Upload 11 files

config.json

@@ -0,0 +1,56 @@
+{
+  "_name_or_path": "chat",
+  "architectures": [
+    "InternLMXComposerForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_InternLM_XComposer.InternLMXComposerConfig",
+    "AutoModel": "modeling_InternLM_XComposer.InternLMXComposerForCausalLM",
+    "AutoModelForCausalLM": "modeling_InternLM_XComposer.InternLMXComposerForCausalLM"
+  },
+  "bias": true,
+  "bos_token_id": 1,
+  "device": "cuda",
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "intern_converted_llm": true,
+  "internlm_lora": {
+    "freeze": false,
+    "learn_param": [
+      "q",
+      "v",
+      "ffn"
+    ],
+    "lora_alpha": 256,
+    "lora_dropout": 0.05,
+    "lora_r": 256
+  },
+  "kqvo_bias": true,
+  "lora_cfg": {
+    "freeze": false,
+    "learn_param": [
+      "q",
+      "v",
+      "ffn"
+    ],
+    "lora_alpha": 256,
+    "lora_dropout": 0.05,
+    "lora_r": 256
+  },
+  "max_position_embeddings": 2048,
+  "model_type": "InternLMXComposer",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_quant": 32,
+  "num_query_token": 64,
+  "pad_token_id": -1,
+  "rms_norm_eps": 1e-05,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.30.2",
+  "use_cache": true,
+  "vocab_size": 103168
+}

configuration_InternLM_XComposer.py

@@ -0,0 +1,62 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+INTERNLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class InternLMXComposerConfig(PretrainedConfig):
+
+    model_type = "InternLMXComposer"
+    _auto_class = "AutoConfig"
+
+    def __init__(
+        self,
+        vocab_size=103168,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=-1,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        bias=True,
+        num_query_token=32,
+        num_quant=32,
+        intern_converted_llm=True,
+        kqvo_bias=True,
+        device='cuda',
+        internlm_lora=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.bias = bias
+        self.num_query_token = num_query_token
+        self.num_quant = num_quant
+        self.internlm_lora = internlm_lora
+        self.kqvo_bias = kqvo_bias
+        self.intern_converted_llm = intern_converted_llm
+        self.device = device
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

modeling_InternLM.py

@@ -0,0 +1,1246 @@
+import math
+from typing import List, Union
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+import torch.utils.checkpoint
+from einops import rearrange
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+
+from .configuration_InternLM_XComposer import InternLMXComposerConfig
+from .modeling_utils import LoRALinear
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "InternLMXComposerConfig"
+
+
+def rotary_embed(x1, x2, cos, sin, conj):
+    x1, x2 = x1.float(), x2.float()
+    if conj:
+        x1, x2 = x1 * cos + x2 * sin, x1 * sin + x2 * cos
+    else:
+        x1, x2 = x1 * cos - x2 * sin, x1 * sin + x2 * cos
+    return x1, x2
+
+
+class LegacyApplyRotaryEmbQKV_(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, qkv, cos, sin, cos_k=None, sin_k=None, interleaved=False):
+        """
+            qkv: (batch_size, seqlen, 3, nheads, headdim)
+            cos, sin: (seqlen, rotary_dim / 2)
+            cos_k, sin_k: (seqlen, rotary_dim / 2), optional
+            interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
+                1st half and 2nd half (GPT-NeoX style).
+        rotary_dim must be <= headdim
+        Apply rotary embedding *inplace* to the first rotary_dim of q and k.
+        """
+        batch, seqlen, three, nheads, headdim = qkv.shape
+        assert three == 3
+        rotary_seqlen, rotary_dim = cos.shape
+        rotary_dim *= 2
+        assert rotary_dim <= headdim
+        assert seqlen <= rotary_seqlen
+        cos_k = cos if cos_k is None else cos_k
+        sin_k = sin if sin_k is None else sin_k
+        assert sin.shape == cos_k.shape == sin_k.shape == (rotary_seqlen, rotary_dim // 2)
+        q_ro = qkv[:, :, 0, :, :rotary_dim]
+        q1, q2 = q_ro.chunk(2, dim=-1) if not interleaved else (q_ro[..., ::2], q_ro[..., 1::2])
+        # rotary_emb.apply_rotary(q1, q2, rearrange(cos[:seqlen], 's d -> s 1 d'),
+        #                         rearrange(sin[:seqlen], 's d -> s 1 d'), q1, q2, False)
+        q1, q2 = rotary_embed(q1, q2, rearrange(cos[:seqlen], 's d -> s 1 d'), rearrange(sin[:seqlen], 's d -> s 1 d'), False)
+        qkv[:, :, 0, :, :rotary_dim] = torch.cat([q1, q2], dim=-1)
+        k_ro = qkv[:, :, 1, :, :rotary_dim]
+        k1, k2 = k_ro.chunk(2, dim=-1) if not interleaved else (k_ro[..., ::2], k_ro[..., 1::2])
+        # rotary_emb.apply_rotary(k1, k2, rearrange(cos_k[:seqlen], 's d -> s 1 d'),
+        #                         rearrange(sin_k[:seqlen], 's d -> s 1 d'), k1, k2, False)
+        k1, k2 = rotary_embed(k1, k2, rearrange(cos_k[:seqlen], 's d -> s 1 d'), rearrange(sin_k[:seqlen], 's d -> s 1 d'), False)
+        qkv[:, :, 1, :, :rotary_dim] = torch.cat([k1, k2], dim=-1)
+        ctx.save_for_backward(cos, sin, cos_k, sin_k)
+        ctx.interleaved = interleaved
+        return qkv
+
+    @staticmethod
+    def backward(ctx, dqkv):
+        cos, sin, cos_k, sin_k = ctx.saved_tensors
+        _, seqlen, _, _, headdim = dqkv.shape
+        rotary_dim = cos.shape[-1]
+        rotary_dim *= 2
+        dq_ro = dqkv[:, :, 0, :, :rotary_dim]
+        dq1, dq2 = (dq_ro.chunk(2, dim=-1) if not ctx.interleaved
+                    else (dq_ro[..., ::2], dq_ro[..., 1::2]))
+        rotary_emb.apply_rotary(dq1, dq2, rearrange(cos[:seqlen], 's d -> s 1 d'),
+                                rearrange(sin[:seqlen], 's d -> s 1 d'), dq1, dq2, True)
+        dk_ro = dqkv[:, :, 1, :, :rotary_dim]
+        dk1, dk2 = (dk_ro.chunk(2, dim=-1) if not ctx.interleaved
+                    else (dk_ro[..., ::2], dk_ro[..., 1::2]))
+        rotary_emb.apply_rotary(dk1, dk2, rearrange(cos_k[:seqlen], 's d -> s 1 d'),
+                                rearrange(sin_k[:seqlen], 's d -> s 1 d'), dk1, dk2, True)
+        return dqkv, None, None, None, None, None
+
+
+class ConvertedInternLMRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim: int, base=10000, scale_base=0, device=None):
+        """ """
+        super().__init__()
+        # Generate and save the inverse frequency buffer (non trainable)
+        inv_freq = 1.0 / (base**(
+            torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.scale_base = scale_base
+        scale = ((torch.arange(0, dim, 2, device=device, dtype=torch.float32) +
+                  0.4 * dim) / (1.4 * dim) if scale_base > 0 else None)
+        self.register_buffer("scale", scale)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+
+    def _update_cos_sin_cache(self, x, indexes):
+        """x: (batch, seqlen, nheads, headdim) or (batch, seqlen, 3, nheads, headdim)"""
+        if not isinstance(indexes, int):
+            seqlen = indexes.max().item() + 1
+        else:
+            seqlen = indexes + 1  # eval_forward
+        # Reset the tables if the sequence length has changed,
+        # or if we're on a new device (possibly due to tracing for instance)
+        if seqlen > self._seq_len_cached or self._cos_cached.device != x.device or self._cos_cached.dtype != x.dtype:
+            self._seq_len_cached = seqlen
+            t = torch.arange(seqlen,
+                             device=x.device,
+                             dtype=self.inv_freq.dtype)
+            # Don't do einsum, it converts fp32 to fp16
+            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(x.dtype)
+                self._sin_cached = torch.sin(freqs).to(x.dtype)
+            else:
+                power = (torch.arange(
+                    seqlen, dtype=self.scale.dtype, device=self.scale.device) -
+                         seqlen // 2) / self.scale_base
+                scale = self.scale.to(device=power.device)**rearrange(
+                    power, "s -> s 1")
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(x.dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(x.dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(x.dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(x.dtype)
+
+    def eval_forward(self, qkv, seqlen_offset=0):
+        """
+        seqlen_offset: can be used in generation where the qkv being passed in is only the last
+        token in the batch.
+        """
+        self._update_cos_sin_cache(qkv, seqlen_offset + qkv.shape[1])
+        if self.scale is None:
+            return legacy_apply_rotary_embed_qkv(
+                qkv, self._cos_cached[seqlen_offset:],
+                self._sin_cached[seqlen_offset:])
+        else:
+            return legacy_apply_rotary_embed_qkv(
+                qkv,
+                self._cos_cached[seqlen_offset:],
+                self._sin_cached[seqlen_offset:],
+                self._cos_k_cached[seqlen_offset:],
+                self._sin_k_cached[seqlen_offset:],
+            )
+
+
+legacy_apply_rotary_embed_qkv = LegacyApplyRotaryEmbQKV_.apply
+
+
+class InternConvertedInternLMAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+    def __init__(self, config: InternLMXComposerConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads}).")
+        if config.lora_cfg is None:
+            self.q_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=config.kqvo_bias)
+            self.k_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=config.kqvo_bias)
+            self.v_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=config.kqvo_bias)
+            self.o_proj = nn.Linear(self.num_heads * self.head_dim,
+                                    self.hidden_size,
+                                    bias=config.kqvo_bias)
+
+        else:
+            lora_cfg = config.lora_cfg
+            if 'q' in lora_cfg['learn_param']:
+                self.q_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=config.kqvo_bias,
+                                         **lora_cfg)
+            else:
+                self.q_proj = nn.Linear(
+                    self.hidden_size,
+                    self.num_heads * self.head_dim,
+                    bias=config.kqvo_bias,
+                )
+            if 'k' in lora_cfg['learn_param']:
+                self.k_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=config.kqvo_bias,
+                                         **lora_cfg)
+            else:
+                self.k_proj = nn.Linear(
+                    self.hidden_size,
+                    self.num_heads * self.head_dim,
+                    bias=config.kqvo_bias,
+                )
+            if 'v' in lora_cfg['learn_param']:
+                self.v_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=config.kqvo_bias,
+                                         **lora_cfg)
+            else:
+                self.v_proj = nn.Linear(
+                    self.hidden_size,
+                    self.num_heads * self.head_dim,
+                    bias=config.kqvo_bias,
+                )
+
+            if 'o' in lora_cfg['learn_param']:
+                self.o_proj = LoRALinear(self.num_heads * self.head_dim,
+                                         self.hidden_size,
+                                         bias=config.kqvo_bias,
+                                         **lora_cfg)
+            else:
+                self.o_proj = nn.Linear(
+                    self.num_heads * self.head_dim,
+                    self.hidden_size,
+                    bias=config.kqvo_bias,
+                )
+
+        self.rotary_emb = ConvertedInternLMRotaryEmbedding(self.head_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        q = query_states
+        k = key_states
+        v = value_states
+
+        qkv = torch.cat([q, k, v], dim=2).contiguous()
+        qkv = qkv.view(bsz, q_len, -1)
+        qkv = rearrange(qkv,
+                        "b s (three h d) -> b s three h d",
+                        three=3,
+                        d=self.head_dim)
+
+        if past_key_value is not None:
+            qkv = self.rotary_emb.eval_forward(
+                qkv, seqlen_offset=past_key_value[0].shape[2])
+        else:
+            qkv = self.rotary_emb.eval_forward(qkv)
+
+        query_states, key_states, value_states = qkv.unbind(2)
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        # cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(
+            2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}")
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(
+                attn_weights,
+                torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights,
+                                             dim=-1,
+                                             dtype=torch.float32).to(
+                                                 query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}")
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class ConvertedLoRALinear(nn.Linear):
+    def __init__(self,
+                 in_features: int,
+                 out_features: int,
+                 bias: bool = True,
+                 device=None,
+                 dtype=None,
+                 lora_r=8,
+                 lora_alpha=16,
+                 lora_dropout=0.05,
+                 **kwargs) -> None:
+        super().__init__(in_features, out_features, bias, device, dtype)
+        self.lora_r = lora_r
+        self.lora_alpha = lora_alpha
+        if lora_dropout > 0.:
+            self.lora_dropout = nn.Dropout(p=lora_dropout)
+        else:
+            self.lora_dropout = lambda x: x
+        self.lora_scaling = self.lora_alpha / self.lora_r
+
+        self.lora_A = nn.Linear(in_features,
+                                self.lora_r,
+                                bias=False,
+                                device=device,
+                                dtype=dtype)
+        self.lora_B = nn.Linear(self.lora_r,
+                                out_features,
+                                bias=False,
+                                device=device,
+                                dtype=dtype)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if hasattr(self, 'lora_A'):
+            # initialize A the same way as the default for nn.Linear and B to zero
+            nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
+            nn.init.zeros_(self.lora_B.weight)
+            # print ("lora weight init {} {}".format(torch.mean(self.lora_A.weight), torch.mean(self.lora_B.weight)))
+
+    def forward(self, x):
+        orig_type = x.dtype
+        res = super().forward(x)
+
+        dim = int(res.shape[-1] // 2)
+
+        r1 = res[..., :dim]
+        r2 = res[..., dim:]
+
+        r1 = r1.float()
+        r2 = r2.float()
+        x_ = x.float()
+
+        tmp = self.lora_B(self.lora_A(
+            self.lora_dropout(x_))) * self.lora_scaling
+        tmp1 = tmp[..., ::2]
+        tmp2 = tmp[..., 1::2]
+
+        r1 += tmp1
+        r2 += tmp2
+
+        r1 = r1.to(orig_type)
+        r2 = r2.to(orig_type)
+
+        res = torch.cat([r1, r2], -1)
+
+        # res += self.lora_B(self.lora_A(
+        #     self.lora_dropout(x))) * self.lora_scaling
+        return res
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: torch.Size,
+                      dtype: torch.dtype,
+                      device: torch.device,
+                      past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len),
+                      torch.tensor(torch.finfo(dtype).min, device=device),
+                      device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([
+            torch.zeros(
+                tgt_len, past_key_values_length, dtype=dtype, device=device),
+            mask
+        ],
+                         dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len,
+                                         tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor,
+                 dtype: torch.dtype,
+                 tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len,
+                                                  src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool),
+                                     torch.finfo(dtype).min)
+
+
+class InternLMRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        InternLMRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1,
+                                                               keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance +
+                                                    self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class InternLMRotaryEmbedding(torch.nn.Module):
+    def __init__(self,
+                 dim,
+                 max_position_embeddings=2048,
+                 base=10000,
+                 device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base
+                          **(torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached,
+                         device=self.inv_freq.device,
+                         dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached",
+                             emb.cos()[None, None, :, :],
+                             persistent=False)
+        self.register_buffer("sin_cached",
+                             emb.sin()[None, None, :, :],
+                             persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached,
+                             device=x.device,
+                             dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached",
+                                 emb.cos()[None, None, :, :],
+                                 persistent=False)
+            self.register_buffer("sin_cached",
+                                 emb.sin()[None, None, :, :],
+                                 persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].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)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    gather_indices = position_ids[:, None, :, None]  # [bs, 1, seq_len, 1]
+    gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
+    cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2,
+                       gather_indices)
+    sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2,
+                       gather_indices)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class InternLMMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int,
+                 hidden_act: str, config: InternLMXComposerConfig):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        if config.lora_cfg is not None and 'ffn' in config.lora_cfg[
+                'learn_param']:
+            lora_cfg = config.lora_cfg
+            self.down_proj = LoRALinear(intermediate_size,
+                                        hidden_size,
+                                        bias=False,
+                                        **lora_cfg)
+            self.up_proj = LoRALinear(hidden_size,
+                                      intermediate_size,
+                                      bias=False,
+                                      **lora_cfg)
+        else:
+            self.down_proj = nn.Linear(intermediate_size,
+                                       hidden_size,
+                                       bias=False)
+            self.up_proj = nn.Linear(hidden_size,
+                                     intermediate_size,
+                                     bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class InternLMAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+    def __init__(self, config: InternLMXComposerConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads}).")
+        if config.lora_cfg is None:
+            self.q_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=False)
+            self.k_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=False)
+            self.v_proj = nn.Linear(self.hidden_size,
+                                    self.num_heads * self.head_dim,
+                                    bias=False)
+            self.o_proj = nn.Linear(self.num_heads * self.head_dim,
+                                    self.hidden_size,
+                                    bias=False)
+        else:
+            lora_cfg = config.lora_cfg
+            if 'q' in lora_cfg['learn_param']:
+                self.q_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=False,
+                                         **lora_cfg)
+            else:
+                self.q_proj = nn.Linear(self.hidden_size,
+                                        self.num_heads * self.head_dim,
+                                        bias=False)
+
+            if 'k' in lora_cfg['learn_param']:
+                self.k_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=False,
+                                         **lora_cfg)
+            else:
+                self.k_proj = nn.Linear(self.hidden_size,
+                                        self.num_heads * self.head_dim,
+                                        bias=False)
+
+            if 'v' in lora_cfg['learn_param']:
+                self.v_proj = LoRALinear(self.hidden_size,
+                                         self.num_heads * self.head_dim,
+                                         bias=False,
+                                         **lora_cfg)
+            else:
+                self.v_proj = nn.Linear(self.hidden_size,
+                                        self.num_heads * self.head_dim,
+                                        bias=False)
+
+            if 'o' in lora_cfg['learn_param']:
+                self.o_proj = LoRALinear(self.num_heads * self.head_dim,
+                                         self.hidden_size,
+                                         bias=False,
+                                         **lora_cfg)
+            else:
+                self.o_proj = nn.Linear(self.num_heads * self.head_dim,
+                                        self.hidden_size,
+                                        bias=False)
+
+        self.rotary_emb = InternLMRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(
+            2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}")
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(
+                attn_weights,
+                torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights,
+                                             dim=-1,
+                                             dtype=torch.float32).to(
+                                                 query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}")
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class InternLMDecoderLayer(nn.Module):
+    def __init__(self, config: InternLMXComposerConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        if hasattr(config,
+                   'intern_converted_llm') and config.intern_converted_llm:
+            self.self_attn = InternConvertedInternLMAttention(config=config)
+        else:
+            self.self_attn = InternLMAttention(config=config)
+        self.mlp = InternLMMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            config=config,
+        )
+        self.input_layernorm = InternLMRMSNorm(config.hidden_size,
+                                               eps=config.rms_norm_eps)
+        self.post_attention_layernorm = InternLMRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor,
+                                                 torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative 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.
+            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`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, )
+
+        if output_attentions:
+            outputs += (self_attn_weights, )
+
+        if use_cache:
+            outputs += (present_key_value, )
+
+        return outputs
+
+
+class InternLMPreTrainedModel(PreTrainedModel):
+    config_class = InternLMXComposerConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["InternLMDecoderLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, InternLMModel):
+            module.gradient_checkpointing = value
+
+
+class InternLMModel(InternLMPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`InternLMDecoderLayer`]
+
+    Args:
+        config: InternLMXComposerConfig
+    """
+    def __init__(self, config: InternLMXComposerConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size,
+                                         self.padding_idx)
+        self.layers = nn.ModuleList([
+            InternLMDecoderLayer(config)
+            for _ in range(config.num_hidden_layers)
+        ])
+        self.norm = InternLMRMSNorm(config.hidden_size,
+                                    eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape,
+                                        inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask,
+                                              inputs_embeds.dtype,
+                                              tgt_len=input_shape[-1]).to(
+                                                  inputs_embeds.device)
+            combined_attention_mask = (expanded_attn_mask
+                                       if combined_attention_mask is None else
+                                       expanded_attn_mask +
+                                       combined_attention_mask)
+
+        return combined_attention_mask
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else
+                                self.config.output_hidden_states)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError(
+                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        if query_embeds is not None:
+            inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)
+            batch_size, seq_length, _ = inputs_embeds.shape
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(past_key_values_length,
+                                        seq_length + past_key_values_length,
+                                        dtype=torch.long,
+                                        device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length_with_past),
+                                        dtype=torch.bool,
+                                        device=inputs_embeds.device)
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds,
+            past_key_values_length)
+
+        hidden_states = 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
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states, )
+
+            past_key_value = past_key_values[
+                idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (
+                    layer_outputs[2 if output_attentions else 1], )
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1], )
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states, )
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v for v in
+                [hidden_states, next_cache, all_hidden_states, all_self_attns]
+                if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class InternLMForCausalLM(InternLMPreTrainedModel):
+    lora_cfg = None  # init in MiniGPT4
+
+    def __init__(self, config):
+        super().__init__(config)
+        # TODO: find a way to explicitly initialize InternLM
+        setattr(config, 'lora_cfg', self.lora_cfg)
+
+        if hasattr(config, 'kqvo_bias'):
+            setattr(config, 'kqvo_bias', config.kqvo_bias)
+        else:
+            setattr(config, 'kqvo_bias', False)
+        self.model = InternLMModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size,
+                                 config.vocab_size,
+                                 bias=False)
+        if hasattr(config, 'ex_size'):
+            self.ex_size = config.ex_size
+        else:
+            self.ex_size = 0
+
+        if hasattr(config, 'sp_id'):
+            self.sp_id = config.sp_id
+        else:
+            self.sp_id = -1
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @classmethod
+    def from_pretrained(cls,
+                        pretrained_model_name_or_path,
+                        llm_cfg=None,
+                        *model_args,
+                        **kwargs):
+        if llm_cfg:
+            if 'torch_dtype' in kwargs:
+                llm_cfg.torch_dtype = kwargs['torch_dtype']
+            if 'load_in_8bit' in kwargs:
+                llm_cfg.load_in_8bit = kwargs['load_in_8bit']
+            if 'device_map' in kwargs:
+                llm_cfg.device_map = kwargs['device_map']
+            return cls._from_config(llm_cfg)
+        else:
+            return super().from_pretrained(pretrained_model_name_or_path,
+                                           *model_args, **kwargs)
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, InternLMForCausalLM
+
+        >>> model = InternLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        ```"""
+
+        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
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            query_embeds=query_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+
+            loss_fct = CrossEntropyLoss(reduce=False)
+            loss_reduce = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            shift_labels = shift_labels.to(shift_logits.device)
+            ###
+            if self.sp_id >= 0:
+                ori_mask = (shift_labels != self.sp_id).float()
+                ori_mask = ori_mask * (shift_labels >= 0).float()
+                local_mask = (shift_labels == self.sp_id).float()
+            else:
+                ori_mask = (shift_labels <
+                            self.config.vocab_size - self.ex_size).float()
+                ori_mask = ori_mask * (shift_labels >= 0).float()
+                local_mask = (shift_labels >=
+                              self.config.vocab_size - self.ex_size).float()
+
+            # Enable model parallelism
+
+            loss = loss_reduce(shift_logits, shift_labels)
+
+            loss_all = loss_fct(shift_logits, shift_labels)
+            loss_o = (loss_all * ori_mask).sum() / ori_mask.sum()
+            if torch.sum(local_mask) == 0:
+                loss_l = loss_o * 0
+            else:
+                loss_l = (loss_all * local_mask).sum() / local_mask.sum()
+
+        if not return_dict:
+            output = (logits, ) + outputs[1:]
+            return (loss, ) + output if loss is not None else output
+
+        if (self.ex_size > 0 or self.sp_id >= 0) and labels is not None:
+            return loss, loss_o, loss_l
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(self,
+                                      input_ids,
+                                      query_embeds=None,
+                                      past_key_values=None,
+                                      attention_mask=None,
+                                      inputs_embeds=None,
+                                      **kwargs):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        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)
+                query_embeds = 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({
+            "position_ids": position_ids,
+            "query_embeds": query_embeds,
+            "past_key_values": past_key_values,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+        })
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(
+                past_state.index_select(0, beam_idx)
+                for past_state in layer_past), )
+        return reordered_past

modeling_InternLM_XComposer.py

@@ -0,0 +1,259 @@
+import copy
+import os
+import sys
+
+dir_path = os.path.dirname(os.path.realpath(__file__))
+sys.path.insert(0, dir_path)
+
+import contextlib
+
+import torch.utils.checkpoint
+from torch.nn import LayerNorm
+from torchvision import transforms
+from torchvision.transforms.functional import InterpolationMode
+from PIL import Image
+
+from .modeling_perceive_sampler import BertConfig, BertLMHeadModel
+from .modeling_vit import *
+from .modeling_InternLM import *
+from .modeling_utils import *
+
+from transformers.utils import logging
+logger = logging.get_logger(__name__)
+
+
+class InternLMXComposerForCausalLM(PreTrainedModel):
+    config_class = InternLMXComposerConfig
+    _auto_class = "AutoModelForCausalLM"
+
+    gen_config = dict(
+        num_beams=5,
+        do_sample=False,
+        min_length=1,
+        repetition_penalty=1.5,
+        length_penalty=1.0,
+        temperature=1.0,
+        max_new_tokens=200,
+    )
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        print('Init VIT ... ', end='')
+        self.visual_encoder = create_eva_vit_g()
+        self.ln_vision = LayerNorm(self.visual_encoder.num_features)
+        print('Done')
+
+        print('Init Perceive Sampler ... ', end='')
+        with all_logging_disabled():
+            self.Qformer, self.query_tokens = self.init_qformer(
+                config.num_query_token, self.visual_encoder.num_features)
+            self.Qformer.bert.embeddings.word_embeddings = None
+            self.Qformer.bert.embeddings.position_embeddings = None
+            for layer in self.Qformer.bert.encoder.layer:
+                layer.output = None
+                layer.intermediate = None
+            self.Qformer.cls = None
+        print('Done')
+
+        print('Init InternLM ... ', end='')
+        self.flag_image_start = nn.Parameter(torch.zeros([1, 1, 4096]))
+        self.flag_image_end = nn.Parameter(torch.zeros([1, 1, 4096]))
+        self.flag_image_start.requires_grad = False
+        self.flag_image_end.requires_grad = False
+
+        internlm_lora = config.internlm_lora
+        self.internlm_lora = internlm_lora
+        setattr(InternLMForCausalLM, 'lora_cfg', internlm_lora)
+
+        if int(torch.__version__[0]) == 1:
+            self.internlm_model = InternLMForCausalLM._from_config(config).to(
+                torch.float16)
+        else:
+            assert int(torch.__version__[0]) == 2
+            # speed up init llm
+            with torch.device('meta'):
+                self.internlm_model = InternLMForCausalLM._from_config(config)
+            self.internlm_model.to_empty(device=config.device).to(torch.float16)
+        for n, m in self.internlm_model.named_modules():
+            if 'lora' in n:
+                m.float()
+
+        self.internlm_proj = nn.Linear(self.Qformer.config.hidden_size,
+                                    self.internlm_model.config.hidden_size)
+        print('Done')
+
+        self.vis_processor = transforms.Compose([
+            transforms.Resize((224, 224),
+                              interpolation=InterpolationMode.BICUBIC),
+            transforms.ToTensor(),
+            transforms.Normalize((0.48145466, 0.4578275, 0.40821073),
+                                 (0.26862954, 0.26130258, 0.27577711)),
+        ])
+
+        self.tokenizer = None
+
+    @property
+    def eoh(self):
+        return self.tokenizer.decode(torch.Tensor([103027]),
+                                     skip_special_tokens=True)
+
+    @property
+    def eoa(self):
+        return self.tokenizer.decode(torch.Tensor([103028]),
+                                     skip_special_tokens=True)
+
+    def maybe_autocast(self, dtype=torch.float16):
+        # if on cpu, don't use autocast
+        # if on gpu, use autocast with dtype if provided, otherwise use torch.float16
+        enable_autocast = self.device != torch.device("cpu")
+
+        if enable_autocast:
+            return torch.cuda.amp.autocast(dtype=dtype)
+        else:
+            return contextlib.nullcontext()
+
+    @classmethod
+    def init_qformer(cls,
+                     num_query_token,
+                     vision_width,
+                     cross_attention_freq=2,
+                     pretrain=True):
+        encoder_config = BertConfig.from_pretrained("bert-base-uncased")
+        encoder_config.encoder_width = vision_width
+        # insert cross-attention layer every other block
+        encoder_config.add_cross_attention = True
+        encoder_config.cross_attention_freq = cross_attention_freq
+        encoder_config.query_length = num_query_token
+        # if pretrain:
+        #     Qformer = BertLMHeadModel.from_pretrained("bert-base-uncased",
+        #                                               config=encoder_config)
+        # else:
+        Qformer = BertLMHeadModel(config=encoder_config)
+        query_tokens = nn.Parameter(
+            torch.zeros(1, num_query_token, encoder_config.hidden_size))
+        query_tokens.data.normal_(mean=0.0,
+                                  std=encoder_config.initializer_range)
+        return Qformer, query_tokens
+
+    def encode_img(self, image):
+        if image is None:
+            return None
+        if isinstance(image, str):
+            image = Image.open(image).convert("RGB")
+            image = self.vis_processor(image).unsqueeze(0).to(self.device)
+        else:
+            assert isinstance(image, torch.Tensor)
+        device = image.device
+        with self.maybe_autocast():
+            image_embeds = self.ln_vision(
+                self.visual_encoder(image)).to(device)
+            image_atts = torch.ones(image_embeds.size()[:-1],
+                                    dtype=torch.long).to(device)
+            query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1,
+                                                    -1)
+            query_output = self.Qformer.bert(
+                query_embeds=query_tokens,
+                encoder_hidden_states=image_embeds,
+                encoder_attention_mask=image_atts,
+                return_dict=True,
+            )
+            inputs_internlm = self.internlm_proj(query_output.last_hidden_state)
+            inputs_internlm = torch.cat([
+                self.flag_image_start.expand(inputs_internlm.shape[0], -1, -1),
+                inputs_internlm,
+                self.flag_image_end.expand(inputs_internlm.shape[0], -1, -1)
+            ],
+                                      dim=1)
+        return inputs_internlm
+
+    def encode_text(self, text, add_special_tokens=False):
+        text_token_ids = self.tokenizer(
+            text,
+            return_tensors='pt',
+            add_special_tokens=add_special_tokens,
+        ).input_ids.to(self.device)
+        text_embeds = self.internlm_model.model.embed_tokens(text_token_ids)
+        return text_embeds
+
+    def decode_text(self, out_embeds):
+        out_text = self.tokenizer.batch_decode(out_embeds,
+                                               skip_special_tokens=True)[0]
+        out_text = out_text.split(self.eoa)[0]
+        return out_text
+
+    def wrap_text(self, user_text, bot_text='', add_special=True):
+        if add_special:
+            eoh = self.eoh
+        else:
+            eoh = ''
+        text = f' <|User|>:{user_text} \n{eoh} <|Bot|>:{bot_text}'
+        return text
+
+    def get_gen_args(self, **kwargs):
+        new_kargs = copy.deepcopy(self.gen_config)
+        new_kargs.update(kwargs)
+        return new_kargs
+
+    def generate(self, text, image=None, **kwargs):
+        text_embeds = self.encode_text(text)
+        img_embeds = self.encode_img(image)
+        prompt_embeds = self.wrap_prompt(text_embeds, img_embeds)
+        out_embeds = self.internlm_model.generate(inputs_embeds=prompt_embeds,
+                                                **self.get_gen_args(**kwargs))
+        out_text = self.decode_text(out_embeds)
+        return out_text
+
+    def chat(self, text, image=None, history=None, **kwargs):
+        text_embeds = self.encode_text(text)
+        img_embeds = self.encode_img(image)
+        prompt_embeds = self.wrap_prompt(text_embeds,
+                                         img_embeds,
+                                         history=history)
+        out_embeds = self.internlm_model.generate(inputs_embeds=prompt_embeds,
+                                                **self.get_gen_args(**kwargs))
+        out_text = self.decode_text(out_embeds)
+
+        # trunc at eoh and eoa
+        clean_out_text_token_ids = self.tokenizer(
+            out_text, return_tensors='pt').input_ids.to(self.device)
+        clean_out_text_embeds = self.internlm_model.model.embed_tokens(
+            clean_out_text_token_ids)
+        clean_prompt_embeds = self.wrap_prompt(text_embeds,
+                                               img_embeds,
+                                               add_special=False)
+        cur_history = torch.cat([clean_prompt_embeds, clean_out_text_embeds],
+                                dim=1)
+        if history is None:
+            history = []
+        history.append(cur_history)
+        return out_text, history
+
+    def wrap_prompt(self,
+                    text_embeds,
+                    img_embeds=None,
+                    history=None,
+                    add_special=True):
+        if add_special:
+            prompt_segs = [' <|User|>:', f'\n{self.eoh} <|Bot|>:']
+        else:
+            prompt_segs = [' <|User|>:', ' <|Bot|>:']  # used in wrap history
+        prompt_seg_embeds = []
+        for i, seg in enumerate(prompt_segs):
+            if history is not None:
+                add_special_tokens = False
+            else:
+                add_special_tokens = i == 0
+            seg_embeds = self.encode_text(
+                seg, add_special_tokens=add_special_tokens)
+            prompt_seg_embeds.append(seg_embeds)
+        if img_embeds is None:
+            img_embeds = text_embeds.new_empty(text_embeds.size(0), 0,
+                                               text_embeds.size(-1))
+        prompt_seg_embeds = [
+            prompt_seg_embeds[0], img_embeds, text_embeds, prompt_seg_embeds[1]
+        ]
+        prompt_embeds = torch.cat(prompt_seg_embeds, dim=1)
+        if history is not None:
+            prompt_embeds = torch.cat([*history, prompt_embeds], dim=1)
+        return prompt_embeds

modeling_perceive_sampler.py

@@ -0,0 +1,1193 @@
+"""
+ * Copyright (c) 2023, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on huggingface code base
+ * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert
+"""
+
+import math
+from typing import Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import Tensor, device
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.models.bert.configuration_bert import BertConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size,
+                                            config.hidden_size,
+                                            padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings,
+                                                config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size,
+                                      eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids",
+            torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.position_embedding_type = getattr(config,
+                                               "position_embedding_type",
+                                               "absolute")
+
+        self.config = config
+
+    def forward(
+        self,
+        input_ids=None,
+        position_ids=None,
+        query_embeds=None,
+        past_key_values_length=0,
+    ):
+        if input_ids is not None:
+            seq_length = input_ids.size()[1]
+        else:
+            seq_length = 0
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length:
+                                             seq_length +
+                                             past_key_values_length].clone()
+
+        if input_ids is not None:
+            embeddings = self.word_embeddings(input_ids)
+            if self.position_embedding_type == "absolute":
+                position_embeddings = self.position_embeddings(position_ids)
+                embeddings = embeddings + position_embeddings
+
+            if query_embeds is not None:
+                embeddings = torch.cat((query_embeds, embeddings), dim=1)
+        else:
+            embeddings = query_embeds
+
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+                config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" %
+                (config.hidden_size, config.num_attention_heads))
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size /
+                                       config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_width, self.all_head_size)
+            self.value = nn.Linear(config.encoder_width, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config,
+                                               "position_embedding_type",
+                                               "absolute")
+        if (self.position_embedding_type == "relative_key"
+                or self.position_embedding_type == "relative_key_query"):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1,
+                self.attention_head_size)
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(
+                self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(
+                self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer,
+                                        key_layer.transpose(-1, -2))
+
+        if (self.position_embedding_type == "relative_key"
+                or self.position_embedding_type == "relative_key_query"):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length,
+                                          dtype=torch.long,
+                                          device=hidden_states.device).view(
+                                              -1, 1)
+            position_ids_r = torch.arange(seq_length,
+                                          dtype=torch.long,
+                                          device=hidden_states.device).view(
+                                              1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = (attention_scores +
+                                    relative_position_scores_query +
+                                    relative_position_scores_key)
+
+        attention_scores = attention_scores / math.sqrt(
+            self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (
+            self.all_head_size, )
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = ((context_layer, attention_probs) if output_attentions else
+                   (context_layer, ))
+
+        outputs = outputs + (past_key_value, )
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size,
+                                      eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(
+            heads)
+        self.self.all_head_size = (self.self.attention_head_size *
+                                   self.self.num_attention_heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,
+                   ) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size,
+                                      eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BertAttention(config)
+        self.layer_num = layer_num
+        if (self.config.add_cross_attention
+                and layer_num % self.config.cross_attention_freq == 0):
+            self.crossattention = BertAttention(
+                config, is_cross_attention=self.config.add_cross_attention)
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+        self.intermediate_query = BertIntermediate(config)
+        self.output_query = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (past_key_value[:2]
+                                    if past_key_value is not None else None)
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                assert (
+                    encoder_hidden_states is not None
+                ), "encoder_hidden_states must be given for cross-attention layers"
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                outputs = (
+                    outputs + cross_attention_outputs[1:-1]
+                )  # add cross attentions if we output attention weights
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text],
+                                         dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output, ) + outputs
+
+        outputs = outputs + (present_key_value, )
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [BertLayer(config, i) for i in range(config.num_hidden_layers)])
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (() if output_attentions
+                                and self.config.add_cross_attention else None)
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states, )
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[
+                i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing",
+                       False) and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value,
+                                      output_attentions, query_length)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1], )
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (
+                    layer_outputs[1], )
+                all_cross_attentions = all_cross_attentions + (
+                    layer_outputs[2], )
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states, )
+
+        if not return_dict:
+            return tuple(v for v in [
+                hidden_states,
+                next_decoder_cache,
+                all_hidden_states,
+                all_self_attentions,
+                all_cross_attentions,
+            ] if v is not None)
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size,
+                                      eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size,
+                                 config.vocab_size,
+                                 bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0,
+                                       std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+class BertModel(BertPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+    def __init__(self, config, add_pooling_layer=False):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: Tensor,
+        input_shape: Tuple[int],
+        device: device,
+        is_decoder: bool,
+        has_query: bool = False,
+    ) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = (seq_ids[None, None, :].repeat(
+                    batch_size, seq_length, 1) <= seq_ids[None, :, None])
+
+                # add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[
+                        1] - causal_mask.shape[1]
+                    if has_query:  # UniLM style attention mask
+                        causal_mask = torch.cat(
+                            [
+                                torch.zeros(
+                                    (batch_size, prefix_seq_len, seq_length),
+                                    device=device,
+                                    dtype=causal_mask.dtype,
+                                ),
+                                causal_mask,
+                            ],
+                            axis=1,
+                        )
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, causal_mask.shape[1],
+                                 prefix_seq_len),
+                                device=device,
+                                dtype=causal_mask.dtype,
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+                extended_attention_mask = (causal_mask[:, None, :, :] *
+                                           attention_mask[:, None, None, :])
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})"
+                .format(input_shape, attention_mask.shape))
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(
+            dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        is_decoder=False,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`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)
+        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:
+            assert (
+                query_embeds is not None
+            ), "You have to specify query_embeds when input_ids is None"
+
+        # past_key_values_length
+        past_key_values_length = (past_key_values[0][0].shape[2] -
+                                  self.config.query_length
+                                  if past_key_values is not None else 0)
+
+        query_length = query_embeds.shape[1] if query_embeds is not None else 0
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            query_embeds=query_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size, seq_length = input_shape
+        device = embedding_output.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)),
+                device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if is_decoder:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask,
+                input_ids.shape,
+                device,
+                is_decoder,
+                has_query=(query_embeds is not None),
+            )
+        else:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask, input_shape, device, is_decoder)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[
+                    0].size()
+            else:
+                (
+                    encoder_batch_size,
+                    encoder_sequence_length,
+                    _,
+                ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size,
+                                    encoder_sequence_length)
+
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [
+                    self.invert_attention_mask(mask)
+                    for mask in encoder_attention_mask
+                ]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape,
+                                                    device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask)
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # 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]
+        head_mask = self.get_head_mask(head_mask,
+                                       self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (self.pooler(sequence_output)
+                         if self.pooler is not None else None)
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids", r"predictions.decoder.bias"
+    ]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=True,
+        reduction="mean",
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        Returns:
+        Example::
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', 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)
+        if labels is not None:
+            use_cache = False
+        if past_key_values is not None:
+            query_embeds = None
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        sequence_output = outputs[0]
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1]:, :]
+
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :
+                                                          -1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(reduction=reduction,
+                                        label_smoothing=0.1)
+            lm_loss = loss_fct(
+                shifted_prediction_scores.view(-1, self.config.vocab_size),
+                labels.view(-1),
+            )
+            if reduction == "none":
+                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)
+
+        if not return_dict:
+            output = (prediction_scores, ) + outputs[2:]
+            return ((lm_loss, ) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(self,
+                                      input_ids,
+                                      query_embeds,
+                                      past=None,
+                                      attention_mask=None,
+                                      **model_kwargs):
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_ids.shape)
+        query_mask = input_ids.new_ones(query_embeds.shape[:-1])
+        attention_mask = torch.cat([query_mask, attention_mask], dim=-1)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids":
+            input_ids,
+            "query_embeds":
+            query_embeds,
+            "attention_mask":
+            attention_mask,
+            "past_key_values":
+            past,
+            "encoder_hidden_states":
+            model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask":
+            model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder":
+            True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(
+                past_state.index_select(0, beam_idx)
+                for past_state in layer_past), )
+        return reordered_past
+
+
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids", r"predictions.decoder.bias"
+    ]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss. 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 in ``[0, ..., config.vocab_size]``
+        """
+
+        return_dict = (return_dict if return_dict is not None else
+                       self.config.use_return_dict)
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1]:, :]
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size),
+                labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores, ) + outputs[2:]
+            return (((masked_lm_loss, ) +
+                     output) if masked_lm_loss is not None else output)
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

modeling_utils.py

@@ -0,0 +1,122 @@
+import logging
+import math
+import os
+from contextlib import contextmanager
+
+import timm.models.hub as timm_hub
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def is_main_process():
+    return get_rank() == 0
+
+
+def download_cached_file(url, check_hash=True, progress=False):
+    """
+    Download a file from a URL and cache it locally. If the file already exists, it is not downloaded again.
+    If distributed, only the main process downloads the file, and the other processes wait for the file to be downloaded.
+    """
+    def get_cached_file_path():
+        # a hack to sync the file path across processes
+        parts = torch.hub.urlparse(url)
+        filename = os.path.basename(parts.path)
+        cached_file = os.path.join(timm_hub.get_cache_dir(), filename)
+
+        return cached_file
+
+    if is_main_process():
+        timm_hub.download_cached_file(url, check_hash, progress)
+
+    if is_dist_avail_and_initialized():
+        dist.barrier()
+
+    return get_cached_file_path()
+
+
+@contextmanager
+def all_logging_disabled(highest_level=logging.CRITICAL):
+    """
+    A context manager that will prevent any logging messages
+    triggered during the body from being processed.
+    :param highest_level: the maximum logging level in use.
+      This would only need to be changed if a custom level greater than CRITICAL
+      is defined.
+    """
+    # two kind-of hacks here:
+    #    * can't get the highest logging level in effect => delegate to the user
+    #    * can't get the current module-level override => use an undocumented
+    #       (but non-private!) interface
+
+    previous_level = logging.root.manager.disable
+
+    logging.disable(highest_level)
+
+    try:
+        yield
+    finally:
+        logging.disable(previous_level)
+
+
+class LoRALinear(nn.Linear):
+    def __init__(self,
+                 in_features: int,
+                 out_features: int,
+                 bias: bool = True,
+                 device=None,
+                 dtype=None,
+                 lora_r=8,
+                 lora_alpha=16,
+                 lora_dropout=0.05,
+                 **kwargs) -> None:
+        super().__init__(in_features, out_features, bias, device, dtype)
+        self.lora_r = lora_r
+        self.lora_alpha = lora_alpha
+        if lora_dropout > 0.:
+            self.lora_dropout = nn.Dropout(p=lora_dropout)
+        else:
+            self.lora_dropout = lambda x: x
+        self.lora_scaling = self.lora_alpha / self.lora_r
+
+        self.lora_A = nn.Linear(in_features,
+                                self.lora_r,
+                                bias=False,
+                                device=device,
+                                dtype=dtype)
+        self.lora_B = nn.Linear(self.lora_r,
+                                out_features,
+                                bias=False,
+                                device=device,
+                                dtype=dtype)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if hasattr(self, 'lora_A'):
+            # initialize A the same way as the default for nn.Linear and B to zero
+            nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
+            nn.init.zeros_(self.lora_B.weight)
+            #print ("lora weight init {} {}".format(torch.mean(self.lora_A.weight), torch.mean(self.lora_B.weight)))
+
+    def forward(self, x):
+        orig_type = x.dtype
+        res = super().forward(x)
+        x = x.float()
+        res += self.lora_B(self.lora_A(
+            self.lora_dropout(x))) * self.lora_scaling
+        return res.to(orig_type)

modeling_vit.py

@@ -0,0 +1,535 @@
+import math
+from functools import partial
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+
+from .modeling_utils import download_cached_file
+
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000,
+        'input_size': (3, 224, 224),
+        'pool_size': None,
+        'crop_pct': .9,
+        'interpolation': 'bicubic',
+        'mean': (0.5, 0.5, 0.5),
+        'std': (0.5, 0.5, 0.5),
+        **kwargs
+    }
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(self,
+                 in_features,
+                 hidden_features=None,
+                 out_features=None,
+                 act_layer=nn.GELU,
+                 drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self,
+                 dim,
+                 num_heads=8,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 attn_drop=0.,
+                 proj_drop=0.,
+                 window_size=None,
+                 attn_head_dim=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] -
+                                          1) * (2 * window_size[1] - 1) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance,
+                            num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h,
+                                                 coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = coords_flatten[:, :,
+                                             None] - coords_flatten[:,
+                                                                    None, :]  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(
+                1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :,
+                            0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = \
+                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
+            relative_position_index[1:, 1:] = relative_coords.sum(
+                -1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer("relative_position_index",
+                                 relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, rel_pos_bias=None):
+        B, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat(
+                (self.q_bias, torch.zeros_like(self.v_bias,
+                                               requires_grad=False),
+                 self.v_bias))
+        # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[
+            2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        if self.relative_position_bias_table is not None:
+            relative_position_bias = \
+                self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                    self.window_size[0] * self.window_size[1] + 1,
+                    self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.permute(
+                2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+        if rel_pos_bias is not None:
+            attn = attn + rel_pos_bias
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 mlp_ratio=4.,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 init_values=None,
+                 act_layer=nn.GELU,
+                 norm_layer=nn.LayerNorm,
+                 window_size=None,
+                 attn_head_dim=None):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(dim,
+                              num_heads=num_heads,
+                              qkv_bias=qkv_bias,
+                              qk_scale=qk_scale,
+                              attn_drop=attn_drop,
+                              proj_drop=drop,
+                              window_size=window_size,
+                              attn_head_dim=attn_head_dim)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(
+            drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim,
+                       hidden_features=mlp_hidden_dim,
+                       act_layer=act_layer,
+                       drop=drop)
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),
+                                        requires_grad=True)
+            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),
+                                        requires_grad=True)
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+    def forward(self, x, rel_pos_bias=None):
+        if self.gamma_1 is None:
+            x = x + self.drop_path(
+                self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            x = x + self.drop_path(self.gamma_1 * self.attn(
+                self.norm1(x), rel_pos_bias=rel_pos_bias))
+            x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] //
+                                                        patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0],
+                            img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans,
+                              embed_dim,
+                              kernel_size=patch_size,
+                              stride=patch_size)
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] -
+                                      1) * (2 * window_size[1] - 1) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance,
+                        num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :,
+                                         None] - coords_flatten[:,
+                                                                None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = \
+            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
+        relative_position_index[1:,
+                                1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer("relative_position_index",
+                             relative_position_index)
+
+        # trunc_normal_(self.relative_position_bias_table, std=.02)
+
+    def forward(self):
+        relative_position_bias = \
+            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1] + 1,
+                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(
+            2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+    def __init__(self,
+                 img_size=224,
+                 patch_size=16,
+                 in_chans=3,
+                 num_classes=1000,
+                 embed_dim=768,
+                 depth=12,
+                 num_heads=12,
+                 mlp_ratio=4.,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.,
+                 norm_layer=nn.LayerNorm,
+                 init_values=None,
+                 use_abs_pos_emb=True,
+                 use_rel_pos_bias=False,
+                 use_shared_rel_pos_bias=False,
+                 use_mean_pooling=True,
+                 init_scale=0.001,
+                 use_checkpoint=False):
+        super().__init__()
+        self.image_size = img_size
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+
+        self.patch_embed = PatchEmbed(img_size=img_size,
+                                      patch_size=patch_size,
+                                      in_chans=in_chans,
+                                      embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(
+                torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(
+                window_size=self.patch_embed.patch_shape, num_heads=num_heads)
+        else:
+            self.rel_pos_bias = None
+        self.use_checkpoint = use_checkpoint
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)
+               ]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList([
+            Block(dim=embed_dim,
+                  num_heads=num_heads,
+                  mlp_ratio=mlp_ratio,
+                  qkv_bias=qkv_bias,
+                  qk_scale=qk_scale,
+                  drop=drop_rate,
+                  attn_drop=attn_drop_rate,
+                  drop_path=dpr[i],
+                  norm_layer=norm_layer,
+                  init_values=init_values,
+                  window_size=self.patch_embed.patch_shape
+                  if use_rel_pos_bias else None) for i in range(depth)
+        ])
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        self.apply(self._init_weights)
+        self.fix_init_weight()
+
+    def fix_init_weight(self):
+        def rescale(param, layer_id):
+            param.div_(math.sqrt(2.0 * layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(
+            self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(
+            batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        rel_pos_bias = self.rel_pos_bias(
+        ) if self.rel_pos_bias is not None else None
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, rel_pos_bias)
+            else:
+                x = blk(x, rel_pos_bias)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        #         x = self.head(x)
+        return x
+
+    def get_intermediate_layers(self, x):
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(
+            batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        features = []
+        rel_pos_bias = self.rel_pos_bias(
+        ) if self.rel_pos_bias is not None else None
+        for blk in self.blocks:
+            x = blk(x, rel_pos_bias)
+            features.append(x)
+
+        return features
+
+
+def interpolate_pos_embed(model, checkpoint_model):
+    if 'pos_embed' in checkpoint_model:
+        pos_embed_checkpoint = checkpoint_model['pos_embed'].float()
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.patch_embed.num_patches
+        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int(
+            (pos_embed_checkpoint.shape[-2] - num_extra_tokens)**0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches**0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" %
+                  (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size,
+                                            embedding_size).permute(
+                                                0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(pos_tokens,
+                                                         size=(new_size,
+                                                               new_size),
+                                                         mode='bicubic',
+                                                         align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            checkpoint_model['pos_embed'] = new_pos_embed
+
+
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def convert_weights_to_fp32(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+    def _convert_weights_to_fp32(l):
+        if hasattr(l, 'weight') and l.weight is not None:
+            if l.weight.dtype == torch.float16:
+                l.weight = l.weight.to(torch.float32)
+        if hasattr(l, 'bias') and l.bias is not None:
+            if l.bias.dtype == torch.float16:
+                l.bias = l.bias.to(torch.float32)
+
+    model.apply(_convert_weights_to_fp32)
+
+
+def create_eva_vit_g(img_size=224,
+                     drop_path_rate=0.4,
+                     use_checkpoint=False,
+                     precision="fp16"):
+    model = VisionTransformer(
+        img_size=img_size,
+        patch_size=14,
+        use_mean_pooling=False,
+        embed_dim=1408,
+        depth=39,
+        num_heads=1408 // 88,
+        mlp_ratio=4.3637,
+        qkv_bias=True,
+        drop_path_rate=drop_path_rate,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        use_checkpoint=use_checkpoint,
+    )
+    # url = "https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth"
+    # cached_file = download_cached_file(url, check_hash=False, progress=True)
+    # state_dict = torch.load(cached_file, map_location="cpu")
+    # interpolate_pos_embed(model, state_dict)
+    #
+    # incompatible_keys = model.load_state_dict(state_dict, strict=False)
+
+    if precision == "fp16":
+        convert_weights_to_fp16(model)
+
+    if precision == "fp32":
+        print('convert ViT weights to fp32')
+        convert_weights_to_fp32(model)
+
+    return model

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<s>",
+  "eos_token": "</s>",
+  "pad_token": "</s>",
+  "unk_token": "<unk>"
+}

tokenization_InternLM_XComposer.py

@@ -0,0 +1,240 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+
+"""Tokenization classes for IntermLM."""
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "./tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {}
+
+
+class InternLMXComposerTokenizer(PreTrainedTokenizer):
+    """
+    Construct a InternLM tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    model_input_names = ["input_ids", "attention_mask"]
+    _auto_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token="</s>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        decode_with_prefix_space=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.decode_with_prefix_space = decode_with_prefix_space
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        self._no_prefix_space_tokens = None
+
+        """ Initialisation"""
+
+    @property
+    def no_prefix_space_tokens(self):
+        if self._no_prefix_space_tokens is None:
+            vocab = self.convert_ids_to_tokens(list(range(self.vocab_size)))
+            self._no_prefix_space_tokens = {i for i, tok in enumerate(vocab) if not tok.startswith("▁")}
+        return self._no_prefix_space_tokens
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    @property
+    def bos_token_id(self) -> Optional[int]:
+        return self.sp_model.bos_id()
+
+    @property
+    def eos_token_id(self) -> Optional[int]:
+        return self.sp_model.eos_id()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def _maybe_add_prefix_space(self, tokens, decoded):
+        if tokens and tokens[0] not in self.no_prefix_space_tokens:
+            return " " + decoded
+        else:
+            return decoded
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        out_string = self.clean_up_tokenization(out_string)
+        out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
+        return out_string[1:]
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if self.add_bos_token:
+            bos_token_ids = [self.bos_token_id]
+        else:
+            bos_token_ids = []
+
+        output = bos_token_ids + token_ids_0
+
+        if token_ids_1 is not None:
+            output = output + token_ids_1
+
+        if self.add_eos_token:
+            output = output + [self.eos_token_id]
+
+        return output
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
+        use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+        """
+        eos = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + eos) * [0]
+        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_InternLM_XComposer.InternLMXComposerTokenizer",
+      null
+    ]
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "tokenizer_class": "InternLMXComposerTokenizer",
+  "unk_token": "<unk>"
+}