Upload folder using huggingface_hub

.ipynb_checkpoints/modeling_dbrx-checkpoint.py

@@ -0,0 +1,1455 @@
+"""PyTorch Dbrx model."""
+
+import math
+import warnings
+from copy import deepcopy
+from functools import partial
+from typing import Any, Callable, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_outputs import (MoeCausalLMOutputWithPast,
+                                           MoeModelOutputWithPast)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import is_flash_attn_2_available, logging
+
+from .configuration_dbrx import DbrxAttentionConfig, DbrxConfig, DbrxFFNConfig
+
+if is_flash_attn_2_available():
+    try:
+        from flash_attn import flash_attn_func, flash_attn_varlen_func
+        from flash_attn.bert_padding import pad_input  # noqa
+        from flash_attn.bert_padding import index_first_axis, unpad_input
+    except:
+        pass
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = 'DbrxConfig'
+
+#############################################################################
+# Copied from LLaMaRotaryEmbedding
+#############################################################################
+
+
+class DbrxRotaryEmbedding(nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 max_position_embeddings: int = 2048,
+                 base: float = 10000.0,
+                 scaling_factor: float = 1.0):
+        super().__init__()
+        self.scaling_factor = scaling_factor
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base**(
+            torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
+        self.register_buffer('inv_freq', inv_freq, persistent=False)
+        # For BC we register cos and sin cached
+        self.max_seq_len_cached = max_position_embeddings
+
+    @torch.no_grad()
+    def forward(
+            self, x: torch.Tensor, position_ids: torch.LongTensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        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)
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., :x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+        unsqueeze_dim: int = 1) -> Tuple[torch.Tensor, torch.Tensor]:
+    """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.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos and
+            sin so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos and sin 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 and sin 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
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """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)
+
+
+#############################################################################
+
+#############################################################################
+# Modified from modeling_mixtral
+#############################################################################
+
+
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor,
+    num_experts: int,
+    top_k: int,
+    attention_mask: Optional[torch.Tensor],
+) -> torch.Tensor:
+    r"""Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts (`int`):
+            Number of experts.
+        top_k (`int`):
+            The number of experts each token is routed to.
+        attention_mask (`torch.Tensor`, None):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return torch.tensor(0.0)
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat(
+            [layer_gate.to(compute_device) for layer_gate in gate_logits],
+            dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits,
+                                                  dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (
+            batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (attention_mask[None, :, :, None, None].expand(
+            (num_hidden_layers, batch_size, sequence_length, top_k,
+             num_experts)).reshape(-1, top_k, num_experts).to(compute_device))
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(
+            expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+                expert_attention_mask, dim=0)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None].expand(
+                (num_hidden_layers, batch_size, sequence_length,
+                 num_experts)).reshape(-1, num_experts).to(compute_device))
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(
+            routing_weights * router_per_expert_attention_mask,
+            dim=0) / torch.sum(router_per_expert_attention_mask, dim=0)
+
+    overall_loss = torch.sum(tokens_per_expert *
+                             router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+#############################################################################
+
+
+def resolve_ffn_act_fn(
+        ffn_act_fn: dict) -> Callable[[torch.Tensor], torch.Tensor]:
+    """Resolve the activation function for the feed-forward network.
+
+    Args:
+        ffn_act_fn (dict): The configuration dictionary for the activation function.
+            The dict config must specify the 'name' of a torch.nn.functional activation
+            function. All of other key values pairs are bound to the function as a partial.
+
+    Returns:
+        Callable[[torch.Tensor], torch.Tensor]: The activation function.
+    """
+    config = deepcopy(ffn_act_fn)
+    name = config.pop('name')
+    if not hasattr(nn.functional, name):
+        raise ValueError(f'Unrecognised activation function name ({name}).')
+    act = getattr(nn.functional, name)
+    return partial(act, **config)
+
+
+#############################################################################
+# Copied from LLaMaAttention
+#############################################################################
+
+
+def _get_unpad_data(attention_mask: torch.Tensor):
+    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,
+    )
+
+
+class DbrxAttention(nn.Module):
+    """Multi-head self attention."""
+
+    def __init__(self,
+                 hidden_size: int,
+                 num_heads: int,
+                 max_position_embeddings: int,
+                 attn_config: DbrxAttentionConfig,
+                 block_idx: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.block_idx = block_idx
+        self.config = attn_config
+        if block_idx is None:
+            logger.warning_once(
+                f'Instantiating {self.__class__.__name__} without passing a `block_idx` is not recommended and will '
+                +
+                'lead to errors during the forward call if caching is used. Please make sure to provide a `block_idx` '
+                + 'when creating this class.')
+
+        self.attn_pdrop = attn_config.attn_pdrop
+        self.clip_qkv = attn_config.clip_qkv
+        self.num_key_value_heads = attn_config.kv_n_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.rope_theta = attn_config.rope_theta
+
+        self.Wqkv = nn.Linear(self.hidden_size,
+                              self.hidden_size +
+                              2 * self.num_key_value_heads * self.head_dim,
+                              bias=False)
+        self.out_proj = nn.Linear(self.hidden_size,
+                                  self.hidden_size,
+                                  bias=False)
+        self.rotary_emb = DbrxRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv_states = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        query_states, key_states, value_states = qkv_states.split(
+            [
+                self.hidden_size,
+                self.num_key_value_heads * self.head_dim,
+                self.num_key_value_heads * self.head_dim,
+            ],
+            dim=2,
+        )
+
+        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)
+
+        past_key_value = getattr(self, 'past_key_value', past_key_value)
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {
+                'sin': sin,
+                'cos': cos,
+                'cache_position': cache_position
+            }
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.block_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)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(
+            2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, :key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights,
+                                             dim=-1,
+                                             dtype=torch.float32).to(
+                                                 query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights,
+                                             p=self.attn_pdrop,
+                                             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.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class DbrxFlashAttention2(DbrxAttention):
+    """Dbrx flash attention module.
+
+    This module inherits from `DbrxAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it
+    calls the public API of flash attention.
+    """
+
+    def __init__(self, *args: Any, **kwargs: Any):
+        if not is_flash_attn_2_available():
+            raise ImportError(
+                'Flash Attention 2 is not available. Please install it with `pip install flash-attn`.'
+            )
+
+        super().__init__(*args, **kwargs)
+
+    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,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        logger.info(
+            'Implicitly setting `output_attentions` to False as it is not supported in Flash Attention.'
+        )
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv_states = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+        query_states, key_states, value_states = qkv_states.split(
+            [
+                self.hidden_size,
+                self.num_key_value_heads * self.head_dim,
+                self.num_key_value_heads * self.head_dim,
+            ],
+            dim=2,
+        )
+
+        # 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)
+
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states, cos, sin)
+
+        past_key_value = getattr(self, 'past_key_value', past_key_value)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {
+                'sin': sin,
+                'cos': cos,
+                'cache_position': cache_position
+            }
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.block_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout
+        # [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attn_pdrop 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. (LlamaRMSNorm handles it correctly)
+        input_dtype = query_states.dtype
+        if input_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 = query_states.dtype
+
+            logger.warning_once(
+                f'The input hidden states seems to be silently casted in float32, this might be '
+                +
+                f'related to the fact you have upcasted embedding or layer norm layers in '
+                + f'float32. We will cast back the input in {target_dtype}.')
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len,
+                                          self.hidden_size).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value  # type: ignore
+
+    def _flash_attention_forward(
+        self,
+        query_states: torch.Tensor,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        attention_mask: Union[torch.LongTensor, None],
+        query_length: int,
+        dropout: float = 0.0,
+        softmax_scale: Optional[float] = None,
+    ):
+        """Use FlashAttention, stripping padding tokens if necessary.
+
+        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.LongTensor | None): 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.
+            query_length (int): The length of the query sequence
+            dropout (float): Attention dropout
+            softmax_scale (float, optional): The scaling of QK^T before applying softmax.
+                Defaults to 1 / sqrt(head_dim)
+        """
+        causal = True
+        # 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
+
+            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,
+            )
+
+            attn_output = pad_input(
+                attn_output_unpad,
+                indices_q,
+                batch_size,
+                query_length,
+            )
+        else:
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states,
+                dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+        return attn_output
+
+    def _upad_input(self, query_layer: torch.Tensor, key_layer: torch.Tensor,
+                    value_layer: torch.Tensor, attention_mask: torch.Tensor,
+                    query_length: int):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(
+            attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads,
+                              head_dim), indices_k)
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads,
+                                head_dim), indices_k)
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.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),
+        )
+
+
+DBRX_ATTENTION_CLASSES = {
+    'eager': DbrxAttention,
+    'flash_attention_2': DbrxFlashAttention2,
+}
+
+
+class DbrxNormAttentionNorm(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        max_position_embeddings: int,
+        resid_pdrop: float,
+        attn_implementation: str,
+        attn_config: DbrxAttentionConfig,
+        block_idx: Optional[int] = None,
+    ):
+        super().__init__()
+        self.block_idx = block_idx
+        self.resid_pdrop = resid_pdrop
+        self.norm_1 = nn.LayerNorm(hidden_size, bias=False)
+        self.attn = DBRX_ATTENTION_CLASSES[attn_implementation](
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            max_position_embeddings=max_position_embeddings,
+            attn_config=attn_config,
+            block_idx=block_idx,
+        )
+        self.norm_2 = nn.LayerNorm(hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[Cache]]:
+
+        residual_states = hidden_states
+        hidden_states = self.norm_1(hidden_states).to(hidden_states.dtype)
+
+        hidden_states, attn_weights, past_key_value = 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,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states,
+                                              p=self.resid_pdrop,
+                                              training=self.training)
+        hidden_states = hidden_states + residual_states
+
+        residual_states = hidden_states
+        hidden_states = self.norm_2(hidden_states).to(hidden_states.dtype)
+
+        return residual_states, hidden_states, attn_weights, past_key_value
+
+
+class DbrxRouter(nn.Module):
+
+    def __init__(self, hidden_size: int, moe_num_experts: int, moe_top_k: int,
+                 moe_jitter_eps: Optional[float],
+                 moe_normalize_expert_weights: Optional[float],
+                 uniform_expert_assignment: bool):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_normalize_expert_weights = moe_normalize_expert_weights
+        self.uniform_expert_assignment = uniform_expert_assignment
+
+        self.layer = nn.Linear(self.hidden_size,
+                               self.moe_num_experts,
+                               bias=False)
+
+    def jitter(self, x: torch.Tensor) -> torch.Tensor:
+        if self.moe_jitter_eps is None:
+            raise RuntimeError('The router does not have moe_jitter_eps set.')
+        low = 1.0 - self.moe_jitter_eps
+        high = 1.0 + self.moe_jitter_eps
+        noise = torch.rand(x.size(), dtype=x.dtype, device=x.device)
+        return low + noise * (high - low)
+
+    def forward(
+            self, x: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
+        if self.training and self.moe_jitter_eps is not None:
+            x = x * self.jitter(x)
+
+        weights = self.layer(x.view(-1,
+                                    x.shape[-1])).softmax(dim=-1,
+                                                          dtype=torch.float32)
+        top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1)
+
+        if self.moe_normalize_expert_weights:
+            top_weights = top_weights / torch.norm(
+                top_weights,
+                p=self.moe_normalize_expert_weights,
+                dim=-1,
+                keepdim=True)
+
+        if self.uniform_expert_assignment:
+            with torch.no_grad():
+                uniform_tensor = torch.arange(
+                    0,
+                    top_experts.numel(),
+                    device=top_experts.device,
+                    dtype=top_experts.dtype) % self.moe_num_experts
+                top_experts = uniform_tensor.reshape(top_experts.shape)
+                # Note, weights and top_weights are not changed
+
+        weights = weights.to(x.dtype)
+        top_weights = top_weights.to(x.dtype)
+        return weights, top_weights, top_experts  # type: ignore
+
+
+class DbrxMLP(nn.Module):
+
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, ffn_act_fn: dict):
+        super().__init__()
+
+        self.w1 = nn.Linear(hidden_size, ffn_hidden_size, bias=False)
+        self.v1 = nn.Linear(hidden_size, ffn_hidden_size, bias=False)
+        self.w2 = nn.Linear(ffn_hidden_size, hidden_size, bias=False)
+        self.activation_fn = resolve_ffn_act_fn(ffn_act_fn)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+
+        return self.w2(self.activation_fn(self.w1(x)) * self.v1(x))
+
+
+class DbrxExperts(nn.Module):
+
+    def __init__(self, hidden_size: int, ffn_hidden_size: int,
+                 moe_num_experts: int, ffn_act_fn: dict):
+        super().__init__()
+        self.moe_num_experts = moe_num_experts
+        self.mlp = nn.ModuleList([DbrxMLP(hidden_size, ffn_hidden_size, ffn_act_fn) for _ in range(moe_num_experts)])
+
+    def forward(self, x: torch.Tensor, weights: torch.Tensor,
+                top_weights: torch.Tensor,
+                top_experts: torch.LongTensor) -> torch.Tensor:
+        bsz, q_len, hidden_size = x.shape
+        x = x.view(-1, hidden_size)
+        out = torch.zeros_like(x)
+
+        expert_mask = nn.functional.one_hot(
+            top_experts, num_classes=self.moe_num_experts).permute(2, 1, 0)
+        for expert_idx in range(0, self.moe_num_experts):
+            topk_idx, token_idx = torch.where(expert_mask[expert_idx])
+            if token_idx.shape[0] == 0:
+                continue
+
+            token_list = token_idx.tolist()
+            topk_list = topk_idx.tolist()
+
+            expert_tokens = x[None, token_list].reshape(-1, hidden_size)
+            expert_out = self.mlp[expert_idx](expert_tokens) * top_weights[token_list, topk_list, None]
+
+            out.index_add_(0, token_idx, expert_out)
+
+        out = out.reshape(bsz, q_len, hidden_size)
+        return out
+
+
+class DbrxFFN(nn.Module):
+
+    def __init__(self, hidden_size: int, ffn_config: DbrxFFNConfig):
+        super().__init__()
+
+        self.router = DbrxRouter(
+            hidden_size,
+            moe_num_experts=ffn_config.moe_num_experts,
+            moe_top_k=ffn_config.moe_top_k,
+            moe_jitter_eps=ffn_config.moe_jitter_eps,
+            moe_normalize_expert_weights=ffn_config.
+            moe_normalize_expert_weights,
+            uniform_expert_assignment=ffn_config.uniform_expert_assignment,
+        )
+
+        self.experts = DbrxExperts(
+            hidden_size=hidden_size,
+            ffn_hidden_size=ffn_config.ffn_hidden_size,
+            moe_num_experts=ffn_config.moe_num_experts,
+            ffn_act_fn=ffn_config.ffn_act_fn,
+        )
+
+    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        weights, top_weights, top_experts = self.router(x)
+        out = self.experts(x, weights, top_weights, top_experts)
+        return out, weights
+
+
+class DbrxBlock(nn.Module):
+
+    def __init__(self, config: DbrxConfig, block_idx: int):
+        super().__init__()
+        self.hidden_size = config.d_model
+        self.resid_pdrop = config.resid_pdrop
+        self.block_idx = block_idx
+        self.norm_attn_norm = DbrxNormAttentionNorm(
+            hidden_size=config.d_model,
+            num_heads=config.n_heads,
+            max_position_embeddings=config.max_seq_len,
+            resid_pdrop=config.resid_pdrop,
+            attn_implementation=config._attn_implementation,
+            attn_config=config.attn_config,
+            block_idx=block_idx,
+        )
+        self.ffn = DbrxFFN(hidden_size=config.d_model,
+                           ffn_config=config.ffn_config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Any,
+    ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, Optional[torch.Tensor]],
+               Tuple[torch.Tensor, Optional[Cache]], Tuple[
+                   torch.Tensor, Optional[torch.Tensor], Optional[Cache]],
+               Tuple[torch.Tensor, Optional[torch.Tensor],
+                     Optional[torch.Tensor]], Tuple[
+                         torch.Tensor, Optional[Cache], Optional[torch.Tensor]],
+               Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache],
+                     Optional[torch.Tensor]],]:
+        """Forward function for DbrxBlock.
+
+        Args:
+            hidden_states (`torch.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            position_ids (`torch.LongTensor`): position ids of shape `(batch, seq_len)`
+            attention_mask (`torch.Tensor`, optional): attention mask of size (batch_size, sequence_length)
+                if flash attention is used or (batch_size, 1, query_sequence_length, key_sequence_length)
+                if default attention is used.
+            past_key_value (`Tuple(torch.Tensor)`, optional): cached past key and value projection states
+            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_router_logits (`bool`, optional): Whether or not to return the router logits.
+            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`).
+            cache_position (`torch.LongTensor`, optional): position ids of the cache
+        """
+        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.`'
+            )
+
+        # Norm + Attention + Norm
+        resid_states, hidden_states, self_attn_weights, present_key_value = self.norm_attn_norm(
+            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,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        # Fully Connected
+        hidden_states, router_logits = self.ffn(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states,
+                                              p=self.resid_pdrop,
+                                              training=self.training)
+        hidden_states = resid_states + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+class DbrxPreTrainedModel(PreTrainedModel):
+    config_class = DbrxConfig
+    base_model_prefix = 'transformer'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['DbrxBlock']
+    _skip_keys_device_placement = ['past_key_values']
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+    _supports_cache_class = True
+
+    def _init_weights(self, module: nn.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_()
+        elif isinstance(module, nn.LayerNorm):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+    def _setup_cache(self, cache_cls: Any, max_batch_size: int,
+                     max_cache_len: int):  # TODO: how to set var type of class?
+        if self.config._attn_implementation == 'flash_attention_2' and cache_cls == StaticCache:
+            raise ValueError(
+                '`static` cache implementation is not compatible with ' +
+                '`attn_implementation==flash_attention_2`. Make sure to use ' +
+                '`spda` in the mean time and open an issue at https://github.com/huggingface/transformers.'
+            )
+
+        for block in self.transformer.blocks:
+            device = block.norm_attn_norm.norm_1.weight.device
+            if hasattr(self.config, '_pre_quantization_dtype'):
+                dtype = self.config._pre_quantization_dtype
+            else:
+                dtype = block.norm_attn_norm.attn.out_proj.weight.dtype
+            block.norm_attn_norm.attn.past_key_value = cache_cls(self.config,
+                                                                 max_batch_size,
+                                                                 max_cache_len,
+                                                                 device=device,
+                                                                 dtype=dtype)
+
+    def _reset_cache(self):
+        for block in self.transformer.blocks:
+            block.norm_attn_norm.attn.past_key_value = None
+
+
+class DbrxModel(DbrxPreTrainedModel):
+    """Transformer decoder consisting of *config.num_hidden_layers*
+
+    [`DbrxBlock`] layers.
+
+    Args:
+        config: DbrxConfig
+    """
+
+    def __init__(self, config: DbrxConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.emb_pdrop = config.emb_pdrop
+
+        self.wte = nn.Embedding(config.vocab_size, config.d_model,
+                                self.padding_idx)
+        self.blocks = nn.ModuleList([
+            DbrxBlock(config, block_idx) for block_idx in range(config.n_layers)
+        ])
+        self.norm_f = nn.LayerNorm(config.d_model, bias=False)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.wte
+
+    def set_input_embeddings(self, value: nn.Embedding):
+        self.wte = value
+
+    def _autocast_input_embeddings(self,
+                                   inputs_embeds: torch.Tensor) -> torch.Tensor:
+        if inputs_embeds.device.type == 'cuda' and torch.is_autocast_enabled():
+            return inputs_embeds.to(dtype=torch.get_autocast_gpu_dtype())
+        elif inputs_embeds.device.type == 'cpu' and torch.is_autocast_cpu_enabled(
+        ):
+            return inputs_embeds.to(dtype=torch.get_autocast_cpu_dtype())
+        else:
+            return inputs_embeds
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        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)
+        output_router_logits = (output_router_logits
+                                if output_router_logits is not None else
+                                self.config.output_router_logits)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                'You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one'
+            )
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                '`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.'
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        inputs_embeds = self._autocast_input_embeddings(
+            inputs_embeds)  # type: ignore
+        inputs_embeds = nn.functional.dropout(inputs_embeds,
+                                              p=self.emb_pdrop,
+                                              training=self.training)
+
+        past_seen_tokens = 0
+        if use_cache:  # kept for BC (cache positions)
+            if not isinstance(past_key_values, StaticCache):
+                past_key_values = DynamicCache.from_legacy_cache(
+                    past_key_values)
+                past_seen_tokens = past_key_values.get_seq_length(  # type: ignore
+                )
+
+        if cache_position is None:
+            if isinstance(past_key_values, StaticCache):
+                raise ValueError(
+                    'cache_position is a required argument when using StaticCache.'
+                )
+            cache_position = torch.arange(  # type: ignore
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device)
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)  # type: ignore
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds,
+                                               cache_position)  # type: ignore
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+
+        for block in self.blocks:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)  # type: ignore
+
+            if self.gradient_checkpointing and self.training:
+                block_outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+            else:
+                block_outputs = block(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = block_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = block_outputs[
+                    2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (block_outputs[1],)  # type: ignore
+
+            if output_router_logits:
+                all_router_logits += (block_outputs[-1],)  # type: ignore
+
+        hidden_states = self.norm_f(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)  # type: ignore
+
+        next_cache = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache()  # type: ignore
+                if isinstance(next_decoder_cache, 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,
+                all_router_logits
+            ] if v is not None)
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+    # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+    # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+    def _update_causal_mask(
+            self, attention_mask: Optional[torch.Tensor],
+            input_tensor: torch.Tensor,
+            cache_position: torch.Tensor) -> Optional[torch.Tensor]:
+        if self.config._attn_implementation == 'flash_attention_2':
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if hasattr(self.blocks[0].norm_attn_norm.attn,
+                   'past_key_value'):  # static cache
+            target_length = self.config.max_position_embeddings
+        else:  # dynamic cache
+            target_length = (attention_mask.shape[-1] if isinstance(
+                attention_mask, torch.Tensor) else cache_position[-1] + 1)
+        target_length = int(target_length)
+
+        causal_mask = torch.full((sequence_length, target_length),
+                                 fill_value=min_dtype,
+                                 dtype=dtype,
+                                 device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(
+            target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None,
+                                  None, :, :].expand(input_tensor.shape[0], 1,
+                                                     -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone(
+            )  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(
+                    0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[
+                    ..., :mask_length].masked_fill(padding_mask, min_dtype)
+            elif attention_mask.dim() == 4:
+                # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
+                # cache. In that case, the 4D attention mask attends to the newest tokens only.
+                if attention_mask.shape[
+                        -2] < cache_position[0] + sequence_length:
+                    offset = cache_position[0]
+                else:
+                    offset = 0
+                mask_shape = attention_mask.shape
+                mask_slice = (attention_mask.eq(0.0)).to(
+                    dtype=dtype) * min_dtype
+                causal_mask[:mask_shape[0], :mask_shape[1],
+                            offset:mask_shape[2] +
+                            offset, :mask_shape[3]] = mask_slice
+
+        if (self.config._attn_implementation == 'sdpa' and
+                attention_mask is not None and
+                attention_mask.device.type == 'cuda'):
+            # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
+            is_tracing = (
+                torch.jit.is_tracing() or
+                isinstance(input_tensor, torch.fx.Proxy) or  # type: ignore
+                (hasattr(torch, '_dynamo') and torch._dynamo.is_compiling()))
+            if not is_tracing and torch.any(attention_mask != 1):
+                # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+                # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+                # Details: https://github.com/pytorch/pytorch/issues/110213
+                causal_mask = AttentionMaskConverter._unmask_unattended(
+                    causal_mask, min_dtype)
+
+        return causal_mask
+
+
+class DbrxForCausalLM(DbrxPreTrainedModel):
+
+    def __init__(self, config: DbrxConfig):
+        super().__init__(config)
+        self.transformer = DbrxModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size,
+                                 config.vocab_size,
+                                 bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.ffn_config.moe_num_experts
+        self.num_experts_per_tok = config.ffn_config.moe_top_k
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.transformer.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Embedding):
+        self.transformer.set_input_embeddings(value)
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder: DbrxModel):
+        self.transformer = decoder
+
+    def get_decoder(self) -> DbrxModel:
+        return self.transformer
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        r"""Forward function for causal language modeling.
+
+        Example:
+        ```python
+        >>> from transformers import AutoTokenizer, DbrxForCausalLM
+
+        >>> model = DbrxForCausalLM.from_pretrained("databricks/dbrx")
+        >>> tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else
+                                self.config.output_hidden_states)
+        output_router_logits = (output_router_logits
+                                if output_router_logits is not None else
+                                self.config.output_router_logits)
+        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.transformer(
+            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,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.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)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None and loss is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(
+                    loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.Tensor,
+            past_key_values: Optional[Cache] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            **kwargs: Any) -> Dict[str, Any]:
+        past_length = 0
+        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 self.generation_config.cache_implementation == 'static':
+            # generation with static cache
+            cache_position = kwargs.get('cache_position', None)
+            if cache_position is None:
+                past_length = 0
+            else:
+                past_length = cache_position[-1] + 1
+            input_ids = input_ids[:, past_length:]
+            position_ids = position_ids[:,
+                                        past_length:] if position_ids is not None else None
+
+        # TODO @gante we should only keep a `cache_position` in generate, and do +=1.
+        # same goes for position ids. Could also help with continued generation.
+        input_length = position_ids.shape[
+            -1] if position_ids is not None else input_ids.shape[-1]
+        cache_position = torch.arange(past_length,
+                                      past_length + input_length,
+                                      device=input_ids.device)
+        position_ids = position_ids.contiguous(
+        ) if position_ids is not None else None
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        model_inputs.update(
+            { # type: ignore
+                'position_ids': position_ids,
+                'cache_position': cache_position,
+                'past_key_values': past_key_values,
+                'use_cache': kwargs.get('use_cache'),
+                'attention_mask': attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values: Cache, beam_idx: torch.LongTensor):
+        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

LICENSE.txt

@@ -0,0 +1,176 @@
+Databricks Open Model License
+
+By using, reproducing, modifying, distributing, performing or displaying
+any portion or element of DBRX or DBRX Derivatives, or otherwise accepting
+the terms of this Agreement, you agree to be bound by this Agreement.
+
+Version Release Date: March 27, 2024
+
+
+Section 1: Definitions
+
+“Agreement” means these terms and conditions that govern the use, reproduction,
+modification, distribution, performance or display of DBRX and/or DBRX
+Derivatives and any terms and conditions incorporated by reference.
+
+“Databricks” or “we” means Databricks, Inc.
+
+“Licensee” or “you” means you, or your employer or any other person or entity
+(if you are entering into this Agreement on such person or entity’s behalf),
+of the age required under applicable laws, rules or regulations to provide
+legal consent and that has legal authority to bind your employer or such other
+person or entity if you are entering in this Agreement on their behalf.
+
+“DBRX Derivatives” means all (i) modifications to DBRX, (ii) works based on
+DBRX and (iii) any other derivative works thereof. Outputs are not deemed DBRX
+Derivatives.
+
+“DBRX” means the foundational large language models and software and
+algorithms, including machine-learning model code, trained model weights,
+inference-enabling code, training-enabling code, fine-tuning enabling code,
+documentation and other elements of the foregoing identified by Databricks at
+https://github.com/databricks/dbrx, regardless of the source that you obtained
+it from.
+
+“Output” means the results of operating DBRX or DBRX Derivatives.
+
+As used in this Agreement, “including” means “including without limitation.”
+
+
+Section 2: License Rights and Conditions on Use and Distribution
+
+2.1 Grant of Rights
+
+You are granted a non-exclusive, worldwide, non-transferable and royalty-free
+limited license under Databricks’ intellectual property or other rights owned
+by Databricks embodied in DBRX to use, reproduce, distribute, copy, modify,
+and create derivative works of DBRX in accordance with the terms of this
+Agreement.
+
+2.2 Reproduction and Distribution
+
+	1. All distributions of DBRX or DBRX Derivatives must be accompanied by a
+    "Notice" text file that contains the following notice: "DBRX is provided
+    under and subject to the Databricks Open Model License, Copyright ©
+    Databricks, Inc. All rights reserved."
+
+	2. If you distribute or make DBRX or DBRX Derivatives available to a third
+    party, you must provide a copy of this Agreement to such third party.
+
+	3. You must cause any modified files that you distribute to carry prominent
+    notices stating that you modified the files.
+
+You may add your own intellectual property statement to your modifications of
+DBRX and, except as set forth in this Section, may provide additional or
+different terms and conditions for use, reproduction, or distribution of DBRX
+or DBRX Derivatives as a whole, provided your use, reproduction, modification,
+distribution, performance, and display of DBRX or DBRX Derivatives otherwise
+complies with the terms and conditions of this Agreement. Any additional or
+different terms and conditions you impose must not conflict with the terms of
+this Agreement and in the event of a conflict, the terms and conditions of this
+Agreement shall govern over any such additional or different terms and conditions.
+
+2.3 Use Restrictions
+
+You will not use DBRX or DBRX Derivatives or any Output to improve any other
+large language model (excluding DBRX or DBRX Derivatives).
+
+You will not use DBRX or DBRX Derivatives:
+
+	1. for any restricted use set forth in the Databricks Open Model Acceptable
+    Use Policy identified at
+    https://www.databricks.com/legal/acceptable-use-policy-open-model
+    ("Acceptable Use Policy"), which is hereby incorporated by reference into
+    this Agreement; or
+
+	2. in violation of applicable laws and regulations.
+
+To the maximum extent permitted by law, Databricks reserves the right to
+restrict (remotely or otherwise) usage of DBRX or DBRX Derivatives that
+Databricks reasonably believes are in violation of this Agreement.
+
+
+Section 3: Additional Commercial Terms
+
+If, on the DBRX version release date, the monthly active users of the products
+or services made available by or for Licensee, or Licensee’s affiliates, is
+greater than 700 million monthly active users in the preceding calendar month,
+you must request a license from Databricks, which we may grant to you in our
+sole discretion, and you are not authorized to exercise any of the rights under
+this Agreement unless or until Databricks otherwise expressly grants you such
+rights.
+
+If you receive DBRX or DBRX Derivatives from a direct or indirect licensee as
+part of an integrated end user product, then this section (Section 3) of the
+Agreement will not apply to you.
+
+
+Section 4: Additional Provisions
+
+4.1 Updates
+
+Databricks may update DBRX from time to time, and you must make reasonable
+efforts to use the latest version of DBRX.
+
+4.2 Intellectual Property
+
+a. No trademark licenses are granted under this Agreement, and in connection
+with DBRX or DBRX Derivatives, neither Databricks nor Licensee may use any name
+or mark owned by or associated with the other or any of its affiliates, except
+as required for reasonable and customary use in describing and redistributing
+DBRX or DBRX Derivatives.
+
+b. Subject to Databricks’ ownership of DBRX and DRBX Derivatives made by or for
+Databricks, with respect to any DBRX Derivatives that are made by you, as
+between you and Databricks, you are and will be the owner of such DBRX
+Derivatives.
+
+c. Databricks claims no ownership rights in Outputs. You are responsible for
+Outputs and their subsequent uses.
+
+d. If you institute litigation or other proceedings against Databricks or any
+entity (including a cross-claim or counterclaim in a lawsuit) alleging that
+DBRX or Outputs or results therefrom, or any portion of any of the foregoing,
+constitutes infringement of intellectual property or other rights owned or
+licensable by you, then any licenses granted to you under this Agreement shall
+terminate as of the date such litigation or claim is filed or instituted. You
+will indemnify and hold harmless Databricks from and against any claim by any
+third party arising out of or related to your use or distribution of DBRX or
+DBRX Derivatives.
+
+4.3 DISCLAIMER OF WARRANTY
+
+UNLESS REQUIRED BY APPLICABLE LAW, DBRX AND ANY OUTPUT AND RESULTS THEREFROM
+ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER
+EXPRESS OR IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE,
+NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU
+ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING OR
+REDISTRIBUTING DBRX OR DBRX DERIVATIVES AND ANY OUTPUT AND ASSUME ANY RISKS
+ASSOCIATED WITH YOUR USE OF DBRX OR DBRX DERIVATIVES AND ANY OUTPUT AND RESULTS.
+
+4.4 LIMITATION OF LIABILITY
+
+IN NO EVENT WILL DATABRICKS OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR
+OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY INDIRECT,
+SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF
+DATABRICKS OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE
+FOREGOING.
+
+4.5 Term and Termination
+
+The term of this Agreement will commence upon your acceptance of this Agreement
+or access to DBRX or DBRX Derivatives and will continue in full force and
+effect until terminated in accordance with the terms and conditions herein.
+Databricks may terminate this Agreement if you are in breach of any term or
+condition of this Agreement. Upon termination of this Agreement, you shall
+delete and cease use of DBRX or any DBRX Derivatives. Sections 1, 4.2(d), 4.3,
+4.4, and 4.6 shall survive the termination of this Agreement.
+
+4.6 Governing Law and Jurisdiction
+
+This Agreement will be governed and construed under the laws of the State of
+California without regard to choice of law principles, and the UN Convention
+on Contracts for the International Sale of Goods does not apply to this
+Agreement. The courts of California shall have exclusive jurisdiction of any
+dispute arising out of this Agreement.

NOTICE.txt

@@ -0,0 +1 @@
+DBRX is provided under and subject to the Databricks Open Model License, Copyright © Databricks, Inc. All rights reserved.

README.md

@@ -0,0 +1,172 @@
+---
+extra_gated_heading: You need to share contact information with Databricks to access this model
+extra_gated_prompt: >-
+
+  ### DBRX Terms of Use
+
+  Use of DBRX is governed by the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and the [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model).
+  
+extra_gated_fields:
+  First Name: text
+  Last Name: text
+  Organization: text
+  Purpose for Base Model Access: text
+  By clicking 'Submit' below, I accept the terms of the license and acknowledge that the information I provide will be collected, stored, processed, and shared in accordance with Databricks' Privacy Notice and I understand I can update my preferences at any time: checkbox
+extra_gated_description: >-
+  The information you provide will be collected, stored, processed, and shared in accordance with Databricks [Privacy Notice](https://www.databricks.com/legal/privacynotice).
+extra_gated_button_content: Submit
+inference: false
+license: other
+license_name: databricks-open-model-license
+license_link: https://www.databricks.com/legal/open-model-license
+---
+
+# DBRX Base
+
+* DBRX Base is a mixture-of-experts (MoE) large language model trained from scratch by Databricks.
+* We are releasing both DBRX Base, a pretrained base model, and DBRX Instruct, a fine-tuned version for few-turn interactions, under [an open license](https://www.databricks.com/legal/open-model-license).
+* This is the repository for DBRX Base. DBRX Instruct can be found [here](https://huggingface.co/databricks/dbrx-instruct).
+* For full details on the DBRX models, please read our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm).
+
+
+## Model Overview
+DBRX is a [transformer-based](https://www.isattentionallyouneed.com/) decoder-only large language model (LLM) that was trained using next-token prediction. 
+It uses a *fine-grained* mixture-of-experts (MoE) architecture with 132B total parameters of which 36B parameters are active on any input. 
+It was pre-trained on 12T tokens of text and code data. 
+Compared to other open MoE models like Mixtral-8x7B and Grok-1, DBRX is fine-grained, meaning it uses a larger number of smaller experts. DBRX has 16 experts and chooses 4, while Mixtral-8x7B and Grok-1 have 8 experts and choose 2. 
+This provides 65x more possible combinations of experts and we found that this improves model quality. 
+DBRX uses rotary position encodings (RoPE), gated linear units (GLU), and grouped query attention (GQA). 
+It uses the GPT-4 tokenizer as provided in the [tiktoken](https://github.com/openai/tiktoken) repository. 
+We made these choices based on exhaustive evaluation and scaling experiments.
+
+DBRX was pretrained on 12T tokens of carefully curated data and a maximum context length of 32K tokens. 
+We estimate that this data is at least 2x better token-for-token than the data we used to pretrain the MPT family of models. 
+This new dataset was developed using the full suite of Databricks tools, including Apache Spark™ and Databricks notebooks for data processing, and Unity Catalog for data management and governance. 
+We used curriculum learning for pretraining, changing the data mix during training in ways we found to substantially improve model quality.
+
+* **Inputs:** DBRX only accepts text-based inputs and accepts a context length of up to 32768 tokens.
+* **Outputs:** DBRX only produces text-based outputs.  
+* **Model Architecture:** More detailed information about DBRX Instruct and DBRX Base can be found in our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm).
+* **License:** [Databricks Open Model License](https://www.databricks.com/legal/open-model-license)
+* **Acceptable Use Policy:** [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model)
+* **Version:** 1.0
+* **Owner:** Databricks, Inc.
+
+
+## Usage
+These are several general ways to use the DBRX models: 
+* DBRX Base and DBRX Instruct are available for download on HuggingFace (see our Quickstart guide below). This is the HF repository for DBRX Base; DBRX Instruct can be found [here](https://huggingface.co/databricks/dbrx-instruct). 
+* The DBRX model repository can be found on GitHub [here](https://github.com/databricks/dbrx). 
+* DBRX Base and DBRX Instruct are available with [Databricks Foundation Model APIs](https://docs.databricks.com/en/machine-learning/foundation-models/index.html) via both *Pay-per-token* and *Provisioned Throughput* endpoints. These are enterprise-ready deployments.
+* For more information on how to fine-tune using LLM-Foundry, please take a look at our LLM pretraining and fine-tuning [documentation](https://github.com/mosaicml/llm-foundry/blob/main/scripts/train/README.md).
+
+
+## Quickstart Guide
+**NOTE: This is DBRX Base, and has not been instruction finetuned. It has not been trained for interactive chat and is only a completion model.**
+If you are looking for the finetuned model, please use [DBRX Instruct](https://huggingface.co/databricks/dbrx-instruct).
+
+Getting started with DBRX models is easy with the `transformers` library. The model requires ~264GB of RAM and the following packages:
+
+```bash
+pip install "transformers>=4.39.2" "tiktoken>=0.6.0"
+```
+
+If you'd like to speed up download time, you can use the `hf_transfer` package as described by Huggingface [here](https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads).
+```bash
+pip install hf_transfer
+export HF_HUB_ENABLE_HF_TRANSFER=1
+```
+
+You will need to request access to this repository to download the model. Once this is granted, 
+[obtain an access token](https://huggingface.co/docs/hub/en/security-tokens) with `read` permission, and supply the token below.
+
+### Run the model on a CPU:
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx-base", trust_remote_code=True, token="hf_YOUR_TOKEN")
+model = AutoModelForCausalLM.from_pretrained("databricks/dbrx-base", device_map="cpu", torch_dtype=torch.bfloat16, trust_remote_code=True, token="hf_YOUR_TOKEN")
+
+input_text = "Databricks was founded in "
+input_ids = tokenizer(input_text, return_tensors="pt")
+
+outputs = model.generate(**input_ids, max_new_tokens=100)
+print(tokenizer.decode(outputs[0]))
+```
+
+### Run the model on multiple GPUs:
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx-base", trust_remote_code=True, token="hf_YOUR_TOKEN")
+model = AutoModelForCausalLM.from_pretrained("databricks/dbrx-base", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True, token="hf_YOUR_TOKEN")
+
+input_text = "Databricks was founded in "
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+outputs = model.generate(**input_ids, max_new_tokens=100)
+print(tokenizer.decode(outputs[0]))
+```
+If your GPU system supports [FlashAttention2](https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2), you can add `attn_implementation=”flash_attention_2”` as a keyword to `AutoModelForCausalLM.from_pretrained()` to achieve faster inference.
+
+
+## Limitations and Ethical Considerations
+### Training Dataset Limitations
+The DBRX models were trained on 12T tokens of text, with a knowledge cutoff date of December 2023.
+
+The training mix used for DBRX contains both natural-language and code examples. The vast majority of our training data is in the English language. We did not test DBRX for non-English proficiency. Therefore, DBRX should be considered a generalist model for text-based use in the English language.
+
+DBRX does not have multimodal capabilities.
+
+### Associated Risks and Recommendations 
+All foundation models are novel technologies that carry various risks, and may output information that is inaccurate, incomplete, biased, or offensive. 
+Users should exercise judgment and evaluate such output for accuracy and appropriateness for their desired use case before using or sharing it. 
+Databricks recommends [using retrieval augmented generation (RAG)](https://www.databricks.com/glossary/retrieval-augmented-generation-rag) in scenarios where accuracy and fidelity are important. 
+We also recommend that anyone using or fine-tuning either DBRX Base or DBRX Instruct perform additional testing around safety in the context of their particular application and domain. 
+
+
+## Intended Uses
+### Intended Use Cases
+The DBRX models are open, general-purpose LLMs intended and licensed for both commercial and research applications. 
+They can be further fine-tuned for various domain-specific natural language and coding tasks. 
+DBRX Base can be used as an off-the-shelf model for text completion for general English-language and coding tasks. 
+
+Please review the Associated Risks section above, as well as the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model) for further information about permissible uses of DBRX Base and its derivatives. 
+
+### Out-of-Scope Use Cases
+DBRX models are not intended to be used out-of-the-box in non-English languages and do not support native code execution, or other forms of function-calling. 
+DBRX models should not be used in any manner that violates applicable laws or regulations or in any other way that is prohibited by the [Databricks Open Model License](https://www.databricks.com/legal/open-model-license) and [Databricks Open Model Acceptable Use Policy](https://www.databricks.com/legal/acceptable-use-policy-open-model). 
+
+
+## Training Stack
+MoE models are complicated to train, and the training of DBRX Base and DBRX Instruct was heavily supported by Databricks’ infrastructure for data processing and large-scale LLM training (e.g., [Composer](https://github.com/mosaicml/composer), [Streaming](https://github.com/mosaicml/streaming), [Megablocks](https://github.com/stanford-futuredata/megablocks), and [LLM Foundry](https://github.com/mosaicml/llm-foundry)). 
+
+Composer is our core library for large-scale training. 
+It provides an optimized training loop, easy [checkpointing](https://docs.mosaicml.com/projects/composer/en/latest/trainer/checkpointing.html) and [logging](https://docs.mosaicml.com/projects/composer/en/latest/trainer/logging.html#wood-logging), 
+[FSDP](https://pytorch.org/docs/stable/fsdp.html)-based [model sharding](https://docs.mosaicml.com/projects/composer/en/latest/notes/distributed_training.html#fullyshardeddataparallel-fsdp), 
+convenient [abstractions](https://docs.mosaicml.com/projects/composer/en/latest/trainer/time.html), extreme customizability via [callbacks](https://docs.mosaicml.com/projects/composer/en/latest/trainer/callbacks.html), and more.
+
+Streaming enables fast, low cost, and scalable training on large datasets from cloud storage. It handles a variety of challenges around deterministic resumption as node counts change, avoiding redundant downloads across devices, high-quality shuffling at scale, sample-level random access, and speed.
+
+Megablocks is a lightweight library for MoE training. Crucially, it supports “dropless MoE,” which avoids inefficient padding and is intended to provide deterministic outputs for a given sequence no matter what other sequences are in the batch.
+
+LLM Foundry ties all of these libraries together to create a simple LLM pretraining, fine-tuning, and inference experience.
+
+DBRX was trained using proprietary optimized versions of the above open source libraries, along with our [LLM training platform](https://www.databricks.com/product/machine-learning/mosaic-ai-training). 
+
+
+## Evaluation
+We find that DBRX outperforms established open-source and open-weight base models on the [Databricks Model Gauntlet](https://www.databricks.com/blog/llm-evaluation-for-icl), the [Hugging Face Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), and HumanEval. 
+The Databricks Model Gauntlet measures performance on more than 30 tasks across six categories: world knowledge, common sense reasoning, language understanding, reading comprehension, symbolic problem solving, and programming. 
+The Hugging Face Open LLM Leaderboard measures the average of ARC-Challenge, HellaSwag, MMLU, TruthfulQA, Winogrande and GSM8k. 
+HumanEval measures coding ability.
+
+Full evaluation details can be found in our [technical blog post](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm). 
+
+
+## Acknowledgements
+The DBRX models were made possible thanks in large part to the open-source community, especially:
+* The [MegaBlocks](https://arxiv.org/abs/2211.15841) library, which established a foundation for our MoE implementation.
+* [PyTorch FSDP](https://arxiv.org/abs/2304.11277), which we built on for distributed training.

config.json

@@ -0,0 +1,38 @@
+{
+  "architectures": [
+    "DbrxForCausalLM"
+  ],
+  "attn_config": {
+    "clip_qkv": 8,
+    "kv_n_heads": 8,
+    "model_type": "",
+    "rope_theta": 500000
+  },
+  "auto_map": {
+    "AutoConfig": "configuration_dbrx.DbrxConfig",
+    "AutoModelForCausalLM": "modeling_dbrx.DbrxForCausalLM"
+  },
+  "d_model": 6144,
+  "emb_pdrop": 0.0,
+  "ffn_config": {
+    "ffn_hidden_size": 10752,
+    "model_type": "",
+    "moe_jitter_eps": 0.01,
+    "moe_loss_weight": 0.05,
+    "moe_num_experts": 16,
+    "moe_top_k": 4
+  },
+  "initializer_range": 0.02,
+  "max_seq_len": 32768,
+  "model_type": "dbrx",
+  "n_heads": 48,
+  "n_layers": 40,
+  "output_router_logits": false,
+  "resid_pdrop": 0.0,
+  "router_aux_loss_coef": 0.05,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.2",
+  "use_cache": true,
+  "vocab_size": 100352
+}

configuration_dbrx.py

@@ -0,0 +1,264 @@
+"""Dbrx configuration."""
+from typing import Any, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DBRX_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class DbrxAttentionConfig(PretrainedConfig):
+    """Configuration class for Dbrx Attention.
+
+    [`DbrxAttention`] class. It is used to instantiate attention layers
+    according to the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        attn_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the attention layers.
+        clip_qkv (`float`, *optional*, defualts to None):
+            If not `None`, clip the queries, keys, and values in the attention layer to this value.
+        kv_n_heads (Optional[int]): For grouped_query_attention only, allow user to specify number of kv heads.
+        rope_theta (float): The base frequency for rope.
+    """
+
+    def __init__(
+        self,
+        attn_pdrop: float = 0,
+        clip_qkv: Optional[float] = None,
+        kv_n_heads: int = 1,
+        rope_theta: float = 10000.0,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.attn_pdrop = attn_pdrop
+        self.clip_qkv = clip_qkv
+        self.kv_n_heads = kv_n_heads
+        self.rope_theta = rope_theta
+
+        for k in ['model_type']:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f'Found unknown {kwargs=}')
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str,
+                        **kwargs: Any) -> 'PretrainedConfig':
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path,
+                                                  **kwargs)
+
+        if config_dict.get('model_type') == 'dbrx':
+            config_dict = config_dict['attn_config']
+
+        if 'model_type' in config_dict and hasattr(
+                cls,
+                'model_type') and config_dict['model_type'] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                +
+                f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.'
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxFFNConfig(PretrainedConfig):
+    """Configuration class for Dbrx FFN.
+
+    [`DbrxFFN`] class. It is used to instantiate feedforward layers according to
+    the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        ffn_act_fn (dict, optional): A dict specifying activation function for the FFN.
+            The dict should have a key 'name' with the value being the name of
+            the activation function along with any additional keyword arguments.
+        ffn_hidden_size (int, optional): The hidden size of the feedforward network.
+        moe_num_experts (int, optional): The number of experts in the mixture of experts layer.
+        moe_top_k (int, optional): The number of experts to use in the mixture of experts layer.
+        moe_jitter_eps (float, optional): The jitter epsilon for the mixture of experts layer.
+        moe_loss_weight (float, optional): The loss weight for the mixture of experts layer.
+        moe_normalize_expert_weights (float, optional): The normalization factor for the expert weights.
+        uniform_expert_assignment (bool, optional): Whether to use uniform expert assignment.
+            This should only be used for benchmarking purposes.
+    """
+
+    def __init__(
+        self,
+        ffn_act_fn: Optional[dict] = None,
+        ffn_hidden_size: int = 3584,
+        moe_num_experts: int = 4,
+        moe_top_k: int = 1,
+        moe_jitter_eps: Optional[float] = None,
+        moe_loss_weight: float = 0.01,
+        moe_normalize_expert_weights: Optional[float] = 1,
+        uniform_expert_assignment: bool = False,
+        **kwargs: Any,
+    ):
+        super().__init__()
+        if ffn_act_fn is None:
+            ffn_act_fn = {'name': 'silu'}
+        self.ffn_act_fn = ffn_act_fn
+        self.ffn_hidden_size = ffn_hidden_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_loss_weight = moe_loss_weight
+        self.moe_normalize_expert_weights = moe_normalize_expert_weights
+        self.uniform_expert_assignment = uniform_expert_assignment
+
+        for k in ['model_type']:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f'Found unknown {kwargs=}')
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str,
+                        **kwargs: Any) -> 'PretrainedConfig':
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path,
+                                                  **kwargs)
+
+        if config_dict.get('model_type') == 'dbrx':
+            config_dict = config_dict['ffn_config']
+
+        if 'model_type' in config_dict and hasattr(
+                cls,
+                'model_type') and config_dict['model_type'] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                +
+                f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.'
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxConfig(PretrainedConfig):
+    """Configuration class for Dbrx.
+
+    [`DbrxModel`]. It is used to instantiate a Dbrx model according to the
+    specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        d_model (`int`, *optional*, defaults to 6144):
+            Dimensionality of the embeddings and hidden states.
+        n_heads (`int`, *optional*, defaults to 48):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_layers (`int`, *optional*, defaults to 40):
+            Number of hidden layers in the Transformer encoder.
+        max_seq_len (`int`, *optional*, defaults to 32768):
+            The maximum sequence length of the model.
+        vocab_size (`int`, *optional*, defaults to 100352):
+            Vocabulary size of the Dbrx model. Defines the maximum number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`DbrxModel`].
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability applied to the attention output before combining with residual.
+        emb_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the embedding layer.
+        attn_config (`dict`, *optional*):
+            A dictionary used to configure the model's attention module.
+        ffn_config (`dict`, *optional*):
+            A dictionary used to configure the model's FFN module.
+        use_cache (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabling this will also
+            allow the model to output the auxiliary loss. See [here]() for more details
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+
+    Example:
+    ```python
+    >>> from transformers import DbrxConfig, DbrxModel
+
+    >>> # Initializing a Dbrx configuration
+    >>> configuration = DbrxConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = DbrxModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = 'dbrx'
+    attribute_map = {
+        'num_attention_heads': 'n_heads',
+        'hidden_size': 'd_model',
+        'num_hidden_layers': 'n_layers',
+        'max_position_embeddings': 'max_seq_len'
+    }
+
+    def __init__(
+        self,
+        d_model: int = 2048,
+        n_heads: int = 16,
+        n_layers: int = 24,
+        max_seq_len: int = 2048,
+        vocab_size: int = 32000,
+        resid_pdrop: float = 0.0,
+        emb_pdrop: float = 0.0,
+        attn_config: Optional[DbrxAttentionConfig] = None,
+        ffn_config: Optional[DbrxFFNConfig] = None,
+        use_cache: bool = True,
+        initializer_range: float = 0.02,
+        output_router_logits: bool = False,
+        router_aux_loss_coef: float = 0.05,
+        **kwargs: Any,
+    ):
+        if attn_config is None:
+            self.attn_config = DbrxAttentionConfig()
+        elif isinstance(attn_config, dict):
+            self.attn_config = DbrxAttentionConfig(**attn_config)
+        else:
+            self.attn_config = attn_config
+
+        if ffn_config is None:
+            self.ffn_config = DbrxFFNConfig()
+        elif isinstance(ffn_config, dict):
+            self.ffn_config = DbrxFFNConfig(**ffn_config)
+        else:
+            self.ffn_config = ffn_config
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.resid_pdrop = resid_pdrop
+        self.emb_pdrop = emb_pdrop
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.output_router_logits = output_router_logits
+        self.router_aux_loss_coef = router_aux_loss_coef
+
+        tie_word_embeddings = kwargs.pop('tie_word_embeddings', False)
+        if tie_word_embeddings:
+            raise ValueError(
+                'tie_word_embeddings is not supported for Dbrx models.')
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": [
+    100257
+  ],
+  "transformers_version": "4.38.2"
+}

huggingface-metadata.txt

@@ -0,0 +1,65 @@
+url: https://huggingface.co/databricks/dbrx-base
+branch: main
+download date: 2024-03-30 00:19:12
+sha256sum:
+    6fc16f714bc5bdae2a9c712ebbd0c60282d51d8102446bdcb42bf73fbcd789cd model-00001-of-00061.safetensors
+    7b33be8a577936012420efcdb9fb42394820083dcc52279ab39f1f469446d92c model-00002-of-00061.safetensors
+    5edb3101b011b9ac53216a8a89344854932cba4a116462d269bafab67b2504ea model-00003-of-00061.safetensors
+    3ab9725625b4be1dbe497733c7da6a826763da7fd4ecb4c7a1b0a01730ad5f00 model-00004-of-00061.safetensors
+    e70596f19ce2b40d833456da403f504af4b33c7a28fc663ece64e4b9b5c023c4 model-00005-of-00061.safetensors
+    6a59eb220f7174521cbb8df029c5665f9bd70afd919e5432461c9d31c2498b41 model-00006-of-00061.safetensors
+    2e77dd3c8552550dd8db66372b1680992d438bbfb6a2950071f1718e1d197ef5 model-00007-of-00061.safetensors
+    49830c115d1c6dd80ee1f187e4c7aa2f44e8565ad8a50ff2474d276e6a7ea436 model-00008-of-00061.safetensors
+    3535faaac65410f8e78bd77bd6ab9c22638a7e753f7d7d73a255b8ee71909ed6 model-00009-of-00061.safetensors
+    388344ef7a364c3cf90ce50d7cb44e477c17233966a8ff84cfa0e998455e5fba model-00010-of-00061.safetensors
+    8f423206b83056355629508621b422e839ef0544c9f6099d3a946ad81ffe13c5 model-00011-of-00061.safetensors
+    46630da164ed5472036c49e1247fd7157f1ca1a3ab25957d27a56dd80f1ec4e9 model-00012-of-00061.safetensors
+    7a09b47e2f13c74b72275226ed5ad56191eea1883328dd5ccc8291a49ab64908 model-00013-of-00061.safetensors
+    7708780cfcba2211f806b2164550ea6982ce7e68dc47db9e4fe42a4727e58d89 model-00014-of-00061.safetensors
+    3c208a59d20af1d5504dd72611def467a79f37d8f047b53c2588d2a594de6190 model-00015-of-00061.safetensors
+    bb09a5c3e813be20033ca5b75308bebc0eea127bbe4d5c254d4877dae9080731 model-00016-of-00061.safetensors
+    9b5dfe0cbdfcd0131aad4605d16339f6ee3e94737bb01b21bcfb54ca4a9fe325 model-00017-of-00061.safetensors
+    627ccca7006bc60d9374a8b0e59d1960fc300cda6a9caa6e1b187ea9c3af58e5 model-00018-of-00061.safetensors
+    2d327dda14ac1c8a0c4f6590a1cd49861164b26aaae1743c0bdf9b0ac62f2b93 model-00019-of-00061.safetensors
+    fc042ddef23192a39cf83b475a3ef29c5dea84a33e1584ec1a90f309f1fd4ba1 model-00020-of-00061.safetensors
+    23143b0dda50c67186262a2d22182fa72c3f9527482d6b693736a43a7c1e5c15 model-00021-of-00061.safetensors
+    cf47ac22655c7a9df91ccb39a5b3b5753d7a5e36a122138329c9a9811e7a2953 model-00022-of-00061.safetensors
+    7bb213db0ee84652be17d60b19ba5eba2c7aa4a0e6e858a2b5e6e26281bbfbf5 model-00023-of-00061.safetensors
+    099d8224582075ddaa2c05d91cf14d12f6d8d0f9ba88a871b0ac5fc8191244eb model-00024-of-00061.safetensors
+    6ad2df7ec70500e020701b80f723af9cb42724c3b66a199453b1480a02f8a199 model-00025-of-00061.safetensors
+    6144ecaada0d9662d8572b23e5116d4ed05a3e634ddcaaf42720db684b0435e5 model-00026-of-00061.safetensors
+    d5d60b11ef83487c2751a9a43043ff2319843653f2c792abdfd604f36d3d0848 model-00027-of-00061.safetensors
+    1921574ab5bda287a6ea20e3ae2d091fde29973059a688944f204dbd7724b147 model-00028-of-00061.safetensors
+    31291c52c7fe8e6ec8465d42767a9ea1603fe64311e18a4d3c36b27a905be62f model-00029-of-00061.safetensors
+    3c99753094273763636069d6d9b2a3474064b38148d57915c2916d2b1e9f0c7a model-00030-of-00061.safetensors
+    33a05e332cae49e0ebe746944f2e4892ff0517876f8e5c06a23f9333ed32e4e0 model-00031-of-00061.safetensors
+    f817a0338220e16a62cce0cbb56cc6859ed85ba2af145ec1a5b336c77ad362dd model-00032-of-00061.safetensors
+    524cff55ac50174e3bc3fab234749217403f106c32d87b3f97af316115eb101a model-00033-of-00061.safetensors
+    f9f51a8c51ab37e4ca31ccc2e4aba30ea206e32ad0b864db7d2404d770000476 model-00034-of-00061.safetensors
+    6b16806ace65db95d602c39d5e372901a265a1e6d3b0d757679cc05aa7fab7e8 model-00035-of-00061.safetensors
+    f2830ff0574e31e39ba6b434187fbfd3b2bb8264a822fe725d93f1fc5348b3d5 model-00036-of-00061.safetensors
+    9ed804ce59a2a8f2c1cff4d5bde7782b866961aa88a5846e4d0250d8f51c84c8 model-00037-of-00061.safetensors
+    c5446321bb918a0e269ccc5a07eb002fca71e0047fd7c08bb7d7dca6709ee999 model-00038-of-00061.safetensors
+    d3e03720f169d8923e3457bad304ad3ba0c75037ceb13bd6716807d01403fdf5 model-00039-of-00061.safetensors
+    b48bee2e8c38c105cc7a8434cc2cf845b3d023d5a849814c17b65ad896fe2c0f model-00040-of-00061.safetensors
+    6d14cabc491a883d8260aa3bea801eb757cb38db51c4081263237fdfc4053400 model-00041-of-00061.safetensors
+    178ea07958e6b2e573025d2bf61dde4cf0638b8dc6e81311c089400d6fa81717 model-00042-of-00061.safetensors
+    b202858f99636de6933dccd1c487d590cfca76b8f9cddee876af04692b38ab80 model-00043-of-00061.safetensors
+    c9c823bc0f84dacb3996d598b4795c400361a9c657f36b29c40b4c7982ad0ed9 model-00044-of-00061.safetensors
+    97ac08f2364a4b67f6a17274c4770637c98d7306eca87ec797e3cd754447b1e3 model-00045-of-00061.safetensors
+    23e05dbc72e799bb043b6497be35cacaebfda43c8057dd04db07e8ac03e74751 model-00046-of-00061.safetensors
+    77e4a2617f18043aecb00d4bb04c41f1ab12eb26f5f78038b53efe70e11edc4c model-00047-of-00061.safetensors
+    6f0226bfb4b796f64deafbbeb80b2c7ea5314b6038afc607bc42b3dbd0a79313 model-00048-of-00061.safetensors
+    ced55f41dc87ca44d1adc2be76144bf0de0045b340eb3231ae3e16bd16a1f6d5 model-00049-of-00061.safetensors
+    6634f09f0cf60e974c23ca533d00f99868e71c05f94ef3e021a851926f5acc6e model-00050-of-00061.safetensors
+    5f3a6f60f6c73f1fb1097ba66d6672bd1d2c24a5c4ec89ebe0e1308f82dccb4e model-00051-of-00061.safetensors
+    7846d13c5af42185217dd4cacabb389dc2fdb8bee06cfa4591f079dbceca3033 model-00052-of-00061.safetensors
+    d898218d5f978b6cc1bff5861d3bb263c5a7c88825e6cdecc5d19be953615b93 model-00053-of-00061.safetensors
+    e9509ff00409d2019de6380d8ffceb718f1b1edc433ffe30ced2b659f6e381db model-00054-of-00061.safetensors
+    074057b7b5436e6dc77571895a15743c6de34d54e69378c9937d5f8017a32a35 model-00055-of-00061.safetensors
+    dfbd182dfc66404837b09cfed5c1ccee2b9f22745e365a556473bed1ca2fb454 model-00056-of-00061.safetensors
+    25a1ca1e6fa15d26b54f20f93cbb3c1be6f5858bd9123e8e68d4afe03d525ded model-00057-of-00061.safetensors
+    d9119d4edf798b355b6625738c540b807916bd605eb21a157d533bd5d4e59fd2 model-00058-of-00061.safetensors
+    95a0fed1c3fa366feae66bdebcef87feced28c54b1081b458bd4503435ddd928 model-00059-of-00061.safetensors
+    32a4d9683495cffbed6ca2e7fa9b7adc47d2d0171984cda92b4535e0e25f6903 model-00060-of-00061.safetensors
+    7d4c98ee4f4f01a06854d8b3a4959484fafa8b7b4371ad25d35a77c880066c43 model-00061-of-00061.safetensors

model-00001-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3c2a81b624cbb921015e97a43a20598b529c193748f8a2a78913016f4c066e7
+size 3523439352

model-00002-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed001899eb392f361397d83bb4e90111ee10677023c68aca7ef92986f00171c1
+size 4404245144

model-00003-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a53fc538e17f263dbafac832259701d71c039ceb0e0a133411bd868f67c1412
+size 4227862560

model-00004-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9f781b7cdfffb7c0ce24de37b45bbcbe6ac74ef0f666429c4faf14e4ac1ce37
+size 4404245144

model-00005-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b95e369dd864c52a4249deb1180a1213ba1aa4ff32236c28c59b008a3373be76
+size 4404245144

model-00006-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8a9a26d88087d1f17f398bad35b00f8c9df6d45684c2d27ed51c6fd8479da0a
+size 4227862560

model-00007-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23a54c86412163b35be1a339b8b414e44f9591eb2b4b8f926d215be0a4925521
+size 4404245144

model-00008-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e62c2e4874bbce04a6aeb208255d57863539740adb4f889c742f9040ba2f3d96
+size 4404245144

model-00009-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f06f6d9014cf96eb3650c99db4045d37c8b7e34c5117e9a2738d400de1658860
+size 4227862560

model-00010-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f85d6cd1c218a5a69d01157f87cdbd50ad097058cbd0bf4eaf3dcca4dc13e8d7
+size 4404245144

model-00011-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6cbb405f375b9a9fb6a7fae49875ec4b71057a83f8809af7485e80a1f8bcad5
+size 4404245144

model-00012-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e56c3d4c9786fcb2eeeddca5f1ffadd4e8a4b64514c807595572ba0fc0f61b76
+size 4227862560

model-00013-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:947da27479f93c092e40bb4436e04b6d73868f85c413a4e5401300eaec099414
+size 4404245144

model-00014-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0af3684c789d505e5800532e7c0b093a0534b31cc4fe3ff7d2cdd70d3af103df
+size 4404245144

model-00015-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55f55d4b560746620978473685d5a6c87a5eb9ac780b97592a77a27e9fd67dce
+size 4227862560

model-00016-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7689e83173a2e0055e7869d7ae681940d55084c4ec0cde3b1ac066ee42342b4
+size 4404245168

model-00017-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6148c0ed255843e70f502c59328e9c40f639e87a4c540708715cdaac823ca474
+size 4404245184

model-00018-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1b24cd6fbcb7458e33aa6db4702aad519a32637b15fda82454747f5e9432e13
+size 4227862592

model-00019-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4da75153d9ed71593a0f26d8751837a3d9d43930025aaf20012045946ace9de
+size 4404245184

model-00020-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b4b005943135f0a19ccd47ec2977305514d0956b59962fa780378cbfa01cd13
+size 4404245184

model-00021-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45bb5601b485a573af20d7b23bad58e0f4146b349644bd8f728b06252102176b
+size 4227862592

model-00022-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186133d4fa899fc798454adf2edf927e3c15dd1ac4325fc38123202d52bea8d8
+size 4404245184

model-00023-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f39b6544788e0f7cbe61e78fc0f5b4e7909702b4debd54d08ce8738b8144713
+size 4404245184

model-00024-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2523a4be387557beb04260b0d08cf8a37d0a87006197c9103f365e3cfad06daf
+size 4227862592

model-00025-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c707561d9b07c3a4ee99c6949760eee338e471f013a581ceb2f32a473280cfaa
+size 4404245184

model-00026-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20c4aeafbbcf1806560fcbfd8fd17d5e15355405f2109cf04354a6d353a09cad
+size 4404245184

model-00027-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41879f07d50975e532be376f7efc62bcb5be7719b9303ef8c807706c43c2ad43
+size 4227862592

model-00028-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da3729c87038162e7f6751435cfe69ac2a9b900102fb8bf9135d6dee46dba5e2
+size 4404245184

model-00029-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85c7a0900e23458a71566d7a396ad00ad20103f7c22895e85c69e29599e50755
+size 4404245184

model-00030-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af610db5f894204a2c2c6891f4644b5ae4f38e429f1b33978d3567981aa61faa
+size 4227862592

model-00031-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8815b02db955ec7b62b7940a60fffaaed34ed066e0a752d083c77f966291f12c
+size 4404245184

model-00032-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f85f2db6f5ee9ef26ec51166dce1b9b439975f112e793b3ef2ec7f243b7d2650
+size 4404245184

model-00033-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa6f00ab09861368f0aaf66ec92bf43b17a70cd8bfb5be6264aa5b21d9a36a25
+size 4227862592

model-00034-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89fbac47b8d8fd6487d15b9c8b7b79578204c3f102fcb6c5dbf1943ac64647a2
+size 4404245184

model-00035-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55fbb3c3bcd978b94ec29434ba6c2741d3687c8fd0c9b204857758566f85f1f6
+size 4404245184

model-00036-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:792058de69a54860e03de6dcd68bb88c758e0e701f9a9054ef95e74b66f641c1
+size 4227862592

model-00037-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e44c4782d6f922d6538f55628604b4b0bbf06999d7c2828b9b62b958ed80db0
+size 4404245184

model-00038-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d3a9421a898bd09894fb6569b09796cab23aa564a043a97565b91534f5ed4fe
+size 4404245184

model-00039-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b597724c9c431c9b66ef7f1c5c98c572e1eb2cc04eab00b4b5578c804f7dabe2
+size 4227862592

model-00040-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c286a04587a4c2798f27e6cb8cb1d8cbaf5b77a6f253f2988f5d1954d02ea718
+size 4404245184

model-00041-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff71194ea96b94d3ada6ded57a5d92d4cd3199e85cda0568b17b5fa93e0347a5
+size 4404245184

model-00042-of-00061.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93347f13c06c95ed4131c03f1bcbade4aaf2fc9e8e8738483467ed482e1ff556
+size 4227862592