Add fixed Raven package for Quasar hybrid generation

raven/__init__.py

@@ -0,0 +1,7 @@
+# -*- coding: utf-8 -*-
+
+from raven.models.raven import RavenConfig, RavenForCausalLM, RavenModel
+
+__all__ = [
+    'RavenConfig', 'RavenForCausalLM', 'RavenModel',
+]

raven/layers/__init__.py

@@ -0,0 +1,5 @@
+# -*- coding: utf-8 -*-
+
+from raven.layers.raven import RavenAttention
+
+__all__ = ['RavenAttention']

raven/layers/raven.py

@@ -0,0 +1,331 @@
+
+
+from __future__ import annotations
+
+import warnings
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import math
+import torch.nn.functional as F
+from einops import rearrange, repeat
+
+from fla.layers.utils import get_unpad_data, index_first_axis, pad_input
+from fla.modules.feature_map import ReLUFeatureMap, SwishFeatureMap, T2RFeatureMap
+from fla.modules.layernorm import rms_norm_linear
+from fla.ops.gsa import chunk_gsa as chunk_raven, fused_recurrent_gsa as fused_recurrent_raven
+
+
+from fla.ops.utils.index import prepare_lens_from_mask
+
+
+from fla.modules import FusedRMSNormGated , RMSNorm, RotaryEmbedding  #Added for RoPE
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+def _max_offset(seqlen_offset):
+    if seqlen_offset is None:
+        return 0
+    if isinstance(seqlen_offset, int):
+        # scalar offset, nothing fancy
+        return seqlen_offset
+    if isinstance(seqlen_offset, torch.Tensor):
+        # tensor of offsets -> take max
+        return int(seqlen_offset.max().item())
+    # list/tuple/other iterables
+    return max(seqlen_offset)
+
+
+class RavenAttention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_k: float = 1.,
+        expand_v: float = 1.,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        num_slots: Optional[int] = None,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        gate_logit_normalizer: int = 8,
+        feature_map: str = 'swish',
+        use_output_gate: bool = False,
+        use_norm: bool = True,
+        layer_idx: Optional[int] = None,
+        scale: Optional[float] = 1.,
+        decay_type: str = 'Mamba2',
+        topk: int = 32,
+        bias_rmm: bool = False,
+        add_gumbel_noise: bool = True,
+        router_score: str = 'sigmoid',
+        router_type: str = 'lin',
+        use_rope: bool = False,
+        **kwargs
+    ) -> RavenAttention:
+        super().__init__()
+
+        self.mode = mode
+        self.decay_type = decay_type
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.num_heads = num_heads
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+        self.key_dim = int(hidden_size * expand_k)
+        self.value_dim = int(hidden_size * expand_v)
+        self.key_dim_per_group = self.key_dim // self.num_kv_groups
+        self.value_dim_per_group = self.value_dim // self.num_kv_groups
+        self.head_k_dim = self.key_dim // self.num_heads
+        self.head_v_dim = self.value_dim // self.num_heads
+        self.topk = topk
+        self.use_output_gate = use_output_gate
+        self.use_rope = use_rope
+        self.gate_logit_normalizer = gate_logit_normalizer
+        self.use_norm = use_norm
+        self.scale = scale
+        self.rope_theta = 10000.
+
+        ##  For Router Design
+        self.bias_rmm = bias_rmm  # For no gumbel router with bias
+        self.add_gumbel_noise = add_gumbel_noise  # For no gumbel router with bias
+        self.router_score = router_score
+        self.router_type = router_type
+
+        if num_slots is None:
+            num_slots = self.head_k_dim
+        self.num_slots = num_slots
+
+        self.layer_idx = layer_idx
+
+        if layer_idx is None:
+            warnings.warn(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.register_module('feature_map', None)
+        if feature_map == 'swish':
+            self.feature_map = SwishFeatureMap()
+        elif feature_map == 'relu':
+            self.feature_map = ReLUFeatureMap()
+        elif feature_map == 't2r':
+            self.feature_map = T2RFeatureMap(self.head_k_dim, self.head_k_dim)
+        else:
+            raise NotImplementedError(f"Feature map `{feature_map}` is not supported now.")
+
+        ## ===== QKV Proj =====
+        self.q_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.key_dim_per_group, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.value_dim_per_group, bias=False)
+
+        ## ===== Forget Gate/ Decay =====
+        if self.decay_type == 'Mamba2':
+            # Decay of Mamba2
+            self.a_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False)
+            A = torch.empty(self.num_heads, dtype=torch.float32).uniform_(0, 16)
+            self.A_log = nn.Parameter(torch.log(A))
+            self.A_log._no_weight_decay = True
+            # hard coded for now
+            dt_min = 0.001
+            dt_max = 0.1
+            dt_init_floor = 1e-4
+            dt = torch.exp(
+                torch.rand(self.num_heads) * (math.log(dt_max) - math.log(dt_min))
+                + math.log(dt_min)
+            )
+            dt = torch.clamp(dt, min=dt_init_floor)
+            # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            self.dt_bias = nn.Parameter(inv_dt)
+            # Just to be explicit. Without this we already don't put wd on dt_bias because of the check
+            # name.endswith("bias") in param_grouping.py
+            self.dt_bias._no_weight_decay = True
+
+        elif self.decay_type == 'GLA':
+            self.f_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.num_slots, bias=False)
+
+        ## ===== Router  =====
+        if self.bias_rmm:
+            self.r_bias =  nn.Parameter(torch.empty( ( self.num_heads , self.num_slots) , dtype=torch.float32))
+
+        if  self.router_type == 'lin':
+            self.r_proj = nn.Linear(self.hidden_size, self.num_heads * self.num_slots , bias=False)
+
+        elif  self.router_type == 'mlp':
+            self.r_proj = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size, bias=True),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, self.num_heads * self.num_slots, bias=False) )
+
+        self.score_fn = (
+            (lambda x: torch.sigmoid(x))
+            if self.router_score == "sigmoid"
+            else (lambda x: torch.softmax(x, dim=-1))
+        )
+
+
+        ## ===== RoPE  =====
+        if self.use_rope:
+            self.rotary = RotaryEmbedding(dim=self.head_k_dim, base=self.rope_theta)   # Added for RoPE
+
+        ## ===== QK Norm  =====
+        self.q_norm = RMSNorm(self.head_k_dim, elementwise_affine, eps=norm_eps)
+        self.k_norm = RMSNorm(self.head_k_dim, elementwise_affine, eps=norm_eps)
+
+        ## ===== Output Layer  =====
+        if self.use_output_gate:
+            self.o_gate_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+            self.o_norm = FusedRMSNormGated(self.head_v_dim, eps=norm_eps)
+        else:
+            self.g_norm = RMSNorm(self.hidden_size, elementwise_affine, eps=norm_eps)
+
+        self.o_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        if attention_mask is not None:
+            assert len(attention_mask.shape) == 2, (
+                "Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
+                "for padding purposes (0 indicating padding). "
+                "Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
+            )
+
+        batch_size, q_len, _ = hidden_states.shape
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+        seqlen_offset, max_seqlen = 0, q_len # Added for RoPE
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            seqlen_offset = past_key_values.get_seq_length(self.layer_idx)   # Added for RoPE
+            last_state = past_key_values[self.layer_idx]
+            max_seqlen = q_len + seqlen_offset  # Added for RoPE
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if attention_mask is not None:
+            indices, cu_seqlens, _ = get_unpad_data(attention_mask[:, -q_len:])
+            hidden_states = index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices).unsqueeze(0) # to deliminate the offsets of padding tokens
+            seqlen_offset = seqlen_offset + prepare_lens_from_mask(attention_mask) - attention_mask.shape[-1]   # Added for RoPE
+            max_seqlen = max(q_len, q_len + _max_offset(seqlen_offset))    # Added for RoPE
+
+        q = rearrange(self.q_proj(hidden_states), '... (h d) -> ... h d', d=self.head_k_dim)
+        k = rearrange(self.k_proj(hidden_states), '... (h d) -> ... h d', d=self.head_k_dim)
+        v = rearrange(self.v_proj(hidden_states), '... (h d) -> ... h d', d=self.head_v_dim)
+        router = rearrange(self.r_proj(hidden_states) , '... (h m) -> ... h m', m=self.num_slots)
+
+        if self.decay_type == 'Mamba2':
+            # Build Mamba2 Decay
+            f = (- self.A_log.float().exp() * F.softplus(self.a_proj(hidden_states).float() + self.dt_bias)).unsqueeze(-1)
+        elif self.decay_type == 'GLA':
+            f = rearrange(self.f_proj(hidden_states) , '... (h m) -> ... h m', m=self.num_slots)
+            f =  F.logsigmoid(f) / self.gate_logit_normalizer
+            if self.num_kv_groups > 1:
+                f = repeat(f, '... h d -> ... (h g) d', g=self.num_kv_groups)
+
+        if self.feature_map is not None:
+            q, k = map(lambda x: self.feature_map(x), (q, k))
+
+        # QK Norm
+        q, k = self.q_norm(q), self.k_norm(k)
+
+        if self.use_rope:
+            assert batch_size == 1, "RoPE is not supported for batch size > 1"
+            max_seqlen = max(max_seqlen, 8192)  # Added for RoPE
+            q, k = self.rotary(q, k, seqlen_offset=seqlen_offset, max_seqlen=max_seqlen, cu_seqlens=cu_seqlens)  #Added for RoPE
+
+        # V Feature map
+        v = F.silu(v)
+
+        # Build RMM Router
+        if self.add_gumbel_noise:
+            if self.training:
+                router = router - torch.empty_like(router).exponential_().log()
+
+        orig_scores = self.score_fn(router)
+        if self.bias_rmm:
+            scores = orig_scores + self.r_bias.float()
+        else:
+            scores = orig_scores
+
+        route_idx = scores.topk(self.topk, dim=-1).indices
+        topk_weights =  torch.gather(orig_scores, dim=-1, index=route_idx)
+
+        if self.router_score == 'sigmoid':
+            topk_weights /= (topk_weights.sum(dim=-1, keepdim=True) + 1e-9)
+
+        s_multihot = torch.zeros_like(router).scatter_(-1, route_idx, topk_weights.to(router.dtype))
+
+        f = (f*s_multihot).to(q.dtype)
+        s = (1-f.exp()).to(q.dtype)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if self.num_kv_groups > 1:
+            k, v = map(lambda x: repeat(x, '... h d -> ... (h g) d', g=self.num_kv_groups), (k, v))
+
+        assert q.shape[-2] == k.shape[-2] == v.shape[-2] == f.shape[-2] == s.shape[-2], (
+            f"Raven head mismatch: q={q.shape}, k={k.shape}, v={v.shape}, f={f.shape}, s={s.shape}"
+        )
+
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_raven(
+                q=q,
+                k=k,
+                v=v,
+                s=s,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                scale=self.scale,
+                cu_seqlens=cu_seqlens,
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_raven(
+                q=q,
+                k=k,
+                v=v,
+                s=s,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                scale=self.scale,
+                cu_seqlens=cu_seqlens,
+            )
+        else:
+            raise NotImplementedError(f"Not supported mode `{mode}`.")
+
+        if past_key_values is not None:
+            past_key_values.update(
+                recurrent_state=recurrent_state,
+                conv_state=None,
+                layer_idx=self.layer_idx,
+                offset=q_len
+            )
+
+
+        if self.use_output_gate:
+            gate_out = rearrange(self.o_gate_proj(hidden_states), '... (h d) -> ... h d', d=self.head_v_dim)
+            o = self.o_norm(F.silu(o), gate_out)
+            o = rearrange(o, '... h d -> ... (h d)')
+            o = self.o_proj(o)
+        else:
+            o = rearrange(o, '... h d -> ... (h d)')
+            o = rms_norm_linear(F.silu(o), self.g_norm.weight, self.g_norm.bias, self.o_proj.weight, self.o_proj.bias)
+
+        if attention_mask is not None:
+            o = pad_input(o.squeeze(0), indices, batch_size, q_len)
+
+        return o, None, past_key_values

raven/models/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

raven/models/raven/__init__.py

@@ -0,0 +1,12 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from raven.models.raven.configuration_raven import RavenConfig
+from raven.models.raven.modeling_raven import RavenForCausalLM, RavenModel
+
+AutoConfig.register(RavenConfig.model_type, RavenConfig, exist_ok=True)
+AutoModel.register(RavenConfig, RavenModel, exist_ok=True)
+AutoModelForCausalLM.register(RavenConfig, RavenForCausalLM, exist_ok=True)
+
+__all__ = ['RavenConfig', 'RavenForCausalLM', 'RavenModel']

raven/models/raven/configuration_raven.py

@@ -0,0 +1,111 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class RavenConfig(PretrainedConfig):
+
+    model_type = 'raven'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        gate_logit_normalizer: Optional[int] = 8,
+        clamp_min: Optional[float] = None,
+        clamp_max: Optional[float] = None,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        num_slots: Optional[int] = 64,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        exapnd_k: float = 1,
+        exapnd_v: float = 1,
+        feature_map: str = 'swish',
+        use_output_gate: bool = False,
+        use_norm: bool = True,
+        max_position_embeddings: int = 2048,
+        hidden_act: str = "swish",
+        decay_type: str = 'Mamba2',
+        bias_rmm: bool = False,
+        add_gumbel_noise: bool = True,
+        router_score: str = 'sigmoid',
+        router_type: str = 'lin',
+        topk = 32,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: Optional[int] = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        initializer_range: float = 0.02,
+        tie_word_embeddings: bool = False,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        use_l2warp: bool = False,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.hidden_size = hidden_size
+        self.gate_logit_normalizer = gate_logit_normalizer
+        self.clamp_min = clamp_min
+        self.clamp_max = clamp_max
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.num_slots = num_slots
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.expand_k = exapnd_k
+        self.expand_v = exapnd_v
+        self.feature_map = feature_map
+        self.use_output_gate = use_output_gate
+        self.use_norm = use_norm
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_act = hidden_act
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.decay_type = decay_type
+        self.topk = topk
+        self.bias_rmm = bias_rmm   
+        self.add_gumbel_noise = add_gumbel_noise   
+        self.router_score = router_score
+        self.router_type = router_type
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.use_l2warp = use_l2warp
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+            attn['use_rope'] = attn.get('use_rope', True)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

raven/models/raven/modeling_raven.py

@@ -0,0 +1,429 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from raven.layers.raven import RavenAttention
+from raven.models.raven.configuration_raven import RavenConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as GSAMLP
+from fla.modules import RMSNorm
+from fla.modules.l2warp import l2_warp
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class RavenBlock(nn.Module):
+    def __init__(self, config: RavenConfig, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                use_rope=config.attn['use_rope'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = RavenAttention(
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                num_kv_heads=config.num_kv_heads,
+                num_slots=config.num_slots,
+                use_short_conv=config.use_short_conv,
+                conv_size=config.conv_size,
+                feature_map=config.feature_map,
+                use_output_gate=config.use_output_gate,
+                use_norm=config.use_norm,
+                gate_fn=config.hidden_act,
+                topk=config.topk,
+                bias_rmm=config.bias_rmm,
+                add_gumbel_noise=config.add_gumbel_noise,
+                router_score=config.router_score,
+                router_type=config.router_type,
+                decay_type = config.decay_type,
+                gate_logit_normalizer=config.gate_logit_normalizer,
+                elementwise_affine=config.elementwise_affine,
+                norm_eps=config.norm_eps,
+                fuse_norm=config.fuse_norm,
+                layer_idx=layer_idx
+            )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = GSAMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class RavenPreTrainedModel(PreTrainedModel):
+
+    config_class = RavenConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['RavenBlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        prenorm_residual_strategy: Optional[str] = None,
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if prenorm_residual_strategy is not None:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                if prenorm_residual_strategy == 'rescale':
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+                elif prenorm_residual_strategy == 'zero':
+                    nn.init.zeros_(p)
+                else:
+                    raise ValueError(f"Invalid prenorm_residual_strategy: {prenorm_residual_strategy}")
+
+
+class RavenModel(RavenPreTrainedModel):
+
+    def __init__(self, config: RavenConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([RavenBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn("`RavenModel` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        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
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class RavenForCausalLM(RavenPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+
+        super().__init__(config)
+        self.model = RavenModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            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()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training and labels is not None
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss(use_l2warp=self.config.use_l2warp)
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            # Enable model parallelism
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+                loss = l2_warp(loss, logits) if self.config.use_l2warp else loss
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )