Upload model

config.json

@@ -0,0 +1,39 @@
+{
+  "activation_function": "silu",
+  "architectures": [
+    "PlusModelForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_nano.NanoConfig",
+    "AutoModelForCausalLM": "modeling_nano.PlusModelForCausalLM"
+  },
+  "bos_token_id": 1,
+  "combined_qkv": true,
+  "eos_token_id": 2,
+  "expanded_lm_head_size": 8192,
+  "expanded_wte_size": 8192,
+  "experimental_full_adaption_rank": null,
+  "ffn": "parallel",
+  "full_adaptation_has_pre_proj": false,
+  "full_adaptation_type": "no",
+  "hidden_size": 1088,
+  "initializer_range": 0.02,
+  "intermediate_size": 4352,
+  "kv_hidden_size": null,
+  "layer_norm_epsilon": 1e-06,
+  "layernorm": "llamarmsnorm",
+  "lm_head_bias": false,
+  "lm_head_projection_bias": false,
+  "max_position_embeddings": 2048,
+  "model_type": "nano",
+  "num_attention_heads": 17,
+  "num_hidden_layers": 14,
+  "pre_proj_dim": null,
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.36.2",
+  "use_bias": false,
+  "use_cache": true,
+  "vocab_size": 32000
+}

configuration_nano.py

@@ -0,0 +1,86 @@
+from collections import OrderedDict
+from typing import Any, List, Mapping, Optional
+
+from transformers import PreTrainedTokenizer, TensorType, is_torch_available
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class NanoConfig(PretrainedConfig):
+    model_type = "nano"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "hidden_size",
+        "max_position_embeddings": "max_position_embeddings",
+        "num_attention_heads": "num_attention_heads",
+        "num_hidden_layers": "num_hidden_layers",
+    }
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        max_position_embeddings=2048,
+        expanded_wte_size=None,
+        expanded_lm_head_size=None,
+        hidden_size=768,
+        kv_hidden_size=None,  # in case you want to use cross-attention
+        num_hidden_layers=10,
+        num_attention_heads=12,
+        intermediate_size=None,
+        activation_function="silu",
+        layer_norm_epsilon=1e-6,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        combined_qkv=True,
+        use_bias=False,
+        lm_head_projection_bias=False,
+        lm_head_bias=False,
+        layernorm="llamarmsnorm", # layernorm, llamarmsnorm
+        rope_scaling=None,
+        rope_theta=10000,
+        ffn="llama-like",
+        experimental_full_adaption_rank = None, # 8
+        full_adaptation_has_pre_proj = True,
+        pre_proj_dim = 1536,
+        full_adaptation_type="no", # "lora", "no", "linear", "linear-r", "linear-ra"
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.pre_proj_dim = pre_proj_dim
+        self.full_adaptation_has_pre_proj = full_adaptation_has_pre_proj
+        self.full_adaptation_type = full_adaptation_type
+        self.tie_word_embeddings = tie_word_embeddings
+        self.experimental_full_adaption_rank = experimental_full_adaption_rank
+        self.ffn = ffn
+        self.rope_theta=rope_theta
+        self.layernorm = layernorm
+        self.rope_scaling=rope_scaling
+        self.lm_head_projection_bias = lm_head_projection_bias
+        self.kv_hidden_size = kv_hidden_size
+        self.lm_head_bias = lm_head_bias
+        self.use_bias = use_bias
+        self.expanded_wte_size = expanded_wte_size
+        self.expanded_lm_head_size = expanded_lm_head_size
+        self.combined_qkv = combined_qkv
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = (
+            intermediate_size if intermediate_size is not None else hidden_size * 4
+        )
+        self.activation_function = activation_function
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.36.2"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ff293e94ee49a78432543259bf7da34c929ad98b92c8486dffd3193b68a393
+size 1090333184

modeling_nano.py

@@ -0,0 +1,694 @@
+import math
+import os
+import random
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from einops import repeat
+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, QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast, TokenClassifierOutput)
+from transformers.modeling_utils import PreTrainedModel, SequenceSummary
+from transformers.utils import (ModelOutput, logging)
+from transformers.utils.model_parallel_utils import (assert_device_map,
+                                                     get_device_map)
+
+from .configuration_nano import NanoConfig
+from transformers.models.llama.modeling_llama import LlamaRMSNorm, LlamaDynamicNTKScalingRotaryEmbedding, LlamaRotaryEmbedding, LlamaLinearScalingRotaryEmbedding
+
+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, unsqueeze_dim=1):
+    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    sin = sin[position_ids].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 NanoAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.head_dim = config.hidden_size // config.num_attention_heads
+        assert (
+            self.head_dim * config.num_attention_heads == config.hidden_size
+        ), "d_model must be divisible by n_head"
+        self.use_bias = config.use_bias
+
+        if not config.combined_qkv or config.kv_hidden_size is not None:
+            self.query = nn.Linear(
+                config.hidden_size, config.hidden_size, bias=self.use_bias
+            )
+            self.key = nn.Linear(
+                config.hidden_size
+                if not config.kv_hidden_size
+                else config.kv_hidden_size,
+                config.hidden_size,
+                bias=self.use_bias,
+            )
+            self.value = nn.Linear(
+                config.hidden_size
+                if not config.kv_hidden_size
+                else config.kv_hidden_size,
+                config.hidden_size,
+                bias=self.use_bias,
+            )
+        else:
+            self.qkv = nn.Linear(
+                config.hidden_size, config.hidden_size * 3, bias=self.use_bias
+            )
+        self.out = nn.Linear(config.hidden_size, config.hidden_size, bias=self.use_bias)
+        self._init_rope()
+
+    def _init_rope(self):
+        if self.config.rope_scaling is None:
+            self.rotary_emb = LlamaRotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.config.max_position_embeddings,
+                base=self.config.rope_theta,
+            )
+        else:
+            scaling_type = self.config.rope_scaling["type"]
+            scaling_factor = self.config.rope_scaling["factor"]
+            if scaling_type == "linear":
+                self.rotary_emb = LlamaLinearScalingRotaryEmbedding(
+                    self.head_dim,
+                    max_position_embeddings=self.config.max_position_embeddings,
+                    scaling_factor=scaling_factor,
+                    base=self.config.rope_theta,
+                )
+            elif scaling_type == "dynamic":
+                self.rotary_emb = LlamaDynamicNTKScalingRotaryEmbedding(
+                    self.head_dim,
+                    max_position_embeddings=self.max_position_embeddings,
+                    scaling_factor=scaling_factor,
+                    base=self.config.rope_theta,
+                )
+            else:
+                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
+
+    def forward(self, x0, x1=None, causal=False, mask=None, position_ids=None, use_cache=True, layer_past=None):
+        batch_size = x0.size(0)
+
+        def split_heads(x):
+            return x.view(
+                batch_size, -1, self.config.num_attention_heads, self.head_dim
+            ).transpose(1, 2)
+
+        if not self.config.combined_qkv:
+            q = split_heads(self.query(x0))
+            k = split_heads(self.key(x1) if x1 is not None else self.key(x0))
+            v = split_heads(
+                self.value(x1 if x1 is not None else x0)
+            )
+        else:
+            q, k, v = self.qkv(x0).chunk(3,-1)
+            q = split_heads(q)
+            k = split_heads(k)
+            v = split_heads(v)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            k = torch.cat((past_key, k), dim=-2)
+            v = torch.cat((past_value, v), dim=-2)
+
+        cos, sin = self.rotary_emb(v, seq_len=v.shape[-2])
+        if self.config.experimental_full_adaption_rank is not None:
+            position_ids = position_ids.repeat_interleave(x0.shape[1]//position_ids.shape[-1],dim=1)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin, position_ids)
+
+        if use_cache is True:
+            present = (k,v)
+        else:
+            present = None
+
+        attn_output = F.scaled_dot_product_attention(
+            q, k, v, attn_mask=None, dropout_p=0.0, is_causal=causal
+        )
+        attn_output = (
+            attn_output.transpose(1, 2)
+            .contiguous()
+            .view(batch_size, -1, self.config.hidden_size)
+        )
+        return self.out(attn_output), present
+
+
+class NanoGLU(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.gate_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.up_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.down_proj = nn.Linear(
+            config.intermediate_size, config.hidden_size, bias=False
+        )
+        self.act_fn = ACT2FN[config.activation_function]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class NanoBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.attn = NanoAttention(config)
+        self.ffn = NanoGLU(config)
+
+        ln_class = LlamaRMSNorm if config.layernorm=="llamarmsnorm" else nn.LayerNorm
+        self.ln1 = ln_class(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.ln2 = ln_class(config.hidden_size, eps=config.layer_norm_epsilon)
+
+    def forward(self, x, mask=None, position_ids=None, use_cache=True, layer_past=None):
+
+        if self.config.ffn == "llamalike":
+            residual = x
+            x = self.ln1(x)
+            attn_out, attn_outs = self.attn(x, causal=True, mask=mask, position_ids=position_ids, use_cache=use_cache, layer_past=layer_past)
+            x = residual + attn_out
+
+            residual = x
+            x = self.ln2(x)
+            x = self.ffn(x)
+            x = residual + x
+        else: # ffn == "parallel"
+            attn_in = self.ln1(x)
+            ffn_in = self.ln2(x)
+
+            attn_out, attn_outs = self.attn(attn_in, causal=True, mask=mask, position_ids=position_ids, use_cache=use_cache, layer_past=layer_past)
+            ffn_out = self.ffn(ffn_in)
+
+            x = x + attn_out + ffn_out
+        
+        if not use_cache: attn_outs = None
+        return (x, attn_outs)
+
+
+
+class NanoPreTrainedModel(PreTrainedModel):
+    config_class = NanoConfig
+    base_model_prefix = "transformer"
+    is_parallelizable = False
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["NanoBlock"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    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 _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, NanoModel):
+            module.gradient_checkpointing = value
+
+class Split(nn.Module):
+    def __init__(self, splits):
+        super().__init__()
+        self.splits=splits
+    def forward(self, x):
+        bs, tokens, _ = x.shape
+        # print("SPLIT X0 SHAPE", x.shape)
+        x = x.view(bs, tokens, self.splits, -1)
+        x = x.permute(0, 1, 2, 3).reshape(bs, tokens * self.splits, -1)
+        # print("SPLIT X1 SHAPE", x.shape)
+        return x
+    
+class Recombine(nn.Module):
+    def __init__(self, splits):
+        super().__init__()
+        self.splits = splits
+    def forward(self, x):
+        bs, _, _ = x.shape
+        # print("RECOMBINE X SHAPE", x.shape)
+        tokens = x.shape[1] // self.splits
+        # print("RECOMBINE TOKENS", tokens, bs)
+        x = x.view(bs, tokens, -1)
+        # print("RECOMBINE X1.SHAPE", x.shape)
+        return x
+
+class Residual(nn.Module):
+    def __init__(self, module, a=None):
+        super().__init__()
+        self.module = module
+        self.a = nn.Parameter(torch.tensor(a, dtype=torch.bfloat16)) if a is not None else None
+    def forward(self, x):
+        return self.module(x) * (self.a if self.a is not None else 1) + x
+
+class LoRA(nn.Module):
+    def __init__(self, d, r, a=1):
+        super().__init__()
+        self.fn_i = nn.Linear(d, r)
+        self.fn_o = nn.Linear(r, d)
+        self.a = nn.Parameter(torch.tensor(a, dtype=self.fn_i.weight.dtype))
+    def forward(self, x):
+        return self.fn_o(self.fn_i(x)) * self.a + x
+    def get_delta_w(self):
+        return torch.mm(self.fn_o.weight, self.fn_i.weight) * self.a
+
+class NanoModel(NanoPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        ln_class = LlamaRMSNorm if config.layernorm=="llamarmsnorm" else nn.LayerNorm
+        
+        if config.experimental_full_adaption_rank is None:
+            if config.expanded_wte_size is not None:
+                self.wte = nn.Sequential(
+                    nn.Embedding(config.vocab_size, config.expanded_wte_size),
+                    nn.Linear(config.expanded_wte_size, config.hidden_size),
+                )
+            else:
+                self.wte = nn.Embedding(config.vocab_size, config.hidden_size)
+        else:
+            assert config.expanded_wte_size is not None, "experimental full adaptation of token embeddings requires expanded_wte_size to be set"
+            # self.wte = nn.Sequential(
+            #     nn.Embedding(config.vocab_size, config.expanded_wte_size),
+            #     LoRA(config.expanded_wte_size, config.experimental_full_adaption_rank),
+            #     Split(config.expanded_wte_size//config.hidden_size)
+            # )
+            self.d_0 = config.expanded_wte_size if (config.full_adaptation_has_pre_proj == False) else config.pre_proj_dim
+            # print("d_0", d_0)
+            self.wte = nn.Sequential(
+                nn.Embedding(config.vocab_size, config.expanded_wte_size),
+                (
+                    nn.Linear(config.expanded_wte_size, config.pre_proj_dim) if config.full_adaptation_has_pre_proj else nn.Identity()
+                ),
+                (
+                    LoRA(self.d_0, config.experimental_full_adaption_rank)
+                    if config.full_adaptation_type == "lora" else
+                    nn.Linear(self.d_0, self.d_0)
+                    if config.full_adaptation_type == "linear" else
+                    Residual(
+                        nn.Linear(self.d_0, self.d_0)
+                    )
+                    if config.full_adaptation_type == "linear-r" else
+                    Residual(
+                        nn.Linear(self.d_0, self.d_0), 1
+                    )
+                    if config.full_adaptation_type == "linear-ra" else
+                    nn.Identity()
+                ),
+                Split(self.d_0//config.hidden_size)
+            )
+        self.h = nn.ModuleList(
+            [NanoBlock(config) for i in range(config.num_hidden_layers)]
+        )
+        self.ln_f = ln_class(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.model_parallel = False
+        self.device_map = None
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.wte[0] if self.config.expanded_wte_size is not None else self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        if self.config.expanded_wte_size is not None:
+            self.wte[0] = new_embeddings
+        else:
+            self.wte = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+        # soooo not all of the params are able to be used, since I just copied this framework from modeling_gpt2            
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_length,
+                input_shape[-1] + past_length,
+                dtype=torch.long,
+                device=device,
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            attention_mask = attention_mask[:, None, None, :]
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+
+        if self.config.add_cross_attention and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_attention_mask = None
+
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+            # print("inputs embeds shape", inputs_embeds.shape)
+        
+        hidden_states = inputs_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        # output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
+        output_shape = (-1,) + (hidden_states.shape[1],) + (hidden_states.size(-1),)
+        # print(output_shape, "output shape")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                # logger.warning_once(
+                #     "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                # )
+                use_cache = False
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                if layer_past is not None:
+                    layer_past = tuple(
+                        past_state.to(hidden_states.device)
+                        for past_state in layer_past
+                    )
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            outputs = block(hidden_states, mask=attention_mask, position_ids=position_ids, use_cache=use_cache, layer_past=layer_past)
+            hidden_states = outputs[0]
+            if use_cache == True:
+                presents = presents + (outputs[1],)
+
+        hidden_states = self.ln_f(hidden_states)
+        hidden_states = hidden_states.view(output_shape)
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, None, all_hidden_states, None, None]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=None,
+            cross_attentions=None,
+        )
+
+class NanoModelForCausalLM(NanoPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = NanoModel(config)
+        if config.experimental_full_adaption_rank is None or config.full_adaptation_type == "no":
+            if (config.expanded_lm_head_size is not None):
+                self.lm_head = nn.Sequential(
+                    nn.Linear(
+                        config.hidden_size, config.expanded_lm_head_size, bias=config.lm_head_projection_bias
+                    ),
+                    nn.Linear(
+                        config.expanded_lm_head_size, config.vocab_size, bias=config.lm_head_bias
+                    ),
+                )
+            else:
+                self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)
+        else:
+            d_0 = config.expanded_lm_head_size if (not config.full_adaptation_has_pre_proj) else config.pre_proj_dim
+            self.lm_head = nn.Sequential(
+                Recombine(d_0//config.hidden_size),
+                nn.Identity() if not config.full_adaptation_has_pre_proj else nn.Linear(d_0, config.expanded_lm_head_size),
+                (
+                    LoRA(config.expanded_lm_head_size, config.experimental_full_adaption_rank)
+                    if config.full_adaptation_type == "lora" else
+                    nn.Linear(config.expanded_lm_head_size, config.expanded_lm_head_size)
+                    if config.full_adaptation_type == "linear" else
+                    Residual(
+                        nn.Linear(config.expanded_lm_head_size, config.expanded_lm_head_size)
+                    )
+                    if config.full_adaptation_type == "linear-r" else
+                    Residual(
+                        nn.Linear(config.expanded_lm_head_size, config.expanded_lm_head_size), 1
+                    )
+                    if config.full_adaptation_type == "linear-ra" else
+                    nn.Identity()
+                ),
+                
+                nn.Linear(config.expanded_lm_head_size, config.vocab_size)
+            )
+        self.model_parallel = False
+        self.device_map = None
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head if (self.config.experimental_full_adaption_rank is None and self.config.expanded_lm_head_size is None) else self.lm_head[-1] 
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
+    ):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past_key_values:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "position_ids": position_ids,
+                "attention_mask": attention_mask,
+                "token_type_ids": token_type_ids,
+            }
+        )
+        return model_inputs
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: 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, CausalLMOutputWithCrossAttentions]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        # print("Hidden states shape", hidden_states.shape)
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(lm_logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            cross_attentions=transformer_outputs.cross_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past_key_values
+        )
+
+
+class VTMModelForCausalLM(NanoModelForCausalLM):
+    _tied_weights_keys = ["lm_head.3.weight"]
+    def __init__(self, config):
+        super().__init__(config)
+
+class VTMPreProjModelForCausalLM(NanoModelForCausalLM):
+    _tied_weights_keys = ["lm_head.3.weight"]
+    def __init__(self, config):
+        super().__init__(config)
+
+class PlusModelForCausalLM(NanoModelForCausalLM):
+    _tied_weights_keys = ["lm_head.1.weight"]
+    def __init__(self, config):
+        super().__init__(config)