Add files using upload-large-folder tool

configs/delta_net_1B.json

@@ -0,0 +1,29 @@
+{
+    "attn": null,
+    "attn_mode": "chunk",
+    "bos_token_id": 1,
+    "conv_size": 4,
+    "eos_token_id": 2,
+    "expand_k": 1,
+    "expand_v": 1,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 2048,
+    "initializer_range": 0.006,
+    "intermediate_size": null,
+    "model_type": "delta_net",
+    "norm_eps": 1e-06,
+    "num_heads": 16,
+    "num_hidden_layers": 24,
+    "pad_token_id": 2,
+    "qk_activation": "silu",
+    "qk_norm": "l2",
+    "tie_word_embeddings": false,
+    "use_beta": true,
+    "use_cache": true,
+    "use_gate": false,
+    "use_output_norm": true,
+    "use_short_conv": true
+}

configs/gla_340M.json

@@ -0,0 +1,24 @@
+{
+  "attn_mode": "chunk",
+  "bos_token_id": 1,
+  "clamp_min": null,
+  "eos_token_id": 2,
+  "expand_k": 0.5,
+  "expand_v": 1,
+  "fuse_cross_entropy": true,
+  "fuse_norm": true,
+  "hidden_act": "swish",
+  "hidden_ratio": 4,
+  "hidden_size": 1024,
+  "initializer_range": 0.006,
+  "intermediate_size": null,
+  "model_type": "gla",
+  "num_heads": 4,
+  "num_hidden_layers": 24,
+  "norm_eps": 1e-06,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "use_gk": true,
+  "use_gv": false,
+  "vocab_size": 32000
+}

configs/gla_7B.json

@@ -0,0 +1,25 @@
+{
+    "attn": null,
+    "attn_mode": "chunk",
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "expand_k": 0.5,
+    "expand_v": 1,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 4096,
+    "initializer_range": 0.006,
+    "intermediate_size": 11008,
+    "model_type": "gla",
+    "norm_eps": 1e-06,
+    "num_heads": 16,
+    "num_hidden_layers": 32,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "use_gk": true,
+    "use_gv": false,
+    "use_output_gate": true,
+    "use_short_conv": false
+}

configs/gsa_340M.json

@@ -0,0 +1,29 @@
+{
+    "bos_token_id": 1,
+    "conv_size": 4,
+    "eos_token_id": 2,
+    "expand_k": 1,
+    "expand_v": 1,
+    "elementwise_affine": false,
+    "feature_map": "swish",
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "gate_logit_normalizer": 4,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 1024,
+    "initializer_range": 0.006,
+    "intermediate_size": null,
+    "model_type": "gsa",
+    "num_heads": 4,
+    "num_hidden_layers": 24,
+    "num_slots": 64,
+    "norm_eps": 1e-06,
+    "share_conv_kernel": true,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "use_norm": true,
+    "use_output_gate": true,
+    "use_rope": false,
+    "use_short_conv": false
+}

configs/hgrn2_340M.json

@@ -0,0 +1,20 @@
+{
+    "attn_mode": "chunk",
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "expand_ratio": 128,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 1024,
+    "initializer_range": 0.006,
+    "intermediate_size": null,
+    "model_type": "hgrn2",
+    "num_heads": 8,
+    "num_hidden_layers": 24,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 32000
+}

configs/mtp_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": "mtp_transformer",
+    "num_heads": 12,
+    "num_hidden_layers": 14,
+    "norm_eps": 1e-06,
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 32000,
+    "n_future_tokens": 4
+}

configs/myopic_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,
+    "use_myopic_loss": true
+}

configs/myopic_transformer_1B.json

@@ -0,0 +1,23 @@
+{
+    "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": 32,
+    "num_kv_heads": null,
+    "pad_token_id": 2,
+    "rope_theta": 10000.0,
+    "tie_word_embeddings": false,
+    "use_myopic_loss": true
+}

configs/myopic_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,
+    "use_myopic_loss": true
+}

configs/transformer_120M.json

@@ -0,0 +1,18 @@
+{
+    "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
+}

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": 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
+}

configs/transformer_7B.json

@@ -0,0 +1,21 @@
+{
+    "attention_bias": false,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "fuse_cross_entropy": true,
+    "fuse_norm": true,
+    "hidden_act": "swish",
+    "hidden_ratio": 4,
+    "hidden_size": 4096,
+    "initializer_range": 0.006,
+    "intermediate_size": 14336,
+    "model_type": "transformer",
+    "norm_eps": 1e-06,
+    "num_heads": 32,
+    "num_hidden_layers": 32,
+    "num_kv_heads": 8,
+    "rope_theta": 10000.0,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "window_size": null
+}

fla/layers/attn.py

@@ -0,0 +1,203 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+import warnings
+from typing import TYPE_CHECKING, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from einops import rearrange
+from transformers.utils import logging
+
+from fla.modules import RMSNorm, RotaryEmbedding
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+except ImportError:
+    warnings.warn(
+        "Flash Attention is not installed. Please install it via `pip install flash-attn --no-build-isolation`",
+        category=ImportWarning
+    )
+    flash_attn_func = None
+
+logger = logging.get_logger(__name__)
+
+
+class Attention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_heads: int = 32,
+        num_kv_heads: Optional[int] = None,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        window_size: Optional[int] = None,
+        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 = self.hidden_size // self.num_heads
+        self.kv_dim = self.num_kv_heads * self.head_dim
+        self.qkv_bias = qkv_bias
+        self.qk_norm = qk_norm
+
+        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.hidden_size, 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.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+        if qk_norm:
+            self.q_norm = RMSNorm(self.head_dim)
+            self.k_norm = RMSNorm(self.head_dim)
+
+        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, q_len, _ = hidden_states.size()
+
+        q, k, v = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(hidden_states)
+
+        q = rearrange(q, '... (h d) -> ... h d', d=self.head_dim)
+        k = rearrange(k, '... (h d) -> ... h d', d=self.head_dim)
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_dim)
+
+        if self.qk_norm:
+            q, k = self.q_norm(q), self.k_norm(k)
+
+        # equivalent to cu_seqlens in `flash_attn`
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+
+        seqlen_offset, max_seqlen = 0, q_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=q_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)
+
+        if flash_attn_func is None:
+            raise ImportError("Please install Flash Attention via `pip install flash-attn --no-build-isolation` first")
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            q, k, v, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(q, k, v, attention_mask, q_len)
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_q, max_seqlen_k = max_seq_lens
+            o = flash_attn_varlen_func(
+                q, k, v,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_q,
+                max_seqlen_k=max_seqlen_k,
+                causal=True,
+                window_size=(-1, -1) if self.window_size is None else (self.window_size-1, 0)
+            )
+            o = pad_input(o, indices_q, batch_size, q_len)
+        elif cu_seqlens is not None:
+            o = flash_attn_varlen_func(
+                q.squeeze(0), k.squeeze(0), v.squeeze(0),
+                cu_seqlens_q=cu_seqlens,
+                cu_seqlens_k=cu_seqlens,
+                max_seqlen_q=max_seqlen,
+                max_seqlen_k=max_seqlen,
+                causal=True,
+                window_size=(-1, -1) if self.window_size is None else (self.window_size-1, 0)
+            ).unsqueeze(0)
+        else:
+            o = flash_attn_func(
+                q, k, v,
+                causal=True,
+                window_size=(-1, -1) if self.window_size is None else (self.window_size-1, 0)
+            )
+        o = o.reshape(batch_size, q_len, -1)
+        o = self.o_proj(o)
+
+        if not output_attentions:
+            attentions = None
+
+        return o, attentions, past_key_values
+
+    def _upad_input(self, q, k, v, attention_mask, q_len):
+        batch_size, seq_len, num_key_value_heads, head_dim = k.shape
+        cache_mask = attention_mask[:, -seq_len:]
+        seqlens = cache_mask.sum(-1, dtype=torch.int32)
+        indices_k = torch.nonzero(cache_mask.flatten(), as_tuple=False).flatten()
+        max_seqlen_k = seqlens.max().item()
+        cu_seqlens_k = F.pad(torch.cumsum(seqlens, dim=0, dtype=torch.int32), (1, 0))
+
+        k = index_first_axis(k.reshape(batch_size * seq_len, num_key_value_heads, head_dim), indices_k)
+        v = index_first_axis(v.reshape(batch_size * seq_len, num_key_value_heads, head_dim), indices_k)
+        if q_len == seq_len:
+            q = index_first_axis(q.reshape(batch_size * seq_len, self.num_heads, head_dim), indices_k)
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_q = max_seqlen_k
+            indices_q = indices_k
+        elif q_len == 1:
+            max_seqlen_q = 1
+            # There is a memcpy here, that is very bad.
+            cu_seqlens_q = torch.arange(batch_size + 1, dtype=torch.int32, device=q.device)
+            indices_q = cu_seqlens_q[:-1]
+            q = q.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -q_len:]
+            q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, attention_mask)
+
+        return q, k, v, indices_q, (cu_seqlens_q, cu_seqlens_k), (max_seqlen_q, max_seqlen_k)

fla/layers/based.py

@@ -0,0 +1,96 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+"""
+Linear attention in Based.
+https://github.com/HazyResearch/zoology/blob/main/zoology/mixers/based.py
+"""
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from fla.modules.feature_map import TaylorFeatureMap
+from fla.ops.based import parallel_based
+from fla.ops.linear_attn import chunk_linear_attn, fused_chunk_linear_attn
+
+
+class BasedLinearAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        feature_dim: int = 16,
+        num_key_value_heads: int = 12,
+        num_heads: int = 12,
+        feature_name: str = "taylor_exp",
+        eps: float = 1e-12,
+        causal: bool = True,
+        mode: str = "parallel",
+    ):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.mode = mode
+        self.feature_name = feature_name
+        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
+        assert self.hidden_size % self.head_dim == 0
+        self.causal = causal
+
+        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()
+        self.feature_map = TaylorFeatureMap(feature_dim)
+        self.eps = eps
+
+    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])
+        if mode == "fused_chunk":
+            q, k = self.feature_map(q), self.feature_map(k)
+            o, _ = fused_chunk_linear_attn(q, k, v, normalize=True, scale=1, head_first=False)
+        elif mode == 'chunk':
+            q, k = self.feature_map(q), self.feature_map(k)
+            o, _ = chunk_linear_attn(q, k, v, normalize=True, scale=1, head_first=False)
+        elif mode == 'parallel':
+            assert q.shape[-1] <= 128
+            o = parallel_based(q, k, v, scale=1, use_norm=True, head_first=False)
+        o = rearrange(o, 'b t h d -> b t (h d)')
+        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)
+        """
+        # hidden_states = hidden_states.transpose(1, 2)
+        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, self.num_heads, self.feature_dim).transpose(1, 2)
+        k = k.view(b, t, self.num_key_value_heads, self.feature_dim).transpose(1, 2)
+        v = v.view(b, t, self.num_key_value_heads, 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/forgetting_attn.py

@@ -0,0 +1,109 @@
+# -*- 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
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from einops import rearrange
+from transformers.utils import logging
+
+from fla.modules import GroupNorm
+from fla.ops.forgetting_attn.parallel import parallel_forgetting_attn
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+logger = logging.get_logger(__name__)
+
+
+class ForgettingAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_heads: int = 32,
+        num_kv_heads: Optional[int] = None,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        window_size: Optional[int] = None,
+        use_output_gate: bool = False,
+        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 = self.hidden_size // self.num_heads
+        self.kv_dim = self.num_kv_heads * self.head_dim
+        self.qkv_bias = qkv_bias
+        self.qk_norm = qk_norm
+
+        self.window_size = window_size
+        self.use_output_gate = use_output_gate
+        self.layer_idx = layer_idx
+
+        self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, 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.f_proj = nn.Linear(self.hidden_size, self.num_heads, bias=True)
+
+        if use_output_gate:
+            self.g_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+        if qk_norm:
+            self.q_norm = GroupNorm(
+                num_groups=self.num_heads,
+                hidden_size=self.hidden_size,
+                is_rms_norm=True,
+            )
+            self.k_norm = GroupNorm(
+                num_groups=self.num_kv_heads,
+                hidden_size=self.kv_dim,
+                is_rms_norm=True,
+            )
+
+    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."
+            )
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        q, k, v = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(hidden_states)
+        f = F.logsigmoid(self.f_proj(hidden_states).float())
+        if self.qk_norm:
+            q, k = self.q_norm(q), self.k_norm(k)
+
+        q = rearrange(q, '... (h d) -> ... h d', d=self.head_dim)
+        k = rearrange(k, '... (h d) -> ... h d', d=self.head_dim)
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_dim)
+
+        o = parallel_forgetting_attn(q, k, v, f, cu_seqlens=cu_seqlens)
+        o = rearrange(o, '... h d -> ... (h d)')
+        if self.use_output_gate:
+            o = self.g_proj(hidden_states).sigmoid() * o
+        o = self.o_proj(o)
+
+        return o, None, past_key_values

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/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/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/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,
+        )

fla/models/delta_net/modeling_delta_net.py

@@ -0,0 +1,415 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.layers.delta_net import DeltaNet
+from fla.models.delta_net.configuration_delta_net import DeltaNetConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as DeltaNetMLP
+from fla.modules import RMSNorm
+
+logger = logging.get_logger(__name__)
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+
+class DeltaNetBlock(nn.Module):
+    def __init__(self, config: DeltaNetConfig, 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 = DeltaNet(
+                mode=config.attn_mode,
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                use_gate=config.use_gate,
+                use_beta=config.use_beta,
+                use_short_conv=config.use_short_conv,
+                use_output_norm=config.use_output_norm,
+                conv_size=config.conv_size,
+                qk_norm=config.qk_norm,
+                qk_activation=config.qk_activation,
+                norm_eps=config.norm_eps,
+                layer_idx=layer_idx
+            )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = DeltaNetMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class DeltaNetPreTrainedModel(PreTrainedModel):
+
+    config_class = DeltaNetConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['DeltaNetBlock']
+    _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 DeltaNetModel(DeltaNetPreTrainedModel):
+
+    def __init__(self, config: DeltaNetConfig):
+        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([DeltaNetBlock(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("`DeltaNetModel` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class DeltaNetForCausalLM(DeltaNetPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = DeltaNetModel(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/forgetting_transformer/__init__.py

@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.forgetting_transformer.configuration_forgetting_transformer import ForgettingTransformerConfig
+from fla.models.forgetting_transformer.modeling_forgetting_transformer import (
+    ForgettingTransformerForCausalLM,
+    ForgettingTransformerModel
+)
+
+AutoConfig.register(ForgettingTransformerConfig.model_type, ForgettingTransformerConfig)
+AutoModel.register(ForgettingTransformerConfig, ForgettingTransformerModel)
+AutoModelForCausalLM.register(ForgettingTransformerConfig, ForgettingTransformerForCausalLM)
+
+
+__all__ = ['ForgettingTransformerConfig', 'ForgettingTransformerForCausalLM', 'ForgettingTransformerModel']

fla/models/forgetting_transformer/modeling_forgetting_transformer.py

@@ -0,0 +1,408 @@
+# -*- 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.forgetting_attn import ForgettingAttention
+from fla.models.forgetting_transformer.configuration_forgetting_transformer import ForgettingTransformerConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as ForgettingTransformerMLP
+from fla.modules import RMSNorm
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+
+logger = logging.get_logger(__name__)
+
+
+class ForgettingTransformerBlock(nn.Module):
+
+    def __init__(self, config: ForgettingTransformerConfig, 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 = ForgettingAttention(
+            hidden_size=config.hidden_size,
+            num_heads=config.num_heads,
+            num_kv_heads=config.num_kv_heads,
+            qkv_bias=config.qkv_bias,
+            qk_norm=config.qk_norm,
+            window_size=config.window_size,
+            use_output_gate=config.use_output_gate,
+            layer_idx=layer_idx
+        )
+
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = ForgettingTransformerMLP(
+            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 ForgettingTransformerPreTrainedModel(PreTrainedModel):
+
+    config_class = ForgettingTransformerConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['ForgettingTransformerBlock']
+    _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)):
+            # 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 ForgettingTransformer 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 ForgettingTransformerModel(ForgettingTransformerPreTrainedModel):
+
+    def __init__(
+        self,
+        config: ForgettingTransformerConfig
+    ) -> ForgettingTransformerModel:
+        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([
+            ForgettingTransformerBlock(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(
+                "`ForgettingTransformerModel` 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 ForgettingTransformerForCausalLM(ForgettingTransformerPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = ForgettingTransformerModel(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
+
+    @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,
+        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
+        logits = None if fuse_linear_and_cross_entropy else self.lm_head(hidden_states[:, -logits_to_keep:])
+
+        loss = None
+        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
+            # Enable model parallelism
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        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/gated_deltanet/__init__.py

@@ -0,0 +1,12 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.gated_deltanet.configuration_gated_deltanet import GatedDeltaNetConfig
+from fla.models.gated_deltanet.modeling_gated_deltanet import GatedDeltaNetForCausalLM, GatedDeltaNetModel
+
+AutoConfig.register(GatedDeltaNetConfig.model_type, GatedDeltaNetConfig)
+AutoModel.register(GatedDeltaNetConfig, GatedDeltaNetModel)
+AutoModelForCausalLM.register(GatedDeltaNetConfig, GatedDeltaNetForCausalLM)
+
+__all__ = ['GatedDeltaNetConfig', 'GatedDeltaNetForCausalLM', 'GatedDeltaNetModel']

fla/models/gated_deltaproduct/__init__.py

@@ -0,0 +1,14 @@
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.gated_deltaproduct.configuration_gated_deltaproduct import GatedDeltaProductConfig
+from fla.models.gated_deltaproduct.modeling_gated_deltaproduct import GatedDeltaProductForCausalLM, GatedDeltaProductModel
+
+AutoConfig.register(GatedDeltaProductConfig.model_type, GatedDeltaProductConfig)
+AutoModel.register(GatedDeltaProductConfig, GatedDeltaProductModel)
+AutoModelForCausalLM.register(GatedDeltaProductConfig, GatedDeltaProductForCausalLM)
+
+__all__ = [
+    "GatedDeltaProductConfig",
+    "GatedDeltaProductForCausalLM",
+    "GatedDeltaProductModel",
+]

fla/models/hgrn/configuration_hgrn.py

@@ -0,0 +1,81 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class HGRNConfig(PretrainedConfig):
+
+    model_type = 'hgrn'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        attn_mode: str = "fused_recurrent",
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 24,
+        expand_ratio: Optional[int] = 1,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        use_lower_bound: bool = True,
+        max_position_embeddings: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        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.attn_mode = attn_mode
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.expand_ratio = expand_ratio
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.use_lower_bound = use_lower_bound
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.elementwise_affine = elementwise_affine
+        self.attn = attn
+        self.norm_eps = norm_eps
+        self.hidden_act = hidden_act
+        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/hgrn2/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.hgrn2.configuration_hgrn2 import HGRN2Config
+from fla.models.hgrn2.modeling_hgrn2 import HGRN2ForCausalLM, HGRN2Model
+
+AutoConfig.register(HGRN2Config.model_type, HGRN2Config)
+AutoModel.register(HGRN2Config, HGRN2Model)
+AutoModelForCausalLM.register(HGRN2Config, HGRN2ForCausalLM)
+
+
+__all__ = ['HGRN2Config', 'HGRN2ForCausalLM', 'HGRN2Model']