Add files using upload-large-folder tool

configs/rectified_transformer_120M.json

@@ -0,0 +1,19 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": false,
+    "hidden_act": "swish",
+    "hidden_size": 768,
+    "initializer_range": 0.02,
+    "max_position_embeddings": 4096,
+    "model_type": "transformer",
+    "num_heads": 12,
+    "num_hidden_layers": 14,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "naive_rectified_attn"
+}

configs/scaled_softpick_transformer_340M.json

@@ -0,0 +1,19 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_size": 1024,
+    "initializer_range": 0.006,
+    "max_position_embeddings": 8192,
+    "model_type": "transformer",
+    "num_heads": 16,
+    "num_hidden_layers": 24,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "parallel_scaled_softpick_attn"
+}

configs/softpick_transformer_120M.json

@@ -0,0 +1,19 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": false,
+    "fuse_norm": false,
+    "hidden_act": "swish",
+    "hidden_size": 768,
+    "initializer_range": 0.02,
+    "max_position_embeddings": 4096,
+    "model_type": "transformer",
+    "num_heads": 12,
+    "num_hidden_layers": 14,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "parallel_softpick_attn"
+}

configs/softpick_transformer_340M.json

@@ -0,0 +1,19 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_size": 1024,
+    "initializer_range": 0.006,
+    "max_position_embeddings": 8192,
+    "model_type": "transformer",
+    "num_heads": 16,
+    "num_hidden_layers": 24,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "parallel_softpick_attn"
+}

configs/softpick_transformer_with_pruning_340M.json

@@ -0,0 +1,63 @@
+{
+  "attention_bias": false,
+  "attn_impl": "parallel_softpick_attn",
+  "bos_token_id": 1,
+  "elementwise_affine": true,
+  "eos_token_id": 2,
+  "fuse_cross_entropy": true,
+  "fuse_norm": true,
+  "fuse_swiglu": true,
+  "hidden_act": "swish",
+  "hidden_ratio": 4,
+  "hidden_size": 1024,
+  "initializer_range": 0.006,
+  "intermediate_size": null,
+  "layer_head_pruned": [
+    [
+      2,
+      1
+    ],
+    [
+      2,
+      7
+    ],
+    [
+      2,
+      12
+    ],
+    [
+      2,
+      13
+    ],
+    [
+      3,
+      5
+    ],
+    [
+      3,
+      13
+    ],
+    [
+      3,
+      14
+    ],
+    [
+      13,
+      6
+    ]
+  ],
+  "max_position_embeddings": 8192,
+  "model_type": "transformer_with_pruning",
+  "norm_eps": 1e-06,
+  "num_heads": 16,
+  "num_hidden_layers": 24,
+  "num_kv_heads": null,
+  "qk_norm": false,
+  "qkv_bias": false,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.51.3",
+  "use_cache": true,
+  "vocab_size": 32000,
+  "window_size": null
+}

configs/stochastic_softpick_transformer_120M.json

@@ -0,0 +1,20 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": false,
+    "fuse_norm": false,
+    "hidden_act": "swish",
+    "hidden_size": 768,
+    "initializer_range": 0.02,
+    "max_position_embeddings": 4096,
+    "model_type": "stochastic_softpick_transformer",
+    "num_heads": 12,
+    "num_hidden_layers": 14,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "parallel_softpick_attn",
+    "stochastic_p": 0.7
+}

configs/transformer_1B.json

@@ -0,0 +1,22 @@
+{
+    "bos_token_id": 1,
+    "elementwise_affine": true,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "fuse_swiglu": true,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 2048,
+    "initializer_range": 0.006,
+    "intermediate_size": null,
+    "max_position_embeddings": 8192,
+    "model_type": "transformer",
+    "norm_eps": 1e-06,
+    "num_heads": 32,
+    "num_hidden_layers": 24,
+    "num_kv_heads": null,
+    "pad_token_id": 2,
+    "rope_theta": 10000.0,
+    "tie_word_embeddings": false
+}

configs/transformer_340M.json

@@ -0,0 +1,18 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": false,
+    "fuse_norm": false,
+    "hidden_act": "swish",
+    "hidden_size": 1024,
+    "initializer_range": 0.006,
+    "max_position_embeddings": 8192,
+    "model_type": "transformer",
+    "num_heads": 16,
+    "num_hidden_layers": 24,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 32000
+}

configs/vanilla_transformer_120M.json

@@ -0,0 +1,19 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": false,
+    "hidden_act": "swish",
+    "hidden_size": 768,
+    "initializer_range": 0.02,
+    "max_position_embeddings": 4096,
+    "model_type": "transformer",
+    "num_heads": 12,
+    "num_hidden_layers": 14,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "attn_impl": "naive_attn"
+}

fla/__init__.py

@@ -0,0 +1,114 @@
+# -*- coding: utf-8 -*-
+
+from fla.layers import (
+    ABCAttention,
+    Attention,
+    BasedLinearAttention,
+    BitAttention,
+    DeltaNet,
+    GatedDeltaNet,
+    GatedDeltaProduct,
+    GatedLinearAttention,
+    GatedSlotAttention,
+    HGRN2Attention,
+    HGRNAttention,
+    LightNetAttention,
+    LinearAttention,
+    MultiScaleRetention,
+    NativeSparseAttention,
+    ReBasedLinearAttention,
+    RWKV6Attention,
+    RWKV7Attention,
+)
+from fla.models import (
+    ABCForCausalLM,
+    ABCModel,
+    BitNetForCausalLM,
+    BitNetModel,
+    DeltaNetForCausalLM,
+    DeltaNetModel,
+    GatedDeltaNetForCausalLM,
+    GatedDeltaNetModel,
+    GatedDeltaProductForCausalLM,
+    GatedDeltaProductModel,
+    GLAForCausalLM,
+    GLAModel,
+    GSAForCausalLM,
+    GSAModel,
+    HGRN2ForCausalLM,
+    HGRN2Model,
+    HGRNForCausalLM,
+    LightNetForCausalLM,
+    LightNetModel,
+    LinearAttentionForCausalLM,
+    LinearAttentionModel,
+    NSAForCausalLM,
+    NSAModel,
+    RetNetForCausalLM,
+    RetNetModel,
+    RWKV6ForCausalLM,
+    RWKV6Model,
+    RWKV7ForCausalLM,
+    RWKV7Model,
+    TransformerForCausalLM,
+    TransformerModel,
+    TransformerWithPruningForCausalLM,
+    TransformerWithPruningModel
+)
+
+__all__ = [
+    'ABCAttention',
+    'Attention',
+    'BasedLinearAttention',
+    'BitAttention',
+    'DeltaNet',
+    'GatedDeltaNet',
+    'GatedDeltaProduct',
+    'GatedLinearAttention',
+    'GatedSlotAttention',
+    'HGRNAttention',
+    'HGRN2Attention',
+    'LightNetAttention',
+    'LinearAttention',
+    'MultiScaleRetention',
+    'NativeSparseAttention',
+    'ReBasedLinearAttention',
+    'RWKV6Attention',
+    'RWKV7Attention',
+    'ABCForCausalLM',
+    'ABCModel',
+    'BitNetForCausalLM',
+    'BitNetModel',
+    'DeltaNetForCausalLM',
+    'DeltaNetModel',
+    'GatedDeltaNetForCausalLM',
+    'GatedDeltaNetModel',
+    'GatedDeltaProductForCausalLM',
+    'GatedDeltaProductModel',
+    'GLAForCausalLM',
+    'GLAModel',
+    'GSAForCausalLM',
+    'GSAModel',
+    'HGRNForCausalLM',
+    'HGRNModel',
+    'HGRN2ForCausalLM',
+    'HGRN2Model',
+    'LightNetForCausalLM',
+    'LightNetModel',
+    'LinearAttentionForCausalLM',
+    'LinearAttentionModel',
+    'NSAForCausalLM',
+    'NSAModel',
+    'RetNetForCausalLM',
+    'RetNetModel',
+    'RWKV6ForCausalLM',
+    'RWKV6Model',
+    'RWKV7ForCausalLM',
+    'RWKV7Model',
+    'TransformerForCausalLM',
+    'TransformerModel',
+    'TransformerWithPruningForCausalLM',
+    'TransformerWithPruningModel',
+]
+
+__version__ = '0.1.2'

fla/layers/delta_net.py

@@ -0,0 +1,291 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torch.nn import functional as F
+
+from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
+from fla.ops.delta_rule import chunk_delta_rule, fused_recurrent_delta_rule
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+def elu_p1(x):
+    return (F.elu(x, 1., False) + 1.).to(x)
+
+
+def sum_norm(x):
+    return (x / x.sum(-1, keepdim=True)).to(x)
+
+
+class DeltaNet(nn.Module):
+    r"""
+    The layer implementaion for [Parallelizing Linear Transformers with the Delta Rule over Sequence Length](https://arxiv.org/abs/2406.06484).  # noqa:
+    DeltaNet was originally proposed in [Linear Transformers Are Secretly Fast Weight Programmers](https://arxiv.org/abs/2102.11174). # noqa
+
+    Args:
+        mode (str, Optional):
+            Which DeltaNet kernel to use.
+            Currently available: `chunk`, `fused_recurrent`, and `fused_chunk`.
+            Default: `chunk`.
+        hidden_size (int, Optional):
+            The hidden size of the input. Default: 1024.
+        expand_k (float, Optional):
+            The expansion ratio for the key dim. Default: 1.0.
+        expand_v (float, Optional):
+            The expansion ratio for the value dim. Default: 1.0.
+        num_heads (int, Optional):
+            The number of heads. Default: 4.
+        use_beta (bool, Optional):
+            Whether to use beta. Default: `True`.
+        use_gate (bool, Optional):
+            Whether to use output gate. Default: `False`.
+        use_short_conv (bool, Optional):
+            Whether to use short convolutions. Default: `True`.
+        conv_size (int, Optional):
+            The kernel size of the short convolution, only used when `use_short_conv` is `True`. Default: 4.
+        conv_bias (bool, Optional):
+            Whether to use bias in the short convolution, only used when `use_short_conv` is `True`. Default: `False`.
+        allow_neg_eigval (bool, Optional):
+            Allow negative eigenvalues. Default: `False`. If set to `True`, the beta will be multiplied by 2.
+            See reference: [Unlocking State-Tracking in Linear RNNs Through Negative Eigenvalues](https://arxiv.org/abs/2411.12537)
+        layer_idx (int, Optional):
+            The index of the layer. Default: None.
+        norm_eps (float, Optional):
+            The epsilon value for the layernorm/rmsnorm layer. Default: 1e-5.
+        qk_activation (str, Optional):
+            The activation function for the query and key. Default: `silu`.
+        qk_norm (str, Optional):
+            The normalization method for the query and key. Default: `l2`.
+    """
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        d_model: int = None,
+        hidden_size: int = 1024,
+        expand_k: float = 1.0,
+        expand_v: float = 1.0,
+        num_heads: int = 4,
+        use_beta: bool = True,
+        use_gate: bool = False,
+        use_short_conv: bool = True,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        allow_neg_eigval: bool = False,
+        layer_idx: int = None,
+        qk_activation: str = 'silu',
+        qk_norm: str = 'l2',
+        norm_eps: float = 1e-5,
+        **kwargs
+    ) -> DeltaNet:
+        super().__init__()
+
+        self.mode = mode
+        self.qk_activation = qk_activation
+        self.qk_norm = qk_norm
+
+        assert self.qk_activation in ['silu', 'relu', 'elu', 'identity']
+        assert self.qk_norm in ['l2', 'sum']
+
+        if d_model is not None:
+            hidden_size = d_model
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.num_heads = num_heads
+        self.use_gate = use_gate
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+        self.allow_neg_eigval = allow_neg_eigval
+
+        self.key_dim = int(hidden_size * expand_k)
+        self.value_dim = int(hidden_size * expand_v)
+        self.head_k_dim = self.key_dim // num_heads
+        self.head_v_dim = self.value_dim // num_heads
+        self.layer_idx = layer_idx
+
+        self.silu = nn.SiLU()
+        if mode == 'fused_chunk':
+            raise NotImplementedError("fused_chunk_delta_rule is now deprecated. Please use `chunk_delta_rule` instead.")
+        assert mode in ['chunk', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+
+        self.use_beta = use_beta
+        if self.use_beta:
+            self.b_proj = nn.Linear(hidden_size, self.num_heads, bias=False)
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(
+                hidden_size=self.key_dim,
+                kernel_size=conv_size,
+                activation='silu' if qk_activation == 'silu' else None
+            )
+            self.k_conv1d = ShortConvolution(
+                hidden_size=self.key_dim,
+                kernel_size=conv_size,
+                activation='silu' if qk_activation == 'silu' else None
+            )
+            self.v_conv1d = ShortConvolution(
+                hidden_size=self.value_dim,
+                kernel_size=conv_size,
+                activation='silu'
+            )
+        else:
+            raise UserWarning(
+                "ShortConvolution is crucial to the performance. "
+                "Do not turn it off, i.e., setting `use_short_conv=False` unless you know what you are doing."
+            )
+        if use_gate:
+            self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+            self.o_norm = FusedRMSNormGated(self.head_v_dim, eps=norm_eps)
+        else:
+            self.o_norm = RMSNorm(self.head_v_dim, eps=norm_eps)
+
+        self.o_proj = nn.Linear(self.value_dim, 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."
+            )
+
+        # change to inference mode.
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            if self.qk_activation == 'silu':
+                q, k = self.silu(q), self.silu(k)
+            v = self.silu(self.v_proj(hidden_states))
+
+        q, k = map(lambda x: rearrange(x, '... (h d) -> ... h d', d=self.head_k_dim), (q, k))
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_v_dim)
+        if self.qk_activation != 'silu':
+            if self.qk_activation == 'relu':
+                q, k = q.relu(), k.relu()
+            elif self.qk_activation == 'elu':
+                q, k = elu_p1(q), elu_p1(k)
+            elif self.qk_activation == 'identity':
+                pass
+            else:
+                raise NotImplementedError
+
+        if self.qk_norm == 'sum':
+            q = sum_norm(q).to(q)
+            k = sum_norm(k).to(k)
+
+        if self.use_beta:
+            beta = self.b_proj(hidden_states).sigmoid()
+        else:
+            beta = q.new_ones(q.shape[0], q.shape[1], q.shape[2])
+
+        if self.allow_neg_eigval:
+            beta = beta * 2.
+
+        # dealing with padding
+        if attention_mask is not None:
+            beta = beta.mul(attention_mask[:, -beta.shape[-2]:, None])
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_delta_rule(
+                q=q,
+                k=k,
+                v=v,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False,
+                use_qk_l2norm_in_kernel=True if self.qk_norm == 'l2' else False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_delta_rule(
+                q=q,
+                k=k,
+                v=v,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False,
+                use_qk_l2norm_in_kernel=True if self.qk_norm == 'l2' else False
+            )
+        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=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        if self.use_gate:
+            g = rearrange(self.g_proj(hidden_states), '... (h d) -> ... h d', d=self.head_v_dim)
+            o = self.o_norm(o, g)
+        else:
+            o = self.o_norm(o)
+        o = rearrange(o, 'b t h d -> b t (h d)')
+        o = self.o_proj(o)
+
+        return o, None, past_key_values

fla/layers/gated_deltanet.py

@@ -0,0 +1,293 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+import math
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torch.nn import functional as F
+
+from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
+from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+@torch.compile
+def elu_p1(x):
+    return (F.elu(x, 1., False) + 1.).to(x)
+
+
+@torch.compile
+def sum_norm(x):
+    return (x / x.sum(-1, keepdim=True)).to(x)
+
+
+class GatedDeltaNet(nn.Module):
+    """
+    The layer implementaion for [Gated Delta Networks: Improving Mamba2 with Delta Rule](https://arxiv.org/abs/2412.06464).  # noqa
+
+    Similar to Mamba2, each layer contains around 6*hidden_size*hidden_size parameters.
+
+    Parameter alloation when use_gate=True:
+        - 0.75 * hidden_size * hidden_size for the q_proj and k_proj each
+        - 1.5 * hidden_size * hidden_size for the v_proj, g_proj and o_proj each
+        - Others are ignorably small.
+        - In total = 0.75 * 2 + 1.5 * 3 = 6 * hidden_size * hidden_size
+    NOTE: num_heads * head_dim = 0.75 * hidden_size, please make sure to set the correct num_heads and head_dim.
+
+    Parameter allocation when use_gate=False:
+        - 1 * hidden_size * hidden_size for the q_proj and k_proj each
+        - 2 * hidden_size * hidden_size for the v_proj and o_proj each
+        - Others are ignorably small.
+        - In total = 1 * 2 + 2 * 2 = 6 * hidden_size * hidden_size
+
+    Args:
+        hidden_size (int, Optional):
+            The hidden size of the input. Default: 2048.
+        expand_v (float, Optional):
+            The expansion ratio for the value dim. Default: 2.0.
+        head_dim (int, Optional):
+            The dimension of each head. Default: 256.
+        num_heads (int, Optional):
+            The number of heads. Default: 4.
+        mode (str, Optional):
+            Which Gated DeltaNet kernel to use.
+            Currently available: `chunk` and `fused_recurrent`.
+            Default: `chunk`.
+        use_beta (bool, Optional):
+            Whether to use beta. Default: `True`.
+        use_gate (bool, Optional):
+            Whether to use output gate. Default: `True`.
+        use_short_conv (bool, Optional):
+            Whether to use short convolutions. Default: `True`.
+        conv_size (int, Optional):
+            The kernel size of the short convolution, only used when `use_short_conv` is `True`. Default: 4.
+        conv_bias (bool, Optional):
+            Whether to use bias in the short convolution, only used when `use_short_conv` is `True`. Default: `False`.
+        layer_idx (int, Optional):
+            The index of the layer. Default: None.
+        norm_eps (float, Optional):
+            The epsilon value for the normalization layer. Default: 1e-5.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        expand_v: float = 2,
+        head_dim: int = 256,
+        num_heads: int = 6,
+        mode: str = 'chunk',
+        use_gate: bool = True,
+        use_short_conv: bool = True,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        layer_idx: int = None,
+        norm_eps: float = 1e-5,
+        **kwargs
+    ) -> GatedDeltaNet:
+        super().__init__()
+
+        self.mode = mode
+
+        self.hidden_size = hidden_size
+        self.expand_v = expand_v
+
+        self.use_gate = use_gate
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.head_dim = head_dim
+        self.num_heads = num_heads
+
+        self.key_dim = int(self.num_heads * self.head_dim)
+        self.value_dim = int(self.key_dim * self.expand_v)
+        self.head_k_dim = head_dim
+        self.head_v_dim = int(head_dim * self.expand_v)
+        self.layer_idx = layer_idx
+
+        # Consistency check: Ensure expand_v produces integer values
+        if not math.isclose(self.key_dim * expand_v, self.value_dim, rel_tol=1e-5):
+            raise ValueError(
+                f"expand_v={expand_v} does not produce an integer value when multiplied by key_dim={self.key_dim}. "
+                f"Resulting value_dim would be {self.key_dim * expand_v}, which is invalid for nn.Linear."
+            )
+        if not math.isclose(head_dim * expand_v, self.head_v_dim, rel_tol=1e-5):
+            raise ValueError(
+                f"expand_v={expand_v} does not produce an integer value when multiplied by head_dim={head_dim}. "
+                f"Resulting head_v_dim would be {head_dim * expand_v}, which is invalid for FusedRMSNormGated."
+            )
+        assert mode in ['chunk', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+        self.a_proj = nn.Linear(hidden_size, self.num_heads, bias=False)
+        self.b_proj = nn.Linear(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
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(
+                hidden_size=self.key_dim,
+                kernel_size=conv_size,
+                activation='silu'
+            )
+            self.k_conv1d = ShortConvolution(
+                hidden_size=self.key_dim,
+                kernel_size=conv_size,
+                activation='silu'
+            )
+            self.v_conv1d = ShortConvolution(
+                hidden_size=self.value_dim,
+                kernel_size=conv_size,
+                activation='silu'
+            )
+        else:
+            raise UserWarning(
+                "ShortConvolution is crucial to the performance. "
+                "Do not turn it off, i.e., setting `use_short_conv=False` unless you know what you are doing."
+            )
+        if use_gate:
+            self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+            self.o_norm = FusedRMSNormGated(self.head_v_dim, eps=norm_eps)
+        else:
+            self.o_norm = RMSNorm(self.head_v_dim, eps=norm_eps)
+        self.o_proj = nn.Linear(self.value_dim, 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."
+            )
+
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+        if self.training:
+            assert mode == 'chunk', "Only chunk mode is supported in training."
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = F.silu(self.q_proj(hidden_states))
+            k = F.silu(self.k_proj(hidden_states))
+            v = F.silu(self.v_proj(hidden_states))
+
+        q, k = map(lambda x: rearrange(x, 'b t (h d) -> b t h d', d=self.head_k_dim), (q, k))
+        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+        beta = self.b_proj(hidden_states).sigmoid()
+        g = -self.A_log.float().exp() * F.softplus(self.a_proj(hidden_states).float() + self.dt_bias)
+
+        # dealing with padding
+        if attention_mask is not None:
+            beta = beta.mul(attention_mask[:, -beta.shape[-2]:, None])
+            g = g.mul(attention_mask[:, -g.shape[-2]:, None])
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'chunk':
+            o, recurrent_state = chunk_gated_delta_rule(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False,
+                use_qk_l2norm_in_kernel=True
+            )
+        elif mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gated_delta_rule(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False,
+                use_qk_l2norm_in_kernel=True
+            )
+        if past_key_values is not None:
+            past_key_values.update(
+                recurrent_state=recurrent_state,
+                conv_state=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        if self.use_gate:
+            g = rearrange(self.g_proj(hidden_states), '... (h d) -> ... h d', d=self.head_v_dim)
+            o = self.o_norm(o, g)
+        else:
+            o = self.o_norm(o)
+        o = rearrange(o, 'b t h d -> b t (h d)')
+        o = self.o_proj(o)
+
+        return o, None, past_key_values

fla/layers/gla.py

@@ -0,0 +1,294 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+
+from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
+from fla.modules.activations import ACT2FN
+from fla.ops.gla import chunk_gla, fused_chunk_gla, fused_recurrent_gla
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class GatedLinearAttention(nn.Module):
+    r"""
+    The layer implementaion for [Gated Linear Attention Transformers with Hardware-Efficient Training](https://arxiv.org/abs/2312.06635).  # noqa
+
+    Args:
+        mode (str, Optional):
+            Which GLA kernel to use.
+            Currently available: `chunk`, `fused_recurrent`, and `fused_chunk`.
+            Default: `chunk`.
+        hidden_size (int, Optional):
+            The hidden size of the input. Default: 1024.
+        expand_k (float, Optional):
+            The expansion ratio for the key dim. Default: 0.5.
+        expand_v (float, Optional):
+            The expansion ratio for the value dim. Default: 1.0.
+        num_heads (int, Optional):
+            The number of heads. Default: 4.
+        num_kv_heads (int, Optional):
+            The number of key/value heads, used for MQA. Default: None.
+        feature_map (str, Optional):
+            Feature map function applied to queries/keys. Default: None.
+        use_short_conv (bool, Optional):
+            Whether to use short convolutions. Default: `False`.
+        conv_size (int, Optional):
+            The kernel size of the short convolution, only used when `use_short_conv` is `True`. Default: 4.
+        conv_bias (bool, Optional):
+            Whether to use bias in the short convolution, only used when `use_short_conv` is `True`. Default: `False`.
+        use_output_gate (bool, Optional):
+            Whether to use output gate. Default: `True`.
+        gate_fn (str, Optional):
+            The activation function for the output gate. Default: `swish`.
+        elementwise_affine (bool, Optional):
+            If `True`, applies elementwise affine to LayerNorm with learnable parameters. Default: `True`.
+        norm_eps (float, Optional):
+            The epsilon value for the layernorm/rmsnorm layer. Default: 1e-5.
+        gate_logit_normalizer (int, Optional):
+            The normalizer for the gate logits, appied after `logsigmoid`. Default: 16.
+        gate_low_rank_dim (int, Optional):
+            The low rank dim for the gate projection. Default: 16.
+        clamp_min (float, Optional):
+            The minimum value for the gate logits. Default: None.
+        fuse_norm (bool, Optional):
+            Whether to fuse the norm and the output gate for better memory footprint. Default: `True`.
+        layer_idx (int, Optional):
+            The index of the layer. Default: None.
+    """
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_k: float = 0.5,
+        expand_v: float = 1.0,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        feature_map: Optional[str] = None,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        use_output_gate: bool = True,
+        gate_fn: str = 'swish',
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        gate_logit_normalizer: int = 16,
+        gate_low_rank_dim: int = 16,
+        clamp_min: Optional[float] = None,
+        fuse_norm: bool = True,
+        layer_idx: int = None,
+    ) -> GatedLinearAttention:
+        super().__init__()
+
+        self.mode = mode
+        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_kv_heads if num_kv_heads is not None else num_heads
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+        self.feature_map_fn = ACT2FN[feature_map] if feature_map is not None else None
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+        self.use_output_gate = use_output_gate
+
+        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.clamp_min = clamp_min
+        self.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_recurrent', 'fused_chunk'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.head_k_dim = self.key_dim // num_heads
+        self.head_v_dim = self.value_dim // num_heads
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim_per_group, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim_per_group, bias=False)
+        if self.use_output_gate:
+            self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
+            self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
+
+        self.gk_proj = nn.Sequential(nn.Linear(hidden_size, gate_low_rank_dim, bias=False),
+                                     nn.Linear(gate_low_rank_dim, self.key_dim_per_group, bias=True))
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+
+        if gate_fn == 'swish' and fuse_norm and use_output_gate:
+            self.g_norm_swish_gate = FusedRMSNormGated(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.fuse_norm_and_gate = True
+        else:
+            self.fuse_norm_and_gate = False
+            self.g_norm = RMSNorm(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.gate_fn = ACT2FN[gate_fn]
+
+        self.gate_logit_normalizer = gate_logit_normalizer
+
+    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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+        gk = self.gk_proj(hidden_states)
+
+        if self.feature_map_fn is not None:
+            q, k = map(self.feature_map_fn, (q, k))
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+        q = rearrange(q, 'b t (h d) -> b t h d', d=self.head_k_dim)
+        if self.num_kv_groups > 1:
+            k, gk = (repeat(x, 'b t (h d) -> b t (h g) d', g=self.num_kv_groups, d=self.head_k_dim) for x in (k, gk))
+            v = repeat(v, 'b t (h d) -> b t (h g) d', g=self.num_kv_groups, d=self.head_v_dim)
+        else:
+            k, gk = (rearrange(x, 'b t (h d) -> b t h d', d=self.head_k_dim) for x in (k, gk))
+            v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+        gk = F.logsigmoid(gk) / self.gate_logit_normalizer
+
+        if self.clamp_min is not None:
+            gk = torch.clamp_min(gk, self.clamp_min)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gla(
+                q=q,
+                k=k,
+                v=v,
+                gk=gk,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'fused_chunk':
+            o, recurrent_state = fused_chunk_gla(
+                q=q,
+                k=k,
+                v=v,
+                g=gk,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_gla(
+                q=q,
+                k=k,
+                v=v,
+                g=gk,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        if self.use_output_gate:
+            g = self.g_proj(hidden_states)
+            if self.fuse_norm_and_gate:
+                g = rearrange(g, 'b t (h d) -> b t h d', d=self.head_v_dim)
+                o = self.g_norm_swish_gate(o, g)
+                o = rearrange(o, 'b t h d -> b t (h d)')
+            else:
+                o = rearrange(self.g_norm(o), 'b t h d -> b t (h d)')
+                o = o * self.gate_fn(g)
+        else:
+            o = rearrange(self.g_norm(o), 'b t h d -> b t (h d)')
+        o = self.o_proj(o)
+
+        return o, None, past_key_values
+
+    def state_size(self, **kwargs) -> int:
+        state_size = self.key_dim * self.head_v_dim
+        for module in self.children():
+            if isinstance(module, ShortConvolution):
+                state_size += module.state_size
+        return state_size

fla/layers/gsa.py

@@ -0,0 +1,227 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+import warnings
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from fla.modules import RMSNorm, ShortConvolution
+from fla.modules.feature_map import ReLUFeatureMap, SwishFeatureMap, T2RFeatureMap
+from fla.modules.layernorm import rms_norm_linear
+from fla.ops.gsa import chunk_gsa, fused_recurrent_gsa
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class GatedSlotAttention(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,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        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.,
+        **kwargs
+    ) -> GatedSlotAttention:
+        super().__init__()
+
+        self.mode = mode
+        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.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.gate_logit_normalizer = gate_logit_normalizer
+
+        self.use_output_gate = use_output_gate
+        self.use_norm = use_norm
+        self.scale = scale
+
+        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.")
+
+        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)
+        self.f_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.num_slots, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
+            self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
+
+        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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+        f = self.f_proj(hidden_states)
+
+        q = rearrange(q, 'b t (h d) -> b t h d', d=self.head_k_dim)
+        k = rearrange(k, 'b t (h d) -> b t h d', d=self.head_k_dim)
+        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+        f = rearrange(f, 'b t (h m) -> b t h m', m=self.num_slots)
+
+        if self.feature_map is not None:
+            q, k = map(lambda x: self.feature_map(x), (q, k))
+        v = F.silu(v)
+
+        f = F.logsigmoid(f) / self.gate_logit_normalizer
+        s = (1 - f.exp()).to(f.dtype)
+        # dealing with left-padding
+        if attention_mask is not None:
+            s = s.mul_(attention_mask[:, -s.shape[1]:, None, None])
+            v = v.mul_(attention_mask[:, -v.shape[1]:, None, None])
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gsa(
+                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,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_gsa(
+                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,
+                head_first=False
+            )
+        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=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        o = rearrange(o, 'b t h d -> b t (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)
+        return o, None, past_key_values
+
+    def state_size(self, *args, **kwargs) -> int:
+        return 2 * self.num_slots * self.hidden_size

fla/layers/hgrn.py

@@ -0,0 +1,168 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# "Hierarchically Gated Recurrent Neural Network for Sequence Modeling" [https://arxiv.org/abs/2311.04823]
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fla.modules import FusedRMSNormGated, ShortConvolution
+from fla.modules.activations import swiglu
+from fla.ops.hgrn import chunk_hgrn, fused_recurrent_hgrn
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class HGRNAttention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_ratio: Optional[int] = 1,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None
+    ) -> HGRNAttention:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+        self.expand_ratio = expand_ratio
+        self.input_dim = int(hidden_size * expand_ratio)
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+
+        self.i_proj = nn.Linear(hidden_size, self.input_dim, bias=False)
+        self.f_proj = nn.Linear(hidden_size, self.input_dim, bias=False)
+        self.g_proj = nn.Linear(hidden_size, self.input_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.input_dim, conv_size, activation=None)
+            self.f_conv1d = ShortConvolution(self.input_dim, conv_size, activation=None)
+            self.i_conv1d = ShortConvolution(self.input_dim, conv_size, activation=None)
+
+        self.g_norm = FusedRMSNormGated(
+            hidden_size=self.input_dim,
+            elementwise_affine=elementwise_affine,
+            eps=norm_eps
+        )
+        self.o_proj = nn.Linear(self.input_dim, 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,
+        lower_bound: Optional[torch.Tensor] = None,
+        **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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if not self.training and hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_i, conv_state_f = None, None
+            if last_state is not None:
+                conv_state_i, conv_state_f = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            i, conv_state_i = self.i_conv1d(
+                x=self.i_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_i,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            f, conv_state_f = self.f_conv1d(
+                x=self.f_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_f,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            i = self.i_proj(hidden_states)
+            f = self.f_proj(hidden_states)
+
+        # the lower bound for the first layer is zero
+        if lower_bound is None or self.layer_idx == 0:
+            i, f = swiglu(i, 1 - f.sigmoid()), F.logsigmoid(f)
+        else:
+            g = lower_bound + (1 - lower_bound) * f.sigmoid()
+            i, f = swiglu(i, 1 - g), g.log()
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            i = i.mul_(attention_mask[:, -i.shape[-2]:, None])
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'chunk':
+            if cu_seqlens is not None:
+                raise NotImplementedError("Chunk mode does not support variable-length sequences.")
+            o, recurrent_state = chunk_hgrn(
+                x=i,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+            )
+        elif mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_hgrn(
+                x=i,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                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=(conv_state_i, conv_state_f) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=i.shape[2]
+            )
+
+        o = self.g_norm(o, self.g_proj(hidden_states))
+        o = self.o_proj(o)
+
+        return o, None, past_key_values
+
+    def state_size(self, **kwargs) -> int:
+        state_size = self.hidden_size
+        for module in self.children():
+            if isinstance(module, ShortConvolution):
+                state_size += module.state_size
+        return state_size

fla/layers/hgrn2.py

@@ -0,0 +1,211 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# "HGRN2: Gated Linear RNNs with State Expansion"[https://arxiv.org/abs/2404.07904]
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from fla.modules import RMSNorm, ShortConvolution
+from fla.modules.activations import swish
+from fla.modules.layernorm import rms_norm_linear
+from fla.ops.gla import chunk_gla, fused_chunk_gla, fused_recurrent_gla
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class HGRN2Attention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        num_heads: Optional[int] = None,
+        expand_ratio: Optional[int] = 128,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None
+    ) -> HGRN2Attention:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+
+        if expand_ratio is None and num_heads is not None:
+            expand_ratio = hidden_size // num_heads
+        elif expand_ratio is not None and num_heads is None:
+            num_heads = hidden_size // expand_ratio
+        elif expand_ratio is None and num_heads is None:
+            raise RuntimeError("One of `expand_ratio` or `num_heads` should be provided.")
+        self.num_heads = num_heads
+        self.expand_ratio = expand_ratio
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.forget_dim = int(self.num_heads * self.expand_ratio)
+        self.input_dim = hidden_size
+        self.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_recurrent', 'fused_chunk'], f"Not suppoerted mode `{mode}`."
+        assert self.forget_dim % num_heads == 0, f"forget dim must be divisible by num_heads of {num_heads}"
+        assert self.input_dim % num_heads == 0, f"input dim must be divisible by num_heads of {num_heads}"
+
+        self.head_f_dim = self.expand_ratio
+        self.head_i_dim = self.hidden_size // num_heads
+
+        self.q_proj = nn.Linear(hidden_size, self.forget_dim, bias=False)
+        self.f_proj = nn.Linear(hidden_size, self.forget_dim, bias=False)
+        self.i_proj = nn.Linear(hidden_size, self.input_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.forget_dim, conv_size, activation=None)
+            self.f_conv1d = ShortConvolution(self.forget_dim, conv_size, activation=None)
+            self.i_conv1d = ShortConvolution(self.input_dim, conv_size, activation=None)
+
+        self.g_norm = RMSNorm(hidden_size=self.hidden_size, elementwise_affine=elementwise_affine, eps=norm_eps)
+        self.o_proj = nn.Linear(self.input_dim, 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,
+        lower_bound: Optional[torch.Tensor] = None,
+        **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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_f, conv_state_i = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_f, conv_state_i = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            f, conv_state_f = self.f_conv1d(
+                x=self.f_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_f,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            i, conv_state_i = self.i_conv1d(
+                x=self.i_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_i,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            f = self.f_proj(hidden_states)
+            i = self.i_proj(hidden_states)
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            i = i.mul_(attention_mask[:, -i.shape[-2]:, None])
+
+        q = swish(q)
+
+        # improve precision
+        f = f.float()
+
+        # the lower bound for the first layer is zero
+        if lower_bound is None or self.layer_idx == 0:
+            k, g = 1 - f.sigmoid(), F.logsigmoid(f)
+        else:
+            g = lower_bound + (1 - lower_bound) * f.sigmoid()
+            k, g = 1 - g, g.log()
+
+        q, k, g = map(lambda x: rearrange(x, '... (h d) -> ... h d', d=self.head_f_dim), (q, k.to(i), g))
+        i = rearrange(i, '... (h d) -> ... h d', d=self.head_i_dim)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gla(
+                q=q,
+                k=k,
+                v=i,
+                gk=g,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'fused_chunk':
+            o, recurrent_state = fused_chunk_gla(
+                q=q,
+                k=k,
+                v=i,
+                g=g,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_gla(
+                q=q,
+                k=k,
+                v=i,
+                g=g,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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=(conv_state_q, conv_state_f, conv_state_i) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        o = rearrange(o, '... h d -> ... (h d)')
+        o = rms_norm_linear(o, self.g_norm.weight, self.g_norm.bias, self.o_proj.weight, self.o_proj.bias)
+        return o, None, past_key_values
+
+    def state_size(self, **kwargs) -> int:
+        state_size = self.forget_dim * self.head_i_dim
+        for module in self.children():
+            if isinstance(module, ShortConvolution):
+                state_size += module.state_size
+        return state_size

fla/layers/lightnet.py

@@ -0,0 +1,210 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# ["You Only Scan Once: Efficient Multi-dimension Sequential Modeling with LightNet"](https://arxiv.org/abs/2405.21022)
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from fla.modules import FusedRMSNormGated, ShortConvolution
+from fla.modules.fused_norm_gate import rms_norm_swish_gate_linear
+from fla.ops.gla import chunk_gla, fused_recurrent_gla
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class LightNetAttention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        num_heads: Optional[int] = None,
+        expand_ratio: Optional[int] = 128,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        gate_low_rank_dim: int = 128,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None
+    ) -> LightNetAttention:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+
+        if expand_ratio is None and num_heads is not None:
+            expand_ratio = hidden_size // num_heads
+        elif expand_ratio is not None and num_heads is None:
+            num_heads = hidden_size // expand_ratio
+        elif expand_ratio is None and num_heads is None:
+            raise RuntimeError("One of `expand_ratio` or `num_heads` should be provided.")
+        self.num_heads = num_heads
+        self.expand_ratio = expand_ratio
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.key_dim = int(self.num_heads * self.expand_ratio)
+        self.value_dim = hidden_size
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_chunk'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.head_f_dim = self.expand_ratio
+        self.head_i_dim = self.hidden_size // num_heads
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation=None)
+            self.k_conv1d = ShortConvolution(self.key_dim, conv_size, activation=None)
+            self.v_conv1d = ShortConvolution(self.value_dim, conv_size, activation=None)
+
+        self.g_proj = nn.Sequential(
+            nn.Linear(hidden_size, gate_low_rank_dim, bias=False),
+            nn.Linear(gate_low_rank_dim, hidden_size, bias=False)
+        )
+        self.g_norm = FusedRMSNormGated(
+            hidden_size=hidden_size,
+            elementwise_affine=elementwise_affine,
+            eps=norm_eps
+        )
+        self.o_proj = nn.Linear(self.value_dim, 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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+
+        q = F.silu(q)
+        q, k = map(lambda x: rearrange(x, '... (h d) -> ... h d', d=self.head_f_dim), (q, k))
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_i_dim)
+        # TODO: this 2 steps took huge amount of time, which should be optimized
+        z = k.float().logcumsumexp(1)
+
+        if cu_seqlens is not None:
+            raise NotImplementedError("LightNet does not support variable-length sequences for now.")
+        k, g = torch.exp(k - z).to(k.dtype), (torch.cat((z[:, :1], z[:, :-1]), 1) - z).to(k.dtype)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gla(
+                q=q,
+                k=k,
+                v=v,
+                gk=g,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_gla(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        o = rms_norm_swish_gate_linear(
+            rearrange(o, 'b t h d -> b t (h d)'),
+            self.g_proj(hidden_states),
+            self.g_norm.weight,
+            self.g_norm.bias,
+            self.o_proj.weight,
+            self.o_proj.bias
+        )
+        return o, None, past_key_values
+
+    def state_size(self, **kwargs) -> int:
+        state_size = self.key_dim * self.head_i_dim
+        for module in self.children():
+            if isinstance(module, ShortConvolution):
+                state_size += module.state_size
+        return state_size

fla/layers/linear_attn.py

@@ -0,0 +1,166 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+
+from fla.modules import RMSNorm
+from fla.modules.feature_map import DPFPFeatureMap, HadamardFeatureMap, HedgehogFeatureMap, T2RFeatureMap
+from fla.ops.linear_attn import chunk_linear_attn, fused_chunk_linear_attn, fused_recurrent_linear_attn
+
+
+class LinearAttention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: str = 1024,
+        expand_k: int = 1.0,
+        expand_v: int = 1.0,
+        num_heads: int = 8,
+        num_kv_heads: Optional[int] = None,
+        feature_map: str = 'elementwise_product',
+        tie_feature_map_qk: bool = False,
+        output_norm: str = 'rmsnorm',
+        norm_q: bool = False,
+        norm_k: bool = False,
+        do_feature_map_norm: bool = False,
+        elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        **kwargs
+    ):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.mode = mode
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_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
+
+        assert mode in ['chunk', 'fused_chunk', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.head_k_dim = self.key_dim // num_heads
+        self.head_v_dim = self.value_dim // num_heads
+        self.do_feature_map_norm = do_feature_map_norm
+
+        if feature_map == 'hedgehog':
+            if tie_feature_map_qk:
+                self.feature_map_q = self.feature_map_k = HedgehogFeatureMap(head_dim=self.head_k_dim)
+            else:
+                self.feature_map_q = HedgehogFeatureMap(head_dim=self.head_k_dim)
+                self.feature_map_k = HedgehogFeatureMap(head_dim=self.head_k_dim)
+
+        elif feature_map == 't2r':
+            if tie_feature_map_qk:
+                self.feature_map_q = self.feature_map_k = T2RFeatureMap(head_dim=self.head_k_dim)
+            else:
+                self.feature_map_q = T2RFeatureMap(head_dim=self.head_k_dim)
+                self.feature_map_k = T2RFeatureMap(head_dim=self.head_k_dim)
+
+        elif feature_map == 'elementwise_product':
+            if tie_feature_map_qk:
+                self.feature_map_q = self.feature_map_k = HadamardFeatureMap(head_dim=self.head_k_dim)
+            else:
+                self.feature_map_q = HadamardFeatureMap(head_dim=self.head_k_dim)
+                self.feature_map_k = HadamardFeatureMap(head_dim=self.head_k_dim)
+
+        elif feature_map == 'dpfp':
+            self.feature_map_q = DPFPFeatureMap(head_dim=self.head_k_dim)
+            self.feature_map_k = DPFPFeatureMap(head_dim=self.head_k_dim)
+
+        elif feature_map == 'elu':
+            def elu(x):
+                return F.elu(x) + 1
+            self.feature_map_q = elu
+            self.feature_map_k = elu
+
+        elif feature_map == 'relu':
+            self.feature_map_q = nn.ReLU()
+            self.feature_map_k = nn.ReLU()
+
+        elif feature_map == 'identity':
+            self.feature_map_q = nn.Identity()
+            self.feature_map_k = nn.Identity()
+        else:
+            raise NotImplementedError(f"Not supported feature map `{feature_map}`.")
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim_per_group, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim_per_group, bias=False)
+
+        if output_norm == 'rmsnorm':
+            self.norm = RMSNorm(hidden_size=self.head_v_dim, elementwise_affine=elementwise_affine, eps=norm_eps)
+        elif output_norm == 'identity':
+            self.norm = nn.Identity()
+        else:
+            raise NotImplementedError(f"Not supported output norm `{output_norm}`.")
+
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+
+        self.norm_q = norm_q
+        self.norm_k = norm_k
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        **kwargs
+    ) -> torch.Tensor:
+        mode = self.mode
+        q = self.q_proj(hidden_states)
+        k = self.k_proj(hidden_states)
+        v = self.v_proj(hidden_states)
+
+        q = rearrange(q, '... (h d) -> ... h d', d=self.head_k_dim)
+        if self.num_kv_groups > 1:
+            k = repeat(k, '... (h d) -> ... (h g) d', d=self.head_k_dim, g=self.num_kv_groups)
+            v = repeat(v, '... (h d) -> ... (h g) d', d=self.head_v_dim, g=self.num_kv_groups)
+        else:
+            k = rearrange(k, '... (h d) -> ... h d', d=self.head_k_dim)
+            v = rearrange(v, '... (h d) -> ... h d', d=self.head_v_dim)
+
+        q = self.feature_map_q(q)
+        k = self.feature_map_k(k)
+
+        if self.norm_q:
+            q = q / (q.sum(-1, True) + 1e-4)
+        if self.norm_k:
+            k = k / (k.sum(-1, True) + 1e-4)
+
+        if mode == 'chunk':
+            o, final_state = chunk_linear_attn(
+                q=q,
+                k=k,
+                v=v,
+                normalize=self.do_feature_map_norm,
+                head_first=False
+            )
+        elif mode == 'fused_chunk':
+            o, final_state = fused_chunk_linear_attn(
+                q=q,
+                k=k,
+                v=v,
+                normalize=self.do_feature_map_norm,
+            )
+        elif mode == 'fused_recurrent':
+            o, final_state = fused_recurrent_linear_attn(
+                q=q,
+                k=k,
+                v=v,
+                normalize=self.do_feature_map_norm,
+            )
+        else:
+            raise NotImplementedError
+        o = self.norm(o)
+        o = rearrange(o, '... h d -> ... (h d)')
+        o = self.o_proj(o)
+        return o

fla/layers/multiscale_retention.py

@@ -0,0 +1,298 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+from transformers.activations import ACT2FN
+
+from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
+from fla.modules.rotary import RotaryEmbedding
+from fla.ops.retention import chunk_retention, fused_chunk_retention, fused_recurrent_retention, parallel_retention
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+class MultiScaleRetention(nn.Module):
+    r"""
+    The layer implementaion for [Retentive Network: A Successor to Transformer for Large Language Models](https://arxiv.org/pdf/2307.08621.pdf).  # noqa
+
+    Args:
+        mode (str, Optional):
+            Which Retention kernel to use.
+            Currently available: `chunk`, `fused_recurrent`, `parallel`, and `fused_chunk`.
+            Default: `chunk`.
+        hidden_size (int, Optional):
+            The hidden size of the input. Default: 1024.
+        expand_k (float, Optional):
+            The expansion ratio for the key dim. Default: 1.0.
+        expand_v (float, Optional):
+            The expansion ratio for the value dim. Default: 2.0.
+        num_heads (int, Optional):
+            The number of heads. Default: 8.
+        num_kv_heads (int, Optional):
+            The number of key/value heads, used for MQA. Default: None.
+        feature_map (str, Optional):
+            Feature map function applied to queries/keys. Default: None.
+        use_short_conv (bool, Optional):
+            Whether to use short convolutions. Default: `False`.
+        conv_size (int, Optional):
+            The kernel size of the short convolution, only used when `use_short_conv` is `True`. Default: 4.
+        conv_bias (bool, Optional):
+            Whether to use bias in the short convolution, only used when `use_short_conv` is `True`. Default: `False`.
+        use_output_gate (bool, Optional):
+            Whether to use output gate. Default: `True`.
+        gate_fn (str, Optional):
+            The activation function for the output gate. Default: `swish`.
+        elementwise_affine (bool, Optional):
+            If `True`, applies elementwise affine to LayerNorm with learnable parameters. Default: `True`.
+        norm_eps (float, Optional):
+            The epsilon value for the layernorm/rmsnorm layer. Default: 1e-5.
+        fuse_norm (bool, Optional):
+            Whether to fuse the norm and the output gate for better memory footprint. Default: `True`.
+        layer_idx (int, Optional):
+            The index of the layer. Default: None.
+    """
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_k: float = 1.0,
+        expand_v: float = 2.0,
+        num_heads: int = 8,
+        num_kv_heads: Optional[int] = None,
+        feature_map: Optional[str] = None,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        use_output_gate: bool = True,
+        gate_fn: str = 'swish',
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        fuse_norm: bool = True,
+        layer_idx: int = None,
+        **kwargs
+    ) -> MultiScaleRetention:
+        super().__init__()
+
+        self.mode = mode
+        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_kv_heads if num_kv_heads is not None else num_heads
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+        self.feature_map_fn = ACT2FN[feature_map] if feature_map is not None else None
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+        self.use_output_gate = use_output_gate
+
+        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.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_chunk', 'parallel', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.head_k_dim = self.key_dim // num_heads
+        self.head_v_dim = self.value_dim // num_heads
+
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim_per_group, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim_per_group, bias=False)
+        if self.use_output_gate:
+            self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
+            self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
+
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+
+        if gate_fn == 'swish' and fuse_norm and use_output_gate:
+            self.g_norm_swish_gate = FusedRMSNormGated(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.fuse_norm_and_gate = True
+        else:
+            self.fuse_norm_and_gate = False
+            self.g_norm = RMSNorm(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.gate_fn = ACT2FN[gate_fn]
+
+        # TODO: fix this issue
+        # https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/ops/triton/rotary.py#L180
+        # Ideally, we would want to support arbitrary d_head_qk
+        assert self.head_k_dim <= 256, "head_k_dim must be less than or equal to 256"
+        self.rotary = RotaryEmbedding(dim=self.head_k_dim)
+
+    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
+    ) -> 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."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+        q = rearrange(q, '... (h d) -> ... h d', d=self.head_k_dim)
+        k = rearrange(k, '... (h d) -> ... h d', d=self.head_k_dim)
+        if self.feature_map_fn is not None:
+            q, k = map(self.feature_map_fn, (q, k))
+
+        seqlen_offset, max_seqlen = 0, q.shape[1]
+        if past_key_values is not None:
+            seqlen_offset = past_key_values.get_seq_length(self.layer_idx)
+            max_seqlen = q.shape[1] + seqlen_offset
+
+            if attention_mask is not None:
+                # to deliminate the offsets of padding tokens
+                seqlen_offset = seqlen_offset + attention_mask.sum(-1) - attention_mask.shape[-1]
+                max_seqlen = q.shape[1] + max(seqlen_offset)
+
+        q, k = self.rotary(q, k, seqlen_offset=seqlen_offset, max_seqlen=max_seqlen, cu_seqlens=cu_seqlens)
+
+        if self.num_kv_groups > 1:
+            k = repeat(k, 'b t h d -> b t (h g) d', g=self.num_kv_groups)
+            v = repeat(v, 'b t (h d) -> b t (h g) d', d=self.head_v_dim, g=self.num_kv_groups)
+        else:
+            v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'chunk':
+            o, recurrent_state = chunk_retention(
+                q=q,
+                k=k,
+                v=v,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'fused_chunk':
+            o, recurrent_state = fused_chunk_retention(
+                q=q,
+                k=k,
+                v=v,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'parallel':
+            o, recurrent_state = parallel_retention(
+                q=q,
+                k=k,
+                v=v,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_retention(
+                q=q,
+                k=k,
+                v=v,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        if self.use_output_gate:
+            g = self.g_proj(hidden_states)
+            if self.fuse_norm_and_gate:
+                g = rearrange(g, 'b t (h d) -> b t h d', d=self.head_v_dim)
+                o = self.g_norm_swish_gate(o, g)
+                o = rearrange(o, 'b t h d -> b t (h d)')
+            else:
+                o = rearrange(self.g_norm(o), 'b t h d -> b t (h d)')
+                o = o * self.gate_fn(g)
+        else:
+            o = rearrange(self.g_norm(o), 'b t h d -> b t (h d)')
+        o = self.o_proj(o)
+
+        return o, None, past_key_values
+
+    def state_size(self, **kwargs) -> int:
+        state_size = self.key_dim * self.head_v_dim
+        for module in self.children():
+            if isinstance(module, ShortConvolution):
+                state_size += module.state_size
+        return state_size

fla/layers/nsa.py

@@ -0,0 +1,138 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from transformers.utils import logging
+
+from fla.modules import RotaryEmbedding
+from fla.ops.nsa.parallel import parallel_nsa
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+logger = logging.get_logger(__name__)
+
+
+class NativeSparseAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_heads: int = 64,
+        num_kv_heads: Optional[int] = 4,
+        head_dim: int = 64,
+        qkv_bias: bool = False,
+        block_size: Optional[int] = 64,
+        block_counts: Optional[Union[torch.LongTensor, int]] = 16,
+        window_size: Optional[int] = 512,
+        rope_theta: Optional[float] = 10000.,
+        max_position_embeddings: Optional[int] = None,
+        layer_idx: int = None
+    ):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        if num_kv_heads is None:
+            self.num_kv_heads = self.num_heads
+        else:
+            self.num_kv_heads = num_kv_heads
+        self.num_kv_groups = num_heads // self.num_kv_heads
+        self.head_dim = head_dim
+        self.kv_dim = self.num_kv_heads * self.head_dim
+        self.qkv_bias = qkv_bias
+
+        self.block_size = block_size
+        self.block_counts = block_counts
+        self.window_size = window_size
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_idx = layer_idx
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=self.qkv_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.kv_dim, bias=self.qkv_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.kv_dim, bias=self.qkv_bias)
+        self.g_proj = nn.Linear(self.hidden_size, self.num_heads * 3, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary = RotaryEmbedding(dim=self.head_dim, base=self.rope_theta)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        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, seq_len, _ = hidden_states.size()
+
+        q = rearrange(self.q_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+        k = rearrange(self.k_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+        v = rearrange(self.v_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+        g = rearrange(self.g_proj(hidden_states), '... (h d) -> ... h d', d=3)
+        g_cmp, g_slc, g_swa = g.sigmoid().unbind(-1)
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+
+        seqlen_offset, max_seqlen = 0, seq_len
+        if past_key_values is not None:
+            seqlen_offset = past_key_values.get_seq_length(self.layer_idx)
+            max_seqlen = q.shape[1] + seqlen_offset
+
+            if attention_mask is not None:
+                # to deliminate the offsets of padding tokens
+                seqlen_offset = seqlen_offset + attention_mask.sum(-1) - attention_mask.shape[-1]
+                max_seqlen = q.shape[1] + max(seqlen_offset)
+
+        if self.max_position_embeddings is not None:
+            max_seqlen = max(max_seqlen, self.max_position_embeddings)
+        q, k = self.rotary(q, k, seqlen_offset=seqlen_offset, max_seqlen=max_seqlen, cu_seqlens=cu_seqlens)
+
+        if past_key_values is not None:
+            cache_has_content = past_key_values.get_seq_length(self.layer_idx) > 0
+            k_cached, v_cached = past_key_values.update(
+                attn_state=(k.flatten(-2, -1), v.flatten(-2, -1)),
+                layer_idx=self.layer_idx,
+                offset=seq_len,
+                cache_kwargs=dict(window_size=self.window_size)
+            )['attn_state']
+            if cache_has_content:
+                k, v = k_cached, v_cached
+                k = rearrange(k, '... (h d) -> ... h d', d=self.head_dim)
+                v = rearrange(v, '... (h d) -> ... h d', d=self.head_dim)
+
+        o = parallel_nsa(
+            q=q,
+            k=k,
+            v=v,
+            g_cmp=g_cmp,
+            g_slc=g_slc,
+            g_swa=g_swa,
+            block_size=self.block_size,
+            block_counts=self.block_counts,
+            window_size=self.window_size,
+            cu_seqlens=cu_seqlens,
+            head_first=False
+        )
+        o = o.reshape(batch_size, seq_len, -1)
+        o = self.o_proj(o)
+
+        if not output_attentions:
+            attentions = None
+
+        return o, attentions, past_key_values

fla/layers/rebased.py

@@ -0,0 +1,133 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+"""
+https://github.com/corl-team/rebased/blob/main/flash_linear_attention/fla/layers/rebased_fast.py
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from fla.modules.feature_map import RebasedFeatureMap
+from fla.ops.linear_attn import chunk_linear_attn, fused_chunk_linear_attn
+from fla.ops.rebased import parallel_rebased
+
+
+class ReBasedLinearAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        l_max: int = 2048,
+        feature_dim: int = 16,
+        num_key_value_heads: int = 16,
+        num_heads: int = 16,
+        use_gamma: Optional[bool] = True,
+        use_beta: Optional[bool] = True,
+        normalize: Optional[bool] = True,
+        causal: bool = True,
+        eps: float = 1e-5,
+        mode: str = "parallel",
+        layer_idx: Optional[int] = None,
+        **kwargs
+    ) -> ReBasedLinearAttention:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.l_max = l_max
+        self.mode = mode
+        assert self.mode in ["fused_chunk", "parallel", 'chunk']
+
+        self.feature_dim = feature_dim
+        self.num_key_value_heads = num_key_value_heads
+        self.num_heads = num_heads
+        self.head_dim = self.hidden_size // self.num_key_value_heads
+        self.use_gamma = use_gamma
+        self.use_beta = use_beta
+        self.normalize = normalize
+        self.causal = causal
+        self.eps = eps
+        self.mode = mode
+        self.layer_idx = layer_idx
+
+        self.feature_map = RebasedFeatureMap(self.feature_dim, use_gamma, use_beta, normalize)
+        self.q_proj = nn.Linear(self.hidden_size, self.feature_dim * self.num_heads, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.feature_dim * self.num_heads, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.dropout = nn.Identity()
+
+    def forward(self, hidden_states: torch.Tensor, **kwargs):
+        mode = self.mode
+        q, k, v = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(hidden_states)
+        q, k, v = map(lambda x: rearrange(x, "... (h d) -> ... h d", d=self.head_dim), [q, k, v])
+        q, k = self.feature_map(q, flatten=(mode != 'parallel')), self.feature_map(k, flatten=(mode != 'parallel'))
+        if mode == "fused_chunk":
+            o = fused_chunk_linear_attn(
+                q=q,
+                k=k,
+                v=v,
+                normalize=True,
+                scale=1,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o = chunk_linear_attn(
+                q=q,
+                k=k,
+                v=v,
+                normalize=True,
+                scale=1,
+                head_first=False
+            )
+        elif mode == 'parallel':
+            assert q.shape[-1] <= 128
+            o = parallel_rebased(
+                q=q,
+                k=k,
+                v=v,
+                eps=self.eps,
+                use_scale=True,
+                use_normalize=True,
+                head_first=False
+            )
+        o = self.o_proj(o)
+        o = self.dropout(o)
+        return o
+
+    # https://github.com/HazyResearch/zoology/blob/main/zoology/mixers/based.py#L119
+    def forward_reference(
+        self,
+        hidden_states: torch.Tensor,
+        filters: torch.Tensor = None,
+        *args,
+        **kwargs
+    ):
+        """
+        x (torch.Tensor): tensor of shape (b, d, t)
+        y (torch.Tensor): tensor of shape (b, d, t)
+        """
+        b, t, _ = hidden_states.size()
+        q, k, v = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(hidden_states)
+
+        q = q.view(b, t, -1, self.feature_dim).transpose(1, 2)
+        k = k.view(b, t, -1, self.feature_dim).transpose(1, 2)
+        v = v.view(b, t, -1, self.head_dim).transpose(1, 2)
+
+        # Linear attention
+        q, k = self.feature_map(q), self.feature_map(k)
+        q, k, v = q.unsqueeze(-2), k.unsqueeze(-2), v.unsqueeze(-1)
+
+        # Compute attention
+        if self.causal:
+            y = ((q * (k * v).cumsum(2)).sum(-1) / ((q * k.cumsum(2)).sum(-1) + self.eps))
+        else:
+            y = ((q * (k * v).sum(2, True)).sum(-1) / ((q * k.sum(2, True)).sum(-1) + self.eps))
+        y = rearrange(y, 'b h t d -> b t (h d)')
+        y = self.o_proj(y.to(hidden_states.dtype))
+        y = self.dropout(y)
+        return y.to(hidden_states.dtype)

fla/layers/rwkv6.py

@@ -0,0 +1,307 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# "Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence"[https://arxiv.org/abs/2404.05892]
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from fla.modules import GroupNorm
+from fla.modules.activations import ACT2FN
+from fla.ops.rwkv6 import chunk_rwkv6, fused_recurrent_rwkv6
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+class RWKV6Attention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_k: float = 0.5,
+        expand_v: float = 1.0,
+        num_heads: int = 4,
+        gate_fn: str = 'swish',
+        proj_low_rank_dim: int = 32,
+        gate_low_rank_dim: int = 64,
+        fuse_norm: bool = True,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None,
+        **kwargs
+    ) -> RWKV6Attention:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.num_heads = num_heads
+        self.proj_low_rank_dim = proj_low_rank_dim
+        self.gate_low_rank_dim = gate_low_rank_dim
+
+        self.key_dim = int(hidden_size * expand_k)
+        self.value_dim = int(hidden_size * expand_v)
+        self.layer_idx = layer_idx
+
+        assert mode in ['chunk', 'fused_recurrent'], f"Not suppoerted mode `{mode}`."
+        assert self.key_dim % num_heads == 0, f"key dim must be divisible by num_heads of {num_heads}"
+        assert self.value_dim % num_heads == 0, f"value dim must be divisible by num_heads of {num_heads}"
+
+        self.head_k_dim = self.key_dim // num_heads
+        self.head_v_dim = self.value_dim // num_heads
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+        self.x_proj = nn.Sequential(
+            LerpLinear(hidden_size, proj_low_rank_dim * 5),
+            nn.Tanh(),
+            nn.Linear(proj_low_rank_dim * 5, hidden_size, bias=False)
+        )
+        self.x_bias = nn.Parameter(torch.zeros(5, hidden_size))
+
+        self.r_proj = DDLerpLinear(hidden_size, self.key_dim)
+        self.w_proj = DDLerpLinear(hidden_size, self.key_dim, low_rank_dim=gate_low_rank_dim)
+        self.k_proj = DDLerpLinear(hidden_size, self.key_dim)
+        self.v_proj = DDLerpLinear(hidden_size, self.value_dim)
+        self.g_proj = DDLerpLinear(hidden_size, self.value_dim)
+        self.bonus = nn.Parameter(torch.zeros(num_heads, self.head_k_dim))
+
+        # TODO: fuse GroupNorm and output gate
+        self.g_norm = GroupNorm(self.num_heads, self.value_dim, elementwise_affine=elementwise_affine, bias=True, eps=norm_eps)
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+        self.gate_fn = ACT2FN[gate_fn]
+
+        self.apply(self._initialize_weights)
+
+    def _initialize_weights(self, module: nn.Module):
+        if getattr(module, "_is_hf_initialized", False):
+            return
+        if isinstance(module, nn.Linear):
+            nn.init.xavier_uniform_(module.weight, gain=2 ** -2.5)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        if isinstance(module, nn.Parameter):
+            nn.init.xavier_uniform_(module, gain=2 ** -2.5)
+        module._is_hf_initialized = True
+
+    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
+    ) -> 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, seq_len, hidden_size = hidden_states.shape
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        if attention_mask is not None:
+            hidden_states = hidden_states.mul_(attention_mask[:, -hidden_states.shape[-2]:, None])
+        if hidden_states.shape[1] == 1 and last_state is not None:
+            shifted = last_state['conv_state'].unsqueeze(1)
+        else:
+            shifted = self.time_shift(hidden_states)
+            if last_state is not None:
+                shifted[:, 0] = last_state['conv_state']
+
+        delta = shifted - hidden_states
+        x = self.x_proj[0](hidden_states, delta).view(batch_size, seq_len, -1, self.proj_low_rank_dim)
+        x = torch.einsum('b t n r, h n r-> b t n h', self.x_proj[1](x), self.x_proj[2].weight.view(hidden_size, 5, -1))
+
+        r, w, k, v, g = x.add_(self.x_bias).unbind(-2)
+        r = self.r_proj(hidden_states, r, delta)
+        w = self.w_proj(hidden_states, w, delta)
+        k = self.k_proj(hidden_states, k, delta)
+        v = self.v_proj(hidden_states, v, delta)
+        g = self.g_proj(hidden_states, g, delta)
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+        r, w, k = map(lambda x: rearrange(x, 'b t (h d) -> b t h d', d=self.head_k_dim), (r, w, k))
+        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+        w = -torch.exp(w)
+        u = self.bonus
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_rwkv6(
+                r=r,
+                k=k,
+                v=v,
+                w=w,
+                u=u,
+                scale=1.,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_rwkv6(
+                q=r,
+                k=k,
+                v=v,
+                g=w,
+                u=u,
+                scale=1.,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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=hidden_states[:, -1],
+                layer_idx=self.layer_idx,
+                offset=r.shape[2]
+            )
+
+        o = self.g_norm(rearrange(o, '... h d -> ... (h d)')) * self.gate_fn(g)
+        o = self.o_proj(o)
+
+        return o, None, past_key_values
+
+
+class LoRA(nn.Module):
+
+    def __init__(
+        self,
+        input_dim: int,
+        output_dim: int,
+        low_rank_dim: int,
+        bias: Optional[bool] = True,
+        activation: Optional[str] = 'tanh'
+    ):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.low_rank_dim = low_rank_dim
+        self.bias = bias
+
+        if activation is None:
+            self.activation = nn.Identity()
+        elif activation == 'sigmoid':
+            self.activation = nn.Sigmoid()
+        elif activation == 'tanh':
+            self.activation = nn.Tanh()
+        elif activation == 'relu':
+            self.activation = nn.ReLU()
+        else:
+            raise ValueError(f"Not supported activation `{activation}`.")
+
+        self.lora = nn.Sequential(
+            nn.Linear(input_dim, low_rank_dim, bias=False),
+            self.activation,
+            nn.Linear(low_rank_dim, output_dim, bias=bias)
+        )
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}("
+        s += f"input_dim={self.input_dim}, low_rank_dim={self.low_rank_dim}, output_dim={self.output_dim}"
+        if not self.bias:
+            s += f", bias={self.bias}"
+        s += ")"
+        return s
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lora(x)
+
+
+class LerpLinear(nn.Module):
+
+    def __init__(
+        self,
+        input_dim: int,
+        output_dim: int,
+        low_rank_dim: Optional[int] = None
+    ):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.low_rank_dim = low_rank_dim
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+        if low_rank_dim is None:
+            self.linear = nn.Linear(input_dim, output_dim, bias=False)
+        else:
+            self.linear = LoRA(input_dim, output_dim, low_rank_dim)
+        self.mu = nn.Parameter(torch.zeros(input_dim))
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.input_dim}, {self.output_dim}"
+        if self.low_rank_dim is not None:
+            s += f", low_rank_dim={self.low_rank_dim}"
+        s += ")"
+        return s
+
+    def forward(self, x: torch.Tensor, delta: Optional[torch.Tensor] = None) -> torch.Tensor:
+        if delta is None:
+            shifted = self.time_shift(x)
+            if len(shifted.shape) == 2:
+                shifted = shifted.unsqueeze(1)
+            delta = shifted - x
+        return self.linear(x + delta * self.mu)
+
+
+class DDLerpLinear(nn.Module):
+
+    def __init__(
+        self,
+        input_dim: int,
+        output_dim: int,
+        low_rank_dim: Optional[int] = None
+    ):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.low_rank_dim = low_rank_dim
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+        if low_rank_dim is None:
+            self.linear = nn.Linear(input_dim, output_dim, bias=False)
+        else:
+            self.linear = LoRA(input_dim, output_dim, low_rank_dim)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.input_dim}, {self.output_dim}"
+        if self.low_rank_dim is not None:
+            s += f", low_rank_dim={self.low_rank_dim}"
+        s += ")"
+        return s
+
+    def forward(self, x: torch.Tensor, mu: torch.Tensor, delta: Optional[torch.Tensor] = None) -> torch.Tensor:
+        if delta is None:
+            shifted = self.time_shift(x)
+            if len(shifted.shape) == 2:
+                shifted = shifted.unsqueeze(1)
+            delta = shifted - x
+        return self.linear(x + delta * mu)

fla/layers/rwkv7.py

@@ -0,0 +1,221 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torch.nn import functional as F
+
+from fla.layers.rwkv6 import LoRA
+from fla.modules import GroupNorm
+from fla.modules.l2norm import l2_norm
+from fla.ops.rwkv7 import chunk_rwkv7, fused_recurrent_rwkv7
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+class RWKV7Attention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        head_dim: Optional[int] = 64,
+        num_heads: Optional[int] = None,
+        decay_low_rank_dim: int = 64,
+        gate_low_rank_dim: int = 128,
+        a_low_rank_dim: int = 64,
+        v_low_rank_dim: int = 16,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None,
+        fuse_norm: bool = False,
+        value_dim: int = None,
+        **kwargs
+    ) -> RWKV7Attention:
+        super().__init__()
+
+        self.mode = mode
+        assert mode in ['chunk', 'fused_recurrent'], f"Not supported mode `{mode}`."
+        self.hidden_size = hidden_size
+
+        self.key_dim = hidden_size
+        self.value_dim = value_dim if value_dim is not None else hidden_size
+        if head_dim is None and num_heads is None:
+            raise ValueError("Either `head_dim` or `num_heads` must be specified.")
+        elif head_dim is not None:
+            self.head_dim = head_dim
+            self.num_heads = int(hidden_size // head_dim)
+        elif num_heads is not None:
+            self.head_dim = int(hidden_size // num_heads)
+            self.num_heads = num_heads
+        self.head_v_dim = int(self.value_dim // self.num_heads)
+
+        self.decay_low_rank_dim = decay_low_rank_dim
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.a_low_rank_dim = a_low_rank_dim
+        self.v_low_rank_dim = v_low_rank_dim
+        self.layer_idx = layer_idx
+        self.fuse_norm = fuse_norm
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        self.x_x = nn.Parameter(torch.zeros(6, hidden_size))
+
+        self.k_k = nn.Parameter(torch.zeros(self.key_dim))
+        self.k_a = nn.Parameter(torch.zeros(self.key_dim))
+        self.r_k = nn.Parameter(torch.zeros(self.num_heads, self.head_dim))
+
+        self.r_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+        self.v_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+
+        self.w_lora = LoRA(hidden_size, self.key_dim, low_rank_dim=decay_low_rank_dim, activation='tanh')
+        if self.layer_idx != 0:
+            self.v_lora = LoRA(hidden_size, self.value_dim, low_rank_dim=v_low_rank_dim, activation=None)
+        self.a_lora = LoRA(hidden_size, self.key_dim, low_rank_dim=a_low_rank_dim, activation=None)
+        self.g_lora = LoRA(hidden_size, self.value_dim, low_rank_dim=gate_low_rank_dim, activation='sigmoid', bias=False)
+
+        if self.fuse_norm:
+            self.g_norm = GroupNorm(
+                num_groups=self.num_heads,
+                hidden_size=self.value_dim,
+                elementwise_affine=elementwise_affine,
+                eps=self.head_dim*norm_eps,
+                bias=True,
+            )
+        else:
+            self.g_norm = nn.GroupNorm(
+                num_groups=self.num_heads,
+                num_channels=self.value_dim,
+                eps=self.head_dim*norm_eps,
+                affine=elementwise_affine
+            )
+
+        self.apply(self._initialize_weights)
+
+    def _initialize_weights(self, module: nn.Module):
+        if getattr(module, "_is_hf_initialized", False):
+            return
+        if isinstance(module, nn.Linear):
+            nn.init.xavier_uniform_(module.weight, gain=2 ** -2.5)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        if isinstance(module, nn.Parameter):
+            nn.init.xavier_uniform_(module, gain=2 ** -2.5)
+        module._is_hf_initialized = True
+
+    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,
+        v_first: torch.Tensor = None,
+        **kwargs
+    ) -> 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, seq_len, _ = hidden_states.shape
+
+        if self.training:
+            # if training, use chunk mode no matter how short the sequence is
+            mode = 'chunk'
+        else:
+            # launching the triton kernel for just one token will actually be slower
+            mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        if attention_mask is not None:
+            hidden_states = hidden_states.mul(attention_mask[:, -hidden_states.shape[-2]:, None])
+        if hidden_states.shape[1] == 1 and last_state is not None:
+            shifted = last_state['conv_state'].unsqueeze(1)
+        else:
+            shifted = self.time_shift(hidden_states)
+            if last_state is not None:
+                shifted[:, 0] = last_state['conv_state']
+
+        # [batch_size, seq_len, hidden_size]
+        delta = shifted - hidden_states
+        xr, xw, xk, xv, xa, xg = hidden_states.addcmul(delta, self.x_x.view(6, 1, 1, -1)).unbind(0)
+
+        r = self.r_proj(xr)
+        # -math.exp(-0.5) = -0.6065306597126334
+        # I think .to(torch.float) is unnecessary here, since we calculate lora in bloat16
+        # when we apply sigmoid, bf16 input will not have numerical issue
+        # FIXME: check if we can remove .to(torch.float)
+        w = -0.6065306597126334 * self.w_lora(xw).to(torch.float).sigmoid()
+
+        k = self.k_proj(xk)
+        v = self.v_proj(xv)
+
+        if self.layer_idx == 0:
+            v_first = v
+        else:
+            v = torch.lerp(v, v_first, self.v_lora(xv).sigmoid())
+        a = self.a_lora(xa).sigmoid()
+        g = self.g_lora(xg)
+
+        if self.fuse_norm:
+            kk = l2_norm(rearrange(k * self.k_k, 'b t (h d) -> b t h d', d=self.head_dim))
+        else:
+            kk = F.normalize(rearrange(k * self.k_k, 'b t (h d) -> b t h d', d=self.head_dim), dim=-1, p=2.0)
+
+        k = k.addcmul(k * (a - 1), self.k_a)
+
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v * attention_mask[:, -v.shape[-2]:, None]
+        r, w, k, a = map(lambda x: rearrange(x, 'b t (h d) -> b t h d', d=self.head_dim), (r, w, k, a))
+        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+
+        rwkv7_fn = chunk_rwkv7 if mode == 'chunk' else fused_recurrent_rwkv7
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        o, recurrent_state = rwkv7_fn(
+            r=r,
+            w=w,
+            k=k,
+            v=v,
+            a=-kk,
+            b=kk * a,
+            scale=1.,
+            initial_state=recurrent_state,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+            head_first=False
+        )
+
+        if past_key_values is not None:
+            past_key_values.update(
+                recurrent_state=recurrent_state,
+                conv_state=hidden_states[:, -1],
+                layer_idx=self.layer_idx,
+                offset=r.shape[1]
+            )
+
+        if self.fuse_norm:
+            o = self.g_norm(rearrange(o, '... h d -> ... (h d)'))
+        else:
+            o = self.g_norm(rearrange(o, 'b t h d -> (b t) (h d)')).view(batch_size, seq_len, -1)
+
+        o = o + ((r * k * self.r_k).sum(-1, keepdim=True) * v).view(batch_size, seq_len, -1)
+        o = self.o_proj(o * g)
+
+        return o, None, past_key_values, v_first

fla/models/__init__.py

@@ -0,0 +1,53 @@
+# -*- coding: utf-8 -*-
+
+from fla.models.abc import ABCConfig, ABCForCausalLM, ABCModel
+from fla.models.bitnet import BitNetConfig, BitNetForCausalLM, BitNetModel
+from fla.models.delta_net import DeltaNetConfig, DeltaNetForCausalLM, DeltaNetModel
+from fla.models.forgetting_transformer import (
+    ForgettingTransformerConfig,
+    ForgettingTransformerForCausalLM,
+    ForgettingTransformerModel
+)
+from fla.models.gated_deltanet import GatedDeltaNetConfig, GatedDeltaNetForCausalLM, GatedDeltaNetModel
+from fla.models.gated_deltaproduct import GatedDeltaProductConfig, GatedDeltaProductForCausalLM, GatedDeltaProductModel
+from fla.models.gla import GLAConfig, GLAForCausalLM, GLAModel
+from fla.models.gsa import GSAConfig, GSAForCausalLM, GSAModel
+from fla.models.hgrn import HGRNConfig, HGRNForCausalLM, HGRNModel
+from fla.models.hgrn2 import HGRN2Config, HGRN2ForCausalLM, HGRN2Model
+from fla.models.lightnet import LightNetConfig, LightNetForCausalLM, LightNetModel
+from fla.models.linear_attn import LinearAttentionConfig, LinearAttentionForCausalLM, LinearAttentionModel
+from fla.models.mamba import MambaConfig, MambaForCausalLM, MambaModel
+from fla.models.mamba2 import Mamba2Config, Mamba2ForCausalLM, Mamba2Model
+from fla.models.nsa import NSAConfig, NSAForCausalLM, NSAModel
+from fla.models.retnet import RetNetConfig, RetNetForCausalLM, RetNetModel
+from fla.models.rwkv6 import RWKV6Config, RWKV6ForCausalLM, RWKV6Model
+from fla.models.rwkv7 import RWKV7Config, RWKV7ForCausalLM, RWKV7Model
+from fla.models.samba import SambaConfig, SambaForCausalLM, SambaModel
+from fla.models.transformer import TransformerConfig, TransformerForCausalLM, TransformerModel
+from fla.models.transformer_with_pruning import TransformerWithPruningConfig, TransformerWithPruningForCausalLM, TransformerWithPruningModel
+from fla.models.stochastic_softpick_transformer import StochasticSoftpickTransformerConfig, StochasticSoftpickTransformerForCausalLM, StochasticSoftpickTransformerModel
+
+__all__ = [
+    'ABCConfig', 'ABCForCausalLM', 'ABCModel',
+    'BitNetConfig', 'BitNetForCausalLM', 'BitNetModel',
+    'DeltaNetConfig', 'DeltaNetForCausalLM', 'DeltaNetModel',
+    'ForgettingTransformerConfig', 'ForgettingTransformerForCausalLM', 'ForgettingTransformerModel',
+    'GatedDeltaNetConfig', 'GatedDeltaNetForCausalLM', 'GatedDeltaNetModel',
+    'GLAConfig', 'GLAForCausalLM', 'GLAModel',
+    'GSAConfig', 'GSAForCausalLM', 'GSAModel',
+    'HGRNConfig', 'HGRNForCausalLM', 'HGRNModel',
+    'HGRN2Config', 'HGRN2ForCausalLM', 'HGRN2Model',
+    'LightNetConfig', 'LightNetForCausalLM', 'LightNetModel',
+    'LinearAttentionConfig', 'LinearAttentionForCausalLM', 'LinearAttentionModel',
+    'MambaConfig', 'MambaForCausalLM', 'MambaModel',
+    'Mamba2Config', 'Mamba2ForCausalLM', 'Mamba2Model',
+    'NSAConfig', 'NSAForCausalLM', 'NSAModel',
+    'RetNetConfig', 'RetNetForCausalLM', 'RetNetModel',
+    'RWKV6Config', 'RWKV6ForCausalLM', 'RWKV6Model',
+    'RWKV7Config', 'RWKV7ForCausalLM', 'RWKV7Model',
+    'SambaConfig', 'SambaForCausalLM', 'SambaModel',
+    'TransformerConfig', 'TransformerForCausalLM', 'TransformerModel',
+    'TransformerWithPruningConfig', 'TransformerWithPruningForCausalLM', 'TransformerWithPruningModel',
+    'GatedDeltaProductConfig', 'GatedDeltaProductForCausalLM', 'GatedDeltaProductModel',
+    'StochasticSoftpickTransformerConfig', 'StochasticSoftpickTransformerForCausalLM', 'StochasticSoftpickTransformerModel'
+]

fla/models/abc/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.abc.configuration_abc import ABCConfig
+from fla.models.abc.modeling_abc import ABCForCausalLM, ABCModel
+
+AutoConfig.register(ABCConfig.model_type, ABCConfig)
+AutoModel.register(ABCConfig, ABCModel)
+AutoModelForCausalLM.register(ABCConfig, ABCForCausalLM)
+
+
+__all__ = ['ABCConfig', 'ABCForCausalLM', 'ABCModel']

fla/models/abc/configuration_abc.py

@@ -0,0 +1,91 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class ABCConfig(PretrainedConfig):
+
+    model_type = 'abc'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        gate_low_rank_dim: int = 16,
+        clamp_min: float = -32,
+        clamp_max: float = 32,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        num_slots: Optional[int] = 64,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        exapnd_k: float = 0.5,
+        exapnd_v: float = 1,
+        hidden_act: str = "swish",
+        max_position_embeddings: int = 2048,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_rope: bool = True,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.hidden_size = hidden_size
+        self.gate_low_rank_dim = gate_low_rank_dim
+        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_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.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_rope = use_rope
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        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.)
+
+        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,
+        )

fla/models/abc/modeling_abc.py

@@ -0,0 +1,418 @@
+# -*- 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.abc import ABCAttention
+from fla.layers.attn import Attention
+from fla.models.abc.configuration_abc import ABCConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as ABCMLP
+from fla.modules import RMSNorm
+
+logger = logging.get_logger(__name__)
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+
+class ABCBlock(nn.Module):
+    def __init__(self, config: ABCConfig, 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'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = ABCAttention(
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                num_slots=config.num_slots,
+                use_short_conv=config.use_short_conv,
+                conv_size=config.conv_size,
+                gate_fn=config.hidden_act,
+                elementwise_affine=config.elementwise_affine,
+                norm_eps=config.norm_eps,
+                use_rope=config.use_rope,
+                clamp_min=config.clamp_min,
+                clamp_max=config.clamp_max,
+                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 = ABCMLP(
+            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)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class ABCPreTrainedModel(PreTrainedModel):
+
+    config_class = ABCConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['ABCBlock']
+    _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] = 'rescale',
+        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 ABCModel(ABCPreTrainedModel):
+
+    def __init__(self, config: ABCConfig):
+        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([ABCBlock(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("`ABCModel` 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,
+                    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 ABCForCausalLM(ABCPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = ABCModel(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
+
+        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()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            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))
+
+        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,
+        )

fla/models/bitnet/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.bitnet.configuration_bitnet import BitNetConfig
+from fla.models.bitnet.modeling_bitnet import BitNetForCausalLM, BitNetModel
+
+AutoConfig.register(BitNetConfig.model_type, BitNetConfig)
+AutoModel.register(BitNetConfig, BitNetModel)
+AutoModelForCausalLM.register(BitNetConfig, BitNetForCausalLM)
+
+
+__all__ = ['BitNetConfig', 'BitNetForCausalLM', 'BitNetModel']

fla/models/bitnet/configuration_bitnet.py

@@ -0,0 +1,67 @@
+# -*- coding: utf-8 -*-
+
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class BitNetConfig(PretrainedConfig):
+
+    model_type = 'bitnet'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 24,
+        num_heads: int = 32,
+        num_kv_heads: int = None,
+        window_size: Optional[int] = None,
+        rope_theta: Optional[float] = 10000.,
+        max_position_embeddings: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        initializer_range: float = 0.006,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.window_size = window_size
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+
+        self.initializer_range = initializer_range
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_cache = use_cache
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        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,
+        )

fla/models/bitnet/modeling_bitnet.py

@@ -0,0 +1,441 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Any, 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.bitattn import BitAttention
+from fla.models.bitnet.configuration_bitnet import BitNetConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, RMSNorm
+from fla.modules.activations import swiglu
+from fla.modules.fused_bitlinear import FusedBitLinear
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class BitNetMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        hidden_ratio: Optional[int] = None,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = 'swish',
+        fuse_swiglu: bool = True
+    ) -> BitNetMLP:
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        # the final number of params is `hidden_ratio * hidden_size^2`
+        # `intermediate_size` is chosen to be a multiple of 256 closest to `2/3 * hidden_size * hidden_ratio`
+        if hidden_ratio is None:
+            hidden_ratio = 4
+        if intermediate_size is None:
+            intermediate_size = int(hidden_size * hidden_ratio * 2 / 3)
+            intermediate_size = 256 * ((intermediate_size + 256 - 1) // 256)
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.fuse_swiglu = fuse_swiglu
+
+        if hidden_act != 'swish':
+            raise ValueError(f'Unsupported hidden_act: {hidden_act}')
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        **kwargs: Unpack[Any]
+    ) -> torch.Tensor:
+        gate, y = self.gate_proj(x), self.up_proj(x)
+        return self.down_proj(swiglu(gate, y))
+
+
+class BitNetBlock(nn.Module):
+
+    def __init__(self, config: BitNetConfig, 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)
+        self.attn = BitAttention(
+            hidden_size=config.hidden_size,
+            num_heads=config.num_heads,
+            num_kv_heads=config.num_kv_heads,
+            window_size=config.window_size,
+            rope_theta=config.rope_theta,
+            max_position_embeddings=config.max_position_embeddings,
+            layer_idx=layer_idx
+        )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = BitNetMLP(
+            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[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs: Unpack[Any]
+    ) -> 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,)
+
+        if output_attentions:
+            outputs += (attentions,)
+
+        if use_cache:
+            outputs += (past_key_values,)
+
+        return outputs
+
+
+class BitNetPreTrainedModel(PreTrainedModel):
+
+    config_class = BitNetConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['BitNetBlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        rescale_prenorm_residual: bool = False,
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d, FusedBitLinear)):
+            # 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 rescale_prenorm_residual:
+            # 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
+                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)
+
+
+class BitNetModel(BitNetPreTrainedModel):
+
+    def __init__(
+        self,
+        config: BitNetConfig
+    ) -> BitNetModel:
+        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([BitNetBlock(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,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Any]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        if output_attentions:
+            warnings.warn(
+                "`BitNetModel` does not support output attention weights now, so `output_attentions` is set 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")
+        elif input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        next_cache = None
+
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    **kwargs
+                )
+            else:
+                layer_outputs = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    **kwargs
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_attns] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class BitNetForCausalLM(BitNetPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = BitNetModel(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
+
+    @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
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Any]
+    ) -> 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,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        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()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            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))
+
+        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,
+        )