        # ).reshape(
+        #     batch_size, seq_length * self.num_heads, self.head_dim
+        # )
+        key = key.permute(0, 2, 1)  # (batch_size, head_dim, seq_length)
+        # value (batch_size, seq_length, head_dim)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            # Concatenate on sequence dimension
+            key = torch.cat((past_key, key), dim=-1)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        if self.reorder_and_upcast_attn:
+            raise NotImplementedError("Reorder and upcast attention not implemented for MQA")
+            attn_output, attn_weights = self._upcast_and_reordered_attn(query, key, value, attention_mask, head_mask)
+        else:
+            attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+
+        attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
+        attn_output = self.c_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)
+
+
+# inherit from gpt_modeling.py, and override `attn` module
+class GPT2CustomBlock(GPT2Block):
+
+    def __init__(self, config: GPT2CustomConfig, layer_idx=None):
+        super().__init__(config, layer_idx)
+        # Override attention module if using multiquery
+        if config.attention_head_type == MULTI_QUERY:
+            self.attn = GPT2MQAttention(config, layer_idx=layer_idx)
+            if config.add_cross_attention:
+                raise NotImplementedError("Cross-attention not implemented for MQA")
+
+
+# inherit from gpt_modeling.py and override `__init__` method
+class GPT2CustomModel(GPT2Model):
+    config_class = GPT2CustomConfig
+    
+    def __init__(self, config):
+        GPT2PreTrainedModel.__init__(self, config)
+
+        self.embed_dim = config.hidden_size
+
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
+
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([GPT2CustomBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
+        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()
+
+
+class GPT2LMHeadCustomModel(GPT2LMHeadModel):
+    config_class = GPT2CustomConfig
+
+    def __init__(self, config):
+        GPT2PreTrainedModel.__init__(self, config)
+        self.transformer = GPT2CustomModel(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()

last-checkpoint/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f97c566b5b347b099859815c57d664d96a266dfabaef6d5851cb0285779db19
+size 8999298713

last-checkpoint/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef4e98a5d26b5fe348f0be8f72b41fd9795a8747c9f5e0e649deb4980cbaf81d
+size 4600336581

last-checkpoint/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26a0cfbb451dff697bffb046244bcaf99faab7209ad569304f25d2aa889095d8
+size 13815

last-checkpoint/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45ff25655ad9115362ea4a876294d11cf1984e796d3d02dff20ef96e1abd3871
+size 13815

last-checkpoint/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fb33b1da2125feaa501f6acf7c67409ff703b0320e8e55bf616863891dd3346
+size 557

last-checkpoint/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbd404efb46db8da5fdab9c606729e0c72c6120c20e22d4b2f62b2ce1db8e12e
+size 627

last-checkpoint/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a70a8fd8575918ccc22e31554743931b9254ca12a0ab397bb4a644c56366bc84
+size 3515

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f052d4abe92e3aa71fb0696f789025fd9684b9f8427d088b12db4204fd7eb6a9
+size 4600336581

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93c306135bed134f36f2dcf224ab2305240315cbf0507b7d7793fe6fab400e32
+size 3515