Initial commit

internvl/model/phi3/configuration_phi3.py

+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License atd
+#
+#     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.
+
+""" Phi-3 model configuration"""
+
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'microsoft/Phi-3-mini-4k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json',
+    'microsoft/Phi-3-mini-128k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json',
+}
+
+
+class Phi3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the
+    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32064):
+            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Phi3Model`].
+        hidden_size (`int`, *optional*, defaults to 3072):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            Dropout probability for mlp outputs.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio after computing the attention scores.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        original_max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model was trained with. This is used to determine the size of the
+            original RoPE embeddings when using long scaling.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon value used for the RMSNorm.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`dict`, *optional*):
+            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
+            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and
+            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
+            divided by the number of attention heads divided by 2.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 32000):
+            The id of the "end-of-sequence" token.
+        pad_token_id (`int`, *optional*, defaults to 32000):
+            The id of the padding token.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If `None`, no sliding window is applied.
+
+    Example:
+
+    ```python
+    >>> from transformers import Phi3Model, Phi3Config
+
+    >>> # Initializing a Phi-3 style configuration
+    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
+
+    >>> # Initializing a model from the configuration
+    >>> model = Phi3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = 'phi3'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        vocab_size=32064,
+        hidden_size=3072,
+        intermediate_size=8192,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attention_dropout=0.0,
+        hidden_act='silu',
+        max_position_embeddings=4096,
+        original_max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        bos_token_id=1,
+        eos_token_id=32000,
+        pad_token_id=32000,
+        sliding_window=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attention_dropout = attention_dropout
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self._rope_scaling_validation()
+        self.sliding_window = sliding_window
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
+            raise ValueError(
+                '`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, '
+                f'got {self.rope_scaling}'
+            )
+        rope_scaling_type = self.rope_scaling.get('type', None)
+        rope_scaling_short_factor = self.rope_scaling.get('short_factor', None)
+        rope_scaling_long_factor = self.rope_scaling.get('long_factor', None)
+        if rope_scaling_type is None or rope_scaling_type not in ['su', 'yarn']:
+            raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")
+        if not (
+            isinstance(rope_scaling_short_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
+            )
+        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
+            )
+        if not (
+            isinstance(rope_scaling_long_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
+            )
+        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
+            )

internvl/model/phi3/modeling_phi3.py

+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" PyTorch Phi-3 model."""
+
+import inspect
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import \
+    _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import (BaseModelOutputWithPast,
+                                           CausalLMOutputWithPast,
+                                           SequenceClassifierOutputWithPast,
+                                           TokenClassifierOutput)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (add_code_sample_docstrings,
+                                add_start_docstrings,
+                                add_start_docstrings_to_model_forward,
+                                is_flash_attn_2_available,
+                                is_flash_attn_greater_or_equal_2_10, logging,
+                                replace_return_docstrings)
+
+from .configuration_phi3 import Phi3Config
+
+logger = logging.get_logger(__name__)
+
+# Transformers scans dependencies in the modeling file, causing issues on conditional loading. The regex only ignores try/catch blocks, but not if statements
+# if is_flash_attn_2_available():
+_flash_supports_window_size = False
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import (index_first_axis, pad_input,  # noqa
+                                         unpad_input)
+
+    _flash_supports_window_size = 'window_size' in list(inspect.signature(flash_attn_func).parameters)
+except ImportError as error:
+    logger.warning(
+        f'`flash-attention` package not found, consider installing for better performance: {error}.'
+    )
+    if not _flash_supports_window_size:
+        logger.warning(
+            "Current `flash-attenton` does not support `window_size`. Either upgrade or use `attn_implementation='eager'`."
+        )
+
+_CHECKPOINT_FOR_DOC = 'microsoft/Phi-3-mini-4k-instruct'
+_CONFIG_FOR_DOC = 'Phi3Config'
+
+PHI3_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    'microsoft/Phi-3-mini-4k-instruct',
+    'microsoft/Phi-3-mini-128k-instruct',
+    # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3
+]
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
+class Phi3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Phi3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
+class Phi3RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        self.register_buffer('inv_freq', None, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if self.inv_freq is None:
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
+            )
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu'
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
+    def __init__(self, dim, config, device=None):
+        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
+
+        self.short_factor = config.rope_scaling['short_factor']
+        self.long_factor = config.rope_scaling['long_factor']
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.original_max_position_embeddings:
+            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
+        else:
+            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
+
+        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
+        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu'
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+
+            scale = self.max_position_embeddings / self.original_max_position_embeddings
+            if scale <= 1.0:
+                scaling_factor = 1.0
+            else:
+                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
+
+            cos = emb.cos() * scaling_factor
+            sin = emb.sin() * scaling_factor
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
+    def __init__(self, dim, config, device=None):
+        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
+
+        self.short_factor = config.rope_scaling['short_factor']
+        self.long_factor = config.rope_scaling['long_factor']
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.original_max_position_embeddings:
+            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
+        else:
+            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
+
+        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
+        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu'
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+
+            scale = self.max_position_embeddings / self.original_max_position_embeddings
+            if scale <= 1.0:
+                scaling_factor = 1.0
+            else:
+                scaling_factor = 0.1 * math.log(scale) + 1.0
+
+            cos = emb.cos() * scaling_factor
+            sin = emb.sin() * scaling_factor
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+        self.activation_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
+
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
+
+        return self.down_proj(up_states)
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Phi3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f'Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will '
+                'lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` '
+                'when creating this class.'
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = config.rope_scaling
+        self.is_causal = True
+
+        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}).'
+            )
+
+        op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
+        self._init_rope()
+
+    def _init_rope(self):
+        if self.rope_scaling is None:
+            self.rotary_emb = Phi3RotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+        else:
+            scaling_type = self.config.rope_scaling['type']
+            if scaling_type == 'su':
+                self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
+            elif scaling_type == 'yarn':
+                self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
+            else:
+                raise ValueError(f'Unknown RoPE scaling type {scaling_type}')
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        logger.warning_once('You are not running the flash-attention implementation, expect numerical differences.')
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f'The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} '
+                    'for auto-regressive decoding with k/v caching, please make sure to initialize the attention class '
+                    'with a layer index.'
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {'sin': sin, 'cos': cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        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
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+
+        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).contiguous()
+        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 Phi3FlashAttention2(Phi3Attention):
+    """
+    Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # Phi3FlashAttention2 attention does not support output_attentions
+
+        if not _flash_supports_window_size:
+            logger.warning_once(
+                "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
+            )
+            raise ValueError('The current flash attention version does not support sliding window attention.')
+
+        output_attentions = False
+
+        if 'padding_mask' in kwargs:
+            warnings.warn(
+                'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`'
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop('padding_mask')
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f'The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} '
+                    'for auto-regressive decoding with k/v caching, please make sure to initialize the attention class '
+                    'with a layer index.'
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        # Because the input can be padded, the absolute sequence length depends on the max position id.
+        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        use_sliding_windows = (
+            _flash_supports_window_size
+            and getattr(self.config, 'sliding_window', None) is not None
+            and kv_seq_len > self.config.sliding_window
+        )
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, 'sliding_window', None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
+                slicing_tokens = 1 - self.config.sliding_window
+
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        f'past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got'
+                        f' {past_key.shape}'
+                    )
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            cache_kwargs = {'sin': sin, 'cos': cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_dropout = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32.
+
+        if query_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, '_pre_quantization_dtype'):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.qkv_proj.weight.dtype
+
+            logger.warning_once(
+                f'The input hidden states seems to be silently casted in float32, this might be related to'
+                f' the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in'
+                f' {target_dtype}.'
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=attn_dropout,
+            use_sliding_windows=use_sliding_windows,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+        use_sliding_windows=False,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            use_sliding_windows (`bool`, *optional*):
+                Whether to activate sliding window attention.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            if not use_sliding_windows:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            if not use_sliding_windows:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+        return attn_output
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+        # On the first iteration we need to properly re-create the padding mask
+        # by slicing it on the proper place
+        if kv_seq_len != attention_mask.shape[-1]:
+            attention_mask_num_tokens = attention_mask.shape[-1]
+            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
+# TODO @Arthur no longer copied from LLama after static cache
+class Phi3SdpaAttention(Phi3Attention):
+    """
+    Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from Phi3Attention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                'Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, '
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                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,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_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.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {'sin': sin, 'cos': cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        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()}'
+                )
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == 'cuda' and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+PHI3_ATTENTION_CLASSES = {
+    'eager': Phi3Attention,
+    'flash_attention_2': Phi3FlashAttention2,
+    'sdpa': Phi3SdpaAttention,
+}
+
+
+class Phi3DecoderLayer(nn.Module):
+    def __init__(self, config: Phi3Config, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+
+        self.mlp = Phi3MLP(config)
+        self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
+        self.post_attention_layernorm = Phi3RMSNorm(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,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        if 'padding_mask' in kwargs:
+            warnings.warn(
+                'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`'
+            )
+        """
+        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.
+            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
+                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+            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
+        attn_outputs, 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 + self.resid_attn_dropout(attn_outputs)
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+PHI3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Phi3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    'The bare Phi-3 model outputting raw hidden-states without any specific head on top.',
+    PHI3_START_DOCSTRING,
+)
+class Phi3PreTrainedModel(PreTrainedModel):
+    config_class = Phi3Config
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['Phi3DecoderLayer']
+    _skip_keys_device_placement = 'past_key_values'
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+    _supports_cache_class = True
+
+    _version = '0.0.5'
+
+    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_()
+
+
+PHI3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    'The bare Phi-3 model outputting raw hidden-states without any specific head on top.',
+    PHI3_START_DOCSTRING,
+)
+class Phi3Model(Phi3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
+
+    Args:
+        config: Phi3Config
+    """
+
+    def __init__(self, config: Phi3Config):
+        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.embed_dropout = nn.Dropout(config.embd_pdrop)
+        self.layers = nn.ModuleList(
+            [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = Phi3RMSNorm(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
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    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,
+        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 input_ids and inputs_embeds at the same time')
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError('You have to specify either input_ids or inputs_embeds')
+
+        past_key_values_length = 0
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    '`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...'
+                )
+                use_cache = False
+
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_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()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self._attn_implementation == 'flash_attention_2' and use_cache:
+            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            if is_padding_right:
+                raise ValueError(
+                    "You are attempting to perform batched generation with padding_side='right'"
+                    ' this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to '
+                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                )
+
+        if self._attn_implementation == 'flash_attention_2':
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+                sliding_window=self.config.sliding_window,
+            )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    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 = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+        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 Phi3ForCausalLM(Phi3PreTrainedModel):
+    _tied_weights_keys = ['lm_head.weight']
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Phi3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
+    def get_decoder(self):
+        return self.model
+
+    # Ignore copy
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    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,
+        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, Phi3ForCausalLM
+
+        >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+
+        >>> prompt = "This is an example script ."
+        >>> 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]
+        'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+        ```"""
+
+        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,
+            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)
+        logits = logits.float()
+
+        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()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        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[:, -input_ids.shape[1] :]
+
+        # 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,
+                'past_key_values': past_key_values,
+                'use_cache': kwargs.get('use_cache'),
+                'attention_mask': attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
+    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.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The [`Phi3Model`] with a sequence classification head on top (linear layer).
+
+    [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
+class Phi3ForSequenceClassification(Phi3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Phi3Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    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,
+        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, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = model_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError('Cannot handle batch sizes > 1 if no padding token is defined.')
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = 'regression'
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = 'single_label_classification'
+                else:
+                    self.config.problem_type = 'multi_label_classification'
+
+            if self.config.problem_type == 'regression':
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == 'single_label_classification':
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == 'multi_label_classification':
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + model_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=model_outputs.past_key_values,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
+class Phi3ForTokenClassification(Phi3PreTrainedModel):
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.model = Phi3Model(config)
+        if hasattr(config, 'classifier_dropout') and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, 'hidden_dropout') and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = model_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            batch_size, seq_length = labels.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (logits,) + model_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )

internvl/patch/__init__.py

+from .llama2_flash_attn_monkey_patch import replace_llama2_attn_with_flash_attn
+from .llama_flash_attn_monkey_patch import replace_llama_attn_with_flash_attn
+from .llama_rmsnorm_monkey_patch import \
+    replace_llama_rmsnorm_with_fused_rmsnorm
+from .pad_data_collator import concat_pad_data_collator, pad_data_collator
+from .train_sampler_patch import replace_train_sampler
+
+__all__ = ['replace_llama_attn_with_flash_attn',
+           'replace_llama_rmsnorm_with_fused_rmsnorm',
+           'replace_llama2_attn_with_flash_attn',
+           'replace_train_sampler',
+           'pad_data_collator',
+           'concat_pad_data_collator']

internvl/patch/llama2_flash_attn_monkey_patch.py

+"""
+This file is copied from: https://github.com/lm-sys/FastChat
+"""
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from flash_attn import __version__ as flash_attn_version
+from flash_attn.bert_padding import pad_input, unpad_input
+from flash_attn.flash_attn_interface import (flash_attn_func,
+                                             flash_attn_varlen_kvpacked_func)
+from transformers.models.llama.modeling_llama import (LlamaAttention,
+                                                      LlamaModel, rotate_half)
+
+
+def apply_rotary_pos_emb(q, k, cos_sin, position_ids):
+    gather_indices = position_ids[:, :, None, None]  # [bsz, seq_len, 1, 1]
+    gather_indices = gather_indices.repeat(
+        1, 1, cos_sin[0].shape[1], cos_sin[0].shape[3]
+    )
+    bsz = gather_indices.shape[0]
+    cos, sin = (
+        torch.gather(x.transpose(1, 2).repeat(bsz, 1, 1, 1), 1, gather_indices)
+        for x in cos_sin
+    )
+    q, k = ((x * cos) + (rotate_half(x) * sin) for x in (q, k))
+    return q, k
+
+
+def forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.Tensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    padding_mask: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if output_attentions:
+        warnings.warn(
+            'Output attentions is not supported for patched `LlamaAttention`, returning `None` instead.'
+        )
+
+    bsz, q_len, _ = hidden_states.size()
+    kv_heads = getattr(self, 'num_key_value_heads', self.num_heads)
+
+    q, k, v = (
+        op(hidden_states).view(bsz, q_len, nh, self.head_dim)
+        for op, nh in (
+            (self.q_proj, self.num_heads),
+            (self.k_proj, kv_heads),
+            (self.v_proj, kv_heads),
+        )
+    )
+    # shape: (b, s, num_heads, head_dim)
+
+    kv_seq_len = k.shape[1]
+    past_kv_len = 0
+    if past_key_value is not None:
+        past_kv_len = past_key_value[0].shape[2]
+        kv_seq_len += past_kv_len
+
+    cos_sin = self.rotary_emb(v, seq_len=kv_seq_len)
+    q, k = apply_rotary_pos_emb(q, k, cos_sin, position_ids)
+
+    if past_key_value is not None:
+        assert (
+            flash_attn_version >= '2.1.0'
+        ), 'past_key_value support requires flash-attn >= 2.1.0'
+        # reuse k, v
+        k = torch.cat([past_key_value[0].transpose(1, 2), k], dim=1)
+        v = torch.cat([past_key_value[1].transpose(1, 2), v], dim=1)
+
+    past_key_value = (k.transpose(1, 2), v.transpose(1, 2)) if use_cache else None
+
+    if attention_mask is None:
+        output = flash_attn_func(q, k, v, 0.0, softmax_scale=None, causal=True).view(
+            bsz, q_len, -1
+        )
+    else:
+        q, indices, cu_q_lens, max_s = unpad_input(q, attention_mask[:, -q_len:])
+        # We can skip concat and call unpad twice but seems better to call unpad only once.
+        kv, _, cu_k_lens, max_k = unpad_input(
+            torch.stack((k, v), dim=2), attention_mask
+        )
+        output_unpad = flash_attn_varlen_kvpacked_func(
+            q,
+            kv,
+            cu_q_lens,
+            cu_k_lens,
+            max_s,
+            max_k,
+            0.0,
+            softmax_scale=None,
+            causal=True,
+        )
+        output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
+        output = pad_input(output_unpad, indices, bsz, q_len)
+
+    return self.o_proj(output), None, past_key_value
+
+
+# Disable the transformation of the attention mask in LlamaModel as flash attention
+# takes a boolean key_padding_mask. Fills in the past kv length for use in forward.
+def _prepare_decoder_attention_mask(
+    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
+):
+    # [bsz, seq_len]
+    if past_key_values_length > 0 and attention_mask is not None:
+        attention_mask = torch.cat(
+            (
+                torch.full(
+                    (input_shape[0], past_key_values_length),
+                    True,
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                ),
+                attention_mask,
+            ),
+            dim=-1,
+        )
+
+    if attention_mask is not None and torch.all(attention_mask):
+        return None  # This uses the faster call when training with full samples
+
+    return attention_mask
+
+
+def replace_llama2_attn_with_flash_attn():
+    cuda_major, cuda_minor = torch.cuda.get_device_capability()
+    if cuda_major < 8:
+        warnings.warn(
+            'Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward.'
+            'ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593'
+        )
+
+    LlamaModel._prepare_decoder_attention_mask = _prepare_decoder_attention_mask
+    LlamaAttention.forward = forward
+
+
+def test():
+    from fastchat.train.llama_flash_attn_monkey_patch import \
+        forward as fastchat_forward
+    from transformers.models.llama.configuration_llama import LlamaConfig
+
+    config = LlamaConfig(
+        hidden_size=1024,
+        intermediate_size=128,
+        num_hidden_layers=1,
+        num_attention_heads=8,
+        max_position_embeddings=16,
+    )
+    device = torch.device('cuda')
+    model = LlamaModel(config)
+    attn = LlamaAttention(config).to(device).half()
+    bsz, hs, seqlen = 2, config.hidden_size, config.max_position_embeddings
+    position_ids = torch.arange(seqlen, dtype=torch.long, device=device).view(
+        -1, seqlen
+    )
+
+    mask = torch.full((bsz, seqlen), True, dtype=torch.bool, device=device)
+    for i in range(4):
+        hidden = torch.rand((bsz, seqlen, hs), dtype=torch.float16, device=device)
+        if i:
+            mask[0, -i:] = False
+            mask[1, :i] = False
+
+        lmask = model._prepare_decoder_attention_mask(mask, hidden.shape[:2], hidden, 0)
+        ref, _, _ = attn.forward(
+            hidden, attention_mask=lmask, position_ids=position_ids
+        )
+
+        fast, _, _ = fastchat_forward(
+            attn, hidden, attention_mask=mask, position_ids=position_ids
+        )
+
+        lmask = _prepare_decoder_attention_mask(
+            model, mask, hidden.shape[:2], hidden, 0
+        )
+        test, _, _ = forward(
+            attn, hidden, attention_mask=lmask, position_ids=position_ids
+        )
+
+        print(f'Mean(abs(ref)) = {torch.mean(torch.abs(ref))}')
+        print(f'Mean(abs(ref - fast)) = {torch.mean(torch.abs(ref - fast))}')
+        print(f'Mean(abs(ref - test)) = {torch.mean(torch.abs(ref - test))}')
+        print(f'Mean(abs(fast - test)) = {torch.mean(torch.abs(fast - test))}')
+        print(f'allclose(fast, test) = {torch.allclose(fast, test)}')
+
+    with torch.no_grad():
+        # Also check that past_kv is handled properly
+        hidden = torch.rand((bsz, seqlen, hs), dtype=torch.float16, device=device)
+        part_len = seqlen // 4
+        assert part_len * 4 == seqlen
+        mask = torch.full((bsz, seqlen), True, dtype=torch.bool, device=device)
+        mask[0, -2:] = False
+        lmask = _prepare_decoder_attention_mask(
+            model, mask, hidden.shape[:2], hidden, 0
+        )
+        oneshot, _, _ = forward(
+            attn, hidden, attention_mask=lmask, position_ids=position_ids
+        )
+        parts = []
+        past_kv, past_kv_len = None, 0
+        for i in range(4):
+            start = part_len * i
+            end = start + part_len
+            hidden_part = hidden[:, start:end, ...]
+            lmask = _prepare_decoder_attention_mask(
+                model,
+                mask[:, start:end],
+                hidden_part.shape[:2],
+                hidden_part,
+                past_kv_len,
+            )
+            part, _, past_kv = forward(
+                attn,
+                hidden_part.clone(),
+                attention_mask=lmask,
+                position_ids=position_ids[:, start:end],
+                past_key_value=past_kv,
+                use_cache=True,
+            )
+            parts.append(part)
+            past_kv_len = past_kv[0].shape[2]
+
+        print(
+            f'allclose(oneshot[:, 0], parts[0]) = {torch.allclose(oneshot[:, :part_len], parts[0])}'
+        )
+        print(
+            f'allclose(oneshot, parts) = {torch.allclose(oneshot, torch.cat(parts, dim=1))}'
+        )
+
+
+if __name__ == '__main__':
+    test()

internvl/patch/llama_flash_attn_monkey_patch.py

+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import transformers
+from torch import nn
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
+
+
+def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = 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]]]:
+    """Input shape: Batch x Time x Channel
+
+    attention_mask: [bsz, q_len]
+    """
+    from einops import rearrange
+    try:  # v1
+        from flash_attn.flash_attn_interface import \
+            flash_attn_unpadded_qkvpacked_func
+    except:  # v2
+        from flash_attn.flash_attn_interface import \
+            flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
+    from flash_attn.bert_padding import pad_input, unpad_input
+
+    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)
+    )
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    assert past_key_value is None, 'past_key_value is not supported'
+
+    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]
+    assert not output_attentions, 'output_attentions is not supported'
+    assert not use_cache, 'use_cache is not supported'
+
+    # Flash attention codes from
+    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
+
+    # transform the data into the format required by flash attention
+    qkv = torch.stack(
+        [query_states, key_states, value_states], dim=2
+    )  # [bsz, nh, 3, q_len, hd]
+    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+    # We have disabled _prepare_decoder_attention_mask in LlamaModel
+    # the attention_mask should be the same as the key_padding_mask
+    key_padding_mask = attention_mask
+
+    if key_padding_mask is None:
+        qkv = rearrange(qkv, 'b s ... -> (b s) ...')
+        max_s = q_len
+        cu_q_lens = torch.arange(
+            0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
+        )
+        output = flash_attn_unpadded_qkvpacked_func(
+            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output = rearrange(output, '(b s) ... -> b s ...', b=bsz)
+    else:
+        nheads = qkv.shape[-2]
+        x = rearrange(qkv, 'b s three h d -> b s (three h d)')
+        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
+        x_unpad = rearrange(
+            x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads
+        )
+        output_unpad = flash_attn_unpadded_qkvpacked_func(
+            x_unpad, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output = rearrange(
+            pad_input(
+                rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices, bsz, q_len
+            ),
+            'b s (h d) -> b s h d',
+            h=nheads,
+        )
+    return self.o_proj(rearrange(output, 'b s h d -> b s (h d)')), None, None
+
+
+# Disable the transformation of the attention mask in LlamaModel as the flash attention
+# requires the attention mask to be the same as the key_padding_mask
+def _prepare_decoder_attention_mask(
+        self, attention_mask, input_shape, inputs_embeds, past_key_values_length
+):
+    # [bsz, seq_len]
+    return attention_mask
+
+
+def forward_2(
+        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
+    )
+
+    assert not output_attentions, 'output_attentions is not supported'
+    assert not use_cache, 'use_cache is not supported'
+    assert past_key_value is None, 'past_key_value is not supported'
+
+    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
+    if self.training:
+        attn_output = F.scaled_dot_product_attention(
+            query_states, key_states, value_states, dropout_p=0.0, is_causal=True
+        )
+        attn_weights = None
+    else:
+        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
+
+
+def replace_llama_attn_with_flash_attn():
+    if hasattr(F, 'scaled_dot_product_attention'):
+        transformers.models.llama.modeling_llama.LlamaAttention.forward = forward_2
+    else:
+        transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
+            _prepare_decoder_attention_mask
+        )
+        transformers.models.llama.modeling_llama.LlamaAttention.forward = forward

internvl/patch/llama_rmsnorm_monkey_patch.py

+import transformers
+
+
+def replace_llama_rmsnorm_with_fused_rmsnorm():
+    try:
+        from functools import partial
+
+        from apex.normalization import FusedRMSNorm
+        LlamaRMSNorm = partial(FusedRMSNorm, eps=1e-6)   # noqa
+        transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
+        print('Discovered apex.normalization.FusedRMSNorm - will use it instead of LlamaRMSNorm')
+    except ImportError:
+        # using the normal LlamaRMSNorm
+        pass
+    except Exception:
+        print('discovered apex but it failed to load, falling back to LlamaRMSNorm')
+        pass

internvl/patch/pad_data_collator.py

+import numpy as np
+import torch
+
+IGNORE_INDEX = -100
+
+
+def pad_data_collator(features, pad_id=0):
+
+    first = features[0]
+    batch = {}
+
+    batch_lens = [feat['input_ids'].shape for feat in features]
+    max_item_length = max(batch_lens)[0]
+    for idx in range(len(features)):
+        feat = features[idx]
+        temp_input_ids = torch.LongTensor([pad_id] * max_item_length)
+        temp_input_ids[:feat['input_ids'].shape[0]] = feat['input_ids']
+        feat['input_ids'] = temp_input_ids
+        temp_labels = torch.LongTensor([IGNORE_INDEX] * max_item_length)
+        temp_labels[:feat['labels'].shape[0]] = feat['labels']
+        feat['labels'] = temp_labels
+        feat['attention_mask'] = feat['input_ids'].ne(pad_id)
+
+    # Special handling for labels.
+    # Ensure that tensor is created with the correct type
+    # (it should be automatically the case, but let's make sure of it.)
+    if 'label' in first and first['label'] is not None:
+        label = first['label'].item() if isinstance(first['label'], torch.Tensor) else first['label']
+        dtype = torch.long if isinstance(label, int) else torch.float
+        batch['labels'] = torch.tensor([f['label'] for f in features], dtype=dtype)
+    elif 'label_ids' in first and first['label_ids'] is not None:
+        if isinstance(first['label_ids'], torch.Tensor):
+            batch['labels'] = torch.stack([f['label_ids'] for f in features])
+        else:
+            dtype = torch.long if isinstance(first['label_ids'][0], int) else torch.float
+            batch['labels'] = torch.tensor([f['label_ids'] for f in features], dtype=dtype)
+
+    # Handling of all other possible keys.
+    # Again, we will use the first element to figure out which key/values are not None for this model.
+    for k, v in first.items():
+        if k not in ('label', 'label_ids') and v is not None and not isinstance(v, str):
+            if isinstance(v, torch.Tensor):
+                batch[k] = torch.stack([f[k] for f in features])
+            elif isinstance(v, np.ndarray):
+                batch[k] = torch.tensor(np.stack([f[k] for f in features]))
+            else:
+                batch[k] = torch.tensor([f[k] for f in features])
+    return batch
+
+
+def concat_pad_data_collator(features, pad_id=0):
+
+    first = features[0]
+    batch = {}
+
+    batch_lens = [feat['input_ids'].shape for feat in features]
+    max_item_length = max(batch_lens)[0]
+    for idx in range(len(features)):
+        feat = features[idx]
+        temp_input_ids = torch.LongTensor([pad_id] * max_item_length)
+        temp_input_ids[:feat['input_ids'].shape[0]] = feat['input_ids']
+        feat['input_ids'] = temp_input_ids
+        temp_labels = torch.LongTensor([IGNORE_INDEX] * max_item_length)
+        temp_labels[:feat['labels'].shape[0]] = feat['labels']
+        feat['labels'] = temp_labels
+        feat['attention_mask'] = feat['input_ids'].ne(pad_id)
+
+    # Special handling for labels.
+    # Ensure that tensor is created with the correct type
+    # (it should be automatically the case, but let's make sure of it.)
+    if 'label' in first and first['label'] is not None:
+        label = first['label'].item() if isinstance(first['label'], torch.Tensor) else first['label']
+        dtype = torch.long if isinstance(label, int) else torch.float
+        batch['labels'] = torch.tensor([f['label'] for f in features], dtype=dtype)
+    elif 'label_ids' in first and first['label_ids'] is not None:
+        if isinstance(first['label_ids'], torch.Tensor):
+            batch['labels'] = torch.stack([f['label_ids'] for f in features])
+        else:
+            dtype = torch.long if isinstance(first['label_ids'][0], int) else torch.float
+            batch['labels'] = torch.tensor([f['label_ids'] for f in features], dtype=dtype)
+
+    # Handling of all other possible keys.
+    # Again, we will use the first element to figure out which key/values are not None for this model.
+    for k, v in first.items():
+        if k not in ('label', 'label_ids', 'pixel_values', 'image_flags') and \
+                v is not None and not isinstance(v, str):
+            if isinstance(v, torch.Tensor):
+                batch[k] = torch.stack([f[k] for f in features])
+            elif isinstance(v, np.ndarray):
+                batch[k] = torch.tensor(np.stack([f[k] for f in features]))
+            else:
+                batch[k] = torch.tensor([f[k] for f in features])
+        if k in ('pixel_values', 'image_flags'):
+            if isinstance(v, torch.Tensor):
+                batch[k] = torch.concat([f[k] for f in features])
+            elif isinstance(v, np.ndarray):
+                batch[k] = torch.concat(np.stack([f[k] for f in features]))
+            else:
+                batch[k] = torch.concat([f[k] for f in features])
+    return batch

internvl/patch/train_sampler_patch.py

+from typing import List, Optional
+
+import torch
+import transformers
+from torch.utils.data import Dataset, Sampler
+from transformers.tokenization_utils_base import BatchEncoding
+from transformers.trainer import (LengthGroupedSampler, RandomSampler,
+                                  has_length)
+from transformers.trainer_pt_utils import logger
+
+
+# copy from https://github.com/haotian-liu/LLaVA/blob/main/llava/train/llava_trainer.py#L38
+def split_to_even_chunks(indices, lengths, num_chunks):
+    """
+    Split a list of indices into `chunks` chunks of roughly equal lengths.
+    """
+
+    if len(indices) % num_chunks != 0:
+        return [indices[i::num_chunks] for i in range(num_chunks)]
+
+    num_indices_per_chunk = len(indices) // num_chunks
+
+    chunks = [[] for _ in range(num_chunks)]
+    chunks_lengths = [0 for _ in range(num_chunks)]
+    for index in indices:
+        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
+        chunks[shortest_chunk].append(index)
+        chunks_lengths[shortest_chunk] += lengths[index]
+        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
+            chunks_lengths[shortest_chunk] = float('inf')
+
+    return chunks
+
+
+# copy from https://github.com/haotian-liu/LLaVA/blob/main/llava/train/llava_trainer.py#L88
+def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    indices = torch.randperm(len(lengths), generator=generator)
+    megabatch_size = world_size * batch_size
+    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
+    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
+    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]
+
+    return [i for megabatch in megabatches for batch in megabatch for i in batch]
+
+
+# modified from https://github.com/haotian-liu/LLaVA/blob/main/llava/train/llava_trainer.py#L99
+class LengthGroupedSampler(Sampler):
+    r"""
+    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
+    keeping a bit of randomness.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        world_size: int,
+        dataset: Optional[Dataset] = None,
+        lengths: Optional[List[int]] = None,
+        model_input_name: Optional[str] = None,
+        generator=None,
+    ):
+        if dataset is None and lengths is None:
+            raise ValueError('One of dataset and lengths must be provided.')
+
+        self.batch_size = batch_size
+        if lengths is None:
+            model_input_name = model_input_name if model_input_name is not None else 'input_ids'
+            if (
+                    not (isinstance(dataset[0], dict) or isinstance(dataset[0], BatchEncoding))
+                    or model_input_name not in dataset[0]
+            ):
+                raise ValueError(
+                    'Can only automatically infer lengths for datasets whose items are dictionaries with an '
+                    f"'{model_input_name}' key."
+                )
+            lengths = [len(feature[model_input_name]) for feature in dataset]
+        elif isinstance(lengths, torch.Tensor):
+            logger.info(
+                'If lengths is a torch.Tensor, LengthGroupedSampler will be slow. Converting lengths to List[int]...'
+            )
+            lengths = lengths.tolist()
+        self.world_size = world_size
+        self.lengths = lengths
+        self.generator = generator
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def __iter__(self):
+        indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        return iter(indices)
+
+
+# patch trainer
+def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+    if self.train_dataset is None or not has_length(self.train_dataset):
+        return None
+    # Build the sampler.
+    if self.args.group_by_length:
+        lengths = []
+        for dataset in self.train_dataset.datasets:
+            lengths = lengths + dataset.length
+        model_input_name = self.tokenizer.model_input_names[0] if self.tokenizer is not None else None
+        return LengthGroupedSampler(
+            self.args.train_batch_size,
+            world_size=self.args.world_size * self.args.gradient_accumulation_steps,
+            # self.args.train_batch_size * self.args.gradient_accumulation_steps,
+            dataset=self.train_dataset,
+            lengths=lengths,
+            model_input_name=model_input_name,
+        )
+    else:
+        return RandomSampler(self.train_dataset)
+
+
+def replace_train_sampler():
+    transformers.Trainer._get_train_sampler = _get_train_sampler
+    print('Replace train sampler!!')

internvl/train/constants.py

+IMG_CONTEXT_TOKEN = '<IMG_CONTEXT>'
+IMG_START_TOKEN = '<img>'
+IMG_END_TOKEN = '</img>'
+QUAD_START_TOKEN = '<quad>'
+QUAD_END_TOKEN = '</quad>'
+REF_START_TOKEN = '<ref>'
+REF_END_TOKEN = '</ref>'
+BOX_START_TOKEN = '<box>'
+BOX_END_TOKEN = '</box>'
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+CLIP_MEAN = (0.4814546, 0.4578275, 0.40821073)
+CLIP_STD = (0.2686295, 0.2613025, 0.2757711)
+SIGLIP_MEAN = (0.5, 0.5, 0.5)
+SIGLIP_STD = (0.5, 0.5, 0.5)

internvl/train/dataset.py

+import io
+
+from transformers.trainer_pt_utils import LabelSmoother
+
+IGNORE_TOKEN_ID = LabelSmoother.ignore_index
+import os
+import random
+from typing import Dict
+
+import cv2
+import imageio
+import numpy as np
+import torch
+import torchvision.transforms as T
+import transformers
+from decord import VideoReader
+from internvl.conversation import get_conv_template
+from PIL import Image
+from torch.utils.data import ConcatDataset, WeightedRandomSampler
+from torchvision.transforms.functional import InterpolationMode
+
+from .constants import (CLIP_MEAN, CLIP_STD, IMAGENET_MEAN, IMAGENET_STD,
+                        IMG_CONTEXT_TOKEN, IMG_END_TOKEN, IMG_START_TOKEN,
+                        SIGLIP_MEAN, SIGLIP_STD)
+
+try:
+    from petrel_client.client import Client
+    from petrel_client.common.config import Config
+except ImportError as E:
+    print('petrel_client is not installed. If you read data locally instead of from ceph, ignore it.')
+import sys
+
+
+def get_frame_indices(num_frames, vlen, sample='rand', fix_start=None, input_fps=1, max_num_frames=-1):
+    if sample in ['rand', 'middle']: # uniform sampling
+        acc_samples = min(num_frames, vlen)
+        # split the video into `acc_samples` intervals, and sample from each interval.
+        intervals = np.linspace(start=0, stop=vlen, num=acc_samples + 1).astype(int)
+        ranges = []
+        for idx, interv in enumerate(intervals[:-1]):
+            ranges.append((interv, intervals[idx + 1] - 1))
+        if sample == 'rand':
+            try:
+                frame_indices = [random.choice(range(x[0], x[1])) for x in ranges]
+            except:
+                frame_indices = np.random.permutation(vlen)[:acc_samples]
+                frame_indices.sort()
+                frame_indices = list(frame_indices)
+        elif fix_start is not None:
+            frame_indices = [x[0] + fix_start for x in ranges]
+        elif sample == 'middle':
+            frame_indices = [(x[0] + x[1]) // 2 for x in ranges]
+        else:
+            raise NotImplementedError
+
+        if len(frame_indices) < num_frames:  # padded with last frame
+            padded_frame_indices = [frame_indices[-1]] * num_frames
+            padded_frame_indices[:len(frame_indices)] = frame_indices
+            frame_indices = padded_frame_indices
+    elif 'fps' in sample:  # fps0.5, sequentially sample frames at 0.5 fps
+        output_fps = float(sample[3:])
+        duration = float(vlen) / input_fps
+        delta = 1 / output_fps  # gap between frames, this is also the clip length each frame represents
+        frame_seconds = np.arange(0 + delta / 2, duration + delta / 2, delta)
+        frame_indices = np.around(frame_seconds * input_fps).astype(int)
+        frame_indices = [e for e in frame_indices if e < vlen]
+        if max_num_frames > 0 and len(frame_indices) > max_num_frames:
+            frame_indices = frame_indices[:max_num_frames]
+            # frame_indices = np.linspace(0 + delta / 2, duration + delta / 2, endpoint=False, num=max_num_frames)
+    else:
+        raise ValueError
+    return frame_indices
+
+
+def read_frames_gif(
+        video_path, num_frames, sample='rand', fix_start=None,
+        client=None, min_num_frames=4
+):
+    if 's3://' in video_path:
+        video_bytes = client.get(video_path)
+        gif = imageio.get_reader(io.BytesIO(video_bytes))
+    else:
+        gif = imageio.get_reader(video_path)
+    vlen = len(gif)
+
+    t_num_frames = np.random.randint(min_num_frames, num_frames + 1)
+    frame_indices = get_frame_indices(
+        t_num_frames, vlen, sample=sample, fix_start=fix_start
+    )
+    frames = []
+    for index, frame in enumerate(gif):
+        if index in frame_indices:
+            frame = cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB).astype(np.uint8)
+            frame = Image.fromarray(frame)
+            frames.append(frame)
+    return frames
+
+
+def read_frames_decord(
+        video_path, num_frames, sample='rand', fix_start=None,
+        client=None, clip=None, min_num_frames=4
+):
+    if 's3://' in video_path:
+        video_bytes = client.get(video_path)
+        video_reader = VideoReader(io.BytesIO(video_bytes), num_threads=1)
+    else:
+        video_reader = VideoReader(video_path, num_threads=1)
+    vlen = len(video_reader)
+    fps = video_reader.get_avg_fps()
+    duration = vlen / float(fps)
+    if clip:
+        start, end = clip
+        duration = end - start
+        vlen = int(duration * fps)
+        start_index = int(start * fps)
+
+    # t_num_frames = min(max(int(duration * sample_fps), min_num_frames), num_frames)
+    t_num_frames = np.random.randint(min_num_frames, num_frames + 1)
+
+    frame_indices = get_frame_indices(
+        t_num_frames, vlen, sample=sample, fix_start=fix_start,
+        input_fps=fps
+    )
+    if clip:
+        frame_indices = [f + start_index for f in frame_indices]
+    frames = video_reader.get_batch(frame_indices).asnumpy()  # (T, H, W, C), np.uint8
+    frames = [Image.fromarray(frames[i]) for i in range(frames.shape[0])]
+    return frames
+
+
+def read_frames_folder(
+        video_path, num_frames, sample='rand', fix_start=None,
+        client=None, clip=None, min_num_frames=4
+):
+    if 's3://' in video_path:
+        image_list = client.list(video_path)
+        frames = []
+        for image in image_list:
+            fp = os.path.join(video_path, image)
+            frame = Image.open(io.BytesIO(client.get(fp)))
+            frames.append(frame)
+    else:
+        image_list = sorted(list(os.listdir(video_path)))
+        frames = []
+        for image in image_list:
+            fp = os.path.join(video_path, image)
+            frame = Image.open(fp).convert('RGB')
+            frames.append(frame)
+    vlen = len(frames)
+
+    t_num_frames = np.random.randint(min_num_frames, num_frames + 1)
+
+    if vlen > t_num_frames:
+        frame_indices = get_frame_indices(
+            t_num_frames, vlen, sample=sample, fix_start=fix_start
+        )
+        frames = [frames[i] for i in frame_indices]
+    return frames
+
+
+class WeightedConcatDataset(ConcatDataset):
+    def __init__(self, datasets, weights):
+        super().__init__(datasets)
+        self.weights = torch.DoubleTensor(weights)
+        self.total_size = sum(len(d) for d in datasets)
+        self.sampler = WeightedRandomSampler(weights=self.weights, num_samples=self.total_size, replacement=True)
+
+    def __iter__(self):
+        return iter(self.sampler)
+
+    def __len__(self):
+        return self.total_size
+
+
+def pil_loader(img_str):
+    buff = io.BytesIO(img_str)
+    img = Image.open(buff)
+    return img.convert('RGB')
+
+
+class TCSLoader(object):
+
+    def __init__(self, conf_path, sc_config_key='sensecore'):
+        print(f'[TCSLoader] config_path: {conf_path}')
+        print('--> before Client(conf_path)')
+        self.client = Client(conf_path)
+        self.sc_config_key = sc_config_key
+        print('--> after Client(conf_path)')
+
+    def __call__(self, fn, image_type='image', max_num_frames=-1, min_num_frames=4, sample='rand', clip=None):
+        if image_type == 'image':
+            img_value_str = self.client.get(fn)
+            img = pil_loader(img_value_str)
+            return img
+
+        elif image_type == 'video':
+            if fn.endswith('/'):
+                frames = read_frames_folder(fn, num_frames=max_num_frames, min_num_frames=min_num_frames,
+                                            client=self.client, sample=sample)
+            elif fn.endswith('.gif'):
+                frames = read_frames_gif(fn, num_frames=max_num_frames, min_num_frames=min_num_frames,
+                                         client=self.client, sample=sample)
+            else:
+                frames = read_frames_decord(fn, num_frames=max_num_frames, min_num_frames=min_num_frames,
+                                            client=self.client, sample=sample, clip=clip)
+            return frames
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+def simulate_jpeg_degradation(quality):
+    def jpeg_degrade(img):
+        with io.BytesIO() as output:
+            img.convert('RGB').save(output, format='JPEG', quality=quality)
+            output.seek(0)  # Move the reading cursor to the start of the stream
+            img_jpeg = Image.open(output).copy()  # Use .copy() to make sure the image is loaded in memory
+        return img_jpeg
+    return jpeg_degrade
+
+
+# Define the JPEG compression quality range, pre-create all JPEG compression functions
+qualities = list(range(75, 101))
+jpeg_degrade_functions = {quality: simulate_jpeg_degradation(quality) for quality in qualities}
+
+
+def build_transform(is_train, input_size, pad2square=False, normalize_type='imagenet'):
+    if normalize_type == 'imagenet':
+        MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
+    elif normalize_type == 'clip':
+        MEAN, STD = CLIP_MEAN, CLIP_STD
+    elif normalize_type == 'siglip':
+        MEAN, STD = SIGLIP_MEAN, SIGLIP_STD
+    else:
+        raise NotImplementedError
+    if is_train:  # use data augumentation
+        transform = T.Compose([
+            T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+            T.RandomChoice([T.Lambda(jpeg_degrade_functions[quality]) for quality in qualities]),
+            T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=MEAN, std=STD)
+        ])
+    else:
+        if pad2square is False:  # now we use this transform function by default
+            transform = T.Compose([
+                T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+                T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+                T.ToTensor(),
+                T.Normalize(mean=MEAN, std=STD)
+            ])
+        else:
+            transform = T.Compose([
+                T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+                T.Lambda(lambda img: expand2square(img, tuple(int(x * 255) for x in MEAN))),
+                T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+                T.ToTensor(),
+                T.Normalize(mean=MEAN, std=STD)
+            ])
+
+    return transform
+
+
+def preprocess(
+        template_name,
+        sources,
+        tokenizer: transformers.PreTrainedTokenizer,
+        num_image_token_list: list,
+        text_only: bool = False,
+        group_by_length: bool = False,
+        use_packed_ds: bool = False,
+        ds_name: str = None,
+        num_image: int = 1
+) -> Dict:
+    conv = get_conv_template(template_name)
+    roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]['from']] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence['from']]
+            assert role == conv.roles[j % 2], f'{i}'
+            conv.append_message(role, sentence['value'])
+        conversations.append(conv.get_prompt())
+
+    if not text_only:
+        new_conversations = []
+        for conversation in conversations:
+            for i in range(num_image):
+                image_tokens = f'{IMG_START_TOKEN}{IMG_CONTEXT_TOKEN * num_image_token_list[i]}{IMG_END_TOKEN}'
+                conversation = conversation.replace('<image>', image_tokens, 1)
+            new_conversations.append(conversation)
+        conversations = new_conversations
+
+    # Tokenize conversations
+    input_ids = tokenizer(
+        conversations,
+        return_tensors='pt',
+        padding=False if group_by_length or use_packed_ds else 'max_length',
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+    ).input_ids
+    targets = input_ids.clone()
+
+    # assert conv.sep_style == SeparatorStyle.ADD_COLON_TWO
+
+    # Mask targets. Only compute loss on the assistant outputs.
+    sep = conv.sep + conv.roles[1] + ': '
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        turns = conversation.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_TOKEN_ID
+        for i, turn in enumerate(turns):
+            if turn == '':
+                break
+            turn_len = len(tokenizer(turn).input_ids)
+
+            parts = turn.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+            # "-2" is hardcoded for the Llama tokenizer to make the offset correct.
+            instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            if i != 0 and not tokenizer.legacy:
+                # The legacy and non-legacy modes handle special tokens differently
+                instruction_len -= 1
+
+            # Ignore the user instructions
+            target[cur_len: cur_len + instruction_len] = IGNORE_TOKEN_ID
+            cur_len += turn_len
+
+            if i != 0 and not tokenizer.legacy:
+                # The legacy and non-legacy modes handle special tokens differently
+                cur_len -= 1
+
+        target[cur_len:] = IGNORE_TOKEN_ID
+
+        if False:  # Inspect and check the correctness of masking
+            z = target.clone()
+            z = torch.where(z == IGNORE_TOKEN_ID, tokenizer.unk_token_id, z)
+            logger.info(tokenizer.decode(z))
+            exit()
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_TOKEN_ID
+                print(
+                    f'WARNING: tokenization mismatch: {cur_len} vs. {total_len}.'
+                    f' #turn = {len(turns) - 1}. (ignored). This dataset is {ds_name}.'
+                )
+                sys.stdout.flush()
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+        attention_mask=input_ids.ne(tokenizer.pad_token_id),
+    )
+
+
+def preprocess_mpt(
+        template_name,
+        sources,
+        tokenizer: transformers.PreTrainedTokenizer,
+        num_image_token_list: list,
+        text_only: bool = False,
+        group_by_length: bool = False,
+        use_packed_ds: bool = False,
+        ds_name: str = None,
+        num_image: int = 1
+) -> Dict:
+    conv = get_conv_template(template_name)
+    roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]['from']] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence['from']]
+            assert role == conv.roles[j % 2], f'{i}'
+            conv.append_message(role, sentence['value'])
+        conversations.append(conv.get_prompt())
+
+    if not text_only:
+        new_conversations = []
+        for conversation in conversations:
+            for i in range(num_image):
+                image_tokens = f'{IMG_START_TOKEN}{IMG_CONTEXT_TOKEN * num_image_token_list[i]}{IMG_END_TOKEN}'
+                conversation = conversation.replace('<image>', image_tokens, 1)
+            new_conversations.append(conversation)
+        conversations = new_conversations
+
+    # Tokenize conversations
+    input_ids = tokenizer(
+        conversations,
+        return_tensors='pt',
+        padding=False if group_by_length or use_packed_ds else 'max_length',
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+    ).input_ids
+    targets = input_ids.clone()
+
+    # Mask targets. Only compute loss on the assistant outputs.
+    sep = conv.sep + conv.roles[1]  # <|im_end|><|im_start|>assistant\n
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        turns = conversation.split(conv.sep)
+        re_turns = [conv.sep.join(turns[:3])]  # system + user + gpt
+        for conv_idx in range(3, len(turns), 2):
+            re_turns.append(conv.sep.join(turns[conv_idx:conv_idx + 2]))  # user + gpt
+        cur_len = 0
+        target[:cur_len] = IGNORE_TOKEN_ID
+        for i, turn in enumerate(re_turns):
+            if turn == '':
+                break
+            turn_len = len(tokenizer(turn).input_ids) + 1
+
+            parts = turn.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+            instruction_len = len(tokenizer(parts[0]).input_ids)
+
+            # Ignore the user instructions
+            target[cur_len: cur_len + instruction_len] = IGNORE_TOKEN_ID
+            # print(f'[question {i}]', tokenizer.decode(input_ids[:, cur_len: cur_len + instruction_len][0]))
+            # print(f'[answer {i}]', tokenizer.decode(input_ids[:, cur_len + instruction_len: cur_len + turn_len][0]))
+            # print(f'[label {i}]', target[cur_len + instruction_len: cur_len + turn_len])
+            cur_len += turn_len
+
+        target[cur_len:] = IGNORE_TOKEN_ID
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_TOKEN_ID
+                print(
+                    f'WARNING: tokenization mismatch: {cur_len} vs. {total_len}.'
+                    f' #turn = {len(turns) - 1}. (ignored). This dataset is {ds_name}.'
+                )
+                sys.stdout.flush()
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+        attention_mask=input_ids.ne(tokenizer.pad_token_id),
+    )
+
+
+def preprocess_phi3(
+        template_name,
+        sources,
+        tokenizer: transformers.PreTrainedTokenizer,
+        num_image_token_list: list,
+        text_only: bool = False,
+        group_by_length: bool = False,
+        use_packed_ds: bool = False,
+        ds_name: str = None,
+        num_image: int = 1
+) -> Dict:
+    conv = get_conv_template(template_name)
+    roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]['from']] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence['from']]
+            assert role == conv.roles[j % 2], f'{i}'
+            conv.append_message(role, sentence['value'])
+        conversations.append(conv.get_prompt())
+
+    if not text_only:
+        new_conversations = []
+        for conversation in conversations:
+            for i in range(num_image):
+                image_tokens = f'{IMG_START_TOKEN}{IMG_CONTEXT_TOKEN * num_image_token_list[i]}{IMG_END_TOKEN}'
+                conversation = conversation.replace('<image>', image_tokens, 1)
+            new_conversations.append(conversation)
+        conversations = new_conversations
+
+    # Tokenize conversations
+    tokenizer.padding_side = 'right'
+    input_ids = tokenizer(
+        conversations,
+        return_tensors='pt',
+        padding=False if group_by_length or use_packed_ds else 'max_length',
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+    ).input_ids
+    targets = input_ids.clone()
+
+    # Mask targets. Only compute loss on the assistant outputs.
+    sep = conv.sep + conv.roles[1]  # <|end|>\n<|assistant|>
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(int(tokenizer.pad_token_id)).sum())
+
+        turns = conversation.split(conv.sep)
+        re_turns = [conv.sep.join(turns[:3])]  # system + user + gpt
+        for conv_idx in range(3, len(turns), 2):
+            re_turns.append(conv.sep.join(turns[conv_idx:conv_idx + 2]))  # user + gpt
+        cur_len = 1
+        target[:cur_len] = IGNORE_TOKEN_ID
+        endoftext_id = tokenizer.convert_tokens_to_ids('<|endoftext|>')
+        target[target == endoftext_id] = IGNORE_TOKEN_ID
+
+        for i, turn in enumerate(re_turns):
+            if turn == '':
+                break
+            if i == 0:
+                turn_len = len(tokenizer(turn).input_ids)
+            else:
+                turn_len = len(tokenizer(turn).input_ids) - 1
+            parts = turn.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if i == 0:
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 1
+            else:
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            # Ignore the user instructions
+            target[cur_len: cur_len + instruction_len] = IGNORE_TOKEN_ID
+            # print(f'[question {i}]', tokenizer.decode(input_ids[:, cur_len: cur_len + instruction_len][0]))
+            # print(f'[answer {i}]', tokenizer.decode(input_ids[:, cur_len + instruction_len: cur_len + turn_len][0]))
+            # print(f'[label {i}]', target[cur_len + instruction_len: cur_len + turn_len])
+            cur_len += turn_len
+
+        target[cur_len:] = IGNORE_TOKEN_ID
+
+        if False:  # Inspect and check the correctness of masking
+            z = target.clone()
+            z = torch.where(z == IGNORE_TOKEN_ID, tokenizer.unk_token_id, z)
+            print(repr(tokenizer.decode(z)))
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_TOKEN_ID
+                print(
+                    f'WARNING: tokenization mismatch: {cur_len} vs. {total_len}.'
+                    f' #turn = {len(turns) - 1}. (ignored). This dataset is {ds_name}.'
+                )
+                sys.stdout.flush()
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+        attention_mask=input_ids.ne(tokenizer.pad_token_id),
+    )
+
+
+def preprocess_internlm(
+        template_name,
+        sources,
+        tokenizer: transformers.PreTrainedTokenizer,
+        num_image_token_list: list,
+        text_only: bool = False,
+        group_by_length: bool = False,
+        use_packed_ds: bool = False,
+        ds_name: str = None,
+        num_image: int = 1
+) -> Dict:
+    conv = get_conv_template(template_name)
+    roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]['from']] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence['from']]
+            assert role == conv.roles[j % 2], f'{i}'
+            sentence['value'] = sentence['value'].strip()
+            conv.append_message(role, sentence['value'])
+        conversations.append(conv.get_prompt())
+
+    if not text_only:
+        new_conversations = []
+        for conversation in conversations:
+            for i in range(num_image):
+                image_tokens = f'{IMG_START_TOKEN}{IMG_CONTEXT_TOKEN * num_image_token_list[i]}{IMG_END_TOKEN}'
+                conversation = conversation.replace('<image>', image_tokens, 1)
+            new_conversations.append(conversation)
+        conversations = new_conversations
+
+    # Tokenize conversations
+    input_ids = tokenizer(
+        conversations,
+        return_tensors='pt',
+        padding=False if group_by_length or use_packed_ds else 'max_length',
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+    ).input_ids
+    targets = input_ids.clone()
+
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())  # 浦语里面 pad_token_id = eos_token_id
+        cur_len = 1
+        target[:cur_len] = IGNORE_TOKEN_ID  # <s>
+        parts = conversation.split(conv.roles[1])  # [UNUSED_TOKEN_146]assistant\n
+        info = parts[0] + conv.roles[1]
+        temp_len = len(tokenizer(info).input_ids) - 1  # 去除tokenizer的<s>
+        target[cur_len: cur_len + temp_len] = IGNORE_TOKEN_ID
+        cur_len = cur_len + temp_len
+
+        for index in range(1, len(parts) - 1):
+            info = parts[index]
+            part1, part2 = info.split(conv.roles[0])
+            temp_len = len(tokenizer(part1).input_ids) - 1
+            cur_len = cur_len + temp_len
+            part = conv.roles[0] + part2 + conv.roles[1]
+            temp_len = len(tokenizer(part).input_ids) - 1
+            target[cur_len: cur_len + temp_len] = IGNORE_TOKEN_ID
+            cur_len = cur_len + temp_len
+        last_info = parts[-1]
+        temp_len = len(tokenizer(last_info).input_ids) - 1
+        cur_len = cur_len + temp_len
+
+        target[cur_len:] = IGNORE_TOKEN_ID
+        if False:  # Inspect and check the correctness of masking
+            z = target.clone()
+            z = torch.where(z == IGNORE_TOKEN_ID, tokenizer.unk_token_id, z)
+            print(repr(tokenizer.decode(z)))
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_TOKEN_ID
+                print(f'WARNING: tokenization mismatch: {cur_len} vs. {total_len}. This dataset is {ds_name}.')
+                sys.stdout.flush()
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+        attention_mask=input_ids.ne(tokenizer.pad_token_id),
+    )
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    # print(f'width: {width}, height: {height}, best_ratio: {best_ratio}')
+    return best_ratio
+
+
+def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images

internvl/train/internvl_chat_finetune.py

+import gc
+import json
+import logging
+import math
+import os
+import random
+import sys
+import traceback
+import warnings
+from copy import deepcopy
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import transformers
+from internvl.dist_utils import init_dist
+from internvl.model.internlm2.modeling_internlm2 import InternLM2ForCausalLM
+from internvl.model.internvl_chat import (InternVisionConfig,
+                                          InternVisionModel,
+                                          InternVLChatConfig,
+                                          InternVLChatModel)
+from internvl.patch import (concat_pad_data_collator,
+                            replace_llama_rmsnorm_with_fused_rmsnorm,
+                            replace_train_sampler)
+from internvl.train.constants import (BOX_END_TOKEN, BOX_START_TOKEN,
+                                      IMG_CONTEXT_TOKEN, IMG_END_TOKEN,
+                                      IMG_START_TOKEN, QUAD_END_TOKEN,
+                                      QUAD_START_TOKEN, REF_END_TOKEN,
+                                      REF_START_TOKEN)
+from internvl.train.dataset import (ConcatDataset, TCSLoader,
+                                    WeightedConcatDataset, build_transform,
+                                    dynamic_preprocess, preprocess,
+                                    preprocess_internlm, preprocess_mpt,
+                                    preprocess_phi3)
+from internvl.train.trainer_monkey_patch import replace_create_optimizer
+from PIL import Image, ImageFile, PngImagePlugin, UnidentifiedImageError
+from torch.utils.data import Dataset
+from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
+                          HfArgumentParser, Trainer, TrainingArguments,
+                          set_seed)
+from transformers.trainer_utils import get_last_checkpoint
+from transformers.utils.logging import (enable_default_handler,
+                                        enable_explicit_format, set_verbosity)
+
+# Apply necessary patches for the transformers library
+replace_llama_rmsnorm_with_fused_rmsnorm()
+replace_train_sampler()
+
+# Try to import petrel_client for image loading, fallback to PIL if unavailable
+try:
+    from petrel_client.client import Client
+    from petrel_client.common.config import Config
+    has_tcs_loader = True
+except ImportError as E:
+    print('petrel_client is not installed. Using PIL to load images.')
+    has_tcs_loader = False
+
+# Set constants for image processing and logging
+IGNORE_INDEX = -100
+Image.MAX_IMAGE_PIXELS = None
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+MaximumDecompressedSize = 1024
+MegaByte = 2 ** 20
+PngImagePlugin.MAX_TEXT_CHUNK = MaximumDecompressedSize * MegaByte
+
+warnings.filterwarnings('ignore')
+logger = logging.getLogger(__name__)
+
+os.environ['TOKENIZERS_PARALLELISM'] = 'true'
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments for specifying model, tokenizer, and configurations.
+    """
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    vision_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    llm_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    mlp_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    freeze_llm: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the LLM decoder.'},
+    )
+    freeze_backbone: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the vision backbone of the model.'},
+    )
+    freeze_mlp: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the MLP layers of the model.'},
+    )
+    unfreeze_vit_layers: int = field(
+        default=0,
+        metadata={'help': 'Specify the number of ViT layers to unfreeze. Default is 0.'},
+    )
+    vision_select_layer: int = field(
+        default=-1,
+        metadata={'help': 'Specify the layer of ViT feature map to use. Default is last layer.'},
+    )
+    use_backbone_lora: int = field(
+        default=0,
+        metadata={'help': 'Set the LoRA adapter rank for the backbone model. Default is 0.'}
+    )
+    use_llm_lora: int = field(
+        default=0,
+        metadata={'help': 'Set the LoRA adapter rank for the LLM. Default is 0.'}
+    )
+    unfreeze_lm_head: bool = field(
+        default=False,
+        metadata={'help': "Set to True to unfreeze the language model's head."},
+    )
+    use_custom_trainer: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to enable the use of a custom trainer.'},
+    )
+    grad_checkpoint: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use gradient checkpointing.'},
+    )
+    drop_path_rate: float = field(
+        default=0.0,
+        metadata={'help': 'Set the drop path rate for the ViT model. Default is 0.'},
+    )
+    ps_version: str = field(
+        default='v2',
+        metadata={'help': 'Specify the version of pixel shuffle implementation. Default is `v1`.'
+                          'Please use `v2` to fix the bug of transposed image.'}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments for specifying data input for training and evaluation.
+    """
+    max_seq_length: Optional[int] = field(
+        default=2048,
+        metadata={
+            'help': (
+                'The maximum total input sequence length after tokenization. Sequences longer '
+                'than this will be truncated, sequences shorter will be padded.'
+            )
+        },
+    )
+    force_image_size: Optional[int] = field(
+        default=448,
+        metadata={'help': 'Set the desired size for the image. Default is 224.'},
+    )
+    down_sample_ratio: Optional[float] = field(
+        default=0.5,
+        metadata={'help': 'Set the desired down-sampling ratio for the image. Default is 1.0.'},
+    )
+    pad2square: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Pad the image to a square shape if set to True.'},
+    )
+    conv_style: Optional[str] = field(
+        default='internlm2-chat', metadata={'help': 'Prompt style for a conversation.'}
+    )
+    meta_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'The path of the meta file of datasets.'},
+    )
+    use_data_resampling: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use data resampling.'},
+    )
+    dynamic_image_size: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use dynamic image size.'},
+    )
+    use_thumbnail: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to add a thumbnail image.'},
+    )
+    min_dynamic_patch: Optional[int] = field(
+        default=1,
+        metadata={'help': 'The minimum number of dynamic patches. Default is 1.'},
+    )
+    max_dynamic_patch: Optional[int] = field(
+        default=12,
+        metadata={'help': 'The maximum number of dynamic patches. Default is 6.'},
+    )
+    normalize_type: Optional[str] = field(
+        default='imagenet',
+        metadata={'help': 'The normalize type for the image. Default is imagenet.'},
+    )
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(
+        self,
+        template_name,
+        meta,
+        tokenizer,
+        tcs_loader,
+        ds_name,
+        num_image_token,
+        image_size=224,
+        is_train=True,
+        pad2square=False,
+        group_by_length=False,
+        dynamic_image_size=False,
+        use_thumbnail=False,
+        min_dynamic_patch=1,
+        max_dynamic_patch=6,
+        min_num_frame=4,  # for video data
+        max_num_frame=12,  # for video data
+        sampling_method='rand',  # for video data
+        repeat_time=1,
+        normalize_type='imagenet',
+        random_seed=0,
+    ):
+        super(LazySupervisedDataset, self).__init__()
+        self.ds_name = ds_name
+        self.tokenizer = tokenizer
+        self.template_name = template_name
+        self.num_image_token = num_image_token
+        logger.info(f'[Dataset] num_image_token: {num_image_token}')
+        logger.info(f'[Dataset] dynamic_image_size: {dynamic_image_size}')
+        logger.info(f'[Dataset] use_thumbnail: {use_thumbnail}')
+        logger.info(f'[Dataset] min_dynamic_patch: {min_dynamic_patch}, max_dynamic_patch: {max_dynamic_patch}')
+
+        self.image_size = image_size
+        self.is_train = is_train
+        self.pad2square = pad2square
+        self.max_num_frame = max_num_frame
+        self.min_num_frame = min_num_frame
+        self.sampling_method = sampling_method
+
+        logger.info('Formatting inputs...Skip in lazy mode')
+        assert meta['annotation'].endswith('jsonl'), f'annotation must be jsonl, but got {meta["annotation"]}'
+
+        with open(meta['annotation'], 'r') as f:
+            self.raw_data = f.readlines()
+            if repeat_time < 1:
+                # If repeat_time is less than 1, select a portion of the data
+                self.raw_data = self.raw_data[:int(len(self.raw_data) * repeat_time)]
+            if repeat_time > 1:
+                assert isinstance(repeat_time, int)
+                # Repeat the list if repeat_time is greater than 1
+                self.raw_data = self.raw_data * repeat_time
+
+        self.rng = np.random.default_rng(seed=random_seed)
+        self.rng.shuffle(self.raw_data)
+
+        gc.collect()
+        self.root = meta['root']
+        self.cached_data_dict = {}
+        self.tcs_loader = tcs_loader
+        self.group_by_length = group_by_length
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.min_dynamic_patch = min_dynamic_patch
+        self.max_dynamic_patch = max_dynamic_patch
+        self.normalize_type = normalize_type
+
+        # If the precomputed length does not exist, roughly estimate the length of
+        # each sample to improve the efficiency of group_by_length.
+        if self.group_by_length:
+            self.conv2length = {}  # Using a dictionary to speed up token length calculation
+            self.length = []
+            for data_item in self.raw_data:
+                data_item = json.loads(data_item)
+                if 'length' in data_item:
+                    token_length = data_item['length']  # Use precomputed length if available
+                else:
+                    # Compute token length using the tokenizer
+                    conversations = '\n'.join([temp['value'] for temp in data_item['conversations']])
+                    str_length = len(conversations)
+                    if str_length not in self.conv2length:
+                        token_length = tokenizer(
+                            conversations, return_tensors='pt', padding=False, truncation=False,
+                        ).input_ids.size(1)
+                        self.conv2length[str_length] = token_length + num_image_token * (
+                                    max_dynamic_patch + use_thumbnail)
+                    else:
+                        token_length = self.conv2length[str_length]
+                self.length.append(token_length)
+        gc.collect()
+
+    def __len__(self):
+        return len(self.raw_data)
+
+    def get_preprocess_function(self):
+        # Select the appropriate preprocessing function based on the template name
+        if self.template_name == 'Hermes-2':
+            preprocess_function = preprocess_mpt
+        elif self.template_name == 'internlm2-chat':
+            preprocess_function = preprocess_internlm
+        elif self.template_name == 'phi3-chat':
+            preprocess_function = preprocess_phi3
+        else:
+            preprocess_function = preprocess
+        return preprocess_function
+
+    def load_image(self, image_path):
+        # Load the image using tcs_loader if available, otherwise use PIL
+        if self.tcs_loader is not None and 's3://' in image_path:
+            return self.tcs_loader(image_path)
+        return Image.open(image_path).convert('RGB')
+
+    def get_image_path(self, image_path):
+        if image_path.startswith('s3://'):  # for ceph
+            image_path = self.root + image_path
+        else:  # for local image
+            image_path = os.path.join(self.root, image_path)
+        return image_path
+
+    def get_transform(self):
+        # Build transformation function
+        transform = build_transform(is_train=self.is_train, input_size=self.image_size,
+                                    pad2square=self.pad2square, normalize_type=self.normalize_type)
+        return transform
+
+    def multi_modal_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Ensure the first conversation contains an image placeholder
+        if '<image>' not in data_item['conversations'][0]['value']:
+            data_item['conversations'][0]['value'] = '<image>\n' + data_item['conversations'][0]['value']
+
+        # Merge the image path
+        image_path = self.get_image_path(data_item['image'])
+
+        # Load the image using tcs_loader if available, otherwise use PIL
+        image = self.load_image(image_path)
+
+        if self.dynamic_image_size:  # If dynamic image size is enabled, preprocess the image dynamically
+            images = dynamic_preprocess(image, min_num=self.min_dynamic_patch, max_num=self.max_dynamic_patch,
+                                        image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+        else:  # Otherwise, use the original image as a single patch
+            images = [image]
+
+        # Apply the transformation to each image and stack the results into a tensor
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+
+        # Ensure that there is only one patch if dynamic image size is not enabled
+        num_patches = pixel_values.size(0)
+        if not self.dynamic_image_size:
+            assert num_patches == 1, f'The number of patches should be 1, but got {num_patches}.'
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, [self.num_image_token * num_patches],
+                                  group_by_length=self.group_by_length, ds_name=self.ds_name)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def multi_modal_multi_image_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        images, num_tiles = [], []
+        num_image = len(data_item['image'])
+        for image_path in data_item['image']:
+            # Merge the image path
+            image_path = self.get_image_path(image_path)
+            # Load the image using tcs_loader if available, otherwise use PIL
+            image = self.load_image(image_path)
+            if self.dynamic_image_size:  # If dynamic image size is enabled, preprocess the image dynamically
+                image = dynamic_preprocess(image, min_num=self.min_dynamic_patch,
+                                           max_num=self.max_dynamic_patch // num_image,
+                                           image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+                images += image
+                num_tiles.append(len(image))
+            else:  # Otherwise, use the original image as a single patch
+                images.append(image)
+                num_tiles.append(1)
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        num_image_tokens = [self.num_image_token * num_tile for num_tile in num_tiles]
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, num_image_tokens, group_by_length=self.group_by_length,
+                                  ds_name=self.ds_name, num_image=num_image)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def video_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Ensure the first conversation contains a video placeholder
+        if '<video>' not in data_item['conversations'][0]['value']:
+            data_item['conversations'][0]['value'] = '<video>\n' + data_item['conversations'][0]['value']
+
+        # Get the video file path
+        video_file = data_item['video']
+        video_path = os.path.join(self.root, video_file)
+
+        # Load the video frames using tcs_loader
+        # TODO: Load videos without using tcsloader.
+        image_list = self.tcs_loader(
+            video_path,
+            image_type='video',
+            max_num_frames=self.max_num_frame,
+            min_num_frames=self.min_num_frame,
+            sample=self.sampling_method,
+            clip=data_item.get('clip', None))
+
+        # Generate special tokens for each video frame
+        special_tokens = '\n'.join(['Frame{}: <image>'.format(i + 1) for i in range(len(image_list))])
+        data_item['conversations'][0]['value'] = data_item['conversations'][0]['value'].replace(
+            '<video>\n', special_tokens)
+
+        # Transform each frame image and stack them into a tensor
+        pixel_values = [transform(image) for image in image_list]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        num_image_tokens = [self.num_image_token] * num_patches
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, num_image_tokens, group_by_length=self.group_by_length,
+                                  ds_name=self.ds_name, num_image=num_patches)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def pure_text_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Create a blank white image
+        image = Image.new('RGB', (224, 224), (255, 255, 255))
+
+        # Dynamically preprocess the image to generate patches
+        images = dynamic_preprocess(image, min_num=self.min_dynamic_patch, max_num=1,
+                                    image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+
+        # Apply the transformation to each image patch and stack them into a tensor
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Ensure there is only one patch
+        assert num_patches == 1, f'The number of patches should be 1, but got {num_patches}.'
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, [self.num_image_token * num_patches], text_only=True,
+                                  group_by_length=self.group_by_length, ds_name=self.ds_name)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([0] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        i = i % len(self.raw_data)
+        while True:
+            try:
+                data_item = json.loads(self.raw_data[i])
+                if 'image' in data_item and len(data_item['image']) != 0:
+                    if type(data_item['image']) == list:
+                        ret = self.multi_modal_multi_image_get_item(data_item)
+                    else:
+                        ret = self.multi_modal_get_item(data_item)
+                elif 'video' in data_item and data_item['video'] is not None and data_item['video'] != '':
+                    ret = self.video_get_item(data_item)
+                else:
+                    ret = self.pure_text_get_item(data_item)
+                break
+            except Exception as e:
+                print(e, self.ds_name, flush=True)
+                if not isinstance(e, UnidentifiedImageError):
+                    traceback.print_exc()
+                data_item = json.loads(self.raw_data[i])
+                if 'image' in data_item:
+                    if type(data_item['image']) == list:
+                        images = [self.root + item for item in data_item['image']]
+                        print(f'Failed to load image: {images}, the dataset is: {self.ds_name}')
+                    else:
+                        if data_item['image'].startswith('s3://'):
+                            data_path = self.root + data_item['image']
+                        else:
+                            data_path = os.path.join(self.root, data_item['image'])
+                        print(f'Failed to load image: {data_path}, the dataset is: {self.ds_name}')
+                elif 'video' in data_item:
+                    data_path = os.path.join(self.root, data_item['video'])
+                    print(f'Failed to load video: {data_path}, the dataset is: {self.ds_name}')
+                i = random.randint(0, len(self.raw_data) - 1)
+        return ret
+
+
+def build_datasets(
+    data_args,
+    tokenizer,
+    tcs_loader,
+    model,
+    group_by_length=False,
+    dynamic_image_size=False,
+    use_thumbnail=False,
+    min_dynamic_patch=1,
+    max_dynamic_patch=12,
+    normalize_type='imagenet',
+):
+    datasets = []
+    lengths = []
+    ds_collections = json.loads(open(data_args.meta_path).read())
+    for ds_idx, ds_name in enumerate(ds_collections.keys()):
+        repeat_time = ds_collections[ds_name]['repeat_time']
+        if 'max_dynamic_patch' in ds_collections[ds_name]:
+            max_num = ds_collections[ds_name]['max_dynamic_patch']
+            logger.info(f'max_dynamic_patch is set to {max_num} according to the meta file')
+        else:
+            max_num = max_dynamic_patch
+        dataset = LazySupervisedDataset(
+            data_args.conv_style, ds_collections[ds_name],
+            tokenizer,
+            tcs_loader,
+            ds_name=ds_name,
+            num_image_token=model.num_image_token,
+            image_size=data_args.force_image_size,
+            is_train=ds_collections[ds_name]['data_augment'],
+            pad2square=data_args.pad2square,
+            group_by_length=group_by_length,
+            dynamic_image_size=dynamic_image_size,
+            use_thumbnail=use_thumbnail,
+            min_dynamic_patch=min_dynamic_patch,
+            max_dynamic_patch=max_num,
+            repeat_time=repeat_time,
+            normalize_type=normalize_type,
+            random_seed=ds_idx,
+        )
+        logger.info(f'Add dataset: {ds_name} with length: {len(dataset)}')
+        datasets.append(dataset)
+        if data_args.use_data_resampling:
+            lengths.append(math.sqrt(len(dataset)))
+        else:
+            lengths.append(len(dataset))
+    if data_args.use_data_resampling:
+        total_length = sum(lengths)
+        weights = [l / total_length for l in lengths]
+        train_dataset = WeightedConcatDataset(datasets, weights)
+    else:
+        train_dataset = ConcatDataset(datasets)
+    return train_dataset
+
+
+def main():
+    # Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # If use DeepSpeed zero3, init_dist must before HfArgumentParser
+    launcher = os.environ.get('LAUNCHER', 'slurm')
+    init_dist(launcher=launcher, backend='nccl')
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith('.json'):
+        # If we pass only one argument to the script, and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    # send_example_telemetry('InternV-Chat', model_args, data_args)
+
+    # Setup logging
+    logging.basicConfig(
+        format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
+        datefmt='%m/%d/%Y %H:%M:%S',
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+
+    if training_args.should_log:
+        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
+        transformers.utils.logging.set_verbosity_info()
+
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    set_verbosity(log_level)
+    enable_default_handler()
+    enable_explicit_format()
+
+    # Log on each process the small summary:
+    logger.warning(
+        f'Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}'
+        + f'distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}'
+    )
+    logger.info(f'Training/evaluation parameters {training_args}')
+
+    # Detecting last checkpoint and eventually continue from last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f'Output directory ({training_args.output_dir}) already exists and is not empty. '
+                'Use --overwrite_output_dir to overcome.'
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f'Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change '
+                'the `--output_dir` or add `--overwrite_output_dir` to train from scratch.'
+            )
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Load pretrained model, tokenizer, and image processor
+    tokenizer_path = model_args.model_name_or_path or model_args.llm_path
+    logger.info(f'Loading Tokenizer: {tokenizer_path}')
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_path, add_eos_token=False, trust_remote_code=True, use_fast=False)
+    tokenizer.tokenizer_path = tokenizer_path
+    tokenizer.model_max_length = data_args.max_seq_length
+    token_list = [IMG_START_TOKEN, IMG_END_TOKEN, IMG_CONTEXT_TOKEN,
+                  QUAD_START_TOKEN, QUAD_END_TOKEN, REF_START_TOKEN,
+                  REF_END_TOKEN, BOX_START_TOKEN, BOX_END_TOKEN]
+    num_new_tokens = tokenizer.add_tokens(token_list, special_tokens=True)
+    img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
+    tcs_loader = TCSLoader('~/petreloss.conf') if has_tcs_loader else None
+
+    if model_args.model_name_or_path is not None:
+        logger.info('Loading InternVLChatModel...')
+        config = InternVLChatConfig.from_pretrained(model_args.model_name_or_path)
+        config.vision_config.drop_path_rate = model_args.drop_path_rate
+        if config.llm_config.model_type == 'internlm2':
+            config.llm_config.attn_implementation = 'flash_attention_2'  # for InternLM
+            logger.info('Using flash_attention_2 for InternLM')
+        else:
+            config.llm_config._attn_implementation = 'flash_attention_2'  # for LLaMA
+            logger.info('Using flash_attention_2 for LLaMA')
+        config.template = data_args.conv_style
+        config.select_layer = model_args.vision_select_layer
+        config.dynamic_image_size = data_args.dynamic_image_size
+        config.use_thumbnail = data_args.use_thumbnail
+        config.ps_version = model_args.ps_version
+        config.min_dynamic_patch = data_args.min_dynamic_patch
+        config.max_dynamic_patch = data_args.max_dynamic_patch
+        model = InternVLChatModel.from_pretrained(
+            model_args.model_name_or_path, torch_dtype=torch.bfloat16, config=config)
+    else:
+        logger.info('Loading ViT-6B...')
+        vision_config = InternVisionConfig.from_pretrained(model_args.vision_path)
+        vision_config.drop_path_rate = model_args.drop_path_rate
+        vision_model = InternVisionModel.from_pretrained(
+            model_args.vision_path, torch_dtype=torch.bfloat16, config=vision_config)
+        logger.info('Loading LLaMA...')
+        llm_config = AutoConfig.from_pretrained(model_args.llm_path, trust_remote_code=True)
+        if llm_config.model_type == 'internlm2':
+            model_type = InternLM2ForCausalLM
+            llm_config.attn_implementation = 'flash_attention_2'  # for InternLM
+            logger.info('Using flash_attention_2 for InternLM')
+        else:
+            model_type = AutoModelForCausalLM
+            llm_config._attn_implementation = 'flash_attention_2'  # for LLaMA
+            logger.info('Using flash_attention_2 for LLaMA')
+        llm = model_type.from_pretrained(
+            model_args.llm_path, torch_dtype=torch.bfloat16,
+            config=llm_config, trust_remote_code=True)
+        logger.info('Building InternVLChatConfig...')
+        internvl_chat_config = InternVLChatConfig(
+            vision_config.to_dict(), llm_config.to_dict(), downsample_ratio=data_args.down_sample_ratio,
+            pad2square=data_args.pad2square, template=data_args.conv_style,
+            select_layer=model_args.vision_select_layer, dynamic_image_size=data_args.dynamic_image_size,
+            use_thumbnail=data_args.use_thumbnail, ps_version=model_args.ps_version,
+            min_dynamic_patch=data_args.min_dynamic_patch, max_dynamic_patch=data_args.max_dynamic_patch)
+        internvl_chat_config.force_image_size = data_args.force_image_size
+        logger.info('Building InternVLChatModel...')
+        model = InternVLChatModel(internvl_chat_config, vision_model, llm)
+    model.img_context_token_id = img_context_token_id
+
+    assert model.config.downsample_ratio == data_args.down_sample_ratio
+
+    if model_args.mlp_path is not None:
+        logger.info('Loading pretrained MLP projector...')
+        state_dict = torch.load(model_args.mlp_path, map_location='cpu')
+        message = model.mlp1.load_state_dict(state_dict)
+        logger.info(message)
+    logger.info('Finished')
+
+    patch_size = model.config.vision_config.patch_size
+    logger.info(f'model.config.force_image_size: {model.config.force_image_size}')
+    logger.info(f'data_args.force_image_size: {data_args.force_image_size}')
+    logger.info(f'model.config.vision_config.image_size: {model.config.vision_config.image_size}')
+    if model.config.vision_config.image_size != data_args.force_image_size:
+        logger.info(f'Resizing position embedding from '
+                    f'{model.config.vision_config.image_size} '
+                    f'to {data_args.force_image_size}...')
+        model.vision_model.resize_pos_embeddings(old_size=model.config.vision_config.image_size,
+                                                 new_size=data_args.force_image_size,
+                                                 patch_size=patch_size)
+        model.config.vision_config.image_size = data_args.force_image_size
+    model.config.force_image_size = data_args.force_image_size
+    model.num_image_token = int((data_args.force_image_size // patch_size) ** 2 * (data_args.down_sample_ratio ** 2))
+
+    if num_new_tokens > 0:
+        model.language_model.resize_token_embeddings(len(tokenizer))
+        output_embeddings = model.language_model.get_output_embeddings().weight.data
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+        model.config.llm_config.vocab_size = len(tokenizer)
+        model.language_model.config.vocab_size = len(tokenizer)
+
+    model.language_model.config.use_cache = False
+    model.vision_model.gradient_checkpointing = True
+    model.vision_model.encoder.gradient_checkpointing = True
+    if model_args.grad_checkpoint:
+        model.language_model._set_gradient_checkpointing()
+
+    train_dataset = build_datasets(
+        data_args, tokenizer, tcs_loader, model, group_by_length=training_args.group_by_length,
+        dynamic_image_size=data_args.dynamic_image_size, use_thumbnail=data_args.use_thumbnail,
+        min_dynamic_patch=data_args.min_dynamic_patch, max_dynamic_patch=data_args.max_dynamic_patch,
+        normalize_type=data_args.normalize_type)
+
+    def _freeze_params(module):
+        for param in module.parameters():
+            param.requires_grad = False
+
+    if model_args.freeze_backbone:
+        # model.vision_model = model.vision_model.eval()
+        _freeze_params(model.vision_model)
+
+    if model_args.freeze_llm:
+        model.language_model = model.language_model.eval()
+        _freeze_params(model.language_model)
+
+    if model_args.unfreeze_lm_head:
+        model.language_model.lm_head.requires_grad = True
+
+    if model_args.use_backbone_lora:
+        model.wrap_backbone_lora(r=model_args.use_backbone_lora, lora_alpha=2 * model_args.use_backbone_lora)
+        model.config.use_backbone_lora = model_args.use_backbone_lora
+
+    if model_args.use_llm_lora:
+        model.wrap_llm_lora(r=model_args.use_llm_lora, lora_alpha=2 * model_args.use_llm_lora)
+        model.config.use_llm_lora = model_args.use_llm_lora
+
+    if model_args.freeze_mlp:
+        _freeze_params(model.mlp1)
+
+    if model_args.unfreeze_vit_layers != 0:
+        layers = model.vision_model.encoder.layers[model_args.unfreeze_vit_layers:]
+        for k, v in layers.named_parameters():
+            logger.info(f'Unfreezing ViT layer: {k}')
+            v.requires_grad = True
+
+    # print trainable parameters
+    if dist.get_rank() == 0:
+        for name, param in model.named_parameters():
+            if param.requires_grad:
+                logger.info(name)
+
+    # set seed for torch dataloaders
+    set_seed(training_args.seed)
+
+    # Initialize our Trainer
+    if model_args.use_custom_trainer:
+        replace_create_optimizer()
+
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset if training_args.do_train else None,
+        eval_dataset=None,
+        tokenizer=tokenizer,
+        data_collator=concat_pad_data_collator
+    )
+
+    # Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+        metrics = train_result.metrics
+        try:
+            metrics['train_samples'] = len(train_dataset)
+        except:
+            metrics['train_samples'] = -1
+
+        trainer.log_metrics('train', metrics)
+        trainer.save_metrics('train', metrics)
+        trainer.save_state()
+
+
+if __name__ == '__main__':
+    main()

internvl/train/internvl_chat_pretrain.py

+import gc
+import json
+import logging
+import math
+import os
+import random
+import sys
+import traceback
+import warnings
+from copy import deepcopy
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import transformers
+from internvl.dist_utils import init_dist
+from internvl.model.internlm2.modeling_internlm2 import InternLM2ForCausalLM
+from internvl.model.internvl_chat import (InternVisionConfig,
+                                          InternVisionModel,
+                                          InternVLChatConfig,
+                                          InternVLChatModel)
+from internvl.patch import (concat_pad_data_collator,
+                            replace_llama_rmsnorm_with_fused_rmsnorm,
+                            replace_train_sampler)
+from internvl.train.constants import (BOX_END_TOKEN, BOX_START_TOKEN,
+                                      IMG_CONTEXT_TOKEN, IMG_END_TOKEN,
+                                      IMG_START_TOKEN, QUAD_END_TOKEN,
+                                      QUAD_START_TOKEN, REF_END_TOKEN,
+                                      REF_START_TOKEN)
+from internvl.train.dataset import (ConcatDataset, TCSLoader,
+                                    WeightedConcatDataset, build_transform,
+                                    dynamic_preprocess, preprocess,
+                                    preprocess_internlm, preprocess_mpt,
+                                    preprocess_phi3)
+from internvl.train.trainer_monkey_patch import replace_create_optimizer
+from PIL import Image, ImageFile, PngImagePlugin, UnidentifiedImageError
+from torch.utils.data import Dataset
+from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
+                          HfArgumentParser, Trainer, TrainingArguments,
+                          set_seed)
+from transformers.trainer_utils import get_last_checkpoint
+from transformers.utils.logging import (enable_default_handler,
+                                        enable_explicit_format, set_verbosity)
+
+# Apply necessary patches for the transformers library
+replace_llama_rmsnorm_with_fused_rmsnorm()
+replace_train_sampler()
+
+# Try to import petrel_client for image loading, fallback to PIL if unavailable
+try:
+    from petrel_client.client import Client
+    from petrel_client.common.config import Config
+    has_tcs_loader = True
+except ImportError as E:
+    print('petrel_client is not installed. Using PIL to load images.')
+    has_tcs_loader = False
+
+# Set constants for image processing and logging
+IGNORE_INDEX = -100
+Image.MAX_IMAGE_PIXELS = None
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+MaximumDecompressedSize = 1024
+MegaByte = 2 ** 20
+PngImagePlugin.MAX_TEXT_CHUNK = MaximumDecompressedSize * MegaByte
+
+warnings.filterwarnings('ignore')
+logger = logging.getLogger(__name__)
+
+os.environ['TOKENIZERS_PARALLELISM'] = 'true'
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments for specifying model, tokenizer, and configurations.
+    """
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    vision_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    llm_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    mlp_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'Path to pretrained model or model identifier from huggingface.co/models'}
+    )
+    freeze_llm: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the LLM decoder.'},
+    )
+    freeze_backbone: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the vision backbone of the model.'},
+    )
+    freeze_mlp: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to freeze the MLP layers of the model.'},
+    )
+    unfreeze_vit_layers: int = field(
+        default=0,
+        metadata={'help': 'Specify the number of ViT layers to unfreeze. Default is 0.'},
+    )
+    vision_select_layer: int = field(
+        default=-1,
+        metadata={'help': 'Specify the layer of ViT feature map to use. Default is last layer.'},
+    )
+    use_backbone_lora: int = field(
+        default=0,
+        metadata={'help': 'Set the LoRA adapter rank for the backbone model. Default is 0.'}
+    )
+    use_llm_lora: int = field(
+        default=0,
+        metadata={'help': 'Set the LoRA adapter rank for the LLM. Default is 0.'}
+    )
+    unfreeze_lm_head: bool = field(
+        default=False,
+        metadata={'help': "Set to True to unfreeze the language model's head."},
+    )
+    use_custom_trainer: bool = field(
+        default=False,
+        metadata={'help': 'Set to True to enable the use of a custom trainer.'},
+    )
+    grad_checkpoint: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use gradient checkpointing.'},
+    )
+    drop_path_rate: float = field(
+        default=0.0,
+        metadata={'help': 'Set the drop path rate for the ViT model. Default is 0.'},
+    )
+    ps_version: str = field(
+        default='v2',
+        metadata={'help': 'Specify the version of pixel shuffle implementation. Default is `v1`.'
+                          'Please use `v2` to fix the bug of transposed image.'}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments for specifying data input for training and evaluation.
+    """
+    max_seq_length: Optional[int] = field(
+        default=2048,
+        metadata={
+            'help': (
+                'The maximum total input sequence length after tokenization. Sequences longer '
+                'than this will be truncated, sequences shorter will be padded.'
+            )
+        },
+    )
+    force_image_size: Optional[int] = field(
+        default=448,
+        metadata={'help': 'Set the desired size for the image. Default is 224.'},
+    )
+    down_sample_ratio: Optional[float] = field(
+        default=0.5,
+        metadata={'help': 'Set the desired down-sampling ratio for the image. Default is 1.0.'},
+    )
+    pad2square: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Pad the image to a square shape if set to True.'},
+    )
+    conv_style: Optional[str] = field(
+        default='internlm2-chat', metadata={'help': 'Prompt style for a conversation.'}
+    )
+    meta_path: Optional[str] = field(
+        default=None,
+        metadata={'help': 'The path of the meta file of datasets.'},
+    )
+    use_data_resampling: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use data resampling.'},
+    )
+    dynamic_image_size: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to use dynamic image size.'},
+    )
+    use_thumbnail: Optional[bool] = field(
+        default=False,
+        metadata={'help': 'Set to True to add a thumbnail image.'},
+    )
+    min_dynamic_patch: Optional[int] = field(
+        default=1,
+        metadata={'help': 'The minimum number of dynamic patches. Default is 1.'},
+    )
+    max_dynamic_patch: Optional[int] = field(
+        default=12,
+        metadata={'help': 'The maximum number of dynamic patches. Default is 6.'},
+    )
+    normalize_type: Optional[str] = field(
+        default='imagenet',
+        metadata={'help': 'The normalize type for the image. Default is imagenet.'},
+    )
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(
+        self,
+        template_name,
+        meta,
+        tokenizer,
+        tcs_loader,
+        ds_name,
+        num_image_token,
+        image_size=224,
+        is_train=True,
+        pad2square=False,
+        group_by_length=False,
+        dynamic_image_size=False,
+        use_thumbnail=False,
+        min_dynamic_patch=1,
+        max_dynamic_patch=6,
+        min_num_frame=4,  # for video data
+        max_num_frame=12,  # for video data
+        sampling_method='rand',  # for video data
+        repeat_time=1,
+        normalize_type='imagenet',
+        random_seed=0,
+    ):
+        super(LazySupervisedDataset, self).__init__()
+        self.ds_name = ds_name
+        self.tokenizer = tokenizer
+        self.template_name = template_name
+        self.num_image_token = num_image_token
+        logger.info(f'[Dataset] num_image_token: {num_image_token}')
+        logger.info(f'[Dataset] dynamic_image_size: {dynamic_image_size}')
+        logger.info(f'[Dataset] use_thumbnail: {use_thumbnail}')
+        logger.info(f'[Dataset] min_dynamic_patch: {min_dynamic_patch}, max_dynamic_patch: {max_dynamic_patch}')
+
+        self.image_size = image_size
+        self.is_train = is_train
+        self.pad2square = pad2square
+        self.max_num_frame = max_num_frame
+        self.min_num_frame = min_num_frame
+        self.sampling_method = sampling_method
+
+        logger.info('Formatting inputs...Skip in lazy mode')
+        assert meta['annotation'].endswith('jsonl'), f'annotation must be jsonl, but got {meta["annotation"]}'
+
+        total_ranks = torch.distributed.get_world_size()
+        current_rank = torch.distributed.get_rank()
+
+        """
+        This section of the code is used to read hundreds of millions of data entries.
+        By using caching and splitting the data according to rank, it ensures fast reading
+        speed and prevents out-of-memory.
+        """
+        # Create a cache directory path
+        basename = os.path.basename(meta['annotation']).replace('.jsonl', '')
+        data_dir = os.path.join(os.path.dirname(meta['annotation']), f'{basename}_temp')
+        os.makedirs(data_dir, exist_ok=True)  # Create the cache directory if it does not exist
+        # Create a temporary path for the current rank
+        temp_path = os.path.join(data_dir, f'{basename}_{current_rank}_of_{total_ranks}.jsonl')
+
+        # Check if the temporary file for the current rank already exists
+        if os.path.exists(temp_path):
+            # If it exists, read the raw data from the file
+            with open(temp_path, 'r') as f:
+                self.raw_data = f.readlines()
+        else:
+            # If it does not exist, read the raw data from the original annotation file
+            with open(meta['annotation'], 'r') as f:
+                self.raw_data = f.readlines()
+
+            # Adjust the raw data based on the repeat_time parameter
+            if repeat_time < 1:
+                self.raw_data = self.raw_data[:int(len(self.raw_data) * repeat_time)]
+            else:
+                self.raw_data = self.raw_data * int(repeat_time)
+
+            # Calculate the total number of lines and distribute lines to each rank
+            total_lines = len(self.raw_data)
+            logger.info(f'total_ranks: {total_ranks}, current_rank: {current_rank}, total_lines: {total_lines}')
+            lines_per_rank = total_lines // total_ranks  # Number of lines each rank should process
+            lines_per_rank = max(1, lines_per_rank)
+
+            # Calculate the start and end line numbers for the current rank
+            start_line = lines_per_rank * current_rank  # Starting line for the current rank
+            end_line = start_line + lines_per_rank  # Ending line for the current rank
+
+            # Assign the appropriate lines to the current rank
+            self.raw_data = self.raw_data[start_line:end_line]
+
+            # Write the raw data for the current rank to the temporary file
+            with open(temp_path, 'w') as f:
+                f.writelines(self.raw_data)
+
+        self.rng = np.random.default_rng(seed=random_seed)
+        self.rng.shuffle(self.raw_data)
+
+        gc.collect()
+        self.root = meta['root']
+        self.cached_data_dict = {}
+        self.tcs_loader = tcs_loader
+        self.group_by_length = group_by_length
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.min_dynamic_patch = min_dynamic_patch
+        self.max_dynamic_patch = max_dynamic_patch
+        self.normalize_type = normalize_type
+
+        # If the precomputed length does not exist, roughly estimate the length of
+        # each sample to improve the efficiency of group_by_length.
+        if self.group_by_length:
+            self.conv2length = {}  # Using a dictionary to speed up token length calculation
+            self.length = []
+            for data_item in self.raw_data:
+                data_item = json.loads(data_item)
+                if 'length' in data_item:
+                    token_length = data_item['length']  # Use precomputed length if available
+                else:
+                    # Compute token length using the tokenizer
+                    conversations = '\n'.join([temp['value'] for temp in data_item['conversations']])
+                    str_length = len(conversations)
+                    if str_length not in self.conv2length:
+                        token_length = tokenizer(
+                            conversations, return_tensors='pt', padding=False, truncation=False,
+                        ).input_ids.size(1)
+                        self.conv2length[str_length] = token_length + num_image_token * (
+                                    max_dynamic_patch + use_thumbnail)
+                    else:
+                        token_length = self.conv2length[str_length]
+                self.length.append(token_length)
+        gc.collect()
+
+    def __len__(self):
+        return len(self.raw_data) * torch.distributed.get_world_size()
+
+    def get_preprocess_function(self):
+        # Select the appropriate preprocessing function based on the template name
+        if self.template_name == 'Hermes-2':
+            preprocess_function = preprocess_mpt
+        elif self.template_name == 'internlm2-chat':
+            preprocess_function = preprocess_internlm
+        elif self.template_name == 'phi3-chat':
+            preprocess_function = preprocess_phi3
+        else:
+            preprocess_function = preprocess
+        return preprocess_function
+
+    def load_image(self, image_path):
+        # Load the image using tcs_loader if available, otherwise use PIL
+        if self.tcs_loader is not None and 's3://' in image_path:
+            return self.tcs_loader(image_path)
+        return Image.open(image_path).convert('RGB')
+
+    def get_image_path(self, image_path):
+        if image_path.startswith('s3://'):  # for ceph
+            image_path = self.root + image_path
+        else:  # for local image
+            image_path = os.path.join(self.root, image_path)
+        return image_path
+
+    def get_transform(self):
+        # Build transformation function
+        transform = build_transform(is_train=self.is_train, input_size=self.image_size,
+                                    pad2square=self.pad2square, normalize_type=self.normalize_type)
+        return transform
+
+    def multi_modal_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Ensure the first conversation contains an image placeholder
+        if '<image>' not in data_item['conversations'][0]['value']:
+            data_item['conversations'][0]['value'] = '<image>\n' + data_item['conversations'][0]['value']
+
+        # Merge the image path
+        image_path = self.get_image_path(data_item['image'])
+
+        # Load the image using tcs_loader if available, otherwise use PIL
+        image = self.load_image(image_path)
+
+        if self.dynamic_image_size:  # If dynamic image size is enabled, preprocess the image dynamically
+            images = dynamic_preprocess(image, min_num=self.min_dynamic_patch, max_num=self.max_dynamic_patch,
+                                        image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+        else:  # Otherwise, use the original image as a single patch
+            images = [image]
+
+        # Apply the transformation to each image and stack the results into a tensor
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+
+        # Ensure that there is only one patch if dynamic image size is not enabled
+        num_patches = pixel_values.size(0)
+        if not self.dynamic_image_size:
+            assert num_patches == 1, f'The number of patches should be 1, but got {num_patches}.'
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, [self.num_image_token * num_patches],
+                                  group_by_length=self.group_by_length, ds_name=self.ds_name)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def multi_modal_multi_image_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        images, num_tiles = [], []
+        num_image = len(data_item['image'])
+        for image_path in data_item['image']:
+            # Merge the image path
+            image_path = self.get_image_path(image_path)
+            # Load the image using tcs_loader if available, otherwise use PIL
+            image = self.load_image(image_path)
+            if self.dynamic_image_size:  # If dynamic image size is enabled, preprocess the image dynamically
+                image = dynamic_preprocess(image, min_num=self.min_dynamic_patch,
+                                           max_num=self.max_dynamic_patch // num_image,
+                                           image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+                images += image
+                num_tiles.append(len(image))
+            else:  # Otherwise, use the original image as a single patch
+                images.append(image)
+                num_tiles.append(1)
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        num_image_tokens = [self.num_image_token * num_tile for num_tile in num_tiles]
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, num_image_tokens, group_by_length=self.group_by_length,
+                                  ds_name=self.ds_name, num_image=num_image)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def video_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Ensure the first conversation contains a video placeholder
+        if '<video>' not in data_item['conversations'][0]['value']:
+            data_item['conversations'][0]['value'] = '<video>\n' + data_item['conversations'][0]['value']
+
+        # Get the video file path
+        video_file = data_item['video']
+        video_path = os.path.join(self.root, video_file)
+
+        # Load the video frames using tcs_loader
+        # TODO: Load videos without using tcsloader.
+        image_list = self.tcs_loader(
+            video_path,
+            image_type='video',
+            max_num_frames=self.max_num_frame,
+            min_num_frames=self.min_num_frame,
+            sample=self.sampling_method,
+            clip=data_item.get('clip', None))
+
+        # Generate special tokens for each video frame
+        special_tokens = '\n'.join(['Frame{}: <image>'.format(i + 1) for i in range(len(image_list))])
+        data_item['conversations'][0]['value'] = data_item['conversations'][0]['value'].replace(
+            '<video>\n', special_tokens)
+
+        # Transform each frame image and stack them into a tensor
+        pixel_values = [transform(image) for image in image_list]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        num_image_tokens = [self.num_image_token] * num_patches
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, num_image_tokens, group_by_length=self.group_by_length,
+                                  ds_name=self.ds_name, num_image=num_patches)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([1] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def pure_text_get_item(self, data_item):
+        # Build transformation function
+        transform = self.get_transform()
+
+        # Create a blank white image
+        image = Image.new('RGB', (224, 224), (255, 255, 255))
+
+        # Dynamically preprocess the image to generate patches
+        images = dynamic_preprocess(image, min_num=self.min_dynamic_patch, max_num=1,
+                                    image_size=self.image_size, use_thumbnail=self.use_thumbnail)
+
+        # Apply the transformation to each image patch and stack them into a tensor
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        num_patches = pixel_values.size(0)
+
+        # Ensure there is only one patch
+        assert num_patches == 1, f'The number of patches should be 1, but got {num_patches}.'
+
+        # Select the appropriate preprocessing function based on the template name
+        preprocess_function = self.get_preprocess_function()
+
+        # Preprocess the conversations and generate the return dictionary
+        ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
+                                  self.tokenizer, [self.num_image_token * num_patches], text_only=True,
+                                  group_by_length=self.group_by_length, ds_name=self.ds_name)
+
+        # Create the final return dictionary
+        ret = dict(
+            input_ids=ret['input_ids'][0],
+            labels=ret['labels'][0],
+            attention_mask=ret['attention_mask'][0],
+            pixel_values=pixel_values,
+            image_flags=torch.tensor([0] * num_patches, dtype=torch.long)
+        )
+        return ret
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        i = i % len(self.raw_data)
+        while True:
+            try:
+                data_item = json.loads(self.raw_data[i])
+                if 'image' in data_item and len(data_item['image']) != 0:
+                    if type(data_item['image']) == list:
+                        ret = self.multi_modal_multi_image_get_item(data_item)
+                    else:
+                        ret = self.multi_modal_get_item(data_item)
+                elif 'video' in data_item and data_item['video'] is not None and data_item['video'] != '':
+                    ret = self.video_get_item(data_item)
+                else:
+                    ret = self.pure_text_get_item(data_item)
+                break
+            except Exception as e:
+                print(e, self.ds_name, flush=True)
+                if not isinstance(e, UnidentifiedImageError):
+                    traceback.print_exc()
+                data_item = json.loads(self.raw_data[i])
+                if 'image' in data_item:
+                    if type(data_item['image']) == list:
+                        images = [self.root + item for item in data_item['image']]
+                        print(f'Failed to load image: {images}, the dataset is: {self.ds_name}')
+                    else:
+                        if data_item['image'].startswith('s3://'):
+                            data_path = self.root + data_item['image']
+                        else:
+                            data_path = os.path.join(self.root, data_item['image'])
+                        print(f'Failed to load image: {data_path}, the dataset is: {self.ds_name}')
+                elif 'video' in data_item:
+                    data_path = os.path.join(self.root, data_item['video'])
+                    print(f'Failed to load video: {data_path}, the dataset is: {self.ds_name}')
+                i = random.randint(0, len(self.raw_data) - 1)
+        return ret
+
+
+def build_datasets(
+    data_args,
+    tokenizer,
+    tcs_loader,
+    model,
+    group_by_length=False,
+    dynamic_image_size=False,
+    use_thumbnail=False,
+    min_dynamic_patch=1,
+    max_dynamic_patch=12,
+    normalize_type='imagenet',
+):
+    datasets = []
+    lengths = []
+    ds_collections = json.loads(open(data_args.meta_path).read())
+    for ds_idx, ds_name in enumerate(ds_collections.keys()):
+        repeat_time = ds_collections[ds_name]['repeat_time']
+        if 'max_dynamic_patch' in ds_collections[ds_name]:
+            max_num = ds_collections[ds_name]['max_dynamic_patch']
+            logger.info(f'max_dynamic_patch is set to {max_num} according to the meta file')
+        else:
+            max_num = max_dynamic_patch
+        dataset = LazySupervisedDataset(
+            data_args.conv_style, ds_collections[ds_name],
+            tokenizer,
+            tcs_loader,
+            ds_name=ds_name,
+            num_image_token=model.num_image_token,
+            image_size=data_args.force_image_size,
+            is_train=ds_collections[ds_name]['data_augment'],
+            pad2square=data_args.pad2square,
+            group_by_length=group_by_length,
+            dynamic_image_size=dynamic_image_size,
+            use_thumbnail=use_thumbnail,
+            min_dynamic_patch=min_dynamic_patch,
+            max_dynamic_patch=max_num,
+            repeat_time=repeat_time,
+            normalize_type=normalize_type,
+            random_seed=ds_idx,
+        )
+        logger.info(f'Add dataset: {ds_name} with length: {len(dataset)}')
+        datasets.append(dataset)
+        if data_args.use_data_resampling:
+            lengths.append(math.sqrt(len(dataset)))
+        else:
+            lengths.append(len(dataset))
+    if data_args.use_data_resampling:
+        total_length = sum(lengths)
+        weights = [l / total_length for l in lengths]
+        train_dataset = WeightedConcatDataset(datasets, weights)
+    else:
+        train_dataset = ConcatDataset(datasets)
+    return train_dataset
+
+
+def main():
+    # Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # If use DeepSpeed zero3, init_dist must before HfArgumentParser
+    launcher = os.environ.get('LAUNCHER', 'slurm')
+    init_dist(launcher=launcher, backend='nccl')
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith('.json'):
+        # If we pass only one argument to the script, and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    # send_example_telemetry('InternV-Chat', model_args, data_args)
+
+    # Setup logging
+    logging.basicConfig(
+        format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
+        datefmt='%m/%d/%Y %H:%M:%S',
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+
+    if training_args.should_log:
+        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
+        transformers.utils.logging.set_verbosity_info()
+
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    set_verbosity(log_level)
+    enable_default_handler()
+    enable_explicit_format()
+
+    # Log on each process the small summary:
+    logger.warning(
+        f'Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}'
+        + f'distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}'
+    )
+    logger.info(f'Training/evaluation parameters {training_args}')
+
+    # Detecting last checkpoint and eventually continue from last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f'Output directory ({training_args.output_dir}) already exists and is not empty. '
+                'Use --overwrite_output_dir to overcome.'
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f'Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change '
+                'the `--output_dir` or add `--overwrite_output_dir` to train from scratch.'
+            )
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Load pretrained model, tokenizer, and image processor
+    tokenizer_path = model_args.model_name_or_path or model_args.llm_path
+    logger.info(f'Loading Tokenizer: {tokenizer_path}')
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_path, add_eos_token=False, trust_remote_code=True, use_fast=False)
+    tokenizer.tokenizer_path = tokenizer_path
+    tokenizer.model_max_length = data_args.max_seq_length
+    token_list = [IMG_START_TOKEN, IMG_END_TOKEN, IMG_CONTEXT_TOKEN,
+                  QUAD_START_TOKEN, QUAD_END_TOKEN, REF_START_TOKEN,
+                  REF_END_TOKEN, BOX_START_TOKEN, BOX_END_TOKEN]
+    num_new_tokens = tokenizer.add_tokens(token_list, special_tokens=True)
+    img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
+    tcs_loader = TCSLoader('~/petreloss.conf') if has_tcs_loader else None
+
+    if model_args.model_name_or_path is not None:
+        logger.info('Loading InternVLChatModel...')
+        config = InternVLChatConfig.from_pretrained(model_args.model_name_or_path)
+        config.vision_config.drop_path_rate = model_args.drop_path_rate
+        if config.llm_config.model_type == 'internlm2':
+            config.llm_config.attn_implementation = 'flash_attention_2'  # for InternLM
+            logger.info('Using flash_attention_2 for InternLM')
+        else:
+            config.llm_config._attn_implementation = 'flash_attention_2'  # for LLaMA
+            logger.info('Using flash_attention_2 for LLaMA')
+        config.template = data_args.conv_style
+        config.select_layer = model_args.vision_select_layer
+        config.dynamic_image_size = data_args.dynamic_image_size
+        config.use_thumbnail = data_args.use_thumbnail
+        config.ps_version = model_args.ps_version
+        config.min_dynamic_patch = data_args.min_dynamic_patch
+        config.max_dynamic_patch = data_args.max_dynamic_patch
+        model = InternVLChatModel.from_pretrained(
+            model_args.model_name_or_path, torch_dtype=torch.bfloat16, config=config)
+    else:
+        logger.info('Loading ViT-6B...')
+        vision_config = InternVisionConfig.from_pretrained(model_args.vision_path)
+        vision_config.drop_path_rate = model_args.drop_path_rate
+        vision_model = InternVisionModel.from_pretrained(
+            model_args.vision_path, torch_dtype=torch.bfloat16, config=vision_config)
+        logger.info('Loading LLaMA...')
+        llm_config = AutoConfig.from_pretrained(model_args.llm_path, trust_remote_code=True)
+        if llm_config.model_type == 'internlm2':
+            model_type = InternLM2ForCausalLM
+            llm_config.attn_implementation = 'flash_attention_2'  # for InternLM
+            logger.info('Using flash_attention_2 for InternLM')
+        else:
+            model_type = AutoModelForCausalLM
+            llm_config._attn_implementation = 'flash_attention_2'  # for LLaMA
+            logger.info('Using flash_attention_2 for LLaMA')
+        llm = model_type.from_pretrained(
+            model_args.llm_path, torch_dtype=torch.bfloat16,
+            config=llm_config, trust_remote_code=True)
+        logger.info('Building InternVLChatConfig...')
+        internvl_chat_config = InternVLChatConfig(
+            vision_config.to_dict(), llm_config.to_dict(), downsample_ratio=data_args.down_sample_ratio,
+            pad2square=data_args.pad2square, template=data_args.conv_style,
+            select_layer=model_args.vision_select_layer, dynamic_image_size=data_args.dynamic_image_size,
+            use_thumbnail=data_args.use_thumbnail, ps_version=model_args.ps_version,
+            min_dynamic_patch=data_args.min_dynamic_patch, max_dynamic_patch=data_args.max_dynamic_patch)
+        internvl_chat_config.force_image_size = data_args.force_image_size
+        logger.info('Building InternVLChatModel...')
+        model = InternVLChatModel(internvl_chat_config, vision_model, llm)
+    model.img_context_token_id = img_context_token_id
+
+    assert model.config.downsample_ratio == data_args.down_sample_ratio
+
+    if model_args.mlp_path is not None:
+        logger.info('Loading pretrained MLP projector...')
+        state_dict = torch.load(model_args.mlp_path, map_location='cpu')
+        message = model.mlp1.load_state_dict(state_dict)
+        logger.info(message)
+    logger.info('Finished')
+
+    patch_size = model.config.vision_config.patch_size
+    logger.info(f'model.config.force_image_size: {model.config.force_image_size}')
+    logger.info(f'data_args.force_image_size: {data_args.force_image_size}')
+    logger.info(f'model.config.vision_config.image_size: {model.config.vision_config.image_size}')
+    if model.config.vision_config.image_size != data_args.force_image_size:
+        logger.info(f'Resizing position embedding from '
+                    f'{model.config.vision_config.image_size} '
+                    f'to {data_args.force_image_size}...')
+        model.vision_model.resize_pos_embeddings(old_size=model.config.vision_config.image_size,
+                                                 new_size=data_args.force_image_size,
+                                                 patch_size=patch_size)
+        model.config.vision_config.image_size = data_args.force_image_size
+    model.config.force_image_size = data_args.force_image_size
+    model.num_image_token = int((data_args.force_image_size // patch_size) ** 2 * (data_args.down_sample_ratio ** 2))
+
+    if num_new_tokens > 0:
+        model.language_model.resize_token_embeddings(len(tokenizer))
+        output_embeddings = model.language_model.get_output_embeddings().weight.data
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+        model.config.llm_config.vocab_size = len(tokenizer)
+        model.language_model.config.vocab_size = len(tokenizer)
+
+    model.language_model.config.use_cache = False
+    model.vision_model.gradient_checkpointing = True
+    model.vision_model.encoder.gradient_checkpointing = True
+    if model_args.grad_checkpoint:
+        model.language_model._set_gradient_checkpointing()
+
+    train_dataset = build_datasets(
+        data_args, tokenizer, tcs_loader, model, group_by_length=training_args.group_by_length,
+        dynamic_image_size=data_args.dynamic_image_size, use_thumbnail=data_args.use_thumbnail,
+        min_dynamic_patch=data_args.min_dynamic_patch, max_dynamic_patch=data_args.max_dynamic_patch,
+        normalize_type=data_args.normalize_type)
+
+    def _freeze_params(module):
+        for param in module.parameters():
+            param.requires_grad = False
+
+    if model_args.freeze_backbone:
+        # model.vision_model = model.vision_model.eval()
+        _freeze_params(model.vision_model)
+
+    if model_args.freeze_llm:
+        model.language_model = model.language_model.eval()
+        _freeze_params(model.language_model)
+
+    if model_args.unfreeze_lm_head:
+        model.language_model.lm_head.requires_grad = True
+
+    if model_args.use_backbone_lora:
+        model.wrap_backbone_lora(r=model_args.use_backbone_lora, lora_alpha=2 * model_args.use_backbone_lora)
+        model.config.use_backbone_lora = model_args.use_backbone_lora
+
+    if model_args.use_llm_lora:
+        model.wrap_llm_lora(r=model_args.use_llm_lora, lora_alpha=2 * model_args.use_llm_lora)
+        model.config.use_llm_lora = model_args.use_llm_lora
+
+    if model_args.freeze_mlp:
+        _freeze_params(model.mlp1)
+
+    if model_args.unfreeze_vit_layers != 0:
+        layers = model.vision_model.encoder.layers[model_args.unfreeze_vit_layers:]
+        for k, v in layers.named_parameters():
+            logger.info(f'Unfreezing ViT layer: {k}')
+            v.requires_grad = True
+
+    # print trainable parameters
+    if dist.get_rank() == 0:
+        for name, param in model.named_parameters():
+            if param.requires_grad:
+                logger.info(name)
+
+    # set seed for torch dataloaders
+    set_seed(training_args.seed)
+
+    # Initialize our Trainer
+    if model_args.use_custom_trainer:
+        replace_create_optimizer()
+
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset if training_args.do_train else None,
+        eval_dataset=None,
+        tokenizer=tokenizer,
+        data_collator=concat_pad_data_collator
+    )
+
+    # Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+        metrics = train_result.metrics
+        try:
+            metrics['train_samples'] = len(train_dataset)
+        except:
+            metrics['train_samples'] = -1
+
+        trainer.log_metrics('train', metrics)
+        trainer.save_metrics('train', metrics)
+        trainer.save_state()
+
+
+if __name__ == '__main__':
+    main()

internvl/train/trainer_monkey_patch.py

+import json
+import os
+
+import torch
+import torch.nn as nn
+import transformers
+from transformers import Trainer, logging
+from transformers.trainer import is_sagemaker_mp_enabled
+
+logger = logging.get_logger(__name__)
+
+
+def get_num_layer_for_vit_and_qllama(var_name, vit_num_max_layer, llama_num_max_layer):
+    if var_name.startswith('internvl.'):
+        var_name = var_name[len('internvl.'):]
+    if var_name in ('query_tokens', 'logit_scale',):
+        return 0
+    if var_name.startswith('clip_projector.'):
+        return vit_num_max_layer
+    if var_name.startswith('clip_projector2.') or var_name.startswith('itm_head.') or \
+            var_name == 'text_projection':
+        return llama_num_max_layer
+    if var_name.startswith('vision_model.'):
+        if 'embeddings.' in var_name:
+            return 0
+        if 'layers.' in var_name:
+            var_name = var_name.split('layers.')[-1]
+            layer_id = int(var_name.split('.')[0])
+            return layer_id + 1
+    if var_name.startswith('qllama.'):
+        if 'embed_tokens' in var_name:
+            return 0
+        if 'layers.' in var_name:
+            var_name = var_name.split('layers.')[-1]
+            layer_id = int(var_name.split('.')[0])
+            return layer_id + 1
+        else:
+            return llama_num_max_layer
+    return 0
+
+
+def param_classification(name):
+    if name.startswith('internvl.'):
+        name = name[len('internvl.'):]
+    if name in ['query_tokens', 'text_projection', 'logit_scale']:
+        return 'qllama'
+    elif name.startswith('vision_model.'):
+        return 'vit'
+    elif name.startswith('qllama.'):
+        return 'qllama'
+    elif name.startswith('clip_projector.'):
+        return 'vit'
+    elif name.startswith('clip_projector2.'):
+        return 'qllama'
+    elif name.startswith('itm_head.'):
+        return 'qllama'
+    else:
+        return 'other'
+
+
+def create_optimizer(self):
+    """
+    Setup the optimizer.
+
+    We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
+    Trainer's init through `optimizers`, or subclass and override this method in a subclass.
+    """
+    opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
+
+    parameter_groups = {}
+    try:  # for stage2 model
+        vit_num_layers = opt_model.config.vision_config.num_hidden_layers + 2
+        qllama_num_layers = opt_model.config.qllama_config.num_hidden_layers + 2
+    except:  # for stage3 model
+        vit_num_layers = opt_model.internvl.config.vision_config.num_hidden_layers + 2
+        qllama_num_layers = opt_model.internvl.config.qllama_config.num_hidden_layers + 2
+    print('vit_num_layers:', vit_num_layers)
+    print('qllama_num_layers:', qllama_num_layers)
+
+    vit_layer_decay_rate = float(os.getenv('VIT_LAYER_DECAY_RATE', 1.0))
+    qllama_layer_decay_rate = float(os.getenv('QLLAMA_LAYER_DECAY_RATE', 1.0))
+    qllama_lr_scale = float(os.getenv('QLLAMA_LR_SCALE', 1.0))
+    print('vit_layer_decay_rate:', vit_layer_decay_rate)
+    print('qllama_layer_decay_rate:', qllama_layer_decay_rate)
+    print('qllama_lr_scale:', qllama_lr_scale)
+
+    for name, param in opt_model.named_parameters():
+        if not param.requires_grad:
+            continue  # frozen weights
+        if len(param.shape) == 1 or name.endswith('.bias'):
+            group_name = 'no_decay'
+            this_weight_decay = 0.
+        else:
+            group_name = 'decay'
+            this_weight_decay = self.args.weight_decay
+
+        cls = param_classification(name)
+        layer_id = get_num_layer_for_vit_and_qllama(name, vit_num_layers, qllama_num_layers)
+        group_name = '%s_layer_%d_%s' % (cls, layer_id, group_name)
+        if group_name not in parameter_groups:
+            if cls == 'vit':
+                scale = vit_layer_decay_rate ** (vit_num_layers - layer_id - 1)
+            elif cls == 'qllama':
+                scale = qllama_layer_decay_rate ** (qllama_num_layers - layer_id - 1)
+                scale = scale * qllama_lr_scale
+            else:
+                scale = 1.0
+            scale = min(1.0, scale)
+            parameter_groups[group_name] = {
+                'weight_decay': this_weight_decay,
+                'params': [],
+                'param_names': [],
+                'lr_scale': scale,
+                'group_name': group_name,
+                'lr': scale * self.args.learning_rate,
+            }
+        parameter_groups[group_name]['params'].append(param)
+        parameter_groups[group_name]['param_names'].append(name)
+
+        rank = torch.distributed.get_rank()
+        if rank == 0:
+            to_display = {}
+            for key in parameter_groups:
+                to_display[key] = {
+                    'param_names': parameter_groups[key]['param_names'],
+                    'lr_scale': parameter_groups[key]['lr_scale'],
+                    'lr': parameter_groups[key]['lr'],
+                    'weight_decay': parameter_groups[key]['weight_decay'],
+                }
+            print('Param groups = %s' % json.dumps(to_display, indent=2))
+
+    optimizer_grouped_parameters = list(parameter_groups.values())
+    optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
+
+    self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+    if optimizer_cls.__name__ == 'Adam8bit':
+        import bitsandbytes
+
+        manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+        skipped = 0
+        for module in opt_model.modules():
+            if isinstance(module, nn.Embedding):
+                skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                logger.info(f'skipped {module}: {skipped / 2 ** 20}M params')
+                manager.register_module_override(module, 'weight', {'optim_bits': 32})
+                logger.debug(f'bitsandbytes: will optimize {module} in fp32')
+        logger.info(f'skipped: {skipped / 2 ** 20}M params')
+
+    if is_sagemaker_mp_enabled():
+        import smdistributed.modelparallel.torch as smp
+        self.optimizer = smp.DistributedOptimizer(self.optimizer)
+
+    return self.optimizer
+
+
+def replace_create_optimizer():
+    print('Replace original create_optimizer with custom create_optimizer')
+    transformers.Trainer.create_optimizer = create_optimizer

internvl_chat.py

+import numpy as np
+import torch
+import torchvision.transforms as T
+from decord import VideoReader, cpu
+from PIL import Image
+from torchvision.transforms.functional import InterpolationMode
+from transformers import AutoModel, AutoTokenizer
+import os
+import math
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+
+def build_transform(input_size):
+    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
+    transform = T.Compose([
+        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+        T.ToTensor(),
+        T.Normalize(mean=MEAN, std=STD)
+    ])
+    return transform
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+
+def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+
+def load_image(image_file, input_size=448, max_num=6):
+    image = Image.open(image_file).convert('RGB')
+    transform = build_transform(input_size=input_size)
+    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
+    pixel_values = [transform(image) for image in images]
+    pixel_values = torch.stack(pixel_values)
+    return pixel_values
+
+def split_model(model_name):
+    device_map = {}
+    world_size = torch.cuda.device_count()
+    num_layers = {
+        'InternVL2-1B': 24, 'InternVL2-2B': 24, 'InternVL2-4B': 32, 'InternVL2-8B': 32,
+        'InternVL2-26B': 48, 'InternVL2-40B': 60, 'InternVL2-Llama3-76B': 80}[model_name]
+    # Since the first GPU will be used for ViT, treat it as half a GPU.
+    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
+    num_layers_per_gpu = [num_layers_per_gpu] * world_size
+    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
+    layer_cnt = 0
+    for i, num_layer in enumerate(num_layers_per_gpu):
+        for j in range(num_layer):
+            device_map[f'language_model.model.layers.{layer_cnt}'] = i
+            layer_cnt += 1
+    device_map['vision_model'] = 0
+    device_map['mlp1'] = 0
+    device_map['language_model.model.tok_embeddings'] = 0
+    device_map['language_model.model.embed_tokens'] = 0
+    device_map['language_model.output'] = 0
+    device_map['language_model.model.norm'] = 0
+    device_map['language_model.lm_head'] = 0
+    device_map[f'language_model.model.layers.{num_layers - 1}'] = 0
+
+    return device_map
+
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3,4,5,6,7"
+path = '/home/wanglch/projects/InternVL/InternVL2-40B'
+device_map = split_model('InternVL2-40B')
+model = AutoModel.from_pretrained(
+    path,
+    torch_dtype=torch.float16,
+    low_cpu_mem_usage=True,
+    trust_remote_code=True,
+    device_map=device_map).eval()
+tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
+
+# set the max number of tiles in `max_num`
+pixel_values = load_image('/home/wanglch/projects/InternVL/internvl_chat/examples/image1.jpg', max_num=12).to(torch.float16).cuda()
+generation_config = dict(max_new_tokens=1024, do_sample=True)
+
+# pure-text conversation (纯文本对话)
+question = 'Hello, who are you?'
+response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
+print(f'User: {question}')
+print(f'Assistant: {response}')
+
+question = 'Can you tell me a story?'
+response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
+print(f'User: {question}')
+print(f'Assistant: {response}')
+
+# single-image single-round conversation (单图单轮对话)
+question = '<image>\nPlease describe the image shortly.'
+response = model.chat(tokenizer, pixel_values, question, generation_config)
+print(f'User: {question}')
+print(f'Assistant: {response}')
+
+# single-image multi-round conversation (单图多轮对话)
+question = '<image>\nPlease describe the image in detail.'
+response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
+print(f'User: {question}')
+print(f'Assistant: {response}')
+
+question = 'Please write a poem according to the image.'
+response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
+print(f'User: {question}')
+print(f'Assistant: {response}')
+

internvl_chat/README.md

+# InternVL-Chat
+
+This folder contains the implementation of the InternVL-Chat.
+
+## 🛠️ Installation
+
+See [INSTALLATION.md](../INSTALLATION.md)
+
+In addition, using this codebase requires executing the following steps:
+
+- Install other requirements:
+
+  ```bash
+  pip install --upgrade pip  # enable PEP 660 support
+  pip install -e .
+  ```
+
+## 📖 Documents
+
+- InternVL 2.0
+
+  - Introduction [\[link\]](https://internvl.readthedocs.io/en/latest/internvl2.0/introduction.html)
+  - Quick Start [\[link\]](https://internvl.readthedocs.io/en/latest/internvl2.0/quick_start.html)
+  - Finetune [\[link\]](https://internvl.readthedocs.io/en/latest/internvl2.0/finetune.html)
+  - Evaluation [\[link\]](https://internvl.readthedocs.io/en/latest/internvl2.0/evaluation.html)
+  - Deployment [\[link\]](https://internvl.readthedocs.io/en/latest/internvl2.0/deployment.html)
+
+- InternVL 1.5
+
+  - Introduction [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.5/introduction.html)
+  - Quick Start [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.5/quick_start.html)
+  - Finetune [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.5/finetune.html)
+  - Evaluation [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.5/evaluation.html)
+  - Deployment [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.5/deployment.html)
+
+- InternVL 1.2
+
+  - Introduction [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.2/introduction.html)
+  - Quick Start [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.2/quick_start.html)
+  - Reproduce [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.2/reproduce.html)
+  - Finetune [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.2/finetune.html)
+  - Evaluation [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.2/evaluation.html)
+
+- InternVL 1.1
+
+  - Introduction [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.1/introduction.html)
+  - Quick Start [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.1/quick_start.html)
+  - Evaluation [\[link\]](https://internvl.readthedocs.io/en/latest/internvl1.1/evaluation.html)

internvl_chat/eval/caption/evaluate_caption.py

+import argparse
+import itertools
+import json
+import os
+import random
+import time
+from functools import partial
+
+import torch
+from internvl.model.internvl_chat import InternVLChatModel
+from internvl.train.dataset import build_transform, dynamic_preprocess
+from PIL import Image
+from pycocoevalcap.eval import COCOEvalCap
+from pycocotools.coco import COCO
+from tqdm import tqdm
+from transformers import AutoTokenizer
+
+ds_collections = {
+    'flickr30k': {
+        'root': 'data/flickr30k/',
+        'annotation': 'data/flickr30k/flickr30k_test_karpathy.json',
+        'max_new_tokens': 30,
+        'min_new_tokens': 8,
+    },
+    'coco': {
+        'root': 'data/coco/',
+        'annotation': ['data/coco/annotations/coco_karpathy_test.json',
+                       'data/coco/annotations/coco_karpathy_test_gt.json'],
+        'max_new_tokens': 30,
+        'min_new_tokens': 8,
+    },
+    'nocaps': {
+        'root': 'data/nocaps/images',
+        'annotation': 'data/nocaps/nocaps_val_4500_captions.json',
+        'max_new_tokens': 30,
+        'min_new_tokens': 8,
+    },
+}
+
+
+class CaptionDataset(torch.utils.data.Dataset):
+
+    def __init__(self, name, root, annotation, prompt, input_size=224, dynamic_image_size=False,
+                 use_thumbnail=False, max_num=6):
+        if name == 'coco':
+            self.images = json.load(open(annotation))
+        else:
+            self.images = json.load(open(annotation))['images']
+        self.name = name
+        self.prompt = prompt
+        self.root = root
+        self.input_size = input_size
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.max_num = max_num
+        self.transform = build_transform(is_train=False, input_size=input_size)
+
+    def __len__(self):
+        return len(self.images)
+
+    def __getitem__(self, idx):
+        if self.name == 'coco':
+            filename = self.images[idx]['image']
+            image_id = int(filename.split('_')[-1].replace('.jpg', ''))
+            image_path = os.path.join(self.root, filename)
+        else:
+            image_id = self.images[idx]['id']
+            if 'file_name' in self.images[idx]:
+                image_path = os.path.join(self.root, self.images[idx]['file_name'])
+            else:
+                image_path = os.path.join(self.root, self.images[idx]['image'])
+
+        image = Image.open(image_path)
+        if self.dynamic_image_size:
+            images = dynamic_preprocess(image, image_size=self.input_size,
+                                        use_thumbnail=self.use_thumbnail,
+                                        max_num=self.max_num)
+        else:
+            images = [image]
+        pixel_values = [self.transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+
+        return {
+            'image_id': image_id,
+            'input_text': self.prompt,
+            'pixel_values': pixel_values
+        }
+
+
+def collate_fn(inputs, tokenizer):
+    pixel_values = torch.cat([_['pixel_values'] for _ in inputs], dim=0)
+    image_ids = [_['image_id'] for _ in inputs]
+    input_texts = [_['input_text'] for _ in inputs]
+    input_tokens = tokenizer(input_texts, return_tensors='pt')
+
+    return pixel_values, image_ids, input_tokens.input_ids, input_tokens.attention_mask
+
+
+class InferenceSampler(torch.utils.data.sampler.Sampler):
+
+    def __init__(self, size):
+        self._size = int(size)
+        assert size > 0
+        self._rank = torch.distributed.get_rank()
+        self._world_size = torch.distributed.get_world_size()
+        self._local_indices = self._get_local_indices(size, self._world_size, self._rank)
+
+    @staticmethod
+    def _get_local_indices(total_size, world_size, rank):
+        shard_size = total_size // world_size
+        left = total_size % world_size
+        shard_sizes = [shard_size + int(r < left) for r in range(world_size)]
+
+        begin = sum(shard_sizes[:rank])
+        end = min(sum(shard_sizes[:rank + 1]), total_size)
+        return range(begin, end)
+
+    def __iter__(self):
+        yield from self._local_indices
+
+    def __len__(self):
+        return len(self._local_indices)
+
+
+def evaluate_chat_model():
+    prompt = 'Provide a one-sentence caption for the provided image.'
+    print('prompt:', prompt)
+    random.seed(args.seed)
+    summaries = []
+
+    for ds_name in args.datasets:
+        annotation = ds_collections[ds_name]['annotation']
+        if type(annotation) == list:
+            annotation = annotation[0]
+        dataset = CaptionDataset(
+            name=ds_name,
+            root=ds_collections[ds_name]['root'],
+            annotation=annotation,
+            prompt=prompt,
+            input_size=image_size,
+            dynamic_image_size=args.dynamic,
+            use_thumbnail=use_thumbnail,
+            max_num=args.max_num
+        )
+        dataloader = torch.utils.data.DataLoader(
+            dataset=dataset,
+            sampler=InferenceSampler(len(dataset)),
+            batch_size=args.batch_size,
+            num_workers=args.num_workers,
+            pin_memory=True,
+            drop_last=False,
+            collate_fn=partial(collate_fn, tokenizer=tokenizer),
+        )
+
+        image_ids, captions = [], []
+        for _, (pixel_values, ids, _, _) in tqdm(enumerate(dataloader)):
+            pixel_values = pixel_values.to(torch.bfloat16).cuda()
+            generation_config = dict(
+                num_beams=args.num_beams,
+                max_new_tokens=ds_collections[ds_name]['max_new_tokens'],
+                min_new_tokens=ds_collections[ds_name]['min_new_tokens'],
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+            )
+            pred = model.chat(
+                tokenizer=tokenizer,
+                pixel_values=pixel_values,
+                question=prompt,
+                generation_config=generation_config,
+                verbose=True
+            )
+            image_ids.extend(ids)
+            captions.extend([pred])
+
+        torch.distributed.barrier()
+
+        world_size = torch.distributed.get_world_size()
+        merged_ids = [None for _ in range(world_size)]
+        merged_captions = [None for _ in range(world_size)]
+        torch.distributed.all_gather_object(merged_ids, image_ids)
+        torch.distributed.all_gather_object(merged_captions, captions)
+
+        merged_ids = [_ for _ in itertools.chain.from_iterable(merged_ids)]
+        merged_captions = [_ for _ in itertools.chain.from_iterable(merged_captions)]
+        average_length = sum(len(x.split()) for x in merged_captions) / len(merged_captions)
+        print(f'Average caption length: {average_length}')
+
+        if torch.distributed.get_rank() == 0:
+            print(f'Evaluating {ds_name} ...')
+
+            results = []
+            for image_id, caption in zip(merged_ids, merged_captions):
+                results.append({
+                    'image_id': int(image_id),
+                    'caption': caption,
+                })
+            time_prefix = time.strftime('%y%m%d%H%M%S', time.localtime())
+            results_file = f'{ds_name}_{time_prefix}.json'
+            results_file = os.path.join(args.out_dir, results_file)
+            json.dump(results, open(results_file, 'w'))
+
+            annotation = ds_collections[ds_name]['annotation']
+            if type(annotation) == list:
+                annotation = annotation[-1]
+            coco = COCO(annotation)
+            coco_result = coco.loadRes(results_file)
+            coco_eval = COCOEvalCap(coco, coco_result)
+            coco_eval.evaluate()
+
+            summary = coco_eval.eval.items()
+            print(summary)
+            summaries.append([args.checkpoint, ds_name, average_length, summary])
+
+        torch.distributed.barrier()
+
+    out_path = '_'.join(args.checkpoint.split('/')[-2:])
+    writer = open(os.path.join(args.out_dir, f'{out_path}.txt'), 'a')
+    print(f"write results to file {os.path.join(args.out_dir, f'{out_path}.txt')}")
+    for summary in summaries:
+        print(summary)
+        writer.write(f'{summary}\n')
+    writer.close()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--checkpoint', type=str, default='')
+    parser.add_argument('--datasets', type=str, default='coco,flickr30k,nocaps')
+    parser.add_argument('--batch-size', type=int, default=1)
+    parser.add_argument('--num-workers', type=int, default=1)
+    parser.add_argument('--num-beams', type=int, default=5)
+    parser.add_argument('--temperature', type=float, default=0.0)
+    parser.add_argument('--out-dir', type=str, default='results')
+    parser.add_argument('--seed', type=int, default=0)
+    parser.add_argument('--dynamic', action='store_true')
+    parser.add_argument('--max-num', type=int, default=6)
+    parser.add_argument('--load-in-8bit', action='store_true')
+    parser.add_argument('--load-in-4bit', action='store_true')
+    parser.add_argument('--auto', action='store_true')
+    args = parser.parse_args()
+
+    if not os.path.exists(args.out_dir):
+        os.makedirs(args.out_dir)
+
+    args.datasets = args.datasets.split(',')
+    print('datasets:', args.datasets)
+    assert args.batch_size == 1, 'Only batch size 1 is supported'
+
+    torch.distributed.init_process_group(
+        backend='nccl',
+        world_size=int(os.getenv('WORLD_SIZE', '1')),
+        rank=int(os.getenv('RANK', '0')),
+    )
+
+    torch.cuda.set_device(int(os.getenv('LOCAL_RANK', 0)))
+
+    if args.auto:
+        os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
+    kwargs = {'device_map': 'auto'} if args.auto else {}
+    tokenizer = AutoTokenizer.from_pretrained(args.checkpoint, trust_remote_code=True, use_fast=False)
+    model = InternVLChatModel.from_pretrained(
+        args.checkpoint, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16,
+        load_in_8bit=args.load_in_8bit, load_in_4bit=args.load_in_4bit, **kwargs).eval()
+    if not args.load_in_8bit and not args.load_in_4bit and not args.auto:
+        model = model.cuda()
+    image_size = model.config.force_image_size or model.config.vision_config.image_size
+    use_thumbnail = model.config.use_thumbnail
+
+    total_params = sum(p.numel() for p in model.parameters()) / 1e9
+    if total_params > 20 or args.dynamic:
+        args.num_beams = 1
+        print(f'[test] total_params: {total_params}B, use num_beams: {args.num_beams}')
+    else:
+        print(f'[test] total_params: {total_params}B')
+    print(f'[test] image_size: {image_size}')
+    print(f'[test] template: {model.config.template}')
+    print(f'[test] dynamic_image_size: {args.dynamic}')
+    print(f'[test] use_thumbnail: {use_thumbnail}')
+    print(f'[test] max_num: {args.max_num}')
+
+    evaluate_chat_model()

internvl_chat/eval/cmmmu/evaluate_cmmmu.py

+import argparse
+import json
+import os
+import random
+
+import torch
+from internvl.model.internvl_chat import InternVLChatModel
+from internvl.train.dataset import build_transform, dynamic_preprocess
+from PIL import Image
+from tqdm import tqdm
+from transformers import AutoTokenizer
+
+ds_collections = {
+    'art_and_design': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_art_and_design.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+    'business': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_business.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+    'health_and_medicine': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_health_and_medicine.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+    'humanities_and_social_sciences': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_humanities_and_social_sciences.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+    'science': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_science.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+    'technology_and_engineering': {
+        'root': 'data/',
+        'annotation': 'data/cmmmu-data/llava_technology_and_engineering.jsonl',
+        'max_new_tokens': 999,
+        'min_new_tokens': 1,
+    },
+}
+
+
+class VQADataset(torch.utils.data.Dataset):
+
+    def __init__(self, root, annotation, input_size=224, dynamic_image_size=False,
+                 use_thumbnail=False, max_num=6):
+        self.root = root
+        self.items = []
+        f = open(annotation)
+        data = f.readlines()
+        for data_line in data:
+            data_line = json.loads(data_line)
+            self.items.append(data_line)
+        self.input_size = input_size
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.max_num = max_num
+        self.transform = build_transform(is_train=False, input_size=input_size)
+
+    def __len__(self):
+        return len(self.items)
+
+    def __getitem__(self, idx):
+        item = self.items[idx]
+        image_path, question = item['image'], item['text']
+        image_path = os.path.join(self.root, image_path)
+        image = Image.open(image_path).convert('RGB')
+        if self.dynamic_image_size:
+            images = dynamic_preprocess(image, image_size=self.input_size,
+                                        use_thumbnail=self.use_thumbnail,
+                                        max_num=self.max_num)
+        else:
+            images = [image]
+        pixel_values = [self.transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+
+        return {
+            'question': question,
+            'pixel_values': pixel_values,
+            'item': item,
+        }
+
+
+def evaluate_chat_model():
+    random.seed(args.seed)
+
+    for ds_name in args.datasets:
+        dataset = VQADataset(
+            root=ds_collections[ds_name]['root'],
+            annotation=ds_collections[ds_name]['annotation'],
+            input_size=image_size,
+            dynamic_image_size=args.dynamic,
+            use_thumbnail=use_thumbnail,
+            max_num=args.max_num
+        )
+
+        print(f'Evaluating {ds_name} ...')
+        results_file = f'{model_id}_{ds_name}.jsonl'
+        results_file = os.path.join(args.out_dir, results_file)
+        writer = open(results_file, 'w')
+
+        for _, data in tqdm(enumerate(dataset)):
+            pixel_value = data['pixel_values']
+            question = data['question']
+            item = data['item']
+            pixel_value = pixel_value.to(torch.bfloat16).cuda()
+            generation_config = dict(
+                num_beams=args.num_beams,
+                max_new_tokens=ds_collections[ds_name]['max_new_tokens'],
+                min_new_tokens=ds_collections[ds_name]['min_new_tokens'],
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+            )
+            pred = model.chat(
+                tokenizer=tokenizer,
+                pixel_values=pixel_value,
+                question=question,
+                generation_config=generation_config,
+                verbose=True
+            )
+            question_id = item['question_id']
+            text = item['text']
+            output = {
+                'question_id': question_id,
+                'prompt': text,
+                'text': pred,
+                'model_id': model_id,
+                'metadata': {}
+            }
+            writer.write(json.dumps(output, ensure_ascii=False) + '\n')
+            writer.flush()
+        print('Results saved to {}'.format(results_file))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--checkpoint', type=str, default='')
+    parser.add_argument('--datasets', type=str, default='')
+    parser.add_argument('--batch-size', type=int, default=1)
+    parser.add_argument('--num-workers', type=int, default=1)
+    parser.add_argument('--num-beams', type=int, default=1)
+    parser.add_argument('--temperature', type=float, default=0.0)
+    parser.add_argument('--out-dir', type=str, default='results')
+    parser.add_argument('--seed', type=int, default=0)
+    parser.add_argument('--dynamic', action='store_true')
+    parser.add_argument('--max-num', type=int, default=6)
+    parser.add_argument('--load-in-8bit', action='store_true')
+    parser.add_argument('--load-in-4bit', action='store_true')
+    parser.add_argument('--auto', action='store_true')
+    args = parser.parse_args()
+
+    if not os.path.exists(args.out_dir):
+        os.makedirs(args.out_dir)
+
+    args.datasets = args.datasets.split(',')
+    print('datasets:', args.datasets)
+    assert args.batch_size == 1, 'Only batch size 1 is supported'
+
+    if args.auto:
+        os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
+    kwargs = {'device_map': 'auto'} if args.auto else {}
+    tokenizer = AutoTokenizer.from_pretrained(args.checkpoint, trust_remote_code=True, use_fast=False)
+    model = InternVLChatModel.from_pretrained(
+        args.checkpoint, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16,
+        load_in_8bit=args.load_in_8bit, load_in_4bit=args.load_in_4bit, **kwargs).eval()
+    if not args.load_in_8bit and not args.load_in_4bit and not args.auto:
+        model = model.cuda()
+    image_size = model.config.force_image_size or model.config.vision_config.image_size
+    use_thumbnail = model.config.use_thumbnail
+
+    total_params = sum(p.numel() for p in model.parameters()) / 1e9
+    if total_params > 20 or args.dynamic:
+        args.num_beams = 1
+        print(f'[test] total_params: {total_params}B, use num_beams: {args.num_beams}')
+    else:
+        print(f'[test] total_params: {total_params}B')
+    print(f'[test] image_size: {image_size}')
+    print(f'[test] template: {model.config.template}')
+    print(f'[test] dynamic_image_size: {args.dynamic}')
+    print(f'[test] use_thumbnail: {use_thumbnail}')
+    print(f'[test] max_num: {args.max_num}')
+
+    model_id = '_'.join(args.checkpoint.split('/')[-2:])
+    evaluate_chat_model()

internvl_chat/eval/llava_bench/eval_gpt_review_bench.py

+import argparse
+import json
+import os
+import time
+
+import openai
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            completion = openai.chat.completions.create(
+                model='gpt-4-0613',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return completion.choices[0].message.content
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+        print(ques, ans1, ans2)
+        inst = image_to_context[ques['image']]
+
+        if isinstance(inst['caption'], list):
+            cap_str = '\n'.join(inst['caption'])
+        else:
+            cap_str = inst['caption']
+
+        category = 'llava_bench_' + json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f'Visual QA category not found in rule file: {category}.'
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx + 1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans1['question_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

internvl_chat/eval/llava_bench/evaluate_llava_bench.py

+import argparse
+import json
+import os
+import random
+
+import torch
+from internvl.model.internvl_chat import InternVLChatModel
+from internvl.train.dataset import build_transform, dynamic_preprocess
+from PIL import Image
+from tqdm import tqdm
+from transformers import AutoTokenizer
+
+ds_collections = {
+    'llava_bench': {
+        'root': 'data/llava-bench-in-the-wild/images',
+        'question': 'data/llava-bench-in-the-wild/questions.jsonl',
+        'max_new_tokens': 1000,
+        'min_new_tokens': 1,
+    },
+}
+
+
+class VQADataset(torch.utils.data.Dataset):
+
+    def __init__(self, root, data, prompt, input_size=224, dynamic_image_size=False,
+                 use_thumbnail=False, max_num=6):
+        self.root = root
+        self.data = open(data).readlines()
+        self.prompt = prompt
+        self.input_size = input_size
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.max_num = max_num
+        self.transform = build_transform(is_train=False, input_size=input_size)
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        data = json.loads(self.data[idx].strip())
+        image, question, question_id, annotation = data['image'], data[
+            'text'], data['question_id'], data.get('answer', None)
+
+        image = os.path.join(self.root, image)
+        image = Image.open(image).convert('RGB')
+        if self.dynamic_image_size:
+            images = dynamic_preprocess(image, image_size=self.input_size,
+                                        use_thumbnail=self.use_thumbnail,
+                                        max_num=self.max_num)
+        else:
+            images = [image]
+        pixel_values = [self.transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        question = question + self.prompt
+        return question_id, question, pixel_values, annotation
+
+
+def evaluate_chat_model():
+    random.seed(args.seed)
+
+    for ds_name in args.datasets:
+        dataset = VQADataset(
+            root=ds_collections[ds_name]['root'],
+            data=ds_collections[ds_name]['question'],
+            prompt=' Please give a detailed answer.',
+            input_size=image_size,
+            dynamic_image_size=args.dynamic,
+            use_thumbnail=use_thumbnail,
+            max_num=args.max_num
+        )
+
+        outputs = []
+        for _, (question_id, question, pixel_values, annotations) in tqdm(enumerate(dataset)):
+            pixel_values = pixel_values.to(torch.bfloat16).cuda()
+            generation_config = dict(
+                num_beams=args.num_beams,
+                max_new_tokens=ds_collections[ds_name]['max_new_tokens'],
+                min_new_tokens=ds_collections[ds_name]['min_new_tokens'],
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+            )
+            pred = model.chat(
+                tokenizer=tokenizer,
+                pixel_values=pixel_values,
+                question=question,
+                generation_config=generation_config,
+                verbose=True
+            )
+            outputs.append({
+                'question_id': question_id,
+                'text': pred,
+                'model_id': model_id,
+                'metadata': {}
+            })
+
+        print(f'Evaluating {ds_name} ...')
+        results_file = 'llava_bench_results.jsonl'
+        results_file = os.path.join(args.out_dir, results_file)
+        writer = open(results_file, 'w')
+        for item in outputs:
+            writer.write(json.dumps(item) + '\n')
+        writer.close()
+        print('Results saved to {}'.format(results_file))
+
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--checkpoint', type=str, default='')
+    parser.add_argument('--datasets', type=str, default='llava_bench')
+    parser.add_argument('--batch-size', type=int, default=1)
+    parser.add_argument('--num-workers', type=int, default=1)
+    parser.add_argument('--num-beams', type=int, default=5)
+    parser.add_argument('--temperature', type=float, default=0.0)
+    parser.add_argument('--out-dir', type=str, default='results')
+    parser.add_argument('--seed', type=int, default=0)
+    parser.add_argument('--dynamic', action='store_true')
+    parser.add_argument('--max-num', type=int, default=6)
+    parser.add_argument('--load-in-8bit', action='store_true')
+    parser.add_argument('--load-in-4bit', action='store_true')
+    parser.add_argument('--auto', action='store_true')
+    args = parser.parse_args()
+
+    if not os.path.exists(args.out_dir):
+        os.makedirs(args.out_dir)
+
+    args.datasets = args.datasets.split(',')
+    print('datasets:', args.datasets)
+    assert args.batch_size == 1, 'Only batch size 1 is supported'
+
+    if args.auto:
+        os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
+    kwargs = {'device_map': 'auto'} if args.auto else {}
+    tokenizer = AutoTokenizer.from_pretrained(args.checkpoint, trust_remote_code=True, use_fast=False)
+    model = InternVLChatModel.from_pretrained(
+        args.checkpoint, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16,
+        load_in_8bit=args.load_in_8bit, load_in_4bit=args.load_in_4bit, **kwargs).eval()
+    if not args.load_in_8bit and not args.load_in_4bit and not args.auto:
+        model = model.cuda()
+    image_size = model.config.force_image_size or model.config.vision_config.image_size
+    use_thumbnail = model.config.use_thumbnail
+
+    total_params = sum(p.numel() for p in model.parameters()) / 1e9
+    if total_params > 20 or args.dynamic:
+        args.num_beams = 1
+        print(f'[test] total_params: {total_params}B, use num_beams: {args.num_beams}')
+    else:
+        print(f'[test] total_params: {total_params}B')
+    print(f'[test] image_size: {image_size}')
+    print(f'[test] template: {model.config.template}')
+    print(f'[test] dynamic_image_size: {args.dynamic}')
+    print(f'[test] use_thumbnail: {use_thumbnail}')
+    print(f'[test] max_num: {args.max_num}')
+
+    model_id = '_'.join(args.checkpoint.split('/')[-2:])
+    evaluate_chat_model()