upload 8 files

config.json

@@ -0,0 +1,46 @@
+{
+  "_name_or_path": "vit-phi-1.5",
+  "activation_function": "gelu_new",
+  "architecture": {
+    "block_cls": "parallel",
+    "mixer": {},
+    "mlp": {
+      "mlp_cls": "mlp"
+    }
+  },
+  "architectures": [
+    "MixFormerVLSequentialForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_mixformer_sequential.MixFormerVLSequentialConfig",
+    "AutoModelForCausalLM": "modeling_mixformer_sequential.MixFormerVLSequentialForCausalLM"
+  },
+  "embd_layer": "default",
+  "embd_pdrop": 0.0,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "mixformer-sequential",
+  "n_embd": 2048,
+  "n_head": 32,
+  "n_inner": null,
+  "n_layer": 24,
+  "n_positions": 2048,
+  "phyagi_version": "0.0.4.dev",
+  "resid_pdrop": 0.0,
+  "rotary_dim": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.32.1",
+  "tokenizer_type": "VitPhiTokenizer",
+  "visual": {
+    "heads": 16,
+    "image_size": 448,
+    "image_start_id": 50470,
+    "layers": 48,
+    "mlp_ratio": 4.9231,
+    "output_dim": 4096,
+    "patch_size": 14,
+    "width": 1664
+  },
+  "vocab_size": 51200
+}

configuration_vitphi.py

@@ -0,0 +1,63 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import math
+from typing import Any, Dict, List, Optional, Union
+
+from transformers import PretrainedConfig
+
+
+class MixFormerVLSequentialConfig(PretrainedConfig):
+    """MixFormer (sequential for DeepSpeed) configuration."""
+
+    model_type = "mixformer-sequential"
+
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+        "input_emb_layer": "embd_layer",  # `input_emb_layer` key is for backward compatibility
+        "blocks": "architecture",  # `blocks` key is for backward compatibility
+    }
+
+    def __init__(
+        self,
+        vocab_size: Optional[int] = 50304,
+        n_positions: Optional[int] = 2048,
+        n_embd: Optional[int] = 1024,
+        n_layer: Optional[int] = 20,
+        n_inner: Optional[int] = None,
+        n_head: Optional[int] = 16,
+        rotary_dim: Optional[int] = 32,
+        activation_function: Optional[str] = "gelu_new",
+        embd_layer: Optional[str] = "default",
+        architecture: Union[Dict[str, Any], List[Dict[str, Any]]] = None,
+        embd_pdrop: Optional[float] = 0.0,
+        resid_pdrop: Optional[float] = 0.0,
+        layer_norm_epsilon: Optional[float] = 1e-5,
+        initializer_range: Optional[float] = 0.02,
+        tie_word_embeddings: Optional[bool] = False,
+        pad_vocab_size_multiple: Optional[int] = 64,
+
+        # vit_hidden_size: Optional[int] = 4096,
+
+        **kwargs
+    ) -> None:
+        self.vocab_size = int(math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple)
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_inner = n_inner
+        self.n_head = n_head
+        self.rotary_dim = min(rotary_dim, n_embd // n_head)
+        self.activation_function = activation_function
+        self.embd_layer = embd_layer
+        self.architecture = architecture
+        self.embd_pdrop = embd_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        # self.vit_hidden_size = vit_hidden_size
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 50256,
+  "transformers_version": "4.32.1"
+
+}

modeling_vitphi.py

@@ -0,0 +1,796 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+# BSD 3-Clause License
+#
+# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of the copyright holder nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from __future__ import annotations
+
+import math
+import copy
+from typing import Any, Dict, Optional, Tuple
+from dataclasses import dataclass, field
+
+import torch
+import torch.nn as nn
+
+from einops import rearrange
+from transformers.activations import ACT2FN
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from .configuration_vitphi import MixFormerVLSequentialConfig
+from .visual import VisionTransformer
+
+
+
+@dataclass
+class InferenceParams:
+    """Inference parameters that are passed to the main model in order
+    to efficienly calculate and store the context during inference.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+    max_sequence_len: int
+    max_batch_size: int
+    sequence_len_offset: int = 0
+    batch_size_offset: int = 0
+    key_value_memory_dict: dict = field(default_factory=dict)
+    fused_ft_kernel: bool = False
+    lengths_per_sample: Optional[torch.Tensor] = None
+
+
+class Embedding(nn.Module):
+    """Token embedding with dropout."""
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+
+    def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.wte(input_ids)
+        hidden_states = self.drop(hidden_states)
+
+        return hidden_states
+
+
+class RotaryEmbedding(nn.Module):
+    """PyTorch implementation of `flash-attn` RotaryEmbedding layer.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+    def __init__(
+            self,
+            dim: int,
+            base: Optional[int] = 10000,
+            scale_base: Optional[float] = None,
+            device: Optional[str] = None,
+            **kwargs,
+    ) -> None:
+        super().__init__()
+
+        if scale_base is not None:
+            raise NotImplementedError
+
+        # Generate and save the inverse frequency buffer (non-trainable)
+        self.dim = dim
+        self.base = base
+        self.scale_base = scale_base
+        self.device = device
+
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            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: torch.FloatTensor, seqlen_offset: Optional[int] = 0) -> None:
+        # Reset the tables if the sequence length has changed,
+        # or if we're on a new device (possibly due to tracing for instance)
+        seqlen = x.shape[1] + seqlen_offset
+
+        # Re-generate the inverse frequency buffer if it's not fp32
+        # (for instance if model.half() was called)
+        if self.inv_freq.dtype != "torch.float32":
+            self.inv_freq = 1.0 / (
+                    self.base ** (torch.arange(0, self.dim, 2, device=self.device, dtype=torch.float32) / self.dim)
+            )
+
+        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=torch.float32)
+
+            # 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, dtype=torch.float32))
+            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 apply_rotary_emb_qkv(
+            self,
+            qkv: torch.FloatTensor,
+            sin: torch.FloatTensor,
+            cos: torch.FloatTensor,
+            sin_k: Optional[torch.FloatTensor] = None,
+            cos_k: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        _, seqlen, three, _, 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_rot = qkv[:, :, 0, :, :rotary_dim]
+        q_pass = qkv[:, :, 0, :, rotary_dim:]
+
+        k_rot = qkv[:, :, 1, :, :rotary_dim]
+        k_pass = qkv[:, :, 1, :, rotary_dim:]
+
+        # Splits the queries and keys in half
+        q1, q2 = q_rot.chunk(2, dim=-1)
+        k1, k2 = k_rot.chunk(2, dim=-1)
+        c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+
+        # Casts to fp32 are necessary to prevent fp16 overflow issues
+        q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
+
+        # Computes the new keys and queries, recasting to original dtype
+        q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
+
+        k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
+
+        return torch.cat(
+            [
+                torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
+                torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+                qkv[:, :, 2:3, :, :],
+            ],
+            axis=2,
+        )
+
+    def forward(self, qkv: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Perform the forward pass.
+        Args:
+            qkv: Query, key and value tensors of shape (batch, seqlen, nheads, headdim) or (batch, seqlen, 3, nheads, headdim).
+            seqlen_offset: Used in generation where the passed `qkv` is only the last token in the batch.
+        Returns:
+            New `qkv` and the cached sinusoids.
+        """
+
+        self._update_cos_sin_cache(qkv, seqlen_offset)
+
+        return self.apply_rotary_emb_qkv(qkv, self._sin_cached[seqlen_offset:], self._cos_cached[seqlen_offset:])
+
+
+def _update_kv_cache(kv, inference_params, layer_idx):
+    """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+    # Pre-allocate memory for key-values for inference.
+    num_heads, head_dim = kv.shape[-2:]
+    if layer_idx not in inference_params.key_value_memory_dict:
+        kv_cache = torch.empty(
+            inference_params.max_batch_size, inference_params.max_sequence_len, 2,
+            num_heads, head_dim, dtype=kv.dtype, device=kv.device
+        )
+        inference_params.key_value_memory_dict[layer_idx] = kv_cache
+    else:
+        kv_cache = inference_params.key_value_memory_dict[layer_idx]
+
+    # Adjust key and value for inference
+    batch_start = inference_params.batch_size_offset
+    batch_end = batch_start + kv.shape[0]
+    sequence_start = inference_params.sequence_len_offset
+    sequence_end = sequence_start + kv.shape[1]
+    assert batch_end <= (kv_cache.shape[0] if kv_cache is not None else v_cache.shape[0])
+    assert sequence_end <= (kv_cache.shape[1] if kv_cache is not None else v_cache.shape[2])
+
+    assert kv_cache is not None
+    kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+    kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
+    return kv
+
+
+class MLP(nn.Module):
+    """Multi-Layer Perceptron.
+    Reference:
+        Attention Is All You Need.
+        https://arxiv.org/pdf/1706.03762.pdf.
+    """
+
+    def __init__(self, config: PretrainedConfig, n_inner: Optional[int] = None, act_fn: Optional[str] = None) -> None:
+        super().__init__()
+
+        act_fn = config.activation_function if act_fn is None else act_fn
+        assert act_fn in ACT2FN.keys(), f"`act_fn` must be one of: {ACT2FN.keys()}."
+
+        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
+        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
+
+        self.fc1 = nn.Linear(config.n_embd, n_inner)
+        self.fc2 = nn.Linear(n_inner, config.n_embd)
+        self.act = ACT2FN[act_fn]
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
+                              error_msgs):
+        old_keys = [prefix + "fc_in.weight", prefix + "fc_out.weight", prefix + "fc_in.bias", prefix + "fc_out.bias"]
+        new_keys = [prefix + "fc1.weight", prefix + "fc2.weight", prefix + "fc1.bias", prefix + "fc2.bias"]
+
+        if all(k in state_dict for k in old_keys) and not all(k in state_dict for k in new_keys):
+            # Older version of `MLP` saved with different key names.
+            for old_key, new_key in zip(old_keys, new_keys):
+                state_dict[new_key] = state_dict.pop(old_key)
+
+        return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
+                                             error_msgs)
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class FusedMLP(nn.Module):
+    """Fused Multi-Layer Perceptron from `flash-attn`.
+    Reference:
+        https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/ops/fused_dense.py.
+    """
+
+    def __init__(self, config: PretrainedConfig, n_inner: Optional[int] = None, act_fn: Optional[str] = None,
+                 raise_on_missing: bool = False) -> None:
+        super().__init__()
+
+        act_fn = config.activation_function if act_fn is None else act_fn
+        assert act_fn in ACT2FN.keys(), f"`act_fn` must be one of: {ACT2FN.keys()}."
+
+        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
+        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
+
+        gelu_activations = ["gelu_new", "gelu_fast", "gelu_approx"]
+        activation = "gelu_approx" if act_fn in gelu_activations else "relu"
+
+        self.mlp = MLP(config, n_inner=n_inner, act_fn=act_fn)
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        return self.mlp(hidden_states)
+
+
+class SelfAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Adapted from https://github.com/Dao-AILab/flash-attention.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, qkv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            qkv: The tensor containing the query, key, and value. (B, S, 3, H, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, S)
+        """
+        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
+        causal = self.causal if causal is None else causal
+        q, k, v = qkv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype,
+                                      device=scores.device)
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, 'b s -> b 1 1 s')
+        if causal:
+            # "triu_tril_cuda_template" not implemented for 'BFloat16'
+            # So we have to construct the mask in float
+            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum('bhts,bshd->bthd', attention_drop, v)
+        return output
+
+
+class CrossAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Adapted from https://github.com/Dao-AILab/flash-attention.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, q, kv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q: The tensor containing the query. (B, Sq, H, D)
+            kv: The tensor containing the key and value. (B, Sk, 2, H, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, Sk)
+        """
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        causal = self.causal if causal is None else causal
+        seqlen_k = kv.shape[1]
+        assert kv.shape[0] == batch_size and kv.shape[3] == q.shape[2] and kv.shape[4] == q.shape[3]
+        k, v = kv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full((batch_size, seqlen_k), -10000.0, dtype=scores.dtype,
+                                      device=scores.device)
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, 'b s -> b 1 1 s')
+        if causal:
+            # "triu_tril_cuda_template" not implemented for 'BFloat16'
+            # So we have to construct the mask in float
+            causal_mask = torch.triu(torch.full((seqlen_q, seqlen_k), -10000.0,
+                                                device=scores.device), 1)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum('bhts,bshd->bthd', attention_drop, v)
+        return output
+
+
+def find_mha_dims(
+        config: PretrainedConfig, n_head: Optional[int] = None, head_dim: Optional[int] = None
+) -> Tuple[int, int]:
+    """Validate and return the number of heads and head dimension for multi-head attention.
+    Args:
+        config: Model configuration.
+        n_head: Number of heads.
+        head_dim: Head dimension.
+    Returns:
+        Number of heads and head dimension.
+    """
+
+    assert all(
+        hasattr(config, attr) for attr in ["n_embd", "n_head"]
+    ), "`config` must have `n_embd` and `n_head` attributes."
+
+    if head_dim is None:
+        assert (
+                config.n_embd % config.n_head == 0
+        ), f"Hidden size ({config.n_embd}) must be divisible by the number of heads ({config.n_head})."
+
+    if n_head is None and head_dim is None:
+        head_dim = config.n_embd // config.n_head
+        n_head = config.n_head
+    elif n_head is None or head_dim is None:
+        raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
+
+    return n_head, head_dim
+
+
+class MHA(nn.Module):
+    """Multi-head attention layer.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+    def __init__(
+            self,
+            config: PretrainedConfig,
+            rotary_dim: Optional[int] = None,
+            n_head: Optional[int] = None,
+            head_dim: Optional[int] = None,
+            bias: Optional[bool] = True,
+            dropout: Optional[float] = 0.0,
+            softmax_scale: Optional[float] = None,
+            causal: Optional[bool] = True,
+            layer_idx: Optional[int] = None,
+            rotary_emb_scale_base: Optional[float] = None,
+            return_residual: Optional[bool] = False,
+            checkpointing: Optional[bool] = False,
+            device: Optional[str] = None,
+            dtype: Optional[torch.dtype] = None,
+            fused_dense: Optional[bool] = True,
+            flash_attn: Optional[bool] = True,
+            cutlass_attn: Optional[bool] = False,
+            flash_rotary: Optional[bool] = True,
+            raise_on_missing: Optional[bool] = False
+    ) -> None:
+        super().__init__()
+
+        factory_kwargs = {"device": device, "dtype": dtype}
+        n_head, head_dim = find_mha_dims(config, n_head, head_dim)
+
+        self.hidden_size = config.n_embd
+        self.n_head = n_head
+        self.head_dim = head_dim
+        self.op_size = n_head * head_dim
+
+        self.causal = causal
+        self.layer_idx = layer_idx
+        self.rotary_emb_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
+        self.fused_dense = fused_dense
+        self.flash_attn = flash_attn
+        self.cutlass_attn = cutlass_attn
+        self.flash_rotary = flash_rotary
+        self.return_residual = return_residual
+        self.checkpointing = checkpointing
+
+        if self.rotary_emb_dim > 0:
+            rotary_kwargs = {"device": device}
+            if rotary_emb_scale_base is not None and rotary_emb_scale_base > 0.0:
+                rotary_kwargs["scale_base"] = rotary_emb_scale_base
+
+            self.rotary_emb = RotaryEmbedding(self.rotary_emb_dim, **rotary_kwargs)
+        else:
+            pass
+
+        self.Wqkv = nn.Linear(self.hidden_size, 3 * self.op_size, bias=bias, **factory_kwargs)
+        self.out_proj = nn.Linear(self.op_size, self.hidden_size, bias=bias, **factory_kwargs)
+
+        self.inner_attn = SelfAttention(causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout)
+        self.inner_cross_attn = CrossAttention(causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout)
+
+    def _update_kv_cache(self, kv: torch.FloatTensor, inference_params: InferenceParams) -> None:
+        """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)
+        Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+        assert self.layer_idx is not None, "Generation requires layer_idx in the constructor"
+
+        return _update_kv_cache(kv, inference_params, self.layer_idx)
+
+    def forward(
+            self,
+            x: torch.FloatTensor,
+            x_kv: Optional[torch.FloatTensor] = None,
+            key_padding_mask: Optional[torch.BoolTensor] = None,
+            cu_seqlens: Optional[torch.LongTensor] = None,
+            max_seqlen: Optional[int] = None,
+            mixer_subset: Optional[torch.LongTensor] = None,
+            past_cache: Optional[InferenceParams] = None,
+            **kwargs
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        """Perform the forward pass.
+        Args:
+            x: (batch, seqlen, hidden_dim) (where hidden_dim = num heads * head dim) if
+                cu_seqlens is None and max_seqlen is None, else (total, hidden_dim) where total
+                is the is the sum of the sequence lengths in the batch.
+            x_kv: (batch, seqlen, hidden_dim), only applicable for cross-attention. If None, use x.
+            key_padding_mask: boolean mask, True means to keep, False means to mask out.
+                (batch, seqlen). Only applicable when not using FlashAttention.
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into x. Only applicable when using
+                FlashAttention.
+            max_seqlen: int. Maximum sequence length in the batch.
+            mixer_subset: for cross-attention only. If not None, will take a subset of x
+                before applying the query projection. Useful for e.g., ViT where we only care
+                about the CLS token in the last layer.
+            past_cache: For generation only.
+        Returns:
+            (batch, seqlen, hidden_dim) if cu_seqlens is None and max_seqlen is None,
+                else (total, hidden_dim) where total is the is the sum of the sequence lengths
+                in the batch.
+        """
+
+        if cu_seqlens is not None:
+            assert max_seqlen is not None
+            assert key_padding_mask is None
+            assert self.flash_attn
+            assert self.rotary_emb_dim == 0
+
+        if key_padding_mask is not None:
+            assert cu_seqlens is None
+            assert max_seqlen is None
+            assert not self.flash_attn
+
+        if past_cache is not None:
+            assert key_padding_mask is None
+            assert cu_seqlens is None and max_seqlen is None
+
+        attn_kwargs = {"key_padding_mask": key_padding_mask}
+
+        assert x_kv is None and mixer_subset is None
+
+        qkv = self.Wqkv(x)
+        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+
+        if past_cache is None:
+            if self.rotary_emb_dim > 0:
+                qkv = self.rotary_emb(qkv)
+                context = self.inner_attn(qkv, **attn_kwargs)
+
+        else:
+            if self.rotary_emb_dim > 0:
+                qkv = self.rotary_emb(qkv, seqlen_offset=past_cache.sequence_len_offset)
+            q = qkv[:, :, 0]
+            kv = self._update_kv_cache(qkv[:, :, 1:], past_cache)
+            # If we're processing the prompt, causal=None (use self.causal).
+            # If we're decoding, then causal=False.
+            causal = None if past_cache.sequence_len_offset == 0 else False
+            context = self.inner_cross_attn(q, kv, causal=causal)
+
+        out = rearrange(context, "... h d -> ... (h d)")
+        out = self.out_proj(out)
+
+        return out if not self.return_residual else (out, x)
+
+
+class ParallelBlock(nn.Module):
+    """Parallel block.
+    This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
+    """
+
+    def __init__(
+            self,
+            config: PretrainedConfig,
+            mixer: Optional[Dict[str, Any]] = None,
+            mlp: Optional[Dict[str, Any]] = None,
+            block_idx: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.block_idx = block_idx
+
+        self.mixer = MHA(config=config, **mixer, layer_idx=block_idx)
+        mlp_cls = mlp.pop('mlp_cls')
+        if mlp_cls == 'fused_mlp':
+            self.mlp = FusedMLP(config=config, **mlp)
+        else:
+            self.mlp = MLP(config=config, **mlp)
+
+    def forward(self, hidden_states: torch.FloatTensor,
+                past_cache: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+        residual = hidden_states
+        hidden_states = self.ln(hidden_states)
+
+        attn_outputs = self.mixer(hidden_states, past_cache=past_cache)
+        if isinstance(attn_outputs, tuple):
+            attn_outputs = attn_outputs[0]
+
+        attn_outputs = self.resid_dropout(attn_outputs)
+        feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
+
+        hidden_states = attn_outputs + feed_forward_hidden_states + residual
+
+        return hidden_states
+
+
+class CausalLMHead(nn.Module):
+    """Causal Language Modeling head.
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+    """
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.linear = nn.Linear(config.n_embd, config.vocab_size)
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.ln(hidden_states)
+        logits = self.linear(hidden_states).to(torch.float32)
+
+        return logits
+
+
+class CausalLMLoss(nn.Module):
+    """Causal Language Modeling loss.
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+    """
+
+    def __init__(self, shift_labels: Optional[bool] = True) -> None:
+        super().__init__()
+
+        self.shift_labels = shift_labels
+        self.loss_fct = nn.CrossEntropyLoss()
+
+    def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
+        if self.shift_labels:
+            logits = logits[..., :-1, :].contiguous()
+            labels = labels[..., 1:].contiguous()
+
+        loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
+
+        return loss
+
+
+class MixFormerSequentialPreTrainedModel(PreTrainedModel):
+    """MixFormer (sequential for DeepSpeed) pre-trained model."""
+
+    config_class = MixFormerVLSequentialConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+
+    def __init__(self, *inputs, **kwargs) -> None:
+        super().__init__(*inputs, **kwargs)
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs) -> Dict[str, Any]:
+        if "use_cache" in kwargs and not kwargs["use_cache"]:
+            return {"input_ids": input_ids}
+
+        if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
+            past_key_values = InferenceParams(
+                max_batch_size=input_ids.shape[0],
+                max_sequence_len=self.config.n_positions,
+                sequence_len_offset=0,
+                batch_size_offset=0,
+                fused_ft_kernel=False,
+                key_value_memory_dict={},
+            )
+        else:
+            # assume past_key_values has cached all but last token in input_ids
+            past_key_values.sequence_len_offset = len(input_ids[0]) - 1
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+        return {"input_ids": input_ids, "past_key_values": past_key_values, **kwargs}
+
+
+class MixFormerVLSequentialForCausalLM(MixFormerSequentialPreTrainedModel):
+    """MixFormer (sequential for DeepSpeed) for Causal Language Modeling."""
+
+    _keys_to_ignore_on_load_missing = [""]
+    _keys_to_ignore_on_load_unexpected = [r"layers\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+    _no_split_modules = ["ParallelBlock"]
+
+    def __init__(self, config: MixFormerVLSequentialConfig) -> None:
+        super().__init__(config)
+
+        modules = [Embedding(config)]
+        block_config = config.architecture
+
+        if not isinstance(block_config, list):
+            block_config = [block_config for _ in range(config.n_layer)]
+
+        if config.n_layer != len(block_config):
+            config.n_layer = len(block_config)
+
+        for block_idx, block in enumerate(block_config):
+            # `block_cls` with `legacy` value is for backward compatibility
+            # `path` key is for backward compatibility
+            block = copy.deepcopy(block) or {"block_cls": "parallel"}
+            block_cls = block.pop("path", None) or block.pop("block_cls", None)
+
+            block["block_idx"] = block_idx
+            modules.append(ParallelBlock(config, **block))
+
+        modules.append(CausalLMHead(config))
+
+        self.layers = nn.Sequential(*modules)
+        self.loss = CausalLMLoss()
+        self.visual = VisionTransformer(**config.visual)
+        self.switcher = nn.Linear(config.visual.output_dim, config.n_embd, bias=False)
+
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.layers[0].wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
+        self.layers[0].wte = new_embeddings
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.layers[-1].linear
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.layers[-1].linear = new_embeddings
+
+    def forward(
+            self, input_ids: torch.LongTensor, labels: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[torch.FloatTensor] = None, **kwargs
+    ) -> CausalLMOutputWithPast:
+        if past_key_values is None and input_ids is not None \
+            and torch.any(input_ids == self.config.visual['image_start_id']):
+            bos_pos = torch.where(input_ids == self.config.visual['image_start_id'])
+            eos_pos = torch.where(input_ids == self.config.visual['image_start_id'] + 1)  # image_end_id = image_start_id + 1
+            assert (bos_pos[0] == eos_pos[0]).all() # 断言batch中的每个样本都有图片的起始和终止符
+            img_pos = torch.stack((bos_pos[0], bos_pos[1], eos_pos[1]), dim=1)
+            images = []
+            for i, a, b in img_pos:
+                image = input_ids[i][a+1: b-1].tolist()
+                image = image[ : image.index(self.config.visual['image_start_id'] + 2)]  # image_pad_id = image_start_id + 2
+                images.append(bytes(image).decode('utf-8'))
+
+            images = self.visual.encode(images)
+            assert images.shape[0] == len(images)
+        else:
+            images = None
+
+        hidden_states = self.layers[0](input_ids)
+        if images is not None:
+            for idx, (i, a, b) in enumerate(img_pos):
+                hidden_states[i][a + 1: b] = self.switcher(images[idx])
+        if not past_key_values:
+            for module in self.layers[1:-1]:
+                hidden_states = module(hidden_states)
+        else:
+            for module in self.layers[1:-1]:
+                hidden_states = module(hidden_states, past_cache=past_key_values)
+        lm_logits = self.layers[-1](hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss(lm_logits, labels)
+
+        return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)

tokenization_vitphi.py

@@ -0,0 +1,574 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Tokenization classes for QWen."""
+
+import base64
+import logging
+import os
+import requests
+import unicodedata
+from typing import Collection, Dict, List, Set, Tuple, Union, Any, Callable, Optional
+
+import tiktoken
+import numpy as np
+from PIL import Image
+from PIL import ImageFont
+from PIL import ImageDraw
+from transformers import PreTrainedTokenizer, AddedToken
+from transformers.utils import try_to_load_from_cache
+
+import matplotlib.colors as mcolors
+from matplotlib.font_manager import FontProperties
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.tiktoken", "ttf": "SimSun.ttf"}
+# FONT_PATH = try_to_load_from_cache("Qwen/Qwen-VL-Chat", "SimSun.ttf")
+FONT_PATH = None
+if FONT_PATH is None:
+    if not os.path.exists("SimSun.ttf"):
+        ttf = requests.get("https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/SimSun.ttf")
+        open("SimSun.ttf", "wb").write(ttf.content)
+    FONT_PATH = "SimSun.ttf"
+
+PAT_STR = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+ENDOFTEXT = "<|endoftext|>"
+# <|endoftext|> 50256
+IMSTART = "<|im_start|>"
+IMEND = "<|im_end|>"
+# as the default behavior is changed to allow special tokens in
+# regular texts, the surface forms of special tokens need to be
+# as different as possible to minimize the impact
+EXTRAS = tuple((f"<|extra_{i}|>" for i in range(205)))
+SPECIAL_TOKENS = (
+                     # ENDOFTEXT,
+                     IMSTART,
+                     IMEND,
+                 ) + EXTRAS
+IMG_TOKEN_SPAN = 256
+
+
+def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
+    with open(tiktoken_bpe_file, "rb") as f:
+        contents = f.read()
+    return {
+        base64.b64decode(token): int(rank)
+        for token, rank in (line.split() for line in contents.splitlines() if line)
+    }
+
+
+def _list_find(
+        input_list: List[Any],
+        candidates: Tuple[Any],
+        start: int = 0,
+):
+    for i in range(start, len(input_list)):
+        if input_list[i] in candidates:
+            return i
+    return -1
+
+
+def _replace_closed_tag(
+        input_tokens: List[Any],
+        start_tags: Union[Any, Tuple[Any]],
+        end_tags: Union[Any, Tuple[Any]],
+        inclusive_replace_func: Callable,
+        exclusive_replace_func: Callable = lambda x: x,
+):
+    if isinstance(start_tags, (str, int)):
+        start_tags = (start_tags,)
+    if isinstance(end_tags, (str, int)):
+        end_tags = (end_tags,)
+    assert len(start_tags) == len(end_tags)
+
+    output_tokens = []
+    end = 0
+    while True:
+        start = _list_find(input_tokens, start_tags, end)
+        if start == -1:
+            break
+        output_tokens.extend(exclusive_replace_func(input_tokens[end: start]))
+        tag_idx = start_tags.index(input_tokens[start])
+        end = _list_find(input_tokens, (end_tags[tag_idx],), start)
+        if end == -1:
+            raise ValueError("Unclosed image token")
+        output_tokens.extend(inclusive_replace_func(input_tokens[start: end + 1]))
+        end += 1
+    output_tokens.extend(exclusive_replace_func(input_tokens[end:]))
+    return output_tokens
+
+
+class VitPhiTokenizer(PreTrainedTokenizer):
+    """VitPhi tokenizer."""
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    def __init__(
+            self,
+            vocab_file,
+            errors="replace",
+            image_start_tag='<img>',
+            image_end_tag='</img>',
+            image_pad_tag='<imgpad>',
+            ref_start_tag='<ref>',
+            ref_end_tag='</ref>',
+            box_start_tag='<box>',
+            box_end_tag='</box>',
+            quad_start_tag='<quad>',
+            quad_end_tag='</quad>',
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.image_start_tag = image_start_tag
+        self.image_end_tag = image_end_tag
+        self.image_pad_tag = image_pad_tag
+        self.ref_start_tag = ref_start_tag
+        self.ref_end_tag = ref_end_tag
+        self.box_start_tag = box_start_tag
+        self.box_end_tag = box_end_tag
+        self.quad_start_tag = quad_start_tag
+        self.quad_end_tag = quad_end_tag
+        self.IMAGE_ST = (
+            ref_start_tag, ref_end_tag,
+            box_start_tag, box_end_tag,
+            quad_start_tag, quad_end_tag,
+            image_start_tag, image_end_tag,
+            image_pad_tag
+        )
+
+        self.errors = errors  # how to handle errors in decoding
+
+        self.mergeable_ranks = _load_tiktoken_bpe(vocab_file)  # type: dict[bytes, int]
+        self.special_tokens = {
+            token: index
+            for index, token in enumerate(
+                SPECIAL_TOKENS + self.IMAGE_ST, start=len(self.mergeable_ranks)
+            )
+        }
+        self.img_start_id = self.special_tokens[self.image_start_tag]
+        self.img_end_id = self.special_tokens[self.image_end_tag]
+        self.img_pad_id = self.special_tokens[self.image_pad_tag]
+        self.ref_start_id = self.special_tokens[self.ref_start_tag]
+        self.ref_end_id = self.special_tokens[self.ref_end_tag]
+        self.box_start_id = self.special_tokens[self.box_start_tag]
+        self.box_end_id = self.special_tokens[self.box_end_tag]
+        self.quad_start_id = self.special_tokens[self.quad_start_tag]
+        self.quad_end_id = self.special_tokens[self.quad_end_tag]
+
+        enc = tiktoken.Encoding(
+            "VitPhi",
+            pat_str=PAT_STR,
+            mergeable_ranks=self.mergeable_ranks,
+            special_tokens=self.special_tokens,
+        )
+        assert (
+                len(self.mergeable_ranks) + len(self.special_tokens) == enc.n_vocab
+        ), f"{len(self.mergeable_ranks) + len(self.special_tokens)} != {enc.n_vocab} in encoding"
+
+        self.decoder = {
+            v: k for k, v in self.mergeable_ranks.items()
+        }  # type: dict[int, bytes|str]
+        self.decoder.update({v: k for k, v in self.special_tokens.items()})
+
+        self.tokenizer = enc  # type: tiktoken.Encoding
+
+        self.eod_id = self.tokenizer.eot_token
+        self.im_start_id = self.special_tokens[IMSTART]
+        self.im_end_id = self.special_tokens[IMEND]
+
+    def __len__(self) -> int:
+        return self.tokenizer.n_vocab
+
+    def get_vocab(self) -> Dict[bytes, int]:
+        return self.mergeable_ranks
+
+    def convert_tokens_to_ids(
+            self, tokens: Union[bytes, str, List[Union[bytes, str]]]
+    ) -> List[int]:
+        ids = []
+        if isinstance(tokens, (str, bytes)):
+            if tokens in self.special_tokens:
+                return self.special_tokens[tokens]
+            else:
+                return self.mergeable_ranks.get(tokens)
+        for token in tokens:
+            if token in self.special_tokens:
+                ids.append(self.special_tokens[token])
+            else:
+                ids.append(self.mergeable_ranks.get(token))
+        return ids
+
+    def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens: bool = False) -> int:
+        if not special_tokens and new_tokens:
+            raise ValueError('Adding regular tokens is not supported')
+        for token in new_tokens:
+            surface_form = token.content if isinstance(token, AddedToken) else token
+            if surface_form not in SPECIAL_TOKENS + self.IMAGE_ST:
+                raise ValueError('Adding unknown special tokens is not supported')
+        return 0
+
+    def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:
+        """
+        Save only the vocabulary of the tokenizer (vocabulary).
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        file_path = os.path.join(save_directory, "vocab.tiktoken")
+        with open(file_path, "w", encoding="utf8") as w:
+            for k, v in self.mergeable_ranks.items():
+                line = base64.b64encode(k).decode("utf8") + " " + str(v) + "\n"
+                w.write(line)
+        return (file_path,)
+
+    def tokenize(
+            self,
+            text: str,
+            allowed_special: Union[Set, str] = "all",
+            disallowed_special: Union[Collection, str] = (),
+            **kwargs,
+    ) -> List[Union[bytes, str]]:
+        """
+        Converts a string in a sequence of tokens.
+
+        Args:
+            text (`str`):
+                The sequence to be encoded.
+            allowed_special (`Literal["all"]` or `set`):
+                The surface forms of the tokens to be encoded as special tokens in regular texts.
+                Default to "all".
+            disallowed_special (`Literal["all"]` or `Collection`):
+                The surface forms of the tokens that should not be in regular texts and trigger errors.
+                Default to an empty tuple.
+
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the underlying model specific encode method.
+
+        Returns:
+            `List[bytes|str]`: The list of tokens.
+        """
+        tokens = []
+        text = unicodedata.normalize("NFC", text)
+
+        # this implementation takes a detour: text -> token id -> token surface forms
+        for t in self.tokenizer.encode(
+                text, allowed_special=allowed_special, disallowed_special=disallowed_special
+        ):
+            tokens.append(self.decoder[t])
+
+        def _encode_imgurl(img_tokens):
+            assert img_tokens[0] == self.image_start_tag and img_tokens[-1] == self.image_end_tag
+            img_tokens = img_tokens[1:-1]
+            img_url = b''.join(img_tokens)
+            out_img_tokens = list(map(self.decoder.get, img_url))
+            if len(out_img_tokens) > IMG_TOKEN_SPAN:
+                raise ValueError("The content in {}..{} is too long".format(
+                    self.image_start_tag, self.image_end_tag))
+            out_img_tokens.extend([self.image_pad_tag] * (IMG_TOKEN_SPAN - len(out_img_tokens)))
+            out_img_tokens = [self.image_start_tag] + out_img_tokens + [self.image_end_tag]
+            return out_img_tokens
+
+        return _replace_closed_tag(tokens, self.image_start_tag, self.image_end_tag, _encode_imgurl)
+
+    def convert_tokens_to_string(self, tokens: List[Union[bytes, str]]) -> str:
+        """
+        Converts a sequence of tokens in a single string.
+        """
+        text = ""
+        temp = b""
+        for t in tokens:
+            if isinstance(t, str):
+                if temp:
+                    text += temp.decode("utf-8", errors=self.errors)
+                    temp = b""
+                text += t
+            elif isinstance(t, bytes):
+                temp += t
+            else:
+                raise TypeError("token should only be of type types or str")
+        if temp:
+            text += temp.decode("utf-8", errors=self.errors)
+        return text
+
+    @property
+    def vocab_size(self):
+        return self.tokenizer.n_vocab
+
+    def _convert_id_to_token(self, index: int) -> Union[bytes, str]:
+        """Converts an id to a token, special tokens included"""
+        if index in self.decoder:
+            return self.decoder[index]
+        raise ValueError("unknown ids")
+
+    def _convert_token_to_id(self, token: Union[bytes, str]) -> int:
+        """Converts a token to an id using the vocab, special tokens included"""
+        if token in self.special_tokens:
+            return self.special_tokens[token]
+        if token in self.mergeable_ranks:
+            return self.mergeable_ranks[token]
+        raise ValueError("unknown token")
+
+    def _tokenize(self, text: str, **kwargs):
+        """
+        Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
+        vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
+
+        Do NOT take care of added tokens.
+        """
+        raise NotImplementedError
+
+    def _decode(
+            self,
+            token_ids: Union[int, List[int]],
+            skip_special_tokens: bool = False,
+            errors: str = None,
+            **kwargs,
+    ) -> str:
+        if isinstance(token_ids, int):
+            token_ids = [token_ids]
+
+        def _decode_imgurl(img_token_ids):
+            assert img_token_ids[0] == self.img_start_id and img_token_ids[-1] == self.img_end_id
+            img_token_ids = img_token_ids[1:-1]
+            img_token_ids = img_token_ids[: img_token_ids.index(self.img_pad_id)]
+            img_url = bytes(img_token_ids).decode('utf-8')
+            return [self.img_start_id] + self.tokenizer.encode(img_url) + [self.img_end_id]
+
+        token_ids = _replace_closed_tag(token_ids, self.img_start_id, self.img_end_id, _decode_imgurl)
+
+        if skip_special_tokens:
+            token_ids = [i for i in token_ids if i < self.eod_id]
+        return self.tokenizer.decode(token_ids, errors=errors or self.errors)
+
+    def to_list_format(self, text: str):
+        text = unicodedata.normalize("NFC", text)
+        token_ids = self.tokenizer.encode(
+            text, allowed_special=set(self.IMAGE_ST + (ENDOFTEXT,)))
+
+        def _encode_vl_info(tokens):
+            if len(tokens) == 0:
+                return []
+            if tokens[0] == self.img_start_id and tokens[-1] == self.img_end_id:
+                key = 'image'
+            elif tokens[0] == self.ref_start_id and tokens[-1] == self.ref_end_id:
+                key = 'ref'
+            elif tokens[0] == self.box_start_id and tokens[-1] == self.box_end_id:
+                key = 'box'
+            elif tokens[0] == self.quad_start_id and tokens[-1] == self.quad_end_id:
+                key = 'quad'
+            else:
+                _tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
+                return [{'text': b''.join(map(_tobytes, map(self.decoder.get, tokens))).decode('utf-8')}]
+            _tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
+            val = b''.join(map(_tobytes, map(self.decoder.get, tokens[1:-1]))).decode('utf-8')
+            return [{key: val}]
+
+        return _replace_closed_tag(
+            token_ids,
+            (self.img_start_id, self.ref_start_id, self.box_start_id, self.quad_start_id),
+            (self.img_end_id, self.ref_end_id, self.box_end_id, self.quad_end_id),
+            _encode_vl_info,
+            _encode_vl_info,
+        )
+
+    def from_list_format(self, list_format: List[Dict]):
+        text = ''
+        num_images = 0
+        for ele in list_format:
+            if 'image' in ele:
+                num_images += 1
+                text += f'Picture {num_images}:'
+                text += self.image_start_tag + ele['image'] + self.image_end_tag
+                text += '\n'
+            elif 'text' in ele:
+                text += ele['text']
+            elif 'box' in ele:
+                if 'ref' in ele:
+                    text += self.ref_start_tag + ele['ref'] + self.ref_end_tag
+                for box in ele['box']:
+                    text += self.box_start_tag + '(%d,%d),(%d,%d)' % (box[0], box[1], box[2], box[3]) + self.box_end_tag
+            else:
+                raise ValueError("Unsupport element: " + str(ele))
+        return text
+
+    def _fetch_latest_picture(self, response, history):
+        if history is None:
+            history = []
+        _history = history + [(response, None)]
+        for q, r in _history[::-1]:
+            for ele in self.to_list_format(q)[::-1]:
+                if 'image' in ele:
+                    return ele['image']
+        return None
+
+    def _fetch_all_box_with_ref(self, text):
+        list_format = self.to_list_format(text)
+        output = []
+        for i, ele in enumerate(list_format):
+            if 'box' in ele:
+                bbox = tuple(map(int, ele['box'].replace('(', '').replace(')', '').split(',')))
+                assert len(bbox) == 4
+                output.append({'box': bbox})
+                if i > 0 and 'ref' in list_format[i - 1]:
+                    output[-1]['ref'] = list_format[i - 1]['ref'].strip()
+        return output
+
+    def draw_bbox_on_latest_picture(
+            self,
+            response,
+            history=None,
+    ) -> Optional[Image.Image]:
+        image = self._fetch_latest_picture(response, history)
+        if image is None:
+            return None
+        if image.startswith("http://") or image.startswith("https://"):
+            image = Image.open(requests.get(image, stream=True).raw).convert("RGB")
+            h, w = image.height, image.width
+        else:
+            image = np.asarray(Image.open(image).convert("RGB"))
+            h, w = image.shape[0], image.shape[1]
+        visualizer = Visualizer(image)
+
+        boxes = self._fetch_all_box_with_ref(response)
+        if not boxes:
+            return None
+        color = random.choice([_ for _ in mcolors.TABLEAU_COLORS.keys()])  # init color
+        for box in boxes:
+            if 'ref' in box:  # random new color for new refexps
+                color = random.choice([_ for _ in mcolors.TABLEAU_COLORS.keys()])
+            x1, y1, x2, y2 = box['box']
+            x1, y1, x2, y2 = (int(x1 / 1000 * w), int(y1 / 1000 * h), int(x2 / 1000 * w), int(y2 / 1000 * h))
+            visualizer.draw_box((x1, y1, x2, y2), alpha=1, edge_color=color)
+            if 'ref' in box:
+                visualizer.draw_text(box['ref'], (x1, y1), color=color, horizontal_alignment="left")
+        return visualizer.output
+
+
+import colorsys
+import logging
+import math
+import numpy as np
+import matplotlib as mpl
+import matplotlib.colors as mplc
+import matplotlib.figure as mplfigure
+import torch
+from matplotlib.backends.backend_agg import FigureCanvasAgg
+from PIL import Image
+import random
+
+logger = logging.getLogger(__name__)
+
+
+class VisImage:
+    def __init__(self, img, scale=1.0):
+        self.img = img
+        self.scale = scale
+        self.width, self.height = img.shape[1], img.shape[0]
+        self._setup_figure(img)
+
+    def _setup_figure(self, img):
+        fig = mplfigure.Figure(frameon=False)
+        self.dpi = fig.get_dpi()
+        # add a small 1e-2 to avoid precision lost due to matplotlib's truncation
+        # (https://github.com/matplotlib/matplotlib/issues/15363)
+        fig.set_size_inches(
+            (self.width * self.scale + 1e-2) / self.dpi,
+            (self.height * self.scale + 1e-2) / self.dpi,
+        )
+        self.canvas = FigureCanvasAgg(fig)
+        # self.canvas = mpl.backends.backend_cairo.FigureCanvasCairo(fig)
+        ax = fig.add_axes([0.0, 0.0, 1.0, 1.0])
+        ax.axis("off")
+        self.fig = fig
+        self.ax = ax
+        self.reset_image(img)
+
+    def reset_image(self, img):
+        img = img.astype("uint8")
+        self.ax.imshow(img, extent=(0, self.width, self.height, 0), interpolation="nearest")
+
+    def save(self, filepath):
+        self.fig.savefig(filepath)
+
+    def get_image(self):
+        canvas = self.canvas
+        s, (width, height) = canvas.print_to_buffer()
+
+        buffer = np.frombuffer(s, dtype="uint8")
+
+        img_rgba = buffer.reshape(height, width, 4)
+        rgb, alpha = np.split(img_rgba, [3], axis=2)
+        return rgb.astype("uint8")
+
+
+class Visualizer:
+    def __init__(self, img_rgb, metadata=None, scale=1.0):
+        self.img = np.asarray(img_rgb).clip(0, 255).astype(np.uint8)
+        self.font_path = FONT_PATH
+        self.output = VisImage(self.img, scale=scale)
+        self.cpu_device = torch.device("cpu")
+
+        # too small texts are useless, therefore clamp to 14
+        self._default_font_size = max(
+            np.sqrt(self.output.height * self.output.width) // 30, 15 // scale
+        )
+
+    def draw_text(
+            self,
+            text,
+            position,
+            *,
+            font_size=None,
+            color="g",
+            horizontal_alignment="center",
+            rotation=0,
+    ):
+        if not font_size:
+            font_size = self._default_font_size
+
+        # since the text background is dark, we don't want the text to be dark
+        color = np.maximum(list(mplc.to_rgb(color)), 0.2)
+        color[np.argmax(color)] = max(0.8, np.max(color))
+
+        x, y = position
+        self.output.ax.text(
+            x,
+            y,
+            text,
+            size=font_size * self.output.scale,
+            fontproperties=FontProperties(fname=self.font_path),
+            bbox={"facecolor": "black", "alpha": 0.8, "pad": 0.7, "edgecolor": "none"},
+            verticalalignment="top",
+            horizontalalignment=horizontal_alignment,
+            color=color,
+            zorder=10,
+            rotation=rotation,
+        )
+        return self.output
+
+    def draw_box(self, box_coord, alpha=0.5, edge_color="g", line_style="-"):
+        x0, y0, x1, y1 = box_coord
+        width = x1 - x0
+        height = y1 - y0
+
+        linewidth = max(self._default_font_size / 4, 1)
+
+        self.output.ax.add_patch(
+            mpl.patches.Rectangle(
+                (x0, y0),
+                width,
+                height,
+                fill=False,
+                edgecolor=edge_color,
+                linewidth=linewidth * self.output.scale,
+                alpha=alpha,
+                linestyle=line_style,
+            )
+        )
+        return self.output
+
+    def get_output(self):
+        return self.output

tokenizer_config.json

@@ -0,0 +1,10 @@
+{
+  "model_max_length": 2048,
+  "tokenizer_class": "VitPhiTokenizer",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_vitphi.VitPhiTokenizer",
+      null
+      ]
+  }
+}

visual.py

@@ -0,0 +1,428 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from collections import OrderedDict
+import math
+import requests
+from io import BytesIO
+from functools import partial
+from PIL import Image
+from typing import Callable, Optional, Sequence, Tuple, List
+import numpy as np
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.init import trunc_normal_
+from torchvision import transforms
+from torchvision.transforms import InterpolationMode
+
+
+def get_abs_pos(abs_pos, tgt_size):
+    # abs_pos: L, C
+    # tgt_size: M
+    # return: M, C
+    src_size = int(math.sqrt(abs_pos.size(0)))
+    tgt_size = int(math.sqrt(tgt_size))
+    dtype = abs_pos.dtype
+
+    if src_size != tgt_size:
+        return F.interpolate(
+            abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
+            size=(tgt_size, tgt_size),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype)
+    else:
+        return abs_pos
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+class Resampler(nn.Module):
+    """
+    A 2D perceiver-resampler network with one cross attention layers by
+        (grid_size**2) learnable queries and 2d sincos pos_emb
+    Outputs:
+        A tensor with the shape of (grid_size**2, embed_dim)
+    """
+
+    def __init__(
+            self,
+            grid_size,
+            embed_dim,
+            num_heads,
+            kv_dim=None,
+            norm_layer=nn.LayerNorm
+    ):
+        super().__init__()
+        self.num_queries = grid_size ** 2
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+
+        self.pos_embed = nn.Parameter(
+            torch.from_numpy(get_2d_sincos_pos_embed(embed_dim, grid_size)).float()
+        ).requires_grad_(False)
+
+        self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
+        trunc_normal_(self.query, std=.02)
+
+        if kv_dim is not None and kv_dim != embed_dim:
+            self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
+        else:
+            self.kv_proj = nn.Identity()
+
+        self.attn = nn.MultiheadAttention(embed_dim, num_heads)
+        self.ln_q = norm_layer(embed_dim)
+        self.ln_kv = norm_layer(embed_dim)
+
+        self.apply(self._init_weights)
+
+    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 forward(self, x, attn_mask=None):
+
+        pos_embed = get_abs_pos(self.pos_embed, x.size(1))
+
+        x = self.kv_proj(x)
+        x = self.ln_kv(x).permute(1, 0, 2)
+
+        N = x.shape[1]
+        q = self.ln_q(self.query)
+        out = self.attn(
+            self._repeat(q, N) + self.pos_embed.unsqueeze(1),
+            x + pos_embed.unsqueeze(1),
+            x,
+            attn_mask=attn_mask)[0]
+        return out.permute(1, 0, 2)
+
+    def _repeat(self, query, N: int):
+        return query.unsqueeze(1).repeat(1, N, 1)
+
+
+class VisualAttention(nn.Module):
+    """self-attention layer class.
+
+    Self-attention layer takes input with size [s, b, h]
+    and returns output of the same size.
+    """
+
+    def __init__(self, embed_dim, num_heads,
+                 bias=True, kdim=None, vdim=None):
+        super(VisualAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+
+        # Per attention head and per partition values.
+        assert embed_dim % num_heads == 0
+        self.hidden_size_per_attention_head = embed_dim // num_heads
+        self.num_attention_heads_per_partition = num_heads
+        self.hidden_size_per_partition = embed_dim
+
+        # Strided linear layer.
+        assert self._qkv_same_embed_dim, 'Only Support SelfAttention Currently'
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim)
+        self.out_proj = nn.Linear(embed_dim, embed_dim)
+        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
+
+    def forward(self, query, key, value, attn_mask=None):
+        # query/key/value: [sq, b, h]
+        sq, b, _ = query.size()
+
+        assert query is key, 'Only Support Self-Attention Currently'
+        sk = sq
+        mixed_x_layer = self.in_proj(query)
+
+        # [sq, b, (np * 3 * hn)] --> [sq, b, np, 3 * hn]
+        new_tensor_shape = mixed_x_layer.size()[:-1] + \
+                           (self.num_attention_heads_per_partition,
+                            3 * self.hidden_size_per_attention_head)
+        mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+
+        # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
+        query_layer, key_layer, value_layer = mixed_x_layer.split(
+            self.hidden_size_per_attention_head, dim=-1)
+
+        # [sq, b, np, hn] -> [sq, b * np, hn]
+        query_layer = query_layer.view(sq,
+                                       b * self.num_attention_heads_per_partition,
+                                       self.hidden_size_per_attention_head).transpose(0, 1)
+        # [sk, b, np, hn] -> [sk, b * np, hn]
+        key_layer = key_layer.view(sk,
+                                   b * self.num_attention_heads_per_partition,
+                                   self.hidden_size_per_attention_head).transpose(0, 1)
+
+        q_scaled = query_layer / self.norm_factor
+        if attn_mask is not None:
+            attention_probs = torch.baddbmm(attn_mask, q_scaled, key_layer.transpose(-2, -1))
+        else:
+            attention_probs = torch.bmm(q_scaled, key_layer.transpose(-2, -1))
+        attention_probs = attention_probs.softmax(dim=-1)
+
+        value_layer = value_layer.view(sk,
+                                       b * self.num_attention_heads_per_partition,
+                                       self.hidden_size_per_attention_head).transpose(0, 1)
+
+        # matmul: [b * np, sq, hn]
+        context_layer = torch.bmm(attention_probs, value_layer)
+
+        # change view [b, np, sq, hn]
+        context_layer = context_layer.view(b,
+                                           self.num_attention_heads_per_partition,
+                                           sq, self.hidden_size_per_attention_head)
+
+        # [b, np, sq, hn] --> [sq, b, np, hn]
+        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+        # [sq, b, np, hn] --> [sq, b, hp]
+        new_context_layer_shape = context_layer.size()[:-2] + \
+                                  (self.hidden_size_per_partition,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        output = self.out_proj(context_layer)
+
+        return output
+
+
+class VisualAttentionBlock(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            n_head: int,
+            mlp_ratio: float = 4.0,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = nn.LayerNorm,
+            is_cross_attention: bool = False,
+    ):
+        super().__init__()
+
+        self.ln_1 = norm_layer(d_model)
+        if is_cross_attention:
+            self.ln_1_kv = norm_layer(d_model)
+
+        self.ln_2 = norm_layer(d_model)
+        mlp_width = int(d_model * mlp_ratio)
+        self.attn = VisualAttention(d_model, n_head)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, mlp_width)),
+            ("gelu", act_layer()),
+            ("c_proj", nn.Linear(mlp_width, d_model))
+        ]))
+
+    def attention(
+            self,
+            q_x: torch.Tensor,
+            k_x: Optional[torch.Tensor] = None,
+            v_x: Optional[torch.Tensor] = None,
+            attn_mask: Optional[torch.Tensor] = None,
+    ):
+        k_x = k_x if k_x is not None else q_x
+        v_x = v_x if v_x is not None else q_x
+
+        attn_mask = attn_mask.to(q_x.dtype) if attn_mask is not None else None
+        return self.attn(q_x, k_x, v_x, attn_mask=attn_mask)
+
+    def forward(
+            self,
+            q_x: torch.Tensor,
+            k_x: Optional[torch.Tensor] = None,
+            v_x: Optional[torch.Tensor] = None,
+            attn_mask: Optional[torch.Tensor] = None,
+    ):
+        k_x = self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
+        v_x = self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None
+
+        x = q_x + self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class TransformerBlock(nn.Module):
+    def __init__(
+            self,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float = 4.0,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = nn.LayerNorm,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+
+        self.resblocks = nn.ModuleList([
+            VisualAttentionBlock(
+                width, heads, mlp_ratio, act_layer=act_layer, norm_layer=norm_layer)
+            for _ in range(layers)
+        ])
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.resblocks[0].mlp.c_fc.weight.dtype
+
+    def get_cast_device(self) -> torch.device:
+        return self.resblocks[0].mlp.c_fc.weight.device
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisionTransformer(nn.Module):
+
+    def __init__(
+            self,
+            image_size: int,
+            patch_size: int,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float,
+            n_queries: int = 256,
+            output_dim: int = 512,
+            **kwargs
+    ):
+        super().__init__()
+        image_height, image_width = self.image_size = (image_size, image_size)
+        patch_height, patch_width = self.patch_size = (patch_size, patch_size)
+        self.grid_size = (image_height // patch_height, image_width // patch_width)
+        self.output_dim = output_dim
+
+        mean = (0.48145466, 0.4578275, 0.40821073)
+        std = (0.26862954, 0.26130258, 0.27577711)
+        self.image_transform = transforms.Compose([
+            transforms.Resize(
+                (image_size, image_size),
+                interpolation=InterpolationMode.BICUBIC
+            ),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ])
+
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        # class embeddings and positional embeddings
+        scale = width ** -0.5
+        self.positional_embedding = nn.Parameter(scale * torch.randn(256, width))
+
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+        act_layer = nn.GELU
+
+        self.ln_pre = norm_layer(width)
+        self.transformer = TransformerBlock(
+            width,
+            layers,
+            heads,
+            mlp_ratio,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+
+        self.attn_pool = Resampler(
+            grid_size=int(math.sqrt(n_queries)),
+            embed_dim=output_dim,
+            num_heads=output_dim // 128,
+            kv_dim=width,
+            norm_layer=norm_layer,
+        )
+        self.ln_post = norm_layer(output_dim)
+        self.proj = nn.Parameter((output_dim ** -0.5) * torch.randn(output_dim, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = x.to(
+            dtype=self.transformer.get_cast_dtype(),
+            device=self.transformer.get_cast_device(),
+        )
+        # to patches
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+        x = x + get_abs_pos(self.positional_embedding, x.size(1))
+
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.attn_pool(x)
+        x = self.ln_post(x)
+        x = x @ self.proj
+
+        return x
+
+    def encode(self, image_paths: List[str]):
+        images = []
+        for image_path in image_paths:
+            if image_path.startswith("http://") or image_path.startswith("https://"):
+                image = Image.open(requests.get(image_path, stream=True).raw)
+            else:
+                image = Image.open(image_path)
+            image = image.convert("RGB")
+            images.append(self.image_transform(image))
+        images = torch.stack(images, dim=0)
+        return self(images)