Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tb/20250712-1101/wandb/run-20250712_110147-top_transformer-top.code.340M.batch16.seqlen4096.context4096.warmup1000.update1.steps100000.lr3e-4.cosine-202507121056/run-top_transformer-top.code.340M.batch16.seqlen4096.context4096.warmup1000.update1.steps100000.lr3e-4.cosine-202507121056.wandb filter=lfs diff=lfs merge=lfs -text

fla/layers/__init__.py

@@ -0,0 +1,44 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from .abc import ABCAttention
+from .attn import Attention
+from .based import BasedLinearAttention
+from .bitattn import BitAttention
+from .delta_net import DeltaNet
+from .forgetting_attn import ForgettingAttention
+from .gated_deltanet import GatedDeltaNet
+from .gated_deltaproduct import GatedDeltaProduct
+from .gla import GatedLinearAttention
+from .gsa import GatedSlotAttention
+from .hgrn import HGRNAttention
+from .hgrn2 import HGRN2Attention
+from .lightnet import LightNetAttention
+from .linear_attn import LinearAttention
+from .multiscale_retention import MultiScaleRetention
+from .nsa import NativeSparseAttention
+from .rebased import ReBasedLinearAttention
+from .rwkv6 import RWKV6Attention
+from .rwkv7 import RWKV7Attention
+
+__all__ = [
+    'ABCAttention',
+    'Attention',
+    'BasedLinearAttention',
+    'BitAttention',
+    'DeltaNet',
+    'ForgettingAttention',
+    'GatedDeltaNet',
+    'GatedDeltaProduct',
+    'GatedLinearAttention',
+    'GatedSlotAttention',
+    'HGRNAttention',
+    'HGRN2Attention',
+    'LightNetAttention',
+    'LinearAttention',
+    'MultiScaleRetention',
+    'NativeSparseAttention',
+    'ReBasedLinearAttention',
+    'RWKV6Attention',
+    'RWKV7Attention',
+]

fla/layers/abc.py

@@ -0,0 +1,218 @@
+# -*- 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
+from einops import rearrange
+
+from fla.modules import FusedRMSNormGated, RMSNorm, RotaryEmbedding, ShortConvolution
+from fla.modules.activations import swiglu, swish
+from fla.ops.abc.chunk import chunk_abc
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+class ABCAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int = 1024,
+        expand_k: float = 0.5,
+        expand_v: float = 1.0,
+        num_heads: int = 4,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        num_slots: Optional[int] = None,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        gate_low_rank_dim: int = 16,
+        gate_logit_normalizer: int = 16,
+        use_rope: bool = True,
+        use_input_gate: bool = False,
+        use_output_gate: bool = True,
+        use_norm: bool = True,
+        clamp_min: Optional[float] = -32,
+        clamp_max: Optional[float] = 32,
+        layer_idx: Optional[int] = None,
+        **kwargs
+    ) -> ABCAttention:
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.num_heads = num_heads
+        self.key_dim = int(self.hidden_size * self.expand_k)
+        self.value_dim = int(self.hidden_size * self.expand_v)
+        self.head_k_dim = self.key_dim // self.num_heads
+        self.head_v_dim = self.value_dim // self.num_heads
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.gate_logit_normalizer = gate_logit_normalizer
+
+        self.use_rope = use_rope
+        self.use_input_gate = use_input_gate
+        self.use_output_gate = use_output_gate
+        self.use_norm = use_norm
+
+        if num_slots is None:
+            num_slots = self.head_k_dim
+        self.num_slots = num_slots
+
+        self.norm_eps = norm_eps
+
+        self.clamp_min = clamp_min
+        self.clamp_max = clamp_max
+        self.layer_idx = layer_idx
+
+        if layer_idx is None:
+            warnings.warn(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.q_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.value_dim, bias=False)
+
+        if use_output_gate:
+            self.g_proj = nn.Linear(self.hidden_size, self.value_dim, bias=False)
+        self.s_proj = nn.Linear(self.hidden_size, self.num_heads * self.num_slots, bias=False)
+        self.o_proj = nn.Linear(self.value_dim, self.hidden_size, 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, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim, conv_size, activation='silu')
+
+        if self.use_norm:
+            if self.use_output_gate:
+                self.g_norm = FusedRMSNormGated(
+                    hidden_size=self.head_v_dim,
+                    elementwise_affine=elementwise_affine,
+                    eps=norm_eps
+                )
+            else:
+                self.g_norm = RMSNorm(
+                    hidden_size=self.head_v_dim,
+                    elementwise_affine=elementwise_affine,
+                    eps=norm_eps
+                )
+
+        if self.use_rope:
+            self.rotary = RotaryEmbedding(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."
+            )
+
+        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 cu_seqlens is not None:
+            raise NotImplementedError("Training with cu_seqlens is not supported yet for ABCAttention")
+        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)
+
+        if self.use_input_gate:
+            q, k, v = map(lambda x: swish(x), (q, k, v))
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+
+        q, k = map(lambda x: rearrange(x, '... (h d) -> ... h d', d=self.head_k_dim), (q, k))
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_v_dim)
+        if self.use_rope:
+            seqlen_offset = 0
+            if past_key_values is not None:
+                seqlen_offset = past_key_values.get_seq_length(self.layer_idx)
+            q, k = self.rotary(q, k, seqlen_offset=seqlen_offset)
+
+        s = rearrange(self.s_proj(hidden_states), '... (h m) -> ... h m', m=self.num_slots)
+        s = s.clamp_(self.clamp_min, self.clamp_max)
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        o, recurrent_state = chunk_abc(
+            q=q,
+            k=k,
+            v=v,
+            s=s,
+            initial_state=recurrent_state,
+            output_final_state=use_cache,
+            head_first=False
+        )
+        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_norm and not self.use_output_gate:
+            o = self.g_norm(o)
+        elif self.use_output_gate:
+            g = rearrange(self.g_proj(hidden_states), '... (h d) -> ... h d', d=self.head_v_dim)
+            o = self.g_norm(o, g) if self.use_norm else swiglu(g, o)
+        o = rearrange(o, '... h d -> ... (h d)')
+        o = self.o_proj(o)
+
+        return o, None, past_key_values
+
+    def state_size(self, seq_len: int = 2048):
+        return 2 * self.num_slots * self.hidden_size

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/bitattn.py

@@ -0,0 +1,192 @@
+# -*- 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 RotaryEmbedding
+from fla.modules.fused_bitlinear import FusedBitLinear
+
+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 BitAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_heads: int = 32,
+        num_kv_heads: Optional[int] = None,
+        window_size: Optional[int] = None,
+        rope_theta: Optional[float] = 10000.,
+        max_position_embeddings: Optional[int] = None,
+        norm_eps: float = 1e-5,
+        layer_idx: int = None
+    ):
+        super().__init__()
+
+        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.hidden_size = hidden_size
+        self.head_dim = self.hidden_size // self.num_heads
+        self.kv_dim = self.num_kv_heads * self.head_dim
+        self.kv_dim = self.num_kv_heads * self.head_dim
+        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 = FusedBitLinear(self.hidden_size, self.hidden_size, bias=False)
+        self.k_proj = FusedBitLinear(self.hidden_size, self.kv_dim, bias=False)
+        self.v_proj = FusedBitLinear(self.hidden_size, self.kv_dim, bias=False)
+        self.o_proj = FusedBitLinear(self.hidden_size, self.hidden_size, bias=False)
+
+        self.rotary = RotaryEmbedding(dim=self.head_dim, base=self.rope_theta)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if attention_mask is not None:
+            assert len(attention_mask.shape) == 2, (
+                "Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
+                "for padding purposes (0 indicating padding). "
+                "Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
+            )
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        q = rearrange(self.q_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+        k = rearrange(self.k_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+        v = rearrange(self.v_proj(hidden_states), '... (h d) -> ... h d', d=self.head_dim)
+
+        # 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/gated_deltaproduct.py

@@ -0,0 +1,351 @@
+from __future__ import annotations
+
+import math
+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 FusedRMSNormSwishGate, RMSNorm, ShortConvolution
+from fla.ops.delta_rule import chunk_delta_rule
+from fla.ops.gated_delta_rule import chunk_gated_delta_rule
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+def elu_p1(x):
+    return (F.elu(x, 1.0, False) + 1.0).to(x)
+
+
+def sum_norm(x):
+    return (x / x.sum(-1, keepdim=True)).to(x)
+
+
+def interleave_multiple_sequences(*sequences):
+    """
+    Interleave multiple sequences together.
+    For example, with sequences [A1, A2], [B1, B2], [C1, C2],
+    returns [A1, B1, C1, A2, B2, C2]
+    """
+    if isinstance(sequences[0], (list, tuple)):
+        sequences = sequences[0]
+
+    if len(sequences) == 1:
+        return sequences[0]
+
+    # All sequences should have the same shape
+    assert all(s.shape == sequences[0].shape for s in sequences)
+
+    # Get the original shape
+    batch_size, seq_len, *rest = sequences[0].shape
+
+    # Stack sequences along a new dimension
+    stacked = torch.stack(sequences, dim=2)
+
+    # Reshape to interleave
+    reshaped = stacked.view(batch_size, seq_len * len(sequences), *rest)
+
+    return reshaped
+
+
+class GatedDeltaProduct(nn.Module):
+    """
+    Generalized version of GatedDoubleDeltaNet that supports arbitrary number of householder transformations.
+    """
+
+    def __init__(
+            self,
+            hidden_size: int = 2048,
+            expand_v: float = 2,
+            head_dim: int = 256,
+            num_heads: int = 6,
+            num_householder: int = 2,  # New parameter for number of householder transformations
+            mode: str = "chunk",
+            use_gate: bool = True,
+            use_forget_gate: bool = True,  # when true Gated DeltaProduct, when false DeltaProduct
+            use_short_conv: bool = True,
+            conv_size: int = 4,
+            conv_bias: bool = False,
+            layer_idx: int | None = None,
+            norm_eps: float = 1e-5,
+            allow_neg_eigval: bool = False,  # when true (Gated) DeltaProduct [-1, 1], when false (Gated) DeltaProduct [0, 1]
+            **kwargs,
+    ) -> None:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+        self.expand_v = expand_v
+        self.use_gate = use_gate
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+        self.head_dim = head_dim
+        self.num_heads = num_heads
+        self.num_householder = num_householder
+        self.allow_neg_eigval = allow_neg_eigval
+        self.use_forget_gate = use_forget_gate
+        self.key_dim = self.num_heads * self.head_dim
+        self.value_dim = int(self.key_dim * self.expand_v)
+        self.head_qk_dim = head_dim
+        self.head_v_dim = int(head_dim * self.expand_v)
+        self.layer_idx = layer_idx
+        self.silu = nn.SiLU()
+        assert mode in ["chunk", "fused_recurrent"], f"Not supported mode `{mode}`."
+        # Create multiple projection layers for each householder transformation
+        self.q_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
+
+        self.k_projs = nn.ModuleList(
+            [
+                nn.Linear(hidden_size, self.key_dim, bias=False)
+                for _ in range(num_householder)
+            ]
+        )
+        self.v_projs = nn.ModuleList(
+            [
+                nn.Linear(hidden_size, self.value_dim, bias=False)
+                for _ in range(num_householder)
+            ]
+        )
+        self.b_projs = nn.ModuleList(
+            [
+                nn.Linear(hidden_size, self.num_heads, bias=False)
+                for _ in range(num_householder)
+            ]
+        )
+        if use_short_conv:
+            self.q_conv1ds = nn.ModuleList(
+                [
+                    ShortConvolution(
+                        hidden_size=self.key_dim,
+                        kernel_size=conv_size,
+                        activation="silu",
+                    )
+                    for _ in range(num_householder)
+                ]
+            )
+            self.k_conv1ds = nn.ModuleList(
+                [
+                    ShortConvolution(
+                        hidden_size=self.key_dim,
+                        kernel_size=conv_size,
+                        activation="silu",
+                    )
+                    for _ in range(num_householder)
+                ]
+            )
+            self.v_conv1ds = nn.ModuleList(
+                [
+                    ShortConvolution(
+                        hidden_size=self.value_dim,
+                        kernel_size=conv_size,
+                        activation="silu",
+                    )
+                    for _ in range(num_householder)
+                ]
+            )
+
+        if self.use_forget_gate:
+            self.a_proj = nn.Linear(hidden_size, self.num_heads, bias=False)
+            A = torch.empty(self.num_heads, dtype=torch.float32).uniform_(0, 16)
+            A_log = torch.log(A)
+            self.A_log = nn.Parameter(A_log)
+            self.A_log._no_weight_decay = True
+
+            # Initialize dt parameters
+            dt_min = 0.001
+            dt_max = 0.1
+            dt_init_floor = 1e-4
+            dt = torch.exp(
+                torch.rand(self.num_heads) * (math.log(dt_max) - math.log(dt_min))
+                + math.log(dt_min)
+            )
+            dt = torch.clamp(dt, min=dt_init_floor)
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            self.dt_bias = nn.Parameter(inv_dt)
+            self.dt_bias._no_weight_decay = True
+
+        if use_gate:
+            self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+            self.o_norm = FusedRMSNormSwishGate(self.head_v_dim, eps=norm_eps)
+        else:
+            self.o_norm = RMSNorm(self.head_v_dim, eps=norm_eps)
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+        self.k_id = torch.nn.Identity()
+        self.apply(self._initialize_weights)
+
+    def _initialize_weights(self, module: nn.Module):
+        if getattr(module, "_is_hf_initialized", False):
+            return
+        if isinstance(module, nn.Linear):
+            nn.init.xavier_uniform_(module.weight, gain=2 ** -2.5)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        module._is_hf_initialized = True
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Cache] = None,
+            use_cache: Optional[bool] = False,
+            output_attentions: Optional[bool] = False,
+            **kwargs: 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)."
+            )
+
+        mode = (
+            "chunk"  # 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+        )
+        if self.training:
+            assert mode == "chunk", "Only chunk mode is supported in training."
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        # Process each householder transformation
+        ks, vs, betas = [], [], []
+        conv_states = []
+
+        for i in range(self.num_householder):
+            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"][
+                        i
+                    ]
+                conv_mask = (
+                    attention_mask[:, -hidden_states.shape[1]:]
+                    if attention_mask is not None
+                    else None
+                )
+
+                k, conv_state_k = self.k_conv1ds[i](
+                    x=self.k_projs[i](hidden_states),
+                    mask=conv_mask,
+                    cache=conv_state_k,
+                    output_final_state=use_cache,
+                )
+                v, conv_state_v = self.v_conv1ds[i](
+                    x=self.v_projs[i](hidden_states),
+                    mask=conv_mask,
+                    cache=conv_state_v,
+                    output_final_state=use_cache,
+                )
+                conv_states.append((conv_state_q, conv_state_k, conv_state_v))
+            else:
+                k = self.silu(self.k_projs[i](hidden_states))
+                v = self.silu(self.v_projs[i](hidden_states))
+
+            ks.append(k)
+            vs.append(v)
+
+            beta = self.b_projs[i](
+                hidden_states
+            ).sigmoid()  # bs, sequence_length, num_heads
+            if attention_mask is not None:
+                beta = beta.mul(attention_mask[:, -hidden_states.shape[1]:, None])
+            if self.allow_neg_eigval:
+                beta = beta * 2
+            betas.append(beta)
+
+        if self.use_short_conv:
+            q, conv_state_q = self.q_conv1ds[0](
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+            )
+        else:
+            q = self.silu(self.q_proj(hidden_states))
+        q = interleave_multiple_sequences(
+            [torch.zeros_like(q)] * (self.num_householder - 1) + [q]
+        )
+        # Interleave all sequences
+        k = interleave_multiple_sequences(ks)
+        v = interleave_multiple_sequences(vs)
+        beta = interleave_multiple_sequences(betas)
+
+        q, k, v = (
+            rearrange(x, "b t (h d) -> b t h d", h=self.num_heads) for x in (q, k, v)
+        )
+
+        recurrent_state = (
+            last_state["recurrent_state"] if last_state is not None else None
+        )
+        offsets = kwargs.get("offsets")
+
+        if mode == "chunk":
+            if self.use_forget_gate:
+                g = -self.A_log.float().exp() * F.softplus(
+                    self.a_proj(hidden_states).float() + self.dt_bias
+                )
+                if attention_mask is not None:
+                    g = g.mul(attention_mask[:, -g.shape[-2]:, None])
+
+                # Interleave g with zeros for non-first transformations
+                g = interleave_multiple_sequences(
+                    [g] + [torch.zeros_like(g)] * (self.num_householder - 1)
+                )
+
+                o, recurrent_state = chunk_gated_delta_rule(
+                    q=q,
+                    k=k,
+                    v=v,
+                    g=g,
+                    beta=beta,
+                    initial_state=recurrent_state,
+                    output_final_state=use_cache,
+                    cu_seqlens=offsets,
+                    head_first=False,
+                    use_qk_l2norm_in_kernel=True
+                )
+            else:
+                o, recurrent_state = chunk_delta_rule(
+                    q=q,
+                    k=k,
+                    v=v,
+                    beta=beta,
+                    initial_state=recurrent_state,
+                    output_final_state=use_cache,
+                    cu_seqlens=offsets,
+                    head_first=False,
+                    use_qk_l2norm_in_kernel=True
+                )
+        else:
+            raise NotImplementedError(f"Not supported mode `{mode}`.")
+
+        # Take every nth element for n householder transformations
+        o = o[:, self.num_householder - 1:: self.num_householder, :]
+
+        if past_key_values is not None:
+            past_key_values.update(
+                recurrent_state=recurrent_state,
+                conv_state=conv_states if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[2],
+            )
+
+        if self.use_gate:
+            g = rearrange(
+                self.g_proj(hidden_states),
+                "... (h d) -> ... h d",
+                h=self.num_heads,
+            )
+            o = self.o_norm(o, g)
+        else:
+            o = self.o_norm(o)
+        o = rearrange(o, "b t h d -> b t (h d)")
+        o = self.o_proj(o)
+
+        return o, None, past_key_values

fla/layers/gla.py

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

fla/layers/gsa.py

@@ -0,0 +1,227 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+import warnings
+from typing import TYPE_CHECKING, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from fla.modules import RMSNorm, ShortConvolution
+from fla.modules.feature_map import ReLUFeatureMap, SwishFeatureMap, T2RFeatureMap
+from fla.modules.layernorm import rms_norm_linear
+from fla.ops.gsa import chunk_gsa, fused_recurrent_gsa
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+    from fla.models.utils import Cache
+
+
+class GatedSlotAttention(nn.Module):
+
+    def __init__(
+        self,
+        mode: str = 'chunk',
+        hidden_size: int = 1024,
+        expand_k: float = 1.,
+        expand_v: float = 1.,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        num_slots: Optional[int] = None,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        gate_logit_normalizer: int = 8,
+        feature_map: str = 'swish',
+        use_output_gate: bool = False,
+        use_norm: bool = True,
+        layer_idx: Optional[int] = None,
+        scale: Optional[float] = 1.,
+        **kwargs
+    ) -> GatedSlotAttention:
+        super().__init__()
+
+        self.mode = mode
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.num_heads = num_heads
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+        self.key_dim = int(hidden_size * expand_k)
+        self.value_dim = int(hidden_size * expand_v)
+        self.key_dim_per_group = self.key_dim // self.num_kv_groups
+        self.value_dim_per_group = self.value_dim // self.num_kv_groups
+        self.head_k_dim = self.key_dim // self.num_heads
+        self.head_v_dim = self.value_dim // self.num_heads
+
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.conv_bias = conv_bias
+
+        self.gate_logit_normalizer = gate_logit_normalizer
+
+        self.use_output_gate = use_output_gate
+        self.use_norm = use_norm
+        self.scale = scale
+
+        if num_slots is None:
+            num_slots = self.head_k_dim
+        self.num_slots = num_slots
+
+        self.layer_idx = layer_idx
+
+        if layer_idx is None:
+            warnings.warn(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.register_module('feature_map', None)
+        if feature_map == 'swish':
+            self.feature_map = SwishFeatureMap()
+        elif feature_map == 'relu':
+            self.feature_map = ReLUFeatureMap()
+        elif feature_map == 't2r':
+            self.feature_map = T2RFeatureMap(self.head_k_dim, self.head_k_dim)
+        else:
+            raise NotImplementedError(f"Feature map `{feature_map}` is not supported now.")
+
+        self.q_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.key_dim_per_group, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.value_dim_per_group, bias=False)
+        self.f_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.num_slots, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
+            self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
+
+        self.g_norm = RMSNorm(self.hidden_size, elementwise_affine, eps=norm_eps)
+        self.o_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        if attention_mask is not None:
+            assert len(attention_mask.shape) == 2, (
+                "Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
+                "for padding purposes (0 indicating padding). "
+                "Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+        f = self.f_proj(hidden_states)
+
+        q = rearrange(q, 'b t (h d) -> b t h d', d=self.head_k_dim)
+        k = rearrange(k, 'b t (h d) -> b t h d', d=self.head_k_dim)
+        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)
+        f = rearrange(f, 'b t (h m) -> b t h m', m=self.num_slots)
+
+        if self.feature_map is not None:
+            q, k = map(lambda x: self.feature_map(x), (q, k))
+        v = F.silu(v)
+
+        f = F.logsigmoid(f) / self.gate_logit_normalizer
+        s = (1 - f.exp()).to(f.dtype)
+        # dealing with left-padding
+        if attention_mask is not None:
+            s = s.mul_(attention_mask[:, -s.shape[1]:, None, None])
+            v = v.mul_(attention_mask[:, -v.shape[1]:, None, None])
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_gsa(
+                q=q,
+                k=k,
+                v=v,
+                s=s,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                scale=self.scale,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'chunk':
+            o, recurrent_state = chunk_gsa(
+                q=q,
+                k=k,
+                v=v,
+                s=s,
+                g=f,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                scale=self.scale,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        else:
+            raise NotImplementedError(f"Not supported mode `{mode}`.")
+
+        if past_key_values is not None:
+            past_key_values.update(
+                recurrent_state=recurrent_state,
+                conv_state=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None,
+                layer_idx=self.layer_idx,
+                offset=q.shape[1]
+            )
+
+        o = rearrange(o, 'b t h d -> b t (h d)')
+        o = rms_norm_linear(F.silu(o), self.g_norm.weight, self.g_norm.bias, self.o_proj.weight, self.o_proj.bias)
+        return o, None, past_key_values
+
+    def state_size(self, *args, **kwargs) -> int:
+        return 2 * self.num_slots * self.hidden_size

fla/layers/hgrn.py

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

fla/layers/hgrn2.py

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

fla/layers/lightnet.py

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

fla/layers/linear_attn.py

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

fla/layers/multiscale_retention.py

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

fla/layers/rebased.py

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

fla/layers/rwkv6.py

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

fla/layers/simple_gla.py

@@ -0,0 +1,261 @@
+# -*- 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
+from einops import rearrange, repeat
+
+from fla.modules import FusedRMSNormGated, RMSNorm, ShortConvolution
+from fla.modules.activations import ACT2FN
+from fla.ops.simple_gla import chunk_simple_gla, fused_recurrent_simple_gla
+
+if TYPE_CHECKING:
+    from fla.models.utils import Cache
+
+
+class SimpleGatedLinearAttention(nn.Module):
+    r"""
+    The layer implementaion for [Gated Linear Attention Transformers with Hardware-Efficient Training](https://arxiv.org/abs/2312.06635).  # noqa
+    This layer calls the simplified GLA kernel in which the gating is head-wise instead of elementwise.
+
+    Args:
+        mode (str, Optional):
+            Which GLA kernel to use.
+            Currently available: `chunk`.
+            Default: `chunk`.
+        hidden_size (int, Optional):
+            The hidden size of the input. Default: 1024.
+        expand_k (float, Optional):
+            The expansion ratio for the key dim. Default: 1.0.
+        expand_v (float, Optional):
+            The expansion ratio for the value dim. Default: 1.0.
+        num_heads (int, Optional):
+            The number of heads. Default: 4.
+        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`.
+        gate_fn (str, Optional):
+            The activation function for the output gate. Default: `swish`.
+        elementwise_affine (bool, Optional):
+            If `True`, applies elementwise affine to LayerNorm with learnable parameters. Default: `True`.
+        norm_eps (float, Optional):
+            The epsilon value for the layernorm/rmsnorm layer. Default: 1e-5.
+        gate_logit_normalizer (int, Optional):
+            The normalizer for the gate logits, appied after `logsigmoid`. Default: 16.
+        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.,
+        expand_v: float = 1.,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        feature_map: Optional[str] = None,
+        use_short_conv: bool = True,
+        conv_size: int = 4,
+        conv_bias: bool = False,
+        gate_fn: str = 'swish',
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-5,
+        gate_logit_normalizer: int = 16,
+        fuse_norm: bool = True,
+        layer_idx: int = None,
+    ) -> SimpleGatedLinearAttention:
+        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.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_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)
+        self.g_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
+
+        if use_short_conv:
+            self.conv_size = conv_size
+            self.q_conv1d = ShortConvolution(self.key_dim, conv_size, activation='silu')
+            self.k_conv1d = ShortConvolution(self.key_dim_per_group, conv_size, activation='silu')
+            self.v_conv1d = ShortConvolution(self.value_dim_per_group, conv_size, activation='silu')
+
+        self.gk_proj = nn.Linear(hidden_size, self.num_heads)
+
+        if gate_fn == 'swish' and fuse_norm:
+            self.g_norm_swish_gate = FusedRMSNormGated(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.fuse_norm_and_gate = True
+        else:
+            self.fuse_norm_and_gate = False
+            self.g_norm = RMSNorm(
+                hidden_size=self.head_v_dim,
+                elementwise_affine=elementwise_affine,
+                eps=norm_eps
+            )
+            self.gate_fn = ACT2FN[gate_fn]
+        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)
+
+        self.gate_logit_normalizer = gate_logit_normalizer
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        if attention_mask is not None:
+            assert len(attention_mask.shape) == 2, (
+                "Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
+                "for padding purposes (0 indicating padding). "
+                "Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
+            )
+
+        # launching the triton kernel for just one token will actually be slower
+        mode = 'fused_recurrent' if hidden_states.shape[1] <= 64 else self.mode
+
+        last_state = None
+        if past_key_values is not None and len(past_key_values) > self.layer_idx:
+            last_state = past_key_values[self.layer_idx]
+
+        cu_seqlens = kwargs.get('cu_seqlens', None)
+        if self.use_short_conv:
+            conv_state_q, conv_state_k, conv_state_v = None, None, None
+            if last_state is not None:
+                conv_state_q, conv_state_k, conv_state_v = last_state['conv_state']
+            conv_mask = attention_mask[:, -hidden_states.shape[1]:] if attention_mask is not None else None
+            q, conv_state_q = self.q_conv1d(
+                x=self.q_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_q,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            k, conv_state_k = self.k_conv1d(
+                x=self.k_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_k,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+            v, conv_state_v = self.v_conv1d(
+                x=self.v_proj(hidden_states),
+                mask=conv_mask,
+                cache=conv_state_v,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens
+            )
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states)
+        gk = self.gk_proj(hidden_states)
+
+        if self.feature_map_fn is not None:
+            q, k = map(self.feature_map_fn, (q, k))
+        # dealing with left-padding
+        if attention_mask is not None:
+            v = v.mul_(attention_mask[:, -v.shape[-2]:, None])
+        q = rearrange(q, '... (h d) -> ... h d', h=self.num_heads)
+        if self.num_kv_groups > 1:
+            k, v = (repeat(x, '... (h d) -> ... (h g) d', h=self.num_kv_heads, g=self.num_kv_groups) for x in (k, v))
+        else:
+            k, v = (rearrange(x, '... (h d) -> ... h d', h=self.num_kv_heads) for x in (k, v))
+        gk = F.logsigmoid(gk) / self.gate_logit_normalizer
+
+        recurrent_state = last_state['recurrent_state'] if last_state is not None else None
+        if mode == 'chunk':
+            o, recurrent_state = chunk_simple_gla(
+                q=q,
+                k=k,
+                v=v,
+                gk=gk,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        elif mode == 'fused_recurrent':
+            o, recurrent_state = fused_recurrent_simple_gla(
+                q=q,
+                k=k,
+                v=v,
+                gk=gk,
+                initial_state=recurrent_state,
+                output_final_state=use_cache,
+                cu_seqlens=cu_seqlens,
+                head_first=False
+            )
+        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]
+            )
+
+        g = self.g_proj(hidden_states)
+        if self.fuse_norm_and_gate:
+            g = rearrange(g, 'b t (h d) -> b t h d', h=self.num_heads)
+            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)
+        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/__init__.py

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

fla/models/utils.py

@@ -0,0 +1,147 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+import transformers
+
+
+class Cache(transformers.cache_utils.Cache):
+    """
+    A cache used for storing hidden states produced by flash linear attention models.
+
+    It stores the states of each layer as the tensor of shape `[batch_size, key_dim, value_dim]`.
+    """
+
+    is_compileable = True
+
+    def __init__(
+        self,
+        seen_tokens: int = 0
+    ) -> Cache:
+        super().__init__()
+
+        self.states: List[Dict[str, Any]] = []
+
+        self._seen_tokens = seen_tokens  # Used in `generate` to keep tally of how many tokens the cache has seen
+
+    def __getitem__(self, layer_idx: int) -> Dict[str, Any]:
+        if layer_idx < len(self):
+            return self.states[layer_idx]
+        else:
+            raise KeyError(f"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}")
+
+    def __iter__(self):
+        for state in self.states:
+            yield state
+
+    def __len__(self):
+        return len(self.states)
+
+    def update(
+        self,
+        recurrent_state: torch.Tensor = None,
+        attn_state: Tuple[torch.Tensor, torch.Tensor] = None,
+        conv_state: Tuple[torch.Tensor] = None,
+        ffn_state: torch.Tensor = None,
+        layer_idx: int = 0,
+        offset: Optional[int] = 1,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        """
+        Updates the cache with the new `recurrent_state`/`attn_state`/`conv_state` for the layer `layer_idx`.
+
+        Args:
+            recurrent_state (`torch.Tensor`, `optional`):
+                The new recurrent state to cache.
+            attn_state (`Tuple[torch.Tensor, torch.Tensor]`, `optional`):
+                The new attention key/value states to cache.
+            conv_state (`Tuple[torch.Tensor]`, `optional`):
+                The new convolution state to cache.
+            layer_idx (`int`, defaults to 0):
+                The index of the layer to cache the states for.
+            offset (`int`, `optional`, defaults to 1):
+                The number of new tokens being processed.
+            cache_kwargs (`Dict[str, Any]`, `optional`):
+                Additional arguments for the cache subclass.
+
+        Return:
+            Dictionary of the updated state.
+        """
+
+        # Update the number of seen tokens
+        if layer_idx == 0:
+            self._seen_tokens += offset
+
+        if attn_state is not None:
+            input_size = attn_state[0].shape[-2]
+            window_size = cache_kwargs.get('window_size', None)
+            if not isinstance(attn_state, Tuple) or len(attn_state) != 2:
+                raise ValueError("`attn_state` must be a tuple of two tensors for key/value states")
+        if len(self.states) <= layer_idx:
+            if attn_state is not None:
+                if window_size is not None and input_size > window_size:
+                    attn_state = (attn_state[0][..., -window_size:, :].contiguous(),
+                                  attn_state[1][..., -window_size:, :].contiguous())
+            state = dict(
+                recurrent_state=recurrent_state,
+                attn_state=attn_state,
+                conv_state=conv_state,
+                ffn_state=ffn_state
+            )
+            self.states.append(state)
+        else:
+            state = self.states[layer_idx]
+            if recurrent_state is not None:
+                state['recurrent_state'] = recurrent_state
+            if attn_state is not None:
+                key_state, value_state = state['attn_state']
+                if window_size is not None and key_state.shape[-2] == window_size:
+                    # DO NOT allocate new memory if the cache is full
+                    # roll the key/value states to the left by `input_size`
+                    key_state = key_state.roll(-input_size, -2)
+                    value_state = value_state.roll(-input_size, -2)
+                    # replace the last `input_size` tokens with the new key/value states
+                    key_state[..., -input_size:, :] = attn_state[0]
+                    value_state[..., -input_size:, :] = attn_state[1]
+                    attn_state = (key_state, value_state)
+                else:
+                    attn_state = (torch.cat([key_state, attn_state[0]], -2),
+                                  torch.cat([value_state, attn_state[1]], -2),)
+                state['attn_state'] = attn_state
+            if conv_state is not None:
+                state['conv_state'] = conv_state
+            if ffn_state is not None:
+                state['ffn_state'] = ffn_state
+
+        return state
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        if len(self.states) <= layer_idx:
+            return 0
+        return self._seen_tokens
+
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states. Cache does not have a maximum length."""
+        return None
+
+    def to_legacy_cache(self) -> Tuple:
+        return tuple(self.states)
+
+    @classmethod
+    @torch.compiler.disable
+    def from_legacy_cache(
+        cls,
+        past_key_values: Optional[Tuple] = None,
+        seen_tokens: int = 0
+    ) -> Cache:
+        """Converts a cache in the legacy cache format into an equivalent `Cache`."""
+
+        cache = cls(seen_tokens)
+        if isinstance(past_key_values, list):
+            for layer_idx in range(len(past_key_values)):
+                cache.states.append(past_key_values[layer_idx])
+        return cache

fla/modules/__init__.py

@@ -0,0 +1,30 @@
+# -*- coding: utf-8 -*-
+
+from fla.modules.convolution import ImplicitLongConvolution, LongConvolution, ShortConvolution
+from fla.modules.fused_bitlinear import BitLinear, FusedBitLinear
+from fla.modules.fused_cross_entropy import FusedCrossEntropyLoss
+from fla.modules.fused_kl_div import FusedKLDivLoss
+from fla.modules.fused_linear_cross_entropy import FusedLinearCrossEntropyLoss
+from fla.modules.fused_linear_listnet_loss import FusedLinearListNetLoss
+from fla.modules.fused_norm_gate import (
+    FusedLayerNormGated,
+    FusedLayerNormSwishGate,
+    FusedLayerNormSwishGateLinear,
+    FusedRMSNormGated,
+    FusedRMSNormSwishGate,
+    FusedRMSNormSwishGateLinear
+)
+from fla.modules.layernorm import GroupNorm, GroupNormLinear, LayerNorm, LayerNormLinear, RMSNorm, RMSNormLinear
+from fla.modules.mlp import GatedMLP
+from fla.modules.rotary import RotaryEmbedding
+
+__all__ = [
+    'ImplicitLongConvolution', 'LongConvolution', 'ShortConvolution',
+    'BitLinear', 'FusedBitLinear',
+    'FusedCrossEntropyLoss', 'FusedLinearCrossEntropyLoss', 'FusedKLDivLoss',
+    'GroupNorm', 'GroupNormLinear', 'LayerNorm', 'LayerNormLinear', 'RMSNorm', 'RMSNormLinear',
+    'FusedLayerNormGated', 'FusedLayerNormSwishGate', 'FusedLayerNormSwishGateLinear',
+    'FusedRMSNormGated', 'FusedRMSNormSwishGate', 'FusedRMSNormSwishGateLinear',
+    'GatedMLP',
+    'RotaryEmbedding'
+]

fla/modules/activations.py

@@ -0,0 +1,471 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Tri Dao, Yu Zhang, Songlin Yang.
+
+import torch
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+
+from fla.ops.utils.op import exp, log
+from fla.utils import autocast_custom_bwd, autocast_custom_fwd, get_multiprocessor_count, input_guard
+
+sigmoid_fwd_codestring = """
+template <typename T> T sigmoid_fwd(T x) {
+    return 1.0f / (1.0f + ::exp(-float(x)));
+}
+"""
+sigmoid_bwd_codestring = """
+template <typename T> T sigmoid_bwd(T x, T g) {
+    float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
+    return float(g) * x_sigmoid * (1.0f - x_sigmoid);
+}
+"""
+
+sigmoid_fwd_jit_fn = torch.cuda.jiterator._create_jit_fn(sigmoid_fwd_codestring)
+sigmoid_bwd_jit_fn = torch.cuda.jiterator._create_jit_fn(sigmoid_bwd_codestring)
+
+
+@torch.compiler.disable
+def sigmoid_fwd(x):
+    return sigmoid_fwd_jit_fn(x)
+
+
+@torch.compiler.disable
+def sigmoid_bwd(x, g):
+    return sigmoid_bwd_jit_fn(x, g)
+
+
+class SigmoidFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x):
+        ctx.save_for_backward(x)
+        return sigmoid_fwd(x)
+
+    @staticmethod
+    def backward(ctx, dout):
+        x, = ctx.saved_tensors
+        return sigmoid_bwd(x, dout)
+
+
+sigmoid = SigmoidFunction.apply
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+    ],
+    key=['D']
+)
+@triton.jit
+def logsigmoid_fwd_kernel(
+    x,
+    y,
+    temperature,
+    T: tl.constexpr,
+    D: tl.constexpr,
+    B: tl.constexpr
+):
+    i = tl.program_id(0)
+    o_i = i * B + tl.arange(0, B)
+    m_i = o_i < T
+
+    b_x = tl.load(x + o_i, mask=m_i, other=0.).to(tl.float32)
+    b_m = tl.minimum(0., b_x)
+    b_z = 1. + exp(-tl.abs(b_x))
+    b_y = (b_m - log(b_z)) / temperature
+    tl.store(y + o_i, b_y.to(y.dtype.element_ty), mask=m_i)
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+    ],
+    key=['D']
+)
+@triton.jit
+def logsigmoid_bwd_kernel(
+    x,
+    dx,
+    dy,
+    temperature,
+    T: tl.constexpr,
+    D: tl.constexpr,
+    B: tl.constexpr
+):
+    i = tl.program_id(0)
+    o_i = i * B + tl.arange(0, B)
+    m_i = o_i < T
+
+    b_x = tl.load(x + o_i, mask=m_i, other=0.).to(tl.float32)
+    b_dy = tl.load(dy + o_i, mask=m_i, other=0.).to(tl.float32)
+    b_dx = b_dy * (1. - tl.sigmoid(b_x)) / temperature
+    tl.store(dx + o_i, b_dx.to(dx.dtype.element_ty), mask=m_i)
+
+
+def logsigmoid_fwd(x: torch.Tensor, temperature: float = 1.) -> torch.Tensor:
+    T, D = x.numel(), x.shape[-1]
+    B = triton.next_power_of_2(triton.cdiv(T, get_multiprocessor_count(x.device.index)))
+    y = torch.empty_like(x)
+    logsigmoid_fwd_kernel[(triton.cdiv(T, B),)](
+        x=x,
+        y=y,
+        temperature=temperature,
+        T=T,
+        D=D,
+        B=B
+    )
+    return y
+
+
+def logsigmoid_bwd(x: torch.Tensor, dy: torch.Tensor, temperature: float = 1.) -> torch.Tensor:
+    T, D = x.numel(), x.shape[-1]
+    B = triton.next_power_of_2(triton.cdiv(T, get_multiprocessor_count(x.device.index)))
+    dx = torch.empty_like(x)
+    logsigmoid_bwd_kernel[(triton.cdiv(T, B),)](
+        x=x,
+        dx=dx,
+        dy=dy,
+        temperature=temperature,
+        T=T,
+        D=D,
+        B=B
+    )
+    return dx
+
+
+class LogSigmoidFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(ctx, x, temperature):
+        ctx.save_for_backward(x,)
+        ctx.temperature = temperature
+        return logsigmoid_fwd(x, temperature)
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, dy):
+        x, = ctx.saved_tensors
+        return logsigmoid_bwd(x, dy, ctx.temperature), None
+
+
+def logsigmoid(x: torch.Tensor, temperature: float = 1.) -> torch.Tensor:
+    return LogSigmoidFunction.apply(x, temperature)
+
+
+swish_fwd_codestring = """
+template <typename T> T swish_fwd(T x) {
+    float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
+    return float(x) * x_sigmoid;
+}
+"""
+swish_bwd_codestring = """
+template <typename T> T swish_bwd(T x, T g) {
+    float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
+    return float(g) * x_sigmoid * (1.0f - float(x) * x_sigmoid + float(x));
+}
+"""
+
+swish_fwd_jit_fn = torch.cuda.jiterator._create_jit_fn(swish_fwd_codestring)
+swish_bwd_jit_fn = torch.cuda.jiterator._create_jit_fn(swish_bwd_codestring)
+
+
+@torch.compiler.disable
+def swish_fwd(x):
+    return swish_fwd_jit_fn(x)
+
+
+@torch.compiler.disable
+def swish_bwd(x, g):
+    return swish_bwd_jit_fn(x, g)
+
+
+class SwishFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x):
+        ctx.save_for_backward(x)
+        return swish_fwd(x)
+
+    @staticmethod
+    def backward(ctx, dout):
+        x, = ctx.saved_tensors
+        return swish_bwd(x, dout)
+
+
+swish = SwishFunction.apply
+
+# 1/sqrt(2*pi)-> 0.3989423
+# 1/sqrt(2)   -> 0.70710678
+# sqrt(2/pi)  -> 0.79788456
+
+
+# this function is tanh approximation of gelu
+# actual gelu is:
+# x * 0.5 * (1.0 + torch.erf(x * 0.70710678))
+@torch.compile
+def bias_gelu(y, bias):
+    x = bias + y
+    return (x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))).to(dtype=y.dtype)
+
+
+# gradient of tanh approximation of gelu
+# gradient of actual gelu is:
+# 0.5 * (1. + torch.erf(x * 0.70710678)) + 0.3989423 * x * torch.exp(-0.5 * x * x)
+@torch.compile
+def bias_gelu_bwd(g, y, bias):
+    """Assume that y has shape (B, D) and bias has shape (D)"""
+    x = bias + y
+    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
+    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
+    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (
+        1 + tanh_out
+    )
+    grad_y = ff * g
+    return grad_y.to(dtype=y.dtype), grad_y.sum(dim=(0), dtype=bias.dtype)
+
+
+class GeLUFunction(torch.autograd.Function):
+
+    @staticmethod
+    # bias is an optional argument
+    def forward(ctx, input, bias):
+        ctx.save_for_backward(input, bias)
+        return bias_gelu(input, bias)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, bias = ctx.saved_tensors
+        tmp = bias_gelu_bwd(grad_output, input, bias)
+        return tmp, tmp
+
+
+bias_gelu_impl = GeLUFunction.apply
+
+
+# this function is tanh approximation of gelu
+# actual gelu is:
+# x * 0.5 * (1.0 + torch.erf(x * 0.70710678))
+@torch.compile
+def gelu_fwd(x):
+    return (x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))).to(dtype=x.dtype)
+
+
+# gradient of tanh approximation of gelu
+# gradient of actual gelu is:
+# 0.5 * (1. + torch.erf(x * 0.70710678)) + 0.3989423 * x * torch.exp(-0.5 * x * x)
+@torch.compile
+def gelu_bwd(g, x):
+    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
+    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
+    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (
+        1 + tanh_out
+    )
+    return (ff * g).to(dtype=x.dtype)
+
+
+class FastGeLUFunction(torch.autograd.Function):
+    @staticmethod
+    # bias is an optional argument
+    def forward(ctx, input):
+        ctx.save_for_backward(input)
+        return gelu_fwd(input)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        (input,) = ctx.saved_tensors
+        tmp = gelu_bwd(grad_output, input)
+        return tmp
+
+
+fast_gelu_impl = FastGeLUFunction.apply
+
+
+@torch.compile
+def relu_bwd(g, x):
+    return torch.where(x >= 0, g, 0.0).to(dtype=x.dtype)
+
+
+@torch.compile
+def sqrelu_fwd(x):
+    r = F.relu(x.float())
+    return (r * r).to(dtype=x.dtype)
+
+
+@torch.compile
+def sqrelu_bwd(g, x):
+    return (2.0 * g * F.relu(x.float())).to(dtype=x.dtype)
+
+
+class SquaredReLUFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, input):
+        ctx.save_for_backward(input)
+        return sqrelu_fwd(input)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, = ctx.saved_tensors
+        return sqrelu_bwd(grad_output, input)
+
+
+sqrelu = SquaredReLUFunction.apply
+
+
+swiglu_fwd_codestring = """
+template <typename T> T swiglu_fwd(T x, T y) {
+    return float(x) * float(y) / (1.0f + ::exp(-float(x)));
+}
+"""
+swiglu_bwd_codestring = """
+template <typename T> T swiglu_bwd(T x, T y, T g, T& dx, T& dy) {
+    float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
+    dx = x_sigmoid * (1 + float(x) * (1.0f - x_sigmoid)) * float(g) * float(y);
+    dy = float(x) * x_sigmoid * float(g);
+}
+"""
+
+swiglu_fwdbwd_codestring = """
+template <typename T> T swiglu_fwdbwd(T x, T y, T g, T& dx, T& dy, T& z) {
+    float x_sigmoid = 1.0f / (1.0f + ::exp(-float(x)));
+    float x_swish = float(x) * x_sigmoid;
+    dx = x_sigmoid * (1 + float(x) * (1.0f - x_sigmoid)) * float(g) * float(y);
+    dy = x_swish * float(g);
+    z = x_swish * float(y);
+}
+"""
+
+
+swiglu_fwd_jit_fn = torch.cuda.jiterator._create_jit_fn(swiglu_fwd_codestring)
+swiglu_bwd_jit_fn = torch.cuda.jiterator._create_multi_output_jit_fn(swiglu_bwd_codestring, num_outputs=2)
+swiglu_fwdbwd_jit_fn = torch.cuda.jiterator._create_multi_output_jit_fn(swiglu_fwdbwd_codestring, num_outputs=3)
+
+
+@torch.compiler.disable
+def swiglu_fwd(x, y):
+    return swiglu_fwd_jit_fn(x, y)
+
+
+@torch.compiler.disable
+def swiglu_bwd(x, y, g):
+    return swiglu_bwd_jit_fn(x, y, g)
+
+
+@torch.compiler.disable
+def swiglu_fwdbwd(x, y, g):
+    return swiglu_fwdbwd_jit_fn(x, y, g)
+
+
+@torch.compile
+def swiglu_fwd_torch(x, y):
+    return (F.silu(x.float()) * y).to(x.dtype)
+
+
+@torch.compile
+def swiglu_bwd_torch(x, y, g):
+    dtype = x.dtype
+    x, y, g = x.float(), y.float(), g.float()
+    x_sigmoid = x.sigmoid()
+    x_swish = x * x_sigmoid
+    dx = x_sigmoid * (1 + x * (1.0 - x_sigmoid)) * g * y
+    dy = x_swish * g
+    return dx.to(dtype), dy.to(dtype)
+
+
+@torch.compile
+def swiglu_fwdbwd_torch(x, y, g):
+    dtype = x.dtype
+    x, y, g = x.float(), y.float(), g.float()
+    x_sigmoid = x.sigmoid()
+    x_swish = x * x_sigmoid
+    dx = x_sigmoid * (1 + x * (1.0 - x_sigmoid)) * g * y
+    dy = x_swish * g
+    z = x_swish * y
+    return dx.to(dtype), dy.to(dtype), z.to(dtype)
+
+
+class SwiGLUFunction(torch.autograd.Function):
+    r"""
+    Swish-Gated Linear Unit (SwiGLU) function.
+
+    .. math::
+        \text{SwiGLU}(x, y) = swish(x) * y = \frac{x}{1 + \exp(-x)} * y
+    """
+
+    @staticmethod
+    def forward(ctx, x, y):
+        ctx.save_for_backward(x, y)
+        if torch.compiler.is_compiling() or isinstance(x, torch.distributed.tensor.DTensor):
+            return swiglu_fwd_torch(x, y)
+        else:
+            return swiglu_fwd(x, y)
+
+    @staticmethod
+    def backward(ctx, dout):
+        x, y = ctx.saved_tensors
+        if torch.compiler.is_compiling() or isinstance(x, torch.distributed.tensor.DTensor):
+            return swiglu_bwd_torch(x, y, dout)
+        else:
+            return swiglu_bwd(x, y, dout)
+
+
+class SwiGLULinearFunction(torch.autograd.Function):
+    r"""
+    Swish-Gated Linear Unit (SwiGLU) function followed by a linear transformation.
+
+    .. math::
+        \text{SwiGLULinear}(x, y, W, b) = (swish(x) * y) W + b
+
+    This simple wrap discards the intermediate results of SwiGLU(x, y) to save memory.
+    """
+
+    @staticmethod
+    @autocast_custom_fwd
+    def forward(ctx, x, y, weight, bias):
+        with torch.no_grad():
+            if torch.compiler.is_compiling() or isinstance(x, torch.distributed.tensor.DTensor):
+                z = swiglu_fwd_torch(x, y)
+            else:
+                z = swiglu_fwd(x, y)
+        out = F.linear(z, weight, bias)
+        # We don't store z, will be recomputed in the backward pass to save memory
+        ctx.save_for_backward(x, y, weight)
+        ctx.linear_bias_is_none = bias is None
+        return out
+
+    @staticmethod
+    @autocast_custom_bwd
+    def backward(ctx, dout, *args):
+        x, y, weight = ctx.saved_tensors
+        dout = dout.reshape(-1, dout.shape[-1])
+        dz = F.linear(dout, weight.t()).view_as(x)
+        with torch.no_grad():
+            if torch.compiler.is_compiling() or isinstance(x, torch.distributed.tensor.DTensor):
+                dx, dy, z = swiglu_fwdbwd_torch(x, y, dz)
+            else:
+                dx, dy, z = swiglu_fwdbwd(x, y, dz)
+        dlinear_weight = torch.einsum("bo,bi->oi", dout, z.reshape(-1, z.shape[-1]))
+        dlinear_bias = None if ctx.linear_bias_is_none else dout.sum(0)
+        return dx, dy, dlinear_weight, dlinear_bias
+
+
+swiglu = SwiGLUFunction.apply
+
+
+swiglu_linear = SwiGLULinearFunction.apply
+
+
+ACT2FN = {
+    'relu': F.relu,
+    'sigmoid': sigmoid,
+    'logsigmoid': logsigmoid,
+    'silu': swish,
+    'swish': swish,
+    'sqrelu': sqrelu,
+    'gelu': fast_gelu_impl,
+    'bias_gelu': bias_gelu_impl,
+}

fla/modules/convolution.py

@@ -0,0 +1,434 @@
+# -*- coding: utf-8 -*-
+
+# from https://github.com/HazyResearch/zoology/blob/main/zoology/mixers/convolution.py
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+from einops import rearrange
+
+from fla.modules.activations import ACT2FN
+from fla.ops.common.utils import prepare_position_ids, prepare_sequence_ids
+from fla.utils import checkpoint, input_guard
+
+try:
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+except ImportError:
+    causal_conv1d_fn = None
+    causal_conv1d_update = None
+
+
+def fft_conv(u, k, dropout_mask, gelu=True, k_rev=None):
+    seqlen = u.shape[-1]
+    fft_size = 2 * seqlen
+    k_f = torch.fft.rfft(k, n=fft_size) / fft_size
+    if k_rev is not None:
+        k_rev_f = torch.fft.rfft(k_rev, n=fft_size) / fft_size
+        k_f = k_f + k_rev_f.conj()
+    u_f = torch.fft.rfft(u.to(dtype=k.dtype), n=fft_size)
+
+    if len(u.shape) > 3:
+        k_f = k_f.unsqueeze(1)
+    y = torch.fft.irfft(u_f * k_f, n=fft_size, norm="forward")[..., :seqlen]
+
+    out = y + u
+    if gelu:
+        out = F.gelu(out)
+    if dropout_mask is not None:
+        return (out * rearrange(dropout_mask, "b H -> b H 1")).to(dtype=u.dtype)
+    else:
+        return out.to(dtype=u.dtype)
+
+
+@checkpoint
+def proj_then_conv1d(
+    x: torch.Tensor,
+    proj_weight: torch.Tensor,
+    conv1d_weight: torch.Tensor,
+    conv1d_bias: Optional[torch.Tensor] = None,
+    cache: Optional[torch.Tensor] = None
+) -> torch.Tensor:
+    # We do matmul and transpose BLH -> HBL at the same time
+    x = rearrange(proj_weight @ rearrange(x, "b t d -> d (b t)"), "d (b t) -> b d t", t=x.shape[-2])
+
+    if causal_conv1d_fn is None:
+        raise ImportError("`causal_conv1d_fn` is not available. Please install `causal-conv1d` first.")
+    if cache is None:
+        x = causal_conv1d_fn(
+            x=x,
+            weight=rearrange(conv1d_weight, "d 1 w -> d w"),
+            bias=conv1d_bias,
+            activation="silu",
+        ).transpose(1, 2)
+    else:
+        assert x.shape[-1] == 1, "Only support decoding with 1 token at a time for now"
+        x = x.squeeze(-1)
+        x = causal_conv1d_update(
+            x=x,
+            weight=rearrange(conv1d_weight, "d 1 w -> d w"),
+            bias=conv1d_bias,
+            cache=cache,
+            activation="silu",
+        )
+    return x
+
+
+@triton.jit
+def causal_conv1d_varlen_states_fwd_kernel(
+    x,
+    cache,
+    offsets,
+    D,
+    W,
+    BD: tl.constexpr,
+    BW: tl.constexpr
+):
+    i_d, i_w, i_n = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    eos = tl.load(offsets + i_n + 1)
+    bos = tl.maximum(tl.load(offsets + i_n), eos - W)
+    o_t = eos - (i_w + 1) * BW + tl.arange(0, BW)
+    o_d = i_d * BD + tl.arange(0, BD)
+    o_w = W - (i_w + 1) * BW + tl.arange(0, BW)
+
+    b_x = tl.load(x + o_t * D + o_d[:, None], mask=(o_t >= bos) & (o_d[:, None] < D), other=0)
+    tl.store(cache + i_n * D*W + o_d[:, None] * W + o_w, b_x, mask=(o_d[:, None] < D) & (o_w >= 0))
+
+
+@input_guard
+def causal_conv1d_varlen_states_fwd(
+    x: torch.Tensor,
+    cache: torch.Tensor,
+    cu_seqlens: torch.Tensor,
+    state_len: int
+) -> torch.Tensor:
+    N, D, W = len(cu_seqlens) - 1, x.shape[-1], state_len
+    cache = torch.empty(N, D, W, dtype=x.dtype, device=x.device) if cache is None else cache
+    BD = min(triton.next_power_of_2(D), 256)
+    BW = min(triton.next_power_of_2(state_len), 16)
+    grid = (triton.cdiv(D, BD), triton.cdiv(W, BW), N)
+    with torch.cuda.device(x.device.index):
+        causal_conv1d_varlen_states_fwd_kernel[grid](
+            x=x,
+            cache=cache,
+            offsets=cu_seqlens,
+            D=D,
+            W=W,
+            BW=BW,
+            BD=BD
+        )
+    return cache
+
+
+class ShortConvolution(nn.Conv1d):
+    """
+    Simple wrapper around `nn.Conv1d` that accepts dimension last.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        kernel_size: int,
+        bias: bool = False,
+        activation: Optional[str] = 'silu',
+        use_fast_conv1d: Optional[bool] = True,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__(
+            in_channels=hidden_size,
+            out_channels=hidden_size,
+            kernel_size=kernel_size,
+            groups=hidden_size,
+            bias=bias,
+            padding=kernel_size - 1,
+            device=device,
+            dtype=dtype,
+        )
+
+        self.hidden_size = hidden_size
+        self.activation = None
+        if activation is not None:
+            assert activation in ['silu', 'swish'], f"Activation `{activation}` not supported yet."
+            self.activation = activation
+
+        if causal_conv1d_fn is None:
+            if use_fast_conv1d:
+                raise RuntimeError(
+                    "Please either install `causal-conv1d>=1.4.0` to enable fast causal short convolution CUDA kernel "
+                    "or set `use_fast_conv1d` to False"
+                )
+            else:
+                warnings.warn(
+                    "The naive Pytorch verison is very slow in practice, "
+                    "please run `pip install causal-conv1d>=1.4.0` to install fast causal short convolution CUDA kernel",
+                    category=ImportWarning
+                )
+        self.use_fast_conv1d = use_fast_conv1d
+
+    def extra_repr(self):
+        s = ('{in_channels}, {out_channels}, kernel_size={kernel_size}'
+             ', stride={stride}')
+        if self.padding != (0,) * len(self.padding):
+            s += ', padding={padding}'
+        if self.dilation != (1,) * len(self.dilation):
+            s += ', dilation={dilation}'
+        if self.output_padding != (0,) * len(self.output_padding):
+            s += ', output_padding={output_padding}'
+        if self.groups != 1:
+            s += ', groups={groups}'
+        if self.bias is None:
+            s += ', bias=False'
+        if self.padding_mode != 'zeros':
+            s += ', padding_mode={padding_mode}'
+        if self.activation is not None:
+            s += ', activation={activation}'
+        if not self.use_fast_conv1d:
+            s += ', use_fast_conv1d={use_fast_conv1d}'
+        return s.format(**self.__dict__)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        cache: Optional[torch.Tensor] = None,
+        output_final_state: bool = False,
+        cu_seqlens: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            x (`torch.Tensor`):
+                Tensor of shape `[B, T, D]`.
+                If `seq_idx` is provided, `B` must be 1.
+            mask (`Optional[torch.Tensor]`):
+                Attention mask dealing with padded positions.
+            cache (`Optional[torch.Tensor]`):
+                Previous cache tensor of shape `[N, D, W]`, where `W` is the kernel size.
+                If provided, the cache is updated **inplace**.
+            output_final_state (Optional[bool]):
+                Whether to output the final state of shape `[N, D, W]`. Default: `False`.
+            cu_seqlens (Optional[torch.LongTensor]):
+                Cumulative sequence lengths for each batch. Used for varlen. Default: `None`.
+                Shape: [B+1]
+
+        Returns:
+            Tensor of shape `[B, T, D]`.
+        """
+
+        B, T, D, W = *x.shape, self.kernel_size[0]
+        N = B if cu_seqlens is None else len(cu_seqlens) - 1
+        if mask is not None:
+            if cu_seqlens is not None:
+                raise ValueError("`mask` and `cu_seqlens` cannot be provided at the same time")
+            x = x.mul_(mask.unsqueeze(-1))
+        if output_final_state and cache is None:
+            cache = x.new_zeros(N, D, W)
+        # during the decoding phase, we assume the batch is composed of sequences of length 1
+        if cache is not None and B * T == N:
+            return self.step(x, cache, cu_seqlens)
+
+        if cache is not None:
+            if cu_seqlens is not None:
+                cache = causal_conv1d_varlen_states_fwd(x, cache, cu_seqlens, W)
+            else:
+                cache[:, :, -min(W, T):].copy_(rearrange(x[..., -min(W, T):, :], 'n w d -> n d w'))
+
+        x = rearrange(x, 'b t d -> b d t')
+        if self.use_fast_conv1d:
+            # Sequence index for each token. Used for varlen.
+            # Suppose a batch consists of two sequences with lengths 3 and 4,
+            # seq_idx=[0, 0, 0, 1, 1, 1, 1] for this batch.
+            # NOTE: No need to provide this arg if `cu_seqlens` is passed.
+            # This arg is just for BC, and will be removed in the future.
+            # [B, T]
+            seq_idx = kwargs.get('seq_idx', None)
+            if cu_seqlens is not None and seq_idx is None:
+                seq_idx = prepare_sequence_ids(prepare_position_ids(cu_seqlens)).to(torch.int32).unsqueeze(0)
+            x = causal_conv1d_fn(
+                x=x,
+                weight=rearrange(self.weight, "d 1 w -> d w"),
+                bias=self.bias,
+                activation=self.activation,
+                seq_idx=seq_idx,
+            )
+        else:
+            if cu_seqlens is not None:
+                raise ValueError("`cu_seqlens` is not supported for the naive Pytorch version")
+            x = self._conv_forward(x, self.weight, self.bias)[..., :x.shape[-1]]
+            if self.activation is not None:
+                x = ACT2FN[self.activation](x)
+        return rearrange(x, "b d t -> b t d"), cache
+
+    def step(
+        self,
+        x: torch.Tensor,
+        cache: torch.Tensor,
+        cu_seqlens: Optional[torch.LongTensor] = None
+    ):
+        shape = x.shape
+        x = x.squeeze(0) if cu_seqlens is not None else x.squeeze(1)
+        if self.use_fast_conv1d:
+            x = causal_conv1d_update(
+                x=x,
+                conv_state=cache,
+                weight=rearrange(self.weight, "d 1 w -> d w"),
+                bias=self.bias,
+                activation=self.activation,
+            )
+        else:
+            dtype = x.dtype
+            # we follow the fast mode that updates the cache in-place
+            cache.copy_(cache.roll(shifts=-1, dims=-1))
+            cache[:, :, -1] = x
+            x = torch.sum(cache * rearrange(self.weight, "d 1 w -> d w"), dim=-1)
+            if self.bias is not None:
+                x = x + self.bias
+            if self.activation is not None:
+                x = ACT2FN[self.activation](x).to(dtype=dtype)
+        return x.view(shape), cache
+
+    @property
+    def state_size(self) -> int:
+        return self.hidden_size * self.kernel_size
+
+
+class LongConvolution(nn.Module):
+    """
+    LongConvolution applies a convolution operation on the input tensor using a fixed
+    filter of length max_len.
+    The filter is learned during training and is applied using FFT convolution.
+    Args:
+        hidden_size (int): The number of expected features in the input and output.
+        max_len (int): The maximum sequence length.
+    Returns:
+        y: [batch_size, seq_len, hidden_size] tensor
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        max_len: int,
+        **kwargs,
+    ):
+        """
+        Initializes the LongConvolution module.
+        Args:
+            hidden_size (int): The number of expected features in the input and output.
+            max_len (int): The maximum sequence length.
+        """
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.filter = nn.Parameter(torch.randn(self.hidden_size, max_len), requires_grad=True)
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        """
+        Applies the LongConvolution operation on the input tensor.
+        Args:
+            x: [batch_size, seq_len, hidden_size] tensor
+        Returns:
+            y: [batch_size, seq_len, hidden_size] tensor
+        """
+        x = x.transpose(1, 2)
+        y = fft_conv(x, self.filter, dropout_mask=None, gelu=False)
+        y = y.transpose(1, 2)
+        return y.to(dtype=x.dtype)
+
+
+class PositionalEmbedding(nn.Module):
+    def __init__(self, emb_dim: int, seq_len: int, **kwargs):
+        """Complex exponential positional embeddings for implicit long convolution filters."""
+        super().__init__()
+
+        self.seq_len = seq_len
+        # The time embedding fed to the filteres is normalized so that t_f = 1
+        t = torch.linspace(0, 1, self.seq_len)[None, :, None]  # 1, L, 1
+
+        if emb_dim > 1:
+            bands = (emb_dim - 1) // 2
+        # To compute the right embeddings we use the "proper" linspace
+        t_rescaled = torch.linspace(0, seq_len - 1, seq_len)[None, :, None]
+        w = 2 * math.pi * t_rescaled / seq_len  # 1, L, 1
+
+        f = torch.linspace(1e-4, bands - 1, bands)[None, None]
+        z = torch.exp(-1j * f * w)
+        z = torch.cat([t, z.real, z.imag], dim=-1)
+        self.z = nn.Parameter(z, requires_grad=False)
+
+    def forward(self, L):
+        return self.z[:, :L]
+
+
+class ImplicitLongConvolution(nn.Module):
+    """
+    Long convolution with implicit filter parameterized by an MLP.
+
+    Args:
+        hidden_size (int):
+            The number of expected features in the input and output.
+        max_len (int):
+            The maximum sequence length.
+        d_emb (Optional[int]):
+            The dimension of the positional embeddings. Must be odd and greater or equal to 3 (time, sine and cosine).
+            Defaults to 3.
+        d_hidden (Optional[int]):
+            The number of features in the hidden layer of the MLP. Defaults to 16.
+
+    Attributes:
+        pos_emb (`PositionalEmbedding`): The positional embedding layer.
+        mlp (`nn.Sequential`): The MLP that parameterizes the implicit filter.
+
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        max_len: int,
+        d_emb: int = 3,
+        d_hidden: int = 16,
+        **kwargs,
+    ):
+        """
+        Long convolution with implicit filter parameterized by an MLP.
+
+
+        """
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.d_emb = d_emb
+
+        assert (
+            d_emb % 2 != 0 and d_emb >= 3
+        ), "d_emb must be odd and greater or equal to 3 (time, sine and cosine)"
+        self.pos_emb = PositionalEmbedding(d_emb, max_len)
+
+        # final linear layer
+        self.mlp = nn.Sequential(
+            nn.Linear(d_emb, d_hidden),
+            torch.nn.ReLU(),
+            nn.Linear(d_hidden, hidden_size),
+        )
+
+    def filter(self, seq_len: int, *args, **kwargs):
+        k = self.mlp(self.pos_emb(seq_len))
+
+        return k.transpose(1, 2)
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        """
+        Args:
+            x: [batch_size, seq_len, hidden_size] tensor
+        Returns:
+            y: [batch_size, seq_len, hidden_size] tensor
+        """
+        x = x.transpose(1, 2)
+        k = self.filter(x.shape[-1])
+        y = fft_conv(x, k, dropout_mask=None, gelu=False)
+
+        y = y.transpose(1, 2)
+        return y.to(dtype=x.dtype)

fla/modules/fused_bitlinear.py

@@ -0,0 +1,638 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# Implementations of BitLinear layer with fused LayerNorm and quantized Linear layer.
+# [The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits](https://arxiv.org/abs/2402.17764)
+# [Scalable MatMul-free Language Modeling](https://arxiv.org/abs/2406.02528)
+
+# Code adapted from https://github.com/ridgerchu/matmulfreellm/
+
+from __future__ import annotations
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+
+from fla.modules.layernorm import RMSNorm
+from fla.utils import get_multiprocessor_count, input_guard, require_version
+
+
+def activation_quant(x):
+    """
+    Per-token quantization to 8 bits. No grouping is needed for quantization.
+
+    Args:
+        x: An activation tensor with shape [n, d].
+
+    Returns:
+        A quantized activation tensor with shape [n, d].
+    """
+    # Compute the scale factor
+    scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
+    # Quantize and then de-quantize the tensor
+    y = (x * scale).round().clamp_(-128, 127) / scale
+    return y
+
+
+def weight_quant(w):
+    """
+    Per-tensor quantization to 1.58 bits. No grouping is needed for quantization.
+
+    Args:
+        w: A weight tensor with shape [d, k].
+
+    Returns:
+        A quantized weight tensor with shape [d, k].
+    """
+    # Compute the scale factor
+    scale = 1.0 / w.abs().mean().clamp_(min=1e-5)
+    # Quantize and then de-quantize the tensor
+    u = (w * scale).round().clamp_(-1, 1) / scale
+    return u
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=1),
+        triton.Config({}, num_warps=2),
+        triton.Config({}, num_warps=4),
+        triton.Config({}, num_warps=8),
+        triton.Config({}, num_warps=16),
+        triton.Config({}, num_warps=32),
+    ],
+    key=["N", "HAS_RESIDUAL", "STORE_RESIDUAL_OUT", "IS_RMS_NORM", "HAS_BIAS"],
+)
+@triton.jit
+def layer_norm_fwd_kernel_quant(
+    X,  # pointer to the input
+    Y,  # pointer to the output
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    RESIDUAL,  # pointer to the residual
+    RESIDUAL_OUT,  # pointer to the residual
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    stride_x_row,  # how much to increase the pointer when moving by 1 row
+    stride_y_row,
+    stride_res_row,
+    stride_res_out_row,
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_RESIDUAL: tl.constexpr,
+    STORE_RESIDUAL_OUT: tl.constexpr,
+    HAS_WEIGHT: tl.constexpr,
+    HAS_BIAS: tl.constexpr
+):
+    # Map the program id to the row of X and Y it should compute.
+    row = tl.program_id(0)
+    X += row * stride_x_row
+    Y += row * stride_y_row
+    if HAS_RESIDUAL:
+        RESIDUAL += row * stride_res_row
+    if STORE_RESIDUAL_OUT:
+        RESIDUAL_OUT += row * stride_res_out_row
+    # Compute mean and variance
+    cols = tl.arange(0, BLOCK_N)
+    x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
+    if HAS_RESIDUAL:
+        residual = tl.load(RESIDUAL + cols, mask=cols < N, other=0.0).to(tl.float32)
+        x += residual
+    if STORE_RESIDUAL_OUT:
+        tl.store(RESIDUAL_OUT + cols, x, mask=cols < N)
+    if not IS_RMS_NORM:
+        mean = tl.sum(x, axis=0) / N
+        tl.store(Mean + row, mean)
+        xbar = tl.where(cols < N, x - mean, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    else:
+        xbar = tl.where(cols < N, x, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    rstd = 1 / tl.sqrt(var + eps)
+    tl.store(Rstd + row, rstd)
+    # Normalize and apply linear transformation
+    mask = cols < N
+    if HAS_WEIGHT:
+        w = tl.load(W + cols, mask=mask).to(tl.float32)
+    if HAS_BIAS:
+        b = tl.load(B + cols, mask=mask).to(tl.float32)
+    x_hat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+
+    y = x_hat * w if HAS_WEIGHT else x_hat
+    if HAS_BIAS:
+        y = y + b
+
+    # Aply quantization to the output
+    scale = 127.0 / tl.maximum(tl.max(tl.abs(y), 0), 1e-5)
+    # Quantize and then de-quantize the tensor
+    y = tl.extra.cuda.libdevice.round(y * scale)
+    y = tl.maximum(tl.minimum(y, 127), -128) / scale
+
+    # Write output
+    tl.store(Y + cols, y, mask=mask)
+
+
+def layer_norm_fwd_quant(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float,
+    residual: torch.Tensor = None,
+    out_dtype: torch.dtype = None,
+    residual_dtype: torch.dtype = None,
+    is_rms_norm: bool = False
+):
+    if residual is not None:
+        residual_dtype = residual.dtype
+    M, N = x.shape
+    # allocate output
+    y = torch.empty_like(x, dtype=x.dtype if out_dtype is None else out_dtype)
+    if residual is not None or (residual_dtype is not None and residual_dtype != x.dtype):
+        residual_out = torch.empty(M, N, device=x.device, dtype=residual_dtype)
+    else:
+        residual_out = None
+    mean = torch.empty((M,), dtype=torch.float32, device=x.device) if not is_rms_norm else None
+    rstd = torch.empty((M,), dtype=torch.float32, device=x.device)
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    # heuristics for number of warps
+    layer_norm_fwd_kernel_quant[(M,)](
+        x,
+        y,
+        weight,
+        bias,
+        residual,
+        residual_out,
+        mean,
+        rstd,
+        x.stride(0),
+        y.stride(0),
+        residual.stride(0) if residual is not None else 0,
+        residual_out.stride(0) if residual_out is not None else 0,
+        N,
+        eps,
+        is_rms_norm,
+        BLOCK_N,
+        residual is not None,
+        residual_out is not None,
+        weight is not None,
+        bias is not None,
+    )
+    # residual_out is None if residual is None and residual_dtype == input_dtype
+    return y, mean, rstd, residual_out if residual_out is not None else x
+
+
+@triton.heuristics({
+    "RECOMPUTE_OUTPUT": lambda args: args["Y"] is not None
+})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=1),
+        triton.Config({}, num_warps=2),
+        triton.Config({}, num_warps=4),
+        triton.Config({}, num_warps=8),
+        triton.Config({}, num_warps=16),
+        triton.Config({}, num_warps=32),
+    ],
+    key=["N", "HAS_DRESIDUAL", "STORE_DRESIDUAL", "IS_RMS_NORM", "HAS_BIAS"],
+)
+@triton.jit
+def layer_norm_bwd_kernel(
+    X,  # pointer to the input
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    Y,  # pointer to the output to be recomputed
+    DY,  # pointer to the output gradient
+    DX,  # pointer to the input gradient
+    DW,  # pointer to the partial sum of weights gradient
+    DB,  # pointer to the partial sum of biases gradient
+    DRESIDUAL,
+    DRESIDUAL_IN,
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    stride_x_row,  # how much to increase the pointer when moving by 1 row
+    stride_y_row,
+    stride_dy_row,
+    stride_dx_row,
+    stride_dres_row,
+    stride_dres_in_row,
+    M,  # number of rows in X
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    rows_per_program,
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_DRESIDUAL: tl.constexpr,
+    STORE_DRESIDUAL: tl.constexpr,
+    HAS_WEIGHT: tl.constexpr,
+    HAS_BIAS: tl.constexpr,
+    RECOMPUTE_OUTPUT: tl.constexpr,
+):
+    # Map the program id to the elements of X, DX, and DY it should compute.
+    row_block_id = tl.program_id(0)
+    row_start = row_block_id * rows_per_program
+    cols = tl.arange(0, BLOCK_N)
+    mask = cols < N
+    X += row_start * stride_x_row
+    if HAS_DRESIDUAL:
+        DRESIDUAL += row_start * stride_dres_row
+    if STORE_DRESIDUAL:
+        DRESIDUAL_IN += row_start * stride_dres_in_row
+    DY += row_start * stride_dy_row
+    DX += row_start * stride_dx_row
+    if RECOMPUTE_OUTPUT:
+        Y += row_start * stride_y_row
+    if HAS_WEIGHT:
+        w = tl.load(W + cols, mask=mask).to(tl.float32)
+        dw = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    if RECOMPUTE_OUTPUT and HAS_BIAS:
+        b = tl.load(B + cols, mask=mask, other=0.0).to(tl.float32)
+    if HAS_BIAS:
+        db = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    row_end = min((row_block_id + 1) * rows_per_program, M)
+    for row in range(row_start, row_end):
+        # Load data to SRAM
+        x = tl.load(X + cols, mask=mask, other=0).to(tl.float32)
+        dy = tl.load(DY + cols, mask=mask, other=0).to(tl.float32)
+        if not IS_RMS_NORM:
+            mean = tl.load(Mean + row)
+        rstd = tl.load(Rstd + row)
+        # Compute dx
+        xhat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+        xhat = tl.where(mask, xhat, 0.0)
+        if RECOMPUTE_OUTPUT:
+            y = xhat * w if HAS_WEIGHT else xhat
+            if HAS_BIAS:
+                y = y + b
+
+            # Aply quantization to the output
+            scale = 127.0 / tl.maximum(tl.max(tl.abs(y), 0), 1e-5)
+            # Quantize and then de-quantize the tensor
+            y = tl.extra.cuda.libdevice.round(y * scale)
+            y = tl.maximum(tl.minimum(y, 127), -128) / scale
+
+            tl.store(Y + cols, y, mask=mask)
+        wdy = dy
+        if HAS_WEIGHT:
+            wdy = dy * w
+            dw += dy * xhat
+        if HAS_BIAS:
+            db += dy
+        if not IS_RMS_NORM:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            c2 = tl.sum(wdy, axis=0) / N
+            dx = (wdy - (xhat * c1 + c2)) * rstd
+        else:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            dx = (wdy - xhat * c1) * rstd
+        if HAS_DRESIDUAL:
+            dres = tl.load(DRESIDUAL + cols, mask=mask, other=0).to(tl.float32)
+            dx += dres
+        # Write dx
+        if STORE_DRESIDUAL:
+            tl.store(DRESIDUAL_IN + cols, dx, mask=mask)
+        tl.store(DX + cols, dx, mask=mask)
+
+        X += stride_x_row
+        if HAS_DRESIDUAL:
+            DRESIDUAL += stride_dres_row
+        if STORE_DRESIDUAL:
+            DRESIDUAL_IN += stride_dres_in_row
+        if RECOMPUTE_OUTPUT:
+            Y += stride_y_row
+        DY += stride_dy_row
+        DX += stride_dx_row
+    if HAS_WEIGHT:
+        tl.store(DW + row_block_id * N + cols, dw, mask=mask)
+    if HAS_BIAS:
+        tl.store(DB + row_block_id * N + cols, db, mask=mask)
+
+
+def layer_norm_bwd(
+    dy: torch.Tensor,
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float,
+    mean: torch.Tensor,
+    rstd: torch.Tensor,
+    dresidual: torch.Tensor = None,
+    has_residual: bool = False,
+    is_rms_norm: bool = False,
+    x_dtype: torch.dtype = None,
+    recompute_output: bool = False,
+):
+    M, N = x.shape
+    # allocate output
+    dx = torch.empty_like(x) if x_dtype is None else torch.empty(M, N, dtype=x_dtype, device=x.device)
+    dresidual_in = torch.empty_like(x) if has_residual and dx.dtype != x.dtype else None
+    y = torch.empty(M, N, dtype=dy.dtype, device=dy.device) if recompute_output else None
+
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    sm_count = get_multiprocessor_count(x.device.index)
+    _dw = torch.empty((sm_count, N), dtype=torch.float32, device=weight.device) if weight is not None else None
+    _db = torch.empty((sm_count, N), dtype=torch.float32, device=bias.device) if bias is not None else None
+    rows_per_program = math.ceil(M / sm_count)
+    grid = (sm_count,)
+    layer_norm_bwd_kernel[grid](
+        x,
+        weight,
+        bias,
+        y,
+        dy,
+        dx,
+        _dw,
+        _db,
+        dresidual,
+        dresidual_in,
+        mean,
+        rstd,
+        x.stride(0),
+        0 if not recompute_output else y.stride(0),
+        dy.stride(0),
+        dx.stride(0),
+        dresidual.stride(0) if dresidual is not None else 0,
+        dresidual_in.stride(0) if dresidual_in is not None else 0,
+        M,
+        N,
+        eps,
+        rows_per_program,
+        is_rms_norm,
+        BLOCK_N,
+        dresidual is not None,
+        dresidual_in is not None,
+        weight is not None,
+        bias is not None,
+    )
+    dw = _dw.sum(0).to(weight.dtype) if weight is not None else None
+    db = _db.sum(0).to(bias.dtype) if bias is not None else None
+    # Don't need to compute dresidual_in separately in this case
+    if has_residual and dx.dtype == x.dtype:
+        dresidual_in = dx
+    return (dx, dw, db, dresidual_in) if not recompute_output else (dx, dw, db, dresidual_in, y)
+
+
+class LayerNormLinearQuantFn(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual=None,
+        eps=1e-6,
+        prenorm=False,
+        residual_in_fp32=False,
+        is_rms_norm=False,
+    ):
+        x_shape_og = x.shape
+        # reshape input data into 2D tensor
+        x = x.reshape(-1, x.shape[-1])
+        if residual is not None:
+            assert residual.shape == x_shape_og
+            residual = residual.reshape(-1, residual.shape[-1])
+        residual_dtype = residual.dtype if residual is not None else (torch.float32 if residual_in_fp32 else None)
+        y, mean, rstd, residual_out = layer_norm_fwd_quant(
+            x,
+            norm_weight,
+            norm_bias,
+            eps,
+            residual,
+            out_dtype=None if not torch.is_autocast_enabled() else torch.get_autocast_gpu_dtype(),
+            residual_dtype=residual_dtype,
+            is_rms_norm=is_rms_norm,
+        )
+        y = y.reshape(x_shape_og)
+        dtype = torch.get_autocast_gpu_dtype() if torch.is_autocast_enabled() else y.dtype
+        linear_weight = weight_quant(linear_weight).to(dtype)
+        linear_bias = linear_bias.to(dtype) if linear_bias is not None else None
+        out = F.linear(y.to(linear_weight.dtype), linear_weight, linear_bias)
+        # We don't store y, will be recomputed in the backward pass to save memory
+        ctx.save_for_backward(residual_out, norm_weight, norm_bias, linear_weight, mean, rstd)
+        ctx.x_shape_og = x_shape_og
+        ctx.eps = eps
+        ctx.is_rms_norm = is_rms_norm
+        ctx.has_residual = residual is not None
+        ctx.prenorm = prenorm
+        ctx.x_dtype = x.dtype
+        ctx.linear_bias_is_none = linear_bias is None
+        return out if not prenorm else (out, residual_out.reshape(x_shape_og))
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, dout, *args):
+        x, norm_weight, norm_bias, linear_weight, mean, rstd = ctx.saved_tensors
+        dout = dout.reshape(-1, dout.shape[-1])
+        dy = F.linear(dout, linear_weight.t())
+        dlinear_bias = None if ctx.linear_bias_is_none else dout.sum(0)
+        assert dy.shape == x.shape
+        if ctx.prenorm:
+            dresidual = args[0]
+            dresidual = dresidual.reshape(-1, dresidual.shape[-1])
+            assert dresidual.shape == x.shape
+        else:
+            dresidual = None
+        dx, dnorm_weight, dnorm_bias, dresidual_in, y = layer_norm_bwd(
+            dy,
+            x,
+            norm_weight,
+            norm_bias,
+            ctx.eps,
+            mean,
+            rstd,
+            dresidual,
+            ctx.has_residual,
+            ctx.is_rms_norm,
+            x_dtype=ctx.x_dtype,
+            recompute_output=True
+        )
+        dlinear_weight = torch.einsum("bo,bi->oi", dout, y)
+        return (
+            dx.reshape(ctx.x_shape_og),
+            dnorm_weight,
+            dnorm_bias,
+            dlinear_weight,
+            dlinear_bias,
+            dresidual_in.reshape(ctx.x_shape_og) if ctx.has_residual else None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+def layer_norm_linear_quant_fn(
+    x,
+    norm_weight,
+    norm_bias,
+    linear_weight,
+    linear_bias,
+    residual=None,
+    eps=1e-6,
+    prenorm=False,
+    residual_in_fp32=False,
+    is_rms_norm=False,
+):
+    return LayerNormLinearQuantFn.apply(
+        x,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        is_rms_norm,
+    )
+
+
+def rms_norm_linear_quant(
+    x: torch.Tensor,
+    norm_weight: torch.Tensor,
+    norm_bias: torch.Tensor,
+    linear_weight: torch.Tensor,
+    linear_bias: torch.Tensor,
+    residual: torch.Tensor = None,
+    eps: float = 1e-5,
+    prenorm: bool = False,
+    residual_in_fp32: bool = False
+):
+    return layer_norm_linear_quant_fn(
+        x=x,
+        norm_weight=norm_weight,
+        norm_bias=norm_bias,
+        linear_weight=linear_weight,
+        linear_bias=linear_bias,
+        residual=residual,
+        eps=eps,
+        prenorm=prenorm,
+        residual_in_fp32=residual_in_fp32,
+        is_rms_norm=True
+    )
+
+
+@require_version("triton>=3.0", "Triton >= 3.0 is required to do online quantization.")
+def bit_linear(x, weight, bias=None, norm_weight=None, norm_bias=None, eps=1e-8):
+    """
+    A functional version of BitLinear that applies quantization to activations and weights.
+
+    Args:
+        x: Input tensor with shape [n, d].
+        weight: Weight tensor with shape [out_features, in_features].
+        bias: Bias tensor with shape [out_features] (optional).
+        norm_weight: Weight tensor for RMS normalization with shape [in_features].
+        norm_bias: Bias tensor for RMS normalization with shape [in_features].
+        eps: A small constant for numerical stability in normalization.
+
+    Returns:
+        Output tensor with shape [n, out_features].
+    """
+    return layer_norm_linear_quant_fn(
+        x,
+        norm_weight,
+        norm_bias,
+        weight,
+        bias,
+        is_rms_norm=True
+    )
+
+
+class BitLinear(nn.Linear):
+    """
+    A custom linear layer that applies quantization on both activations and weights.
+    This is primarily for training; kernel optimization is needed for efficiency in deployment.
+    """
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        bias: bool = False,
+        norm_eps: float = 1e-8
+    ):
+        """
+        Initializes the BitLinear layer.
+
+        Args:
+            in_features: Size of each input sample.
+            out_features: Size of each output sample.
+            bias: If set to False, the layer will not learn an additive bias. Default: True.
+        """
+        # Initialize the superclass nn.Linear with the given parameters
+        super(BitLinear, self).__init__(in_features, out_features, bias=bias)
+
+        self.norm = RMSNorm(in_features, eps=norm_eps)
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}({super().extra_repr()}, norm_eps={self.norm.eps})"
+
+    def forward(self, x):
+        """
+        Overrides the forward pass to include quantization.
+
+        Args:
+            x: An input tensor with shape [n, d].
+
+        Returns:
+            An output tensor with shape [n, d].
+        """
+        # Weight tensor
+        w = self.weight
+
+        # Apply RMS normalization to the input
+        x_norm = self.norm(x)
+
+        # Apply quantization to both activations and weights
+        # Uses Straight-Through Estimator (STE) trick with .detach() for gradient flow
+        x_quant = x_norm + (activation_quant(x_norm) - x_norm).detach()
+        w_quant = w + (weight_quant(w) - w).detach()
+        # Perform linear operation with quantized values
+        y = F.linear(x_quant, w_quant)
+
+        return y
+
+
+class FusedBitLinear(BitLinear):
+    """
+    A custom linear layer that applies quantization on both activations and weights.
+    This is primarily for training; kernel optimization is needed for efficiency in deployment.
+    """
+
+    def __init__(self, in_features, out_features, bias=False):
+        """
+        Initializes the BitLinear layer.
+
+        Args:
+            in_features: Size of each input sample.
+            out_features: Size of each output sample.
+            bias: If set to False, the layer will not learn an additive bias. Default: True.
+        """
+        # Initialize the superclass nn.Linear with the given parameters
+        super(FusedBitLinear, self).__init__(in_features, out_features, bias=bias)
+
+    def forward(self, x):
+        return layer_norm_linear_quant_fn(
+            x,
+            self.norm.weight,
+            self.norm.bias,
+            self.weight,
+            self.bias,
+            is_rms_norm=True
+        )

fla/modules/fused_cross_entropy.py

@@ -0,0 +1,419 @@
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2023, Tri Dao.
+
+from typing import Any, Tuple
+
+import torch
+import torch.nn as nn
+import triton
+import triton.language as tl
+
+from fla.ops.utils.op import exp, log
+from fla.utils import input_guard
+
+# `all_gather_into_tensor` and `reduce_scatter_tensor` are new placeholders for
+# `_all_gather_base` and `_reduce_scatter_base`. They require the most recent
+# version of PyTorch. The following 2 lines are for backward compatibility with
+# older PyTorch.
+if "all_gather_into_tensor" not in dir(torch.distributed):
+    torch.distributed.all_gather_into_tensor = torch.distributed._all_gather_base
+
+
+@triton.heuristics({
+    "HAS_SMOOTHING": lambda args: args["label_smoothing"] > 0.0,
+})
+@triton.jit
+def cross_entropy_fwd_kernel(
+    loss_ptr,  # data ptrs
+    lse_ptr,
+    z_loss_ptr,
+    logits_ptr,
+    labels_ptr,
+    label_smoothing,
+    logit_scale,
+    lse_square_scale,
+    ignore_index,
+    total_classes,
+    class_start_idx,  # Useful for tensor parallel when each rank only has a subset of classes
+    n_cols,  # shapes
+    n_rows,
+    logits_row_stride,  # strides
+    BLOCK_SIZE: tl.constexpr,
+    HAS_SMOOTHING: tl.constexpr,
+    # if SPLIT (e.g. tensor parallel), don't include the LSE in the loss since it's not the final LSE
+    SPLIT: tl.constexpr,
+):
+    row_idx = tl.program_id(0)
+    col_block_idx = tl.program_id(1)
+    logits_ptr = logits_ptr + row_idx * logits_row_stride.to(tl.int64)
+    col_offsets = col_block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    label_idx = tl.load(labels_ptr + row_idx)
+    logits = tl.load(logits_ptr + col_offsets, mask=col_offsets < n_cols, other=-float("inf"))
+    logits = logits.to(tl.float32) * logit_scale
+    max_logits = tl.max(logits, 0)
+    if HAS_SMOOTHING:
+        sum_logits = tl.sum(tl.where(col_offsets < n_cols, logits, 0.0), 0)
+    lse = log(tl.sum(exp(logits - max_logits), 0)) + max_logits
+    tl.store(lse_ptr + col_block_idx * n_rows + row_idx, lse)
+    if label_idx == ignore_index:
+        loss = 0.0
+        z_loss = 0.0
+    else:
+        label_idx -= class_start_idx
+        if label_idx >= col_block_idx * BLOCK_SIZE and label_idx < min(
+            n_cols, (col_block_idx + 1) * BLOCK_SIZE
+        ):
+            logits_label = tl.load(logits_ptr + label_idx) * logit_scale
+            if HAS_SMOOTHING:
+                loss = (
+                    (lse if not SPLIT else 0.0)
+                    - label_smoothing * sum_logits / total_classes
+                    - (1 - label_smoothing) * logits_label
+                )
+            else:
+                loss = (lse if not SPLIT else 0.0) - logits_label
+        else:
+            # If label is out of bounds, we set the CE loss to 0.0. But we still want the label_smoothing loss
+            if HAS_SMOOTHING:
+                loss = label_smoothing * ((lse if not SPLIT else 0.0) - sum_logits / total_classes)
+            else:
+                loss = 0.0
+        if not SPLIT:
+            z_loss = lse_square_scale * lse * lse
+            loss += z_loss
+        else:
+            z_loss = 0.0
+    tl.store(loss_ptr + col_block_idx * n_rows + row_idx, loss)
+    if not SPLIT:
+        tl.store(z_loss_ptr + col_block_idx * n_rows + row_idx, z_loss)
+
+
+@triton.heuristics({
+    "HAS_SMOOTHING": lambda args: args["label_smoothing"] > 0.0,
+})
+@triton.jit
+def cross_entropy_bwd_kernel(
+    dlogits_ptr,  # data ptrs
+    dloss_ptr,
+    logits_ptr,
+    lse_ptr,
+    labels_ptr,
+    label_smoothing,
+    logit_scale,
+    lse_square_scale,
+    ignore_index,
+    total_classes,
+    class_start_idx,  # Useful for tensor parallel when each rank only has a subset of classes
+    n_cols,  # shapes
+    logits_row_stride,  # strides
+    dlogits_row_stride,
+    dloss_row_stride,
+    BLOCK_SIZE: tl.constexpr,
+    HAS_SMOOTHING: tl.constexpr,
+):
+    row_idx = tl.program_id(0)
+    col_block_idx = tl.program_id(1)
+    logits_ptr = logits_ptr + row_idx * logits_row_stride.to(tl.int64)
+    dlogits_ptr = dlogits_ptr + row_idx * dlogits_row_stride.to(tl.int64)
+    col_offsets = col_block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    label_idx = tl.load(labels_ptr + row_idx)
+    if label_idx != ignore_index:
+        dloss = tl.load(dloss_ptr + row_idx * dloss_row_stride)
+    else:
+        dloss = 0.0
+    logits = tl.load(logits_ptr + col_offsets, mask=col_offsets < n_cols, other=-float("inf")).to(
+        tl.float32
+    ) * logit_scale
+    lse = tl.load(lse_ptr + row_idx)
+    probs = exp(logits - lse)
+    probs += 2.0 * lse_square_scale * lse * probs
+    label_idx -= class_start_idx
+    if HAS_SMOOTHING:
+        smooth_negative = label_smoothing / total_classes
+        probs = tl.where(col_offsets == label_idx, probs - (1 - label_smoothing), probs) - smooth_negative
+    else:
+        probs = tl.where(col_offsets == label_idx, probs - 1.0, probs)
+    tl.store(dlogits_ptr + col_offsets, (dloss * logit_scale) * probs, mask=col_offsets < n_cols)
+
+
+def fused_cross_entropy_forward(
+    logits: torch.Tensor,
+    target: torch.Tensor,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    lse_square_scale: float = 0.0,
+    ignore_index: int = -100,
+    process_group=None,
+):
+    n_rows, n_cols = logits.shape
+    assert target.shape == (n_rows,)
+    world_size = 1 if process_group is None else torch.distributed.get_world_size(process_group)
+    total_classes = world_size * n_cols
+    rank = 0 if process_group is None else torch.distributed.get_rank(process_group)
+    class_start_idx = rank * n_cols
+
+    if logits.stride(-1) != 1:
+        logits = logits.contiguous()
+    # Set these similar to https://github.com/openai/triton/blob/main/python/tutorials/02-fused-softmax.py
+    MAX_BLOCK_SIZE = 64 * 1024
+    BLOCK_SIZE = min(triton.next_power_of_2(n_cols), MAX_BLOCK_SIZE)
+    num_warps = (
+        4
+        if BLOCK_SIZE < 2048
+        else (8 if BLOCK_SIZE < 8192 else (16 if BLOCK_SIZE < 128 * 1024 else 32))
+    )
+    # We may split the lse computation across multiple blocks, then do a reduction
+    # lse(local_lse) to get the final LSE. This is faster for large n_cols (e.g., > 64k)
+    # where having just one thread block processing more than 64k elements is slow.
+    split = world_size > 1 or n_cols > MAX_BLOCK_SIZE
+    n_splits = (n_cols + BLOCK_SIZE - 1) // BLOCK_SIZE
+    loss_shape = (n_splits, n_rows) if n_splits > 1 else (n_rows,)
+    losses = torch.empty(*loss_shape, dtype=torch.float, device=logits.device)
+    lse = torch.empty(*loss_shape, dtype=torch.float, device=logits.device)
+    z_losses = torch.empty(*loss_shape, dtype=torch.float, device=logits.device)
+
+    cross_entropy_fwd_kernel[(n_rows, n_splits)](
+        losses,  # data ptrs
+        lse,
+        z_losses,
+        logits,
+        target,
+        label_smoothing,
+        logit_scale,
+        lse_square_scale,
+        ignore_index,
+        total_classes,
+        class_start_idx,
+        n_cols,  # shapes
+        n_rows,
+        logits.stride(0),  # strides
+        BLOCK_SIZE=BLOCK_SIZE,  # constants
+        num_warps=num_warps,
+        SPLIT=split
+    )
+
+    if split:
+        # If there's no label_smoothing, if target are in the vocab of this partition, losses contains
+        # - predicted logit, and 0 otherwise.
+        # If there's label_smoothing=0.1, for target in the vocab of this partition, losses contains
+        # -0.9 * predicted logit - 0.1 * sum logit / total_classes.
+        # For target not in the vocab of this partition, losses contains
+        # -0.1 * sum logit / total_classes.
+        if n_splits > 1:
+            lse = torch.logsumexp(lse, dim=0)
+            losses = losses.sum(dim=0)
+        if world_size > 1:
+            lse_allgather = torch.empty(world_size, n_rows, dtype=lse.dtype, device=lse.device)
+            torch.distributed.all_gather_into_tensor(lse_allgather, lse, group=process_group)
+            handle_losses = torch.distributed.all_reduce(
+                losses, op=torch.distributed.ReduceOp.SUM, group=process_group, async_op=True
+            )
+            lse = torch.logsumexp(lse_allgather, dim=0)
+            handle_losses.wait()
+        # After the allreduce, if there's no label_smoothing, the total losses are - predicted_logit,
+        # we just have to add the (global) lse.
+        # If there's label_smoothing=0.1, the total losses are
+        # -0.9 * predicted_logit - 0.1 * sum logit / total_classes.
+        # Again, we just have to add the (global) lse.
+        losses += lse
+        if lse_square_scale != 0.0:
+            z_losses = lse_square_scale * lse.square()
+            z_losses.masked_fill_(target == ignore_index, 0.0)
+            losses += z_losses
+        else:
+            z_losses = torch.zeros_like(losses)
+        losses.masked_fill_(target == ignore_index, 0.0)
+
+    return losses, z_losses, lse, total_classes, class_start_idx
+
+
+class CrossEntropyLossFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        logits,
+        target,
+        label_smoothing=0.0,
+        logit_scale=1.0,
+        lse_square_scale=0.0,
+        ignore_index=-100,
+        inplace_backward=False,
+        process_group=None,
+    ):
+        losses, z_losses, lse, total_classes, class_start_idx = fused_cross_entropy_forward(
+            logits,
+            target,
+            label_smoothing,
+            logit_scale,
+            lse_square_scale,
+            ignore_index,
+            process_group,
+        )
+        ctx.save_for_backward(logits, lse, target)
+        ctx.mark_non_differentiable(z_losses)
+        ctx.label_smoothing = label_smoothing
+        ctx.logit_scale = logit_scale
+        ctx.lse_square_scale = lse_square_scale
+        ctx.ignore_index = ignore_index
+        ctx.total_classes = total_classes
+        ctx.class_start_idx = class_start_idx
+        ctx.inplace_backward = inplace_backward
+
+        return losses, z_losses
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, grad_losses, grad_z_losses):
+        del grad_z_losses  # z_losses are only for logging.
+
+        logits, lse, target = ctx.saved_tensors
+        dlogits = logits if ctx.inplace_backward else torch.empty_like(logits)
+        n_rows, n_cols = logits.shape
+        BLOCK_SIZE = min(triton.next_power_of_2(n_cols), 4 * 1024)
+        num_warps = 4 if BLOCK_SIZE < 2048 else (8 if BLOCK_SIZE < 8192 else 16)
+        def grid(META): return (n_rows, triton.cdiv(n_cols, META["BLOCK_SIZE"]))  # noqa
+        cross_entropy_bwd_kernel[grid](
+            dlogits,  # data ptrs
+            grad_losses,
+            logits,
+            lse,
+            target,
+            ctx.label_smoothing,
+            ctx.logit_scale,
+            ctx.lse_square_scale,
+            ctx.ignore_index,
+            ctx.total_classes,
+            ctx.class_start_idx,
+            n_cols,  # shapes
+            logits.stride(0),  # strides
+            dlogits.stride(0),
+            grad_losses.stride(0),
+            BLOCK_SIZE=BLOCK_SIZE,  # constants
+            num_warps=num_warps,
+        )
+        return dlogits, None, None, None, None, None, None, None, None
+
+
+def cross_entropy_loss(
+    logits: torch.Tensor,
+    target: torch.Tensor,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    lse_square_scale: float = 0.0,
+    ignore_index=-100,
+    inplace_backward: bool = False,
+    process_group=None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Arguments:
+        logits: [batch, vocab_size]
+        target: [batch,]
+        label_smoothing: float
+        logit_scale: float.
+            Multiply logits by this scale before calculating the loss.
+        lse_square_scale: float.
+            If > 0, we add lse_square_scale * lse(logits) ^ 2 to the loss.
+            This is also referred to as "z-loss".
+        ignore_index: int.
+            If target == ignore_index, the loss is set to 0.0.
+        inplace_backward: bool.
+            If True, we do the backward pass in-place by modifying the logits.
+            This saves memory.
+        process_group:
+            if not None, we're doing Tensor Parallel: each process is responsible for
+            one part of the vocab. The loss will be aggregated across processes.
+    Returns:
+        losses: [batch,], float
+        z_losses: [batch,], float
+    """
+    return CrossEntropyLossFunction.apply(
+        logits,
+        target,
+        label_smoothing,
+        logit_scale,
+        lse_square_scale,
+        ignore_index,
+        inplace_backward,
+        process_group,
+    )
+
+
+class FusedCrossEntropyLoss(nn.Module):
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        reduction: str = "mean",
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        lse_square_scale: float = 0.0,
+        inplace_backward: bool = False,
+        process_group: Any = None,
+        return_z_loss: bool = False,
+    ):
+        """
+        Arguments:
+            ignore_index: int. If target == ignore_index, the loss is set to 0.0.
+            label_smoothing: float
+            lse_square_scale: float. If > 0, we add lse_square_scale * lse(logits) ^ 2 to the loss.
+                This is also referred to as "z-loss".
+            inplace_backward: bool. If True, we do the backward pass in-place by modifying the logits.
+                This saves memory.
+            process_group: if not None, we're doing Tensor Parallel: each process is responsible for
+                one part of the vocab. The loss will be aggregated across processes.
+            return_z_loss: bool. If True, we return the component of the loss contributed by
+                the lse_square_scale value. This value is only for logging and does not support
+                backprop.
+        """
+        super().__init__()
+        if reduction not in ["mean", "none", "sum"]:
+            raise NotImplementedError("Only support reduction = 'mean' or 'none' or 'sum'")
+        self.ignore_index = ignore_index
+        self.reduction = reduction
+        self.label_smoothing = label_smoothing
+        self.logit_scale = logit_scale
+        self.lse_square_scale = lse_square_scale
+        self.inplace_backward = inplace_backward
+        self.process_group = process_group
+        self.return_z_loss = return_z_loss
+
+    def forward(self, input, target):
+        """
+        Arguments:
+            input: (batch, vocab_size)
+            target: (batch,)
+        Returns:
+            losses: (batch,) if reduction is 'none', else (1,), dtype float
+            z_loss: (batch,) if reduction is 'none', else (1,), dtype float (if self.return_z_loss)
+        """
+        assert input.is_cuda and target.is_cuda, "Only support CUDA tensors"
+        loss, z_loss = cross_entropy_loss(
+            input,
+            target,
+            label_smoothing=self.label_smoothing,
+            logit_scale=self.logit_scale,
+            lse_square_scale=self.lse_square_scale,
+            ignore_index=self.ignore_index,
+            inplace_backward=self.inplace_backward,
+            process_group=self.process_group,
+        )
+        if self.reduction == "mean":
+            loss = loss.sum() / (target != self.ignore_index).sum()
+        elif self.reduction == "sum":
+            loss = loss.sum()
+        else:
+            loss = loss
+
+        if not self.return_z_loss:
+            return loss
+
+        if self.reduction == "mean":
+            z_loss = z_loss.sum() / (target != self.ignore_index).sum()
+        elif self.reduction == "sum":
+            z_loss = z_loss.sum()
+        else:
+            z_loss = z_loss
+
+        return loss, z_loss

fla/modules/fused_kl_div.py

@@ -0,0 +1,323 @@
+# -*- coding: utf-8 -*-
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+
+from fla.ops.utils.op import exp, log
+from fla.utils import input_guard
+
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576
+# https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+MAX_FUSED_SIZE = 65536 // 2
+
+
+@triton.jit
+def kl_div_kernel(
+    logits,
+    target_logits,
+    loss,
+    s_logits,
+    s_loss,
+    reduction: tl.constexpr,
+    N: tl.constexpr,
+    V: tl.constexpr,
+    BV: tl.constexpr
+):
+    # https://github.com/triton-lang/triton/issues/1058
+    # If N*V is too large, i_n * stride will overflow out of int32, so we convert to int64
+    i_n = tl.program_id(0).to(tl.int64)
+
+    logits += i_n * s_logits
+    target_logits += i_n * s_logits
+
+    # m is the max value. use the notation from the paper
+    sm = float('-inf')
+    tm = float('-inf')
+    # d is the sum. use the notation from the paper
+    sd, td = 0.0, 0.0
+
+    NV = tl.cdiv(V, BV)
+    for iv in range(0, NV):
+        o_x = iv * BV + tl.arange(0, BV)
+        # for student
+        b_sl = tl.load(logits + o_x, mask=o_x < V, other=float('-inf'))
+        b_sm = tl.max(b_sl)
+        m_new = tl.maximum(sm, b_sm)
+        sd = sd * exp(sm - m_new) + tl.sum(exp(b_sl - m_new))
+        sm = m_new
+        # for teacher
+        b_tl = tl.load(target_logits + o_x, mask=o_x < V, other=float('-inf'))
+        b_tm = tl.max(b_tl)
+        m_new = tl.maximum(tm, b_tm)
+        td = td * exp(tm - m_new) + tl.sum(exp(b_tl - m_new))
+        tm = m_new
+
+    b_loss = 0.
+    # KL(y_true || y) = exp(y_true) * (log(y_true) - log(y))
+    for iv in range(0, NV):
+        o_x = iv * BV + tl.arange(0, BV)
+        b_sl = tl.load(logits + o_x, mask=o_x < V, other=float('-inf'))
+        b_tl = tl.load(target_logits + o_x, mask=o_x < V, other=float('-inf'))
+        b_sp_log = b_sl - sm - log(sd)
+        b_tp_log = b_tl - tm - log(td)
+        b_sp = exp(b_sp_log)
+        b_tp = exp(b_tp_log)
+        b_kl = tl.where(o_x < V, b_tp * (b_tp_log - b_sp_log), 0)
+        b_dl = -b_tp + b_sp
+        b_loss += tl.sum(b_kl)
+        if reduction == 'batchmean':
+            b_dl = b_dl / N
+        tl.store(logits + o_x, b_dl, mask=o_x < V)
+
+    # Normalize the loss by the number of elements if reduction is 'batchmean'
+    if reduction == 'batchmean':
+        b_loss = b_loss / N
+
+    tl.store(loss + i_n * s_loss, b_loss)
+
+
+@triton.jit
+def elementwise_mul_kernel(
+    x,
+    g,
+    N: tl.constexpr,
+    B: tl.constexpr
+):
+    """
+    This function multiplies each element of the tensor pointed by x with the value pointed by g.
+    The multiplication is performed in-place on the tensor pointed by x.
+
+    Parameters:
+    x:
+        Pointer to the input tensor.
+    g:
+        Pointer to the gradient output value.
+    N (int):
+        The number of columns in the input tensor.
+    B (int):
+        The block size for Triton operations.
+    """
+
+    # Get the program ID and convert it to int64 to avoid overflow
+    i_x = tl.program_id(0).to(tl.int64)
+    o_x = i_x * B + tl.arange(0, B)
+
+    # Load the gradient output value
+    b_g = tl.load(g)
+    b_x = tl.load(x + o_x, mask=o_x < N)
+    tl.store(x + o_x, b_x * b_g, mask=o_x < N)
+
+
+def fused_kl_div_forward(
+    x: torch.Tensor,
+    target_x: torch.Tensor,
+    weight: torch.Tensor,
+    target_weight: torch.Tensor,
+    reduction: str = 'batchmean'
+):
+    device = x.device
+
+    # ideally, we would like to achieve the same memory consumption as [N, H],
+    # so the expected chunk size should be:
+    # NC = ceil(V / H)
+    # C = ceil(N / NC)
+    # for ex: N = 4096*4, V = 32000, H = 4096 ==> NC = 8, C = ceil(N / NC) = 2048
+    N, H, V = *x.shape, weight.shape[0]
+    BV = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    # TODO: in real cases, we may need to limit the number of chunks NC to
+    # ensure the precisions of accumulated gradients
+    NC = min(8, triton.cdiv(V, H))
+    C = triton.next_power_of_2(triton.cdiv(N, NC))
+    NC = triton.cdiv(N, C)
+
+    dx = torch.zeros_like(x, device=device)
+    dw = torch.zeros_like(weight, device=device) if weight is not None else None
+    # we use fp32 for loss accumulator
+    loss = torch.zeros(N, dtype=torch.float32, device=device)
+
+    for ic in range(NC):
+        start, end = ic * C, min((ic + 1) * C, N)
+        # [C, N]
+        c_sx = x[start:end]
+        c_tx = target_x[start:end]
+        # when doing matmul, use the original precision
+        # [C, V]
+        c_sl = F.linear(c_sx, weight)
+        c_tl = F.linear(c_tx, target_weight)
+
+        # unreduced loss
+        c_loss = loss[start:end]
+
+        # Here we calculate the gradient of c_sx in place so we can save memory.
+        kl_div_kernel[(c_sx.shape[0],)](
+            logits=c_sl,
+            target_logits=c_tl,
+            loss=c_loss,
+            s_logits=c_sl.stride(-2),
+            s_loss=c_loss.stride(-1),
+            reduction=reduction,
+            N=N,
+            V=V,
+            BV=BV,
+            num_warps=32
+        )
+
+        # gradient of logits is computed in-place by the above triton kernel and is of shape: C x V
+        # thus dx[start: end] should be of shape: C x H
+        # additionally, since we are chunking the inputs, observe that the loss and gradients are calculated only
+        # on `n_non_ignore` tokens. However, the gradient of the input should be calculated for all tokens.
+        # Thus, we need an additional scaling factor of (n_non_ignore/total) to scale the gradients.
+        # [C, H]
+
+        dx[start:end] = torch.mm(c_sl, weight)
+
+        if weight is not None:
+            torch.addmm(input=dw, mat1=c_sl.t(), mat2=c_sx, out=dw)
+
+    loss = loss.sum()
+    return loss, dx, dw
+
+
+def fused_kl_div_backward(
+    do: torch.Tensor,
+    dx: torch.Tensor,
+    dw: torch.Tensor
+):
+    # If cross entropy is the last layer, do is 1.0. Skip the mul to save time
+    if torch.ne(do, torch.tensor(1.0, device=do.device)):
+        # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+        # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+        N, H = dx.shape
+        B = min(MAX_FUSED_SIZE, triton.next_power_of_2(H))
+
+        elementwise_mul_kernel[(triton.cdiv(N * H, B),)](
+            x=dx,
+            g=do,
+            N=N*H,
+            B=B,
+            num_warps=32,
+        )
+
+        # handle dw
+        if dw is not None:
+            V, H = dw.shape
+            elementwise_mul_kernel[(triton.cdiv(V * H, B),)](
+                x=dw,
+                g=do,
+                N=V*H,
+                B=B,
+                num_warps=32,
+            )
+
+    return dx, dw
+
+
+class FusedKLDivLossFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        target_x: torch.Tensor,
+        weight: torch.Tensor,
+        target_weight: torch.Tensor,
+        reduction: str
+    ):
+        loss, dx, dw = fused_kl_div_forward(
+            x=x,
+            target_x=target_x,
+            weight=weight,
+            target_weight=target_weight,
+            reduction=reduction
+        )
+        ctx.save_for_backward(dx, dw)
+        return loss
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, do):
+        dx, dw = ctx.saved_tensors
+        dx, dw = fused_kl_div_backward(do, dx, dw)
+        return dx, None, dw, None, None
+
+
+def fused_kl_div_loss(
+    x: torch.Tensor,
+    target_x: torch.Tensor,
+    weight: torch.Tensor,
+    target_weight: torch.Tensor,
+    reduction: str = 'batchmean'
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target_x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        target_weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        reduction:
+            Specifies the reduction to apply to the output: 'batchmean'. Default: 'batchmean'.
+    Returns:
+        loss
+    """
+    return FusedKLDivLossFunction.apply(
+        x,
+        target_x,
+        weight,
+        target_weight,
+        reduction
+    )
+
+
+class FusedKLDivLoss(nn.Module):
+
+    def __init__(
+        self,
+        reduction: str = 'batchmean'
+    ):
+        """
+        Args:
+            reduction:
+                Specifies the reduction to apply to the output: 'batchmean'. Default: 'batchmean'.
+        """
+        super().__init__()
+
+        assert reduction in ['batchmean'], f"reduction: {reduction} is not supported"
+
+        self.reduction = reduction
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        target_x: torch.Tensor,
+        weight: torch.Tensor,
+        target_weight: torch.Tensor
+    ):
+        """
+        Args:
+            x (torch.Tensor): [batch_size * seq_len, hidden_size]
+            target_x (torch.Tensor): [batch_size * seq_len, hidden_size]
+            weight (torch.Tensor): [vocab_size, hidden_size]
+                where `vocab_size` is the number of classes.
+            target_weight (torch.Tensor): [vocab_size, hidden_size]
+                where `vocab_size` is the number of classes.
+        Returns:
+            loss
+        """
+        loss = fused_kl_div_loss(
+            x=x,
+            target_x=target_x,
+            weight=weight,
+            target_weight=target_weight,
+            reduction=self.reduction
+        )
+        return loss

fla/modules/fused_linear_cross_entropy.py

@@ -0,0 +1,570 @@
+# -*- coding: utf-8 -*-
+
+# Code adapted from
+# https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/fused_linear_cross_entropy.py
+
+from functools import partial
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+from torch.distributed import DeviceMesh
+from torch.distributed.tensor import DTensor, Replicate, Shard, distribute_module
+from torch.distributed.tensor.parallel import ParallelStyle
+
+from fla.ops.utils import logsumexp_fwd
+from fla.ops.utils.op import exp
+from fla.utils import input_guard
+
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576
+# https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+MAX_FUSED_SIZE = 65536 // 2
+
+
+@triton.jit
+def cross_entropy_kernel(
+    logits,
+    lse,
+    target,
+    loss,
+    total,
+    ignore_index,
+    label_smoothing: tl.constexpr,
+    logit_scale: tl.constexpr,
+    reduction: tl.constexpr,
+    V: tl.constexpr,
+    BV: tl.constexpr
+):
+    """
+    This kernel computes both cross entropy loss and the gradient of the input.
+    We only consider hard label + mean reduction for now.
+    Please refer to https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html for the math.
+
+    Args:
+        logits:
+            Pointer to logits tensor.
+        lse:
+            Pointer to logsumexp tensor.
+        target: Pointer to target tensor.
+        loss:
+            Pointer to tensor to store the loss.
+        V (int):
+            The number of columns in the input tensor.
+        total (int):
+            The number of non-ignored classes.
+        ignore_index (int):
+            The index to ignore in the target.
+        label_smoothing (float):
+            The amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        reduction (str):
+            The string for the reduction to apply
+        BV (int):
+            The block size for vocab.
+    """
+
+    # https://github.com/triton-lang/triton/issues/1058
+    # If B*T*V is too large, i_n * stride will overflow out of int32, so we convert to int64
+    i_n = tl.program_id(0).to(tl.int64)
+    NV = tl.cdiv(V, BV)
+
+    # 1. Load target first because if the target is ignore_index, we can return right away
+    b_y = tl.load(target + i_n)
+
+    # 2. locate the start index
+    logits += i_n * V
+
+    if b_y == ignore_index:
+        # set all x as 0
+        for i in range(0, V, BV):
+            o_v = i + tl.arange(0, BV)
+            tl.store(logits + o_v, 0.0, mask=o_v < V)
+        return
+
+    # Online softmax: 2 loads + 1 store (compared with 3 loads + 1 store for the safe softmax)
+    # Refer to Algorithm 3 in the paper: https://arxiv.org/pdf/1805.02867
+
+    # 3. [Online softmax] first pass: compute logsumexp
+    # we did this in anouter kernel
+    b_l = tl.load(logits + b_y) * logit_scale
+    b_lse = tl.load(lse + i_n)
+
+    # 4. Calculate the loss
+    # loss = lse - logits_l
+    b_loss = b_lse - b_l
+
+    # Label smoothing is a general case of normal cross entropy
+    # See the full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issue-2503665310
+    b_z = 0.0
+    eps = label_smoothing / V
+
+    # We need tl.debug_barrier() as mentioned in
+    # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
+    tl.debug_barrier()
+
+    # 5. [Online Softmax] Second pass: compute gradients
+    # For 'mean' reduction, gradients are normalized by number of non-ignored elements
+    # dx_y = (softmax(x_y) - 1) / N
+    # dx_i = softmax(x_i) / N, i != y
+    # For label smoothing:
+    # dx_i = (softmax(x_y) - label_smoothing / V) / N, i != y
+    # dx_y = (softmax(x_y) - label_smoothing / V - (1 - label_smoothing)) / N
+    #      = dx_i - (1 - label_smoothing) / N
+    for iv in range(0, NV):
+        o_v = iv * BV + tl.arange(0, BV)
+        b_logits = tl.load(logits + o_v, mask=o_v < V, other=float('-inf')) * logit_scale
+        if label_smoothing > 0:
+            # scale X beforehand to avoid overflow
+            b_z += tl.sum(tl.where(o_v < V, -eps * b_logits, 0.0))
+        b_p = (exp(b_logits - b_lse) - eps) * logit_scale
+        if reduction == "mean":
+            b_p = b_p / total
+        tl.store(logits + o_v, b_p, mask=o_v < V)
+
+        tl.debug_barrier()
+
+    # Orginal loss = H(q, p),  with label smoothing regularization = H(q', p) and (label_smoothing / V) = eps
+    # H(q', p) = (1 - label_smoothing) * H(q, p) + label_smoothing * H(u, p)
+    #          = (1 - label_smoothing) * H(q, p) + eps * sum(logsoftmax(x_i))
+    # By using m (global max of xi) and d (sum of e^(xi-m)), we can simplify as:
+    #          = (1 - label_smoothing) * H(q, p) + (-sum(x_i * eps) + label_smoothing * (m + logd))
+    # Refer to H(q', p) in section 7 of the paper:
+    # https://arxiv.org/pdf/1512.00567
+    # pytorch:
+    # https://github.com/pytorch/pytorch/blob/2981534f54d49fa3a9755c9b0855e7929c2527f0/aten/src/ATen/native/LossNLL.cpp#L516
+    # See full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issuecomment-2333753087
+    if label_smoothing > 0:
+        b_loss = b_loss * (1 - label_smoothing) + (b_z + label_smoothing * b_lse)
+
+    # 6. Specially handle the i==y case where `dx_y = (softmax(x_y) - (1 - label_smoothing) / N`
+    b_l = tl.load(logits + b_y)
+
+    # Normalize the loss by the number of non-ignored elements if reduction is "mean"
+    if reduction == 'mean':
+        b_loss = b_loss / total
+        b_l += (label_smoothing - 1) / total * logit_scale
+    else:
+        b_l += (label_smoothing - 1) * logit_scale
+
+    tl.store(loss + i_n, b_loss)
+    tl.store(logits + b_y, b_l)
+
+
+@triton.jit
+def elementwise_mul_kernel(
+    x,
+    g,
+    N: tl.constexpr,
+    B: tl.constexpr
+):
+    """
+    This function multiplies each element of the tensor pointed by x with the value pointed by g.
+    The multiplication is performed in-place on the tensor pointed by x.
+
+    Parameters:
+    x:
+        Pointer to the input tensor.
+    g:
+        Pointer to the gradient output value.
+    N (int):
+        The number of columns in the input tensor.
+    B (int):
+        The block size for Triton operations.
+    """
+
+    # Get the program ID and convert it to int64 to avoid overflow
+    i_x = tl.program_id(0).to(tl.int64)
+    o_x = i_x * B + tl.arange(0, B)
+
+    # Load the gradient output value
+    b_g = tl.load(g)
+    b_x = tl.load(x + o_x, mask=o_x < N)
+    tl.store(x + o_x, b_x * b_g, mask=o_x < N)
+
+
+def fused_linear_cross_entropy_forward(
+    x: torch.Tensor,
+    target: torch.LongTensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    ignore_index: int = -100,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+):
+    device = x.device
+    # inputs have shape: [N, H]
+    # materialized activations will have shape: [N, V]
+    # the increase in memory = [N, V]
+    # reduction can be achieved by partitioning the number of tokens N into smaller chunks.
+
+    # ideally, we would like to achieve the same memory consumption as [N, H],
+    # so the expected chunk size should be:
+    # NC = ceil(V / H)
+    # C = ceil(N / NC)
+    # for ex: N = 4096*4, V = 32000, H = 4096 ==> NC = 8, C = ceil(N / NC) = 2048
+    N, H, V = *x.shape, weight.shape[0]
+    BV = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    # TODO: in real cases, we may need to limit the number of chunks NC to
+    # ensure the precisions of accumulated gradients
+    NC = min(num_chunks, triton.cdiv(V, H))
+    C = triton.next_power_of_2(triton.cdiv(N, NC))
+    NC = triton.cdiv(N, C)
+
+    # [N, H]
+    dx = torch.zeros_like(x, device=device)
+    # [V, H]
+    dw = torch.zeros_like(weight, device=device, dtype=torch.float) if weight is not None else None
+    # [V]
+    db = torch.zeros_like(bias, device=device, dtype=torch.float) if bias is not None else None
+    # [N]
+    loss = torch.zeros(N, device=device, dtype=torch.float)
+
+    total = target.ne(ignore_index).sum().item()
+
+    for ic in range(NC):
+        start, end = ic * C, min((ic + 1) * C, N)
+        # [C, N]
+        c_x = x[start:end]
+        # when doing matmul, use the original precision
+        # [C, V]
+        c_logits = F.linear(c_x, weight, bias)
+        c_target = target[start:end]
+        # [C]
+        # keep lse in fp32 to maintain precision
+        c_lse = logsumexp_fwd(c_logits, scale=logit_scale, dtype=torch.float)
+
+        # unreduced loss
+        c_loss = loss[start:end]
+
+        # Here we calculate the gradient of c_logits in place so we can save memory.
+        cross_entropy_kernel[(c_logits.shape[0],)](
+            logits=c_logits,
+            lse=c_lse,
+            target=c_target,
+            loss=c_loss,
+            total=total,
+            ignore_index=ignore_index,
+            label_smoothing=label_smoothing,
+            logit_scale=logit_scale,
+            reduction=reduction,
+            V=V,
+            BV=BV,
+            num_warps=32
+        )
+
+        # gradient of logits is computed in-place by the above triton kernel and is of shape: C x V
+        # thus dx should be of shape: C x H
+        dx[start:end] = torch.mm(c_logits, weight)
+
+        # keep dw in fp32 to maintain precision
+        if weight is not None:
+            dw += c_logits.t() @ c_x
+
+        if bias is not None:
+            torch.add(input=db, other=c_logits.sum(0), out=db)
+
+    loss = loss.sum()
+    if dw is not None:
+        dw = dw.to(weight)
+    if db is not None:
+        db = db.to(bias)
+    return loss, dx, dw, db
+
+
+def fused_linear_cross_entropy_backward(
+    do: torch.Tensor,
+    dx: torch.Tensor,
+    dw: torch.Tensor,
+    db: torch.Tensor
+):
+    # If cross entropy is the last layer, do is 1.0. Skip the mul to save time
+    if torch.ne(do, torch.tensor(1.0, device=do.device)):
+        # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+        # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+        N, H = dx.shape
+        B = min(MAX_FUSED_SIZE, triton.next_power_of_2(H))
+
+        elementwise_mul_kernel[(triton.cdiv(N * H, B),)](
+            x=dx,
+            g=do,
+            N=N*H,
+            B=B,
+            num_warps=32,
+        )
+
+        # handle dw
+        if dw is not None:
+            V, H = dw.shape
+            elementwise_mul_kernel[(triton.cdiv(V * H, B),)](
+                x=dw,
+                g=do,
+                N=V*H,
+                B=B,
+                num_warps=32,
+            )
+
+        if db is not None:
+            V = db.shape[0]
+            elementwise_mul_kernel[(triton.cdiv(V, B),)](
+                x=db,
+                g=do,
+                N=V,
+                B=B,
+                num_warps=32,
+            )
+    return dx, dw, db
+
+
+class FusedLinearCrossEntropyFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        target: torch.LongTensor,
+        weight: torch.Tensor,
+        bias: torch.Tensor = None,
+        ignore_index: int = -100,
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        """
+        Fusing the last linear layer with cross-entropy loss
+            Reference: https://github.com/mgmalek/efficient_cross_entropy
+
+        Handle the forward and backward pass of the final linear layer via cross-entropy loss by avoiding
+        the materialization of the large logits tensor. Since Cross Entropy Loss is the last layer, we can
+        compute the gradient at the forward pass. By doing so, we don't have to store the x and target
+        for the backward pass.
+
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target (torch.LongTensor): [batch_size * seq_len]
+            where each value is in [0, vocab_size).
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        bias (Optional[torch.Tensor]): [vocab_size]
+            where `vocab_size` is the number of classes.
+        ignore_index:
+            the index to ignore in the target.
+        label_smoothing:
+            the amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        logit_scale: float = 1.0,
+            A scaling factor applied to the logits. Default: 1.0
+        num_chunks: int
+            The number of chunks to split the input tensor into for processing.
+            This can help optimize memory usage and computation speed.
+            Default: 8
+        reduction:
+            Specifies the reduction to apply to the output: 'mean' | 'sum'.
+            'mean': the weighted mean of the output is taken,
+            'sum': the output will be summed.
+            Default: 'mean'.
+        """
+        loss, dx, dw, db = fused_linear_cross_entropy_forward(
+            x,
+            target,
+            weight,
+            bias,
+            ignore_index,
+            label_smoothing,
+            logit_scale,
+            num_chunks,
+            reduction
+        )
+        # downcast to dtype and store for backward
+        ctx.save_for_backward(
+            dx.detach(),
+            dw.detach() if weight is not None else None,
+            db.detach() if bias is not None else None,
+        )
+        return loss
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, do):
+        dx, dw, db = ctx.saved_tensors
+        dx, dw, db = fused_linear_cross_entropy_backward(do, dx, dw, db)
+        return dx, None, dw, db, None, None, None, None, None
+
+
+def fused_linear_cross_entropy_loss(
+    x: torch.Tensor,
+    target: torch.LongTensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    ignore_index: int = -100,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target (torch.LongTensor): [batch_size * seq_len]
+            where each value is in [0, vocab_size).
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        bias (Optional[torch.Tensor]): [vocab_size]
+            where `vocab_size` is the number of classes.
+        ignore_index: int.
+            If target == ignore_index, the loss is set to 0.0.
+        label_smoothing: float
+        logit_scale: float
+            A scaling factor applied to the logits. Default: 1.0
+        num_chunks: int
+            The number of chunks to split the input tensor into for processing.
+            This can help optimize memory usage and computation speed.
+            Default: 8
+        reduction:
+            Specifies the reduction to apply to the output: 'mean' | 'sum'.
+            'mean': the weighted mean of the output is taken,
+            'sum': the output will be summed.
+            Default: 'mean'.
+    Returns:
+        losses: [batch,], float
+    """
+    return FusedLinearCrossEntropyFunction.apply(
+        x,
+        target,
+        weight,
+        bias,
+        ignore_index,
+        label_smoothing,
+        logit_scale,
+        num_chunks,
+        reduction
+    )
+
+
+class FusedLinearCrossEntropyLoss(nn.Module):
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        """
+        Args:
+            ignore_index: int.
+                If target == ignore_index, the loss is set to 0.0.
+            label_smoothing: float
+            logit_scale: float
+                A scaling factor applied to the logits. Default: 1.0
+            num_chunks: int
+                The number of chunks to split the input tensor into for processing.
+                This can help optimize memory usage and computation speed.
+                Default: 8
+            reduction:
+                Specifies the reduction to apply to the output: 'mean' | 'sum'.
+                'mean': the weighted mean of the output is taken,
+                'sum': the output will be summed.
+                Default: 'mean'.
+        """
+        super().__init__()
+
+        assert reduction in ["mean", "sum"], f"reduction: {reduction} is not supported"
+
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.logit_scale = logit_scale
+        self.num_chunks = num_chunks
+        self.reduction = reduction
+
+    @torch.compiler.disable
+    def forward(
+        self,
+        x: torch.Tensor,
+        target: torch.LongTensor,
+        weight: torch.Tensor,
+        bias: Optional[torch.Tensor] = None
+    ):
+        """
+        Args:
+            x (torch.Tensor): [batch_size, seq_len, hidden_size]
+            target (torch.LongTensor): [batch_size, seq_len]
+                where each value is in [0, V).
+            weight (torch.Tensor): [vocab_size, hidden_size]
+                where `vocab_size` is the number of classes.
+            bias (Optional[torch.Tensor]): [vocab_size]
+                where `vocab_size` is the number of classes.
+        Returns:
+            loss
+        """
+        loss = fused_linear_cross_entropy_loss(
+            x.view(-1, x.shape[-1]),
+            target.view(-1),
+            weight=weight,
+            bias=bias,
+            ignore_index=self.ignore_index,
+            label_smoothing=self.label_smoothing,
+            logit_scale=self.logit_scale,
+            num_chunks=self.num_chunks,
+            reduction=self.reduction
+        )
+        return loss
+
+
+class LinearLossParallel(ParallelStyle):
+    def __init__(
+        self,
+        *,
+        sequence_dim: int = 1,
+        use_local_output: bool = False,
+    ):
+        super().__init__()
+
+        self.sequence_sharding = (Shard(sequence_dim),)
+        self.use_local_output = use_local_output
+
+    @staticmethod
+    def _prepare_input_fn(sequence_sharding, mod, inputs, device_mesh):
+        x, target, weight, bias = inputs
+
+        if not isinstance(x, DTensor):
+            # assume the input passed in already sharded on the sequence dim and create the DTensor
+            x = DTensor.from_local(x, device_mesh, sequence_sharding)
+        if x.placements != sequence_sharding:
+            x = x.redistribute(placements=sequence_sharding, async_op=True)
+        if not isinstance(target, DTensor):
+            target = DTensor.from_local(target, device_mesh, [Replicate()])
+        if target.placements != sequence_sharding:
+            target = target.redistribute(placements=sequence_sharding, async_op=True)
+
+        if not isinstance(weight, DTensor):
+            weight = DTensor.from_local(weight, device_mesh, [Replicate()])
+        if weight.placements != [Replicate()]:
+            # we replicate the weight/bias in FLCE
+            weight = weight.redistribute(placements=[Replicate()], async_op=True)
+
+        if bias is not None and not isinstance(bias, DTensor):
+            bias = DTensor.from_local(bias, device_mesh, [Replicate()])
+        if bias is not None and bias.placements != [Replicate()]:
+            bias = bias.redistribute(placements=[Replicate()], async_op=True)
+
+        return x.to_local(), target.to_local(), weight.to_local(), bias.to_local() if bias is not None else bias
+
+    @staticmethod
+    def _prepare_output_fn(use_local_output, mod, outputs, device_mesh):
+        return outputs.to_local() if use_local_output else outputs
+
+    def _apply(self, module: nn.Module, device_mesh: DeviceMesh) -> nn.Module:
+        return distribute_module(
+            module,
+            device_mesh,
+            partition_fn=None,
+            input_fn=partial(self._prepare_input_fn, self.sequence_sharding),
+            output_fn=partial(self._prepare_output_fn, self.use_local_output)
+        )

fla/modules/fused_linear_listnet_loss.py

@@ -0,0 +1,427 @@
+# -*- coding: utf-8 -*-
+
+# Code adapted from
+# https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/fused_linear_cross_entropy.py
+
+from functools import partial
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+from torch.distributed import DeviceMesh
+from torch.distributed.tensor import DTensor, Replicate, Shard, distribute_module
+from torch.distributed.tensor.parallel import ParallelStyle
+
+from fla.ops.utils import logsumexp_fwd
+from fla.ops.utils.op import exp
+from fla.utils import input_guard
+
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576
+# https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+MAX_FUSED_SIZE = 65536 // 2
+
+@triton.jit
+def listnet_kernel(
+    logits,
+    targets,  # Now full target distributions
+    lse_logits,
+    lse_targets,
+    loss,
+    total,
+    ignore_index,
+    logit_scale: tl.constexpr,
+    reduction: tl.constexpr,
+    V: tl.constexpr,
+    BV: tl.constexpr
+):
+    i_n = tl.program_id(0).to(tl.int64)
+    NV = tl.cdiv(V, BV)
+    
+    # Pointers to current token's data
+    logits_ptr = logits + i_n * V
+    targets_ptr = targets + i_n * V
+    loss_ptr = loss + i_n
+    
+    # Compute prediction softmax
+    b_lse_logits = tl.load(lse_logits + i_n)
+    b_lse_targets = tl.load(lse_targets + i_n)
+    b_loss = 0.0
+    
+    # Compute gradient: softmax(pred) - softmax(target)
+    for iv in range(0, NV):
+        o_v = iv * BV + tl.arange(0, BV)
+        mask = o_v < V
+        
+        # Load target and compute softmax
+        t_val = tl.load(targets_ptr + o_v, mask=mask, other=0.0)
+        p_target = tl.exp(t_val - b_lse_targets)
+        
+        # Load logits and compute softmax
+        l_val = tl.load(logits_ptr + o_v, mask=mask, other=0.0) * logit_scale
+        l_val_minus_lse = l_val - b_lse_logits
+        p_pred = tl.exp(l_val_minus_lse)
+        
+        # Gradient calculation
+        grad_val = p_pred - p_target
+        if reduction == "mean":
+            grad_val = grad_val / total
+        grad_val = tl.where(b_lse_targets == float('inf'), 0.0, grad_val)
+        tl.store(logits_ptr + o_v, grad_val, mask=mask)
+        
+        # Cross-entropy loss
+        # instead of: b_loss -= tl.sum(p_target * tl.log(p_pred), axis=0)
+        b_loss -= tl.sum(p_target * l_val_minus_lse, axis=0)
+    
+    tl.store(loss_ptr, b_loss)
+
+@triton.jit
+def elementwise_mul_kernel(
+    x,
+    g,
+    N: tl.constexpr,
+    B: tl.constexpr
+):
+    """
+    This function multiplies each element of the tensor pointed by x with the value pointed by g.
+    The multiplication is performed in-place on the tensor pointed by x.
+
+    Parameters:
+    x:
+        Pointer to the input tensor.
+    g:
+        Pointer to the gradient output value.
+    N (int):
+        The number of columns in the input tensor.
+    B (int):
+        The block size for Triton operations.
+    """
+
+    # Get the program ID and convert it to int64 to avoid overflow
+    i_x = tl.program_id(0).to(tl.int64)
+    o_x = i_x * B + tl.arange(0, B)
+
+    # Load the gradient output value
+    b_g = tl.load(g)
+    b_x = tl.load(x + o_x, mask=o_x < N)
+    tl.store(x + o_x, b_x * b_g, mask=o_x < N)
+
+
+def fused_linear_listnet_forward(
+    x: torch.Tensor,
+    targets: torch.Tensor,  # Float tensor [N, V]
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    ignore_index: int = -100,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+):
+    N, H, V = *x.shape, weight.shape[0]
+    BV = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    NC = min(num_chunks, triton.cdiv(V, H))
+    C = triton.next_power_of_2(triton.cdiv(N, NC))
+    NC = triton.cdiv(N, C)
+
+    # Initialize outputs
+    dx = torch.zeros_like(x)
+    dw = torch.zeros_like(weight, dtype=torch.float) if weight is not None else None
+    db = torch.zeros_like(bias, dtype=torch.float) if bias is not None else None
+    loss = torch.zeros(N, device=x.device, dtype=torch.float)
+    total = N  # All tokens considered
+
+    for ic in range(NC):
+        start, end = ic * C, min((ic + 1) * C, N)
+        c_x = x[start:end]
+        c_logits = F.linear(c_x, weight, bias)
+        c_targets = targets[start:end]
+        c_lse_logits = logsumexp_fwd(c_logits, scale=logit_scale, dtype=torch.float)
+        c_lse_targets = logsumexp_fwd(c_targets, dtype=torch.float).nan_to_num(nan=float("inf"))
+        c_loss = loss[start:end]
+
+        # Call ListNet kernel
+        listnet_kernel[(c_logits.shape[0],)](
+            logits=c_logits,
+            targets=c_targets,  # Full target distributions
+            lse_logits=c_lse_logits,
+            lse_targets=c_lse_targets,
+            loss=c_loss,
+            total=total,
+            ignore_index=ignore_index,
+            logit_scale=logit_scale,
+            reduction=reduction,
+            V=V,
+            BV=BV,
+            num_warps=32
+        )
+
+        # Backward through linear layer
+        dx[start:end] = torch.mm(c_logits, weight)
+        if weight is not None:
+            dw += c_logits.t() @ c_x
+        if bias is not None:
+            db += c_logits.sum(0)
+
+    loss = loss.sum()
+    if reduction == "mean":
+        loss = loss / total
+        
+    return loss, dx, dw, db
+
+
+def fused_linear_listnet_backward(
+    do: torch.Tensor,
+    dx: torch.Tensor,
+    dw: torch.Tensor,
+    db: torch.Tensor
+):
+    # If cross entropy is the last layer, do is 1.0. Skip the mul to save time
+    if torch.ne(do, torch.tensor(1.0, device=do.device)):
+        # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+        # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+        N, H = dx.shape
+        B = min(MAX_FUSED_SIZE, triton.next_power_of_2(H))
+
+        elementwise_mul_kernel[(triton.cdiv(N * H, B),)](
+            x=dx,
+            g=do,
+            N=N*H,
+            B=B,
+            num_warps=32,
+        )
+
+        # handle dw
+        if dw is not None:
+            V, H = dw.shape
+            elementwise_mul_kernel[(triton.cdiv(V * H, B),)](
+                x=dw,
+                g=do,
+                N=V*H,
+                B=B,
+                num_warps=32,
+            )
+
+        if db is not None:
+            V = db.shape[0]
+            elementwise_mul_kernel[(triton.cdiv(V, B),)](
+                x=db,
+                g=do,
+                N=V,
+                B=B,
+                num_warps=32,
+            )
+    return dx, dw, db
+
+
+class FusedLinearListNetFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        targets: torch.Tensor,  # Float targets
+        weight: torch.Tensor,
+        bias: torch.Tensor = None,
+        ignore_index: int = -100,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        loss, dx, dw, db = fused_linear_listnet_forward(
+            x, targets, weight, bias, ignore_index, 
+            logit_scale, num_chunks, reduction
+        )
+        ctx.save_for_backward(dx, dw, db)
+        return loss
+
+    @staticmethod
+    def backward(ctx, do):
+        dx, dw, db = ctx.saved_tensors
+        dx, dw, db = fused_linear_listnet_backward(do, dx, dw, db)
+        return dx, None, dw, db, None, None, None, None
+
+
+def fused_linear_listnet_loss(
+    x: torch.Tensor,
+    target: torch.LongTensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    ignore_index: int = -100,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target (torch.LongTensor): [batch_size * seq_len]
+            where each value is in [0, vocab_size).
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        bias (Optional[torch.Tensor]): [vocab_size]
+            where `vocab_size` is the number of classes.
+        ignore_index: int.
+            If target == ignore_index, the loss is set to 0.0.
+        label_smoothing: float
+        logit_scale: float
+            A scaling factor applied to the logits. Default: 1.0
+        num_chunks: int
+            The number of chunks to split the input tensor into for processing.
+            This can help optimize memory usage and computation speed.
+            Default: 8
+        reduction:
+            Specifies the reduction to apply to the output: 'mean' | 'sum'.
+            'mean': the weighted mean of the output is taken,
+            'sum': the output will be summed.
+            Default: 'mean'.
+    Returns:
+        losses: [batch,], float
+    """
+    return FusedLinearListNetFunction.apply(
+        x,
+        target,
+        weight,
+        bias,
+        ignore_index,
+        logit_scale,
+        num_chunks,
+        reduction
+    )
+
+
+class FusedLinearListNetLoss(nn.Module):
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        """
+        Args:
+            ignore_index: int.
+                If target == ignore_index, the loss is set to 0.0.
+            label_smoothing: float
+            logit_scale: float
+                A scaling factor applied to the logits. Default: 1.0
+            num_chunks: int
+                The number of chunks to split the input tensor into for processing.
+                This can help optimize memory usage and computation speed.
+                Default: 8
+            reduction:
+                Specifies the reduction to apply to the output: 'mean' | 'sum'.
+                'mean': the weighted mean of the output is taken,
+                'sum': the output will be summed.
+                Default: 'mean'.
+        """
+        super().__init__()
+
+        assert reduction in ["mean", "sum"], f"reduction: {reduction} is not supported"
+
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.logit_scale = logit_scale
+        self.num_chunks = num_chunks
+        self.reduction = reduction
+
+    @torch.compiler.disable
+    def forward(
+        self,
+        x: torch.Tensor,
+        target: torch.LongTensor,
+        weight: torch.Tensor,
+        bias: Optional[torch.Tensor] = None
+    ):
+        """
+        Args:
+            x (torch.Tensor): [batch_size, seq_len, hidden_size]
+            target (torch.LongTensor): [batch_size, seq_len]
+                where each value is in [0, V).
+            weight (torch.Tensor): [vocab_size, hidden_size]
+                where `vocab_size` is the number of classes.
+            bias (Optional[torch.Tensor]): [vocab_size]
+                where `vocab_size` is the number of classes.
+        Returns:
+            loss
+        """
+        loss = fused_linear_listnet_loss(
+            x.view(-1, x.shape[-1]),
+            target.view(-1, target.shape[-1]),
+            weight=weight,
+            bias=bias,
+            ignore_index=self.ignore_index,
+            label_smoothing=self.label_smoothing,
+            logit_scale=self.logit_scale,
+            num_chunks=self.num_chunks,
+            reduction=self.reduction
+        )
+        return loss
+
+
+class LinearLossParallel(ParallelStyle):
+    def __init__(
+        self,
+        *,
+        sequence_dim: int = 1,
+        use_local_output: bool = False,
+    ):
+        super().__init__()
+
+        self.sequence_sharding = (Shard(sequence_dim),)
+        self.use_local_output = use_local_output
+
+    @staticmethod
+    def _prepare_input_fn(sequence_sharding, mod, inputs, device_mesh):
+        x, target, weight, bias = inputs
+
+        if not isinstance(x, DTensor):
+            # assume the input passed in already sharded on the sequence dim and create the DTensor
+            x = DTensor.from_local(x, device_mesh, sequence_sharding)
+        if x.placements != sequence_sharding:
+            x = x.redistribute(placements=sequence_sharding, async_op=True)
+        if not isinstance(target, DTensor):
+            target = DTensor.from_local(target, device_mesh, [Replicate()])
+        if target.placements != sequence_sharding:
+            target = target.redistribute(placements=sequence_sharding, async_op=True)
+
+        if not isinstance(weight, DTensor):
+            weight = DTensor.from_local(weight, device_mesh, [Replicate()])
+        if weight.placements != [Replicate()]:
+            # we replicate the weight/bias in FLCE
+            weight = weight.redistribute(placements=[Replicate()], async_op=True)
+
+        if bias is not None and not isinstance(bias, DTensor):
+            bias = DTensor.from_local(bias, device_mesh, [Replicate()])
+        if bias is not None and bias.placements != [Replicate()]:
+            bias = bias.redistribute(placements=[Replicate()], async_op=True)
+
+        return x.to_local(), target.to_local(), weight.to_local(), bias.to_local() if bias is not None else bias
+
+    @staticmethod
+    def _prepare_output_fn(use_local_output, mod, outputs, device_mesh):
+        return outputs.to_local() if use_local_output else outputs
+
+    def _apply(self, module: nn.Module, device_mesh: DeviceMesh) -> nn.Module:
+        return distribute_module(
+            module,
+            device_mesh,
+            partition_fn=None,
+            input_fn=partial(self._prepare_input_fn, self.sequence_sharding),
+            output_fn=partial(self._prepare_output_fn, self.use_local_output)
+        )
+
+# Naive ListNet loss function implementation
+def list_net_loss(y_pred, y_true):
+    """
+    ListNet loss introduced in "Learning to Rank: From Pairwise Approach to Listwise Approach".
+    :param y_pred: predictions from the model, shape [*, slate_length]
+    :param y_true: ground truth labels, shape [*, slate_length]
+    :return: loss value, a torch.Tensor
+    """
+    return torch.mean(-torch.sum(F.softmax(y_true, dim=-1).nan_to_num(nan=0) * F.log_softmax(y_pred, dim=-1), dim=-1))

fla/modules/fused_norm_gate.py

@@ -0,0 +1,995 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from __future__ import annotations
+
+import math
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+
+from fla.utils import get_multiprocessor_count, input_guard
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+        for num_stages in [2, 3, 4]
+    ],
+    key=['N', 'HAS_RESIDUAL', 'STORE_RESIDUAL_OUT', 'IS_RMS_NORM', 'HAS_BIAS'],
+)
+@triton.jit
+def layer_norm_gated_fwd_kernel(
+    X,  # pointer to the input
+    G,  # pointer to the gate
+    Y,  # pointer to the output
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    RESIDUAL,  # pointer to the residual
+    RESIDUAL_OUT,  # pointer to the residual
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    ACTIVATION: tl.constexpr,
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_RESIDUAL: tl.constexpr,
+    STORE_RESIDUAL_OUT: tl.constexpr,
+    HAS_WEIGHT: tl.constexpr,
+    HAS_BIAS: tl.constexpr
+):
+    # Map the program id to the row of X and Y it should compute.
+    row = tl.program_id(0)
+    X += row * N
+    Y += row * N
+    G += row * N
+    if HAS_RESIDUAL:
+        RESIDUAL += row * N
+    if STORE_RESIDUAL_OUT:
+        RESIDUAL_OUT += row * N
+    # Compute mean and variance
+    cols = tl.arange(0, BLOCK_N)
+    x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
+    if HAS_RESIDUAL:
+        residual = tl.load(RESIDUAL + cols, mask=cols < N, other=0.0).to(tl.float32)
+        x += residual
+    if STORE_RESIDUAL_OUT:
+        tl.store(RESIDUAL_OUT + cols, x, mask=cols < N)
+    if not IS_RMS_NORM:
+        mean = tl.sum(x, axis=0) / N
+        tl.store(Mean + row, mean)
+        xbar = tl.where(cols < N, x - mean, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    else:
+        xbar = tl.where(cols < N, x, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    rstd = 1 / tl.sqrt(var + eps)
+    tl.store(Rstd + row, rstd)
+    # Normalize and apply linear transformation
+    mask = cols < N
+    if HAS_WEIGHT:
+        w = tl.load(W + cols, mask=mask).to(tl.float32)
+    if HAS_BIAS:
+        b = tl.load(B + cols, mask=mask).to(tl.float32)
+    x_hat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+    y = x_hat * w if HAS_WEIGHT else x_hat
+    if HAS_BIAS:
+        y = y + b
+
+    # Swish output gate
+    g = tl.load(G + cols, mask=cols < N, other=0.0).to(tl.float32)
+    if ACTIVATION == 'swish':
+        y = y * g * tl.sigmoid(g)
+    elif ACTIVATION == 'silu':
+        y = y * g * tl.sigmoid(g)
+    elif ACTIVATION == 'sigmoid':
+        y = y * tl.sigmoid(g)
+
+    # Write output
+    tl.store(Y + cols, y, mask=mask)
+
+
+def layer_norm_gated_fwd(
+    x: torch.Tensor,
+    g: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    activation: str = 'swish',
+    eps: float = 1e-5,
+    residual: torch.Tensor = None,
+    out_dtype: torch.dtype = None,
+    residual_dtype: torch.dtype = None,
+    is_rms_norm: bool = False
+):
+    if residual is not None:
+        residual_dtype = residual.dtype
+    M, N = x.shape
+    if residual is not None:
+        assert residual.shape == (M, N)
+    if weight is not None:
+        assert weight.shape == (N,)
+    if bias is not None:
+        assert bias.shape == (N,)
+    # allocate output
+    y = torch.empty_like(x, dtype=x.dtype if out_dtype is None else out_dtype)
+    if residual is not None or (residual_dtype is not None and residual_dtype != x.dtype):
+        residual_out = torch.empty(M, N, device=x.device, dtype=residual_dtype)
+    else:
+        residual_out = None
+    mean = torch.empty((M,), dtype=torch.float, device=x.device) if not is_rms_norm else None
+    rstd = torch.empty((M,), dtype=torch.float, device=x.device)
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    # heuristics for number of warps
+
+    layer_norm_gated_fwd_kernel[(M,)](
+        x,
+        g,
+        y,
+        weight,
+        bias,
+        residual,
+        residual_out,
+        mean,
+        rstd,
+        N,
+        eps,
+        ACTIVATION=activation,
+        IS_RMS_NORM=is_rms_norm,
+        BLOCK_N=BLOCK_N,
+        HAS_RESIDUAL=residual is not None,
+        STORE_RESIDUAL_OUT=residual_out is not None,
+        HAS_WEIGHT=weight is not None,
+        HAS_BIAS=bias is not None,
+    )
+    # residual_out is None if residual is None and residual_dtype == input_dtype
+    return y, mean, rstd, residual_out if residual_out is not None else x
+
+
+@triton.heuristics({
+    'RECOMPUTE_OUTPUT': lambda args: args["Y"] is not None
+})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+        for num_stages in [2, 3, 4]
+    ],
+    key=['N', 'HAS_DRESIDUAL', 'STORE_DRESIDUAL', 'IS_RMS_NORM', 'HAS_BIAS'],
+)
+@triton.jit
+def layer_norm_gated_bwd_kernel(
+    X,  # pointer to the input
+    G,  # pointer to the gate
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    Y,  # pointer to the output to be recomputed
+    DY,  # pointer to the output gradient
+    DX,  # pointer to the input gradient
+    DG,  # pointer to the gate gradient
+    DW,  # pointer to the partial sum of weights gradient
+    DB,  # pointer to the partial sum of biases gradient
+    DRESIDUAL,
+    DRESIDUAL_IN,
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    M,  # number of rows in X
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    rows_per_program,
+    ACTIVATION: tl.constexpr,
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_DRESIDUAL: tl.constexpr,
+    STORE_DRESIDUAL: tl.constexpr,
+    HAS_WEIGHT: tl.constexpr,
+    HAS_BIAS: tl.constexpr,
+    RECOMPUTE_OUTPUT: tl.constexpr,
+):
+    # Map the program id to the elements of X, DX, and DY it should compute.
+    row_block_id = tl.program_id(0)
+    row_start = row_block_id * rows_per_program
+    cols = tl.arange(0, BLOCK_N)
+    mask = cols < N
+    X += row_start * N
+    G += row_start * N
+    if HAS_DRESIDUAL:
+        DRESIDUAL += row_start * N
+    if STORE_DRESIDUAL:
+        DRESIDUAL_IN += row_start * N
+    DY += row_start * N
+    DX += row_start * N
+    DG += row_start * N
+    if RECOMPUTE_OUTPUT:
+        Y += row_start * N
+    if HAS_WEIGHT:
+        w = tl.load(W + cols, mask=mask).to(tl.float32)
+        dw = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    if HAS_BIAS:
+        b = tl.load(B + cols, mask=mask, other=0.0).to(tl.float32)
+    if HAS_BIAS:
+        db = tl.zeros((BLOCK_N,), dtype=tl.float32)
+
+    row_end = min((row_block_id + 1) * rows_per_program, M)
+    for row in range(row_start, row_end):
+        # Load data to SRAM
+        x = tl.load(X + cols, mask=mask, other=0).to(tl.float32)
+        g = tl.load(G + cols, mask=mask, other=0).to(tl.float32)
+        dy = tl.load(DY + cols, mask=mask, other=0).to(tl.float32)
+
+        if not IS_RMS_NORM:
+            mean = tl.load(Mean + row)
+        rstd = tl.load(Rstd + row)
+        # Compute dx
+        xhat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+        xhat = tl.where(mask, xhat, 0.0)
+
+        y = xhat * w if HAS_WEIGHT else xhat
+        if HAS_BIAS:
+            y = y + b
+        if RECOMPUTE_OUTPUT:
+            tl.store(Y + cols, y, mask=mask)
+
+        sigmoid_g = tl.sigmoid(g)
+        if ACTIVATION == 'swish':
+            dg = dy * y * (sigmoid_g + g * sigmoid_g * (1 - sigmoid_g))
+            dy = dy * g * sigmoid_g
+        elif ACTIVATION == 'silu':
+            dg = dy * y * (sigmoid_g + g * sigmoid_g * (1 - sigmoid_g))
+            dy = dy * g * sigmoid_g
+        elif ACTIVATION == 'sigmoid':
+            dg = dy * y * sigmoid_g * (1 - sigmoid_g)
+            dy = dy * sigmoid_g
+        wdy = dy
+        if HAS_WEIGHT:
+            wdy = dy * w
+            dw += dy * xhat
+        if HAS_BIAS:
+            db += dy
+        if not IS_RMS_NORM:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            c2 = tl.sum(wdy, axis=0) / N
+            dx = (wdy - (xhat * c1 + c2)) * rstd
+        else:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            dx = (wdy - xhat * c1) * rstd
+        if HAS_DRESIDUAL:
+            dres = tl.load(DRESIDUAL + cols, mask=mask, other=0).to(tl.float32)
+            dx += dres
+        # Write dx
+        if STORE_DRESIDUAL:
+            tl.store(DRESIDUAL_IN + cols, dx, mask=mask)
+        tl.store(DX + cols, dx, mask=mask)
+        tl.store(DG + cols, dg, mask=mask)
+
+        X += N
+        G += N
+        if HAS_DRESIDUAL:
+            DRESIDUAL += N
+        if STORE_DRESIDUAL:
+            DRESIDUAL_IN += N
+        if RECOMPUTE_OUTPUT:
+            Y += N
+        DY += N
+        DX += N
+        DG += N
+    if HAS_WEIGHT:
+        tl.store(DW + row_block_id * N + cols, dw, mask=mask)
+    if HAS_BIAS:
+        tl.store(DB + row_block_id * N + cols, db, mask=mask)
+
+
+def layer_norm_gated_bwd(
+    dy: torch.Tensor,
+    x: torch.Tensor,
+    g: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    activation: str = 'swish',
+    eps: float = 1e-5,
+    mean: torch.Tensor = None,
+    rstd: torch.Tensor = None,
+    dresidual: torch.Tensor = None,
+    has_residual: bool = False,
+    is_rms_norm: bool = False,
+    x_dtype: torch.dtype = None,
+    recompute_output: bool = False,
+):
+    M, N = x.shape
+    assert dy.shape == (M, N)
+    if dresidual is not None:
+        assert dresidual.shape == (M, N)
+    if weight is not None:
+        assert weight.shape == (N,)
+    if bias is not None:
+        assert bias.shape == (N,)
+    # allocate output
+    dx = torch.empty_like(x) if x_dtype is None else torch.empty(M, N, dtype=x_dtype, device=x.device)
+    dg = torch.empty_like(g) if x_dtype is None else torch.empty(M, N, dtype=x_dtype, device=x.device)
+    dresidual_in = torch.empty_like(x) if has_residual and dx.dtype != x.dtype else None
+    y = torch.empty(M, N, dtype=dy.dtype, device=dy.device) if recompute_output else None
+
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    sm_count = get_multiprocessor_count(x.device.index)
+    dw = torch.empty((sm_count, N), dtype=torch.float, device=weight.device) if weight is not None else None
+    db = torch.empty((sm_count, N), dtype=torch.float, device=bias.device) if bias is not None else None
+    rows_per_program = math.ceil(M / sm_count)
+    grid = (sm_count,)
+    layer_norm_gated_bwd_kernel[grid](
+        x,
+        g,
+        weight,
+        bias,
+        y,
+        dy,
+        dx,
+        dg,
+        dw,
+        db,
+        dresidual,
+        dresidual_in,
+        mean,
+        rstd,
+        M,
+        N,
+        eps,
+        rows_per_program,
+        ACTIVATION=activation,
+        IS_RMS_NORM=is_rms_norm,
+        BLOCK_N=BLOCK_N,
+        HAS_DRESIDUAL=dresidual is not None,
+        STORE_DRESIDUAL=dresidual_in is not None,
+        HAS_WEIGHT=weight is not None,
+        HAS_BIAS=bias is not None,
+    )
+    dw = dw.sum(0).to(weight.dtype) if weight is not None else None
+    db = db.sum(0).to(bias.dtype) if bias is not None else None
+    # Don't need to compute dresidual_in separately in this case
+    if has_residual and dx.dtype == x.dtype:
+        dresidual_in = dx
+    return (dx, dg, dw, db, dresidual_in) if not recompute_output else (dx, dg, dw, db, dresidual_in, y)
+
+
+class LayerNormGatedFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        weight: torch.Tensor,
+        bias: torch.Tensor,
+        activation: str,
+        residual: Optional[torch.Tensor] = None,
+        eps: float = 1e-6,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False,
+        is_rms_norm: bool = False,
+    ):
+        x_shape_og = x.shape
+        g_shape_og = g.shape
+        # reshape input data into 2D tensor
+        x = x.reshape(-1, x.shape[-1])
+        g = g.reshape(-1, g.shape[-1])
+        if residual is not None:
+            assert residual.shape == x_shape_og
+            residual = residual.reshape(-1, residual.shape[-1])
+        residual_dtype = (
+            residual.dtype
+            if residual is not None
+            else (torch.float if residual_in_fp32 else None)
+        )
+        y, mean, rstd, residual_out = layer_norm_gated_fwd(
+            x=x,
+            g=g,
+            weight=weight,
+            bias=bias,
+            activation=activation,
+            eps=eps,
+            residual=residual,
+            residual_dtype=residual_dtype,
+            is_rms_norm=is_rms_norm
+        )
+        ctx.save_for_backward(residual_out, g, weight, bias, mean, rstd)
+        ctx.x_shape_og = x_shape_og
+        ctx.g_shape_og = g_shape_og
+        ctx.activation = activation
+        ctx.eps = eps
+        ctx.is_rms_norm = is_rms_norm
+        ctx.has_residual = residual is not None
+        ctx.prenorm = prenorm
+        ctx.x_dtype = x.dtype
+        y = y.reshape(x_shape_og)
+        return y if not prenorm else (y, residual_out.reshape(x_shape_og))
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, dy, *args):
+        x, g, weight, bias, mean, rstd = ctx.saved_tensors
+        dy = dy.reshape(-1, dy.shape[-1])
+        assert dy.shape == x.shape
+        if ctx.prenorm:
+            dresidual = args[0]
+            dresidual = dresidual.reshape(-1, dresidual.shape[-1])
+            assert dresidual.shape == x.shape
+        else:
+            dresidual = None
+        dx, dg, dw, db, dresidual_in = layer_norm_gated_bwd(
+            dy=dy,
+            x=x,
+            g=g,
+            weight=weight,
+            bias=bias,
+            activation=ctx.activation,
+            eps=ctx.eps,
+            mean=mean,
+            rstd=rstd,
+            dresidual=dresidual,
+            has_residual=ctx.has_residual,
+            is_rms_norm=ctx.is_rms_norm,
+            x_dtype=ctx.x_dtype,
+        )
+        return (
+            dx.reshape(ctx.x_shape_og),
+            dg.reshape(ctx.g_shape_og),
+            dw,
+            db,
+            None,
+            dresidual_in.reshape(ctx.x_shape_og) if ctx.has_residual else None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+class LayerNormGatedLinearFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        norm_weight: torch.Tensor,
+        norm_bias: torch.Tensor,
+        linear_weight: torch.Tensor,
+        linear_bias: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+        eps: float = 1e-6,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False,
+        is_rms_norm: bool = False,
+    ):
+        x_shape_og = x.shape
+        g_shape_og = g.shape
+        # reshape input data into 2D tensor
+        x = x.reshape(-1, x.shape[-1])
+        g = g.reshape(-1, g.shape[-1])
+        if residual is not None:
+            assert residual.shape == x_shape_og
+            residual = residual.reshape(-1, residual.shape[-1])
+        residual_dtype = (
+            residual.dtype
+            if residual is not None
+            else (torch.float if residual_in_fp32 else None)
+        )
+        y, mean, rstd, residual_out = layer_norm_gated_fwd(
+            x=x,
+            g=g,
+            weight=norm_weight,
+            bias=norm_bias,
+            eps=eps,
+            residual=residual,
+            residual_dtype=residual_dtype,
+            is_rms_norm=is_rms_norm
+        )
+        y = y.reshape(x_shape_og)
+        dtype = torch.get_autocast_gpu_dtype() if torch.is_autocast_enabled() else y.dtype
+        linear_weight = linear_weight.to(dtype)
+        linear_bias = linear_bias.to(dtype) if linear_bias is not None else None
+        out = F.linear(y.to(linear_weight.dtype), linear_weight, linear_bias)
+        # We don't store y, will be recomputed in the backward pass to save memory
+        ctx.save_for_backward(residual_out, g, norm_weight, norm_bias, linear_weight, mean, rstd)
+        ctx.x_shape_og = x_shape_og
+        ctx.g_shape_og = g_shape_og
+        ctx.eps = eps
+        ctx.is_rms_norm = is_rms_norm
+        ctx.has_residual = residual is not None
+        ctx.prenorm = prenorm
+        ctx.x_dtype = x.dtype
+        ctx.linear_bias_is_none = linear_bias is None
+        return out if not prenorm else (out, residual_out.reshape(x_shape_og))
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, dout, *args):
+        x, g, norm_weight, norm_bias, linear_weight, mean, rstd = ctx.saved_tensors
+        dout = dout.reshape(-1, dout.shape[-1])
+        dy = F.linear(dout, linear_weight.t())
+        dlinear_bias = None if ctx.linear_bias_is_none else dout.sum(0)
+        assert dy.shape == x.shape
+        if ctx.prenorm:
+            dresidual = args[0]
+            dresidual = dresidual.reshape(-1, dresidual.shape[-1])
+            assert dresidual.shape == x.shape
+        else:
+            dresidual = None
+        dx, dg, dnorm_weight, dnorm_bias, dresidual_in, y = layer_norm_gated_bwd(
+            dy=dy,
+            x=x,
+            g=g,
+            norm_weight=norm_weight,
+            norm_bias=norm_bias,
+            eps=ctx.eps,
+            mean=mean,
+            rstd=rstd,
+            dresidual=dresidual,
+            has_residual=ctx.has_residual,
+            is_rms_norm=ctx.is_rms_norm,
+            x_dtype=ctx.x_dtype,
+            recompute_output=True,
+        )
+        dlinear_weight = torch.einsum("bo,bi->oi", dout, y)
+        return (
+            dx.reshape(ctx.x_shape_og),
+            dg.reshape(ctx.g_shape_og),
+            dnorm_weight,
+            dnorm_bias,
+            dlinear_weight,
+            dlinear_bias,
+            dresidual_in.reshape(ctx.x_shape_og) if ctx.has_residual else None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+def layer_norm_gated(
+    x: torch.Tensor,
+    g: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    activation: str = 'swish',
+    residual: Optional[torch.Tensor] = None,
+    prenorm: bool = False,
+    residual_in_fp32: bool = False,
+    eps: float = 1e-6
+):
+    return LayerNormGatedFunction.apply(
+        x,
+        g,
+        weight,
+        bias,
+        activation,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        False
+    )
+
+
+def rms_norm_gated(
+    x: torch.Tensor,
+    g: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    activation: str = 'swish',
+    residual: Optional[torch.Tensor] = None,
+    prenorm: bool = False,
+    residual_in_fp32: bool = False,
+    eps: float = 1e-6
+):
+    return LayerNormGatedFunction.apply(
+        x,
+        g,
+        weight,
+        bias,
+        activation,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        True
+    )
+
+
+def layer_norm_swish_gate_linear(
+    x: torch.Tensor,
+    g: torch.Tensor,
+    norm_weight: torch.Tensor,
+    norm_bias: torch.Tensor,
+    linear_weight: torch.Tensor,
+    linear_bias: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+    prenorm: bool = False,
+    residual_in_fp32: bool = False,
+    eps: float = 1e-6
+):
+    return LayerNormGatedLinearFunction.apply(
+        x,
+        g,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        False
+    )
+
+
+def rms_norm_swish_gate_linear(
+    x,
+    g: torch.Tensor,
+    norm_weight: torch.Tensor,
+    norm_bias: torch.Tensor,
+    linear_weight: torch.Tensor,
+    linear_bias: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+    prenorm: bool = False,
+    residual_in_fp32: bool = False,
+    eps: float = 1e-6
+):
+    return LayerNormGatedLinearFunction.apply(
+        x,
+        g,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        True
+    )
+
+
+class FusedLayerNormGated(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        bias: bool = False,
+        activation: str = 'swish',
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedLayerNormGated:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.elementwise_affine = elementwise_affine
+        self.eps = eps
+        self.activation = activation
+
+        if self.activation not in ['swish', 'silu', 'sigmoid']:
+            raise ValueError(f"Unsupported activation: {self.activation}")
+
+        self.register_parameter("weight", None)
+        self.register_parameter("bias", None)
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+            if bias:
+                self.bias = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+            if self.bias is not None:
+                nn.init.zeros_(self.bias)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.hidden_size}"
+        if not self.elementwise_affine:
+            s += f", elementwise_affine={self.elementwise_affine}"
+        s += f", eps={self.eps}"
+        s += f", activation={self.activation}"
+        s += ")"
+        return s
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False
+    ) -> torch.Tensor:
+        return layer_norm_gated(
+            x,
+            g,
+            self.weight,
+            self.bias,
+            self.activation,
+            residual=residual,
+            eps=self.eps,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32
+        )
+
+
+class FusedRMSNormGated(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        activation: str = 'swish',
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedRMSNormGated:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.elementwise_affine = elementwise_affine
+        self.eps = eps
+        self.activation = activation
+
+        if self.activation not in ['swish', 'silu', 'sigmoid']:
+            raise ValueError(f"Unsupported activation: {self.activation}")
+
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+        else:
+            self.register_parameter("weight", None)
+        self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.hidden_size}"
+        if not self.elementwise_affine:
+            s += f", elementwise_affine={self.elementwise_affine}"
+        s += f", eps={self.eps}"
+        s += f", activation={self.activation}"
+        s += ")"
+        return s
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False
+    ) -> torch.Tensor:
+        return rms_norm_gated(
+            x,
+            g,
+            self.weight,
+            self.bias,
+            self.activation,
+            residual=residual,
+            eps=self.eps,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32
+        )
+
+
+class FusedLayerNormSwishGate(FusedLayerNormGated):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        bias: bool = False,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedLayerNormSwishGate:
+        super().__init__(
+            hidden_size=hidden_size,
+            elementwise_affine=elementwise_affine,
+            bias=bias,
+            eps=eps,
+            device=device,
+            dtype=dtype
+        )
+
+
+class FusedRMSNormSwishGate(FusedRMSNormGated):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedRMSNormSwishGate:
+        super().__init__(
+            hidden_size=hidden_size,
+            elementwise_affine=elementwise_affine,
+            eps=eps,
+            device=device,
+            dtype=dtype
+        )
+
+
+class FusedLayerNormGatedLinear(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedLayerNormGatedLinear:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.elementwise_affine = elementwise_affine
+        self.eps = eps
+
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+        else:
+            self.register_parameter("weight", None)
+        self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.hidden_size}"
+        if not self.elementwise_affine:
+            s += f", elementwise_affine={self.elementwise_affine}"
+        s += f", eps={self.eps}"
+        s += ")"
+        return s
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        weight: Optional[torch.Tensor] = None,
+        bias: Optional[torch.Tensor] = None,
+        residual: Optional[torch.Tensor] = None,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False
+    ) -> torch.Tensor:
+        return layer_norm_swish_gate_linear(
+            x,
+            g,
+            self.weight,
+            self.bias,
+            weight,
+            bias,
+            residual=residual,
+            eps=self.eps,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32
+        )
+
+
+class FusedLayerNormSwishGateLinear(FusedLayerNormGatedLinear):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedLayerNormSwishGateLinear:
+        super().__init__(
+            hidden_size=hidden_size,
+            elementwise_affine=elementwise_affine,
+            eps=eps,
+            device=device,
+            dtype=dtype
+        )
+
+
+class FusedRMSNormGatedLinear(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedRMSNormGatedLinear:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.elementwise_affine = elementwise_affine
+        self.eps = eps
+
+        self.register_parameter("weight", None)
+        self.register_parameter("bias", None)
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}({self.hidden_size}"
+        if not self.elementwise_affine:
+            s += f", elementwise_affine={self.elementwise_affine}"
+        s += f", eps={self.eps}"
+        s += ")"
+        return s
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        weight: Optional[torch.Tensor] = None,
+        bias: Optional[torch.Tensor] = None,
+        residual: Optional[torch.Tensor] = None,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False
+    ) -> torch.Tensor:
+        return rms_norm_swish_gate_linear(
+            x,
+            g,
+            self.weight,
+            self.bias,
+            weight,
+            bias,
+            residual=residual,
+            eps=self.eps,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32
+        )
+
+
+class FusedRMSNormSwishGateLinear(FusedRMSNormGatedLinear):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> FusedRMSNormSwishGateLinear:
+        super().__init__(
+            hidden_size=hidden_size,
+            elementwise_affine=elementwise_affine,
+            eps=eps,
+            device=device,
+            dtype=dtype
+        )

fla/modules/grpo.py

@@ -0,0 +1,396 @@
+# -*- coding: utf-8 -*-
+
+# https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py
+"""
+# Get the per-token log probabilities for the completions for the model and the reference model
+    def _get_per_token_logps(self, model, input_ids, attention_mask, logits_to_keep):
+        # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
+        logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits
+        logits = logits[:, :-1, :]  # (B, L-1, V), exclude the last logit: it corresponds to the next token pred
+
+        input_ids = input_ids[:, -logits_to_keep:]
+        # For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
+        # See https://github.com/huggingface/trl/issues/2770
+        logits = logits[:, -logits_to_keep:]
+        return selective_log_softmax(logits, input_ids)  #  compute logprobs for the input tokens
+
+    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+        if return_outputs:
+            raise ValueError("The GRPOTrainer does not support returning outputs")
+        # Compute the per-token log probabilities for the model
+
+        prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
+        completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"]
+        input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
+        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
+        logits_to_keep = completion_ids.size(1)  # we only need to compute the logits for the completion tokens
+
+        per_token_logps = self._get_per_token_logps(model, input_ids, attention_mask, logits_to_keep)
+
+        # Compute the KL divergence between the model and the reference model
+        ref_per_token_logps = inputs["ref_per_token_logps"]
+        per_token_kl = torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1
+
+        # x - x.detach() allows for preserving gradients from x
+        advantages = inputs["advantages"]
+        per_token_loss = torch.exp(per_token_logps - per_token_logps.detach()) * advantages.unsqueeze(1)
+        per_token_loss = -(per_token_loss - self.beta * per_token_kl)
+        loss = ((per_token_loss * completion_mask).sum(dim=1) / completion_mask.sum(dim=1)).mean()
+
+        # Log the metrics
+        completion_length = self.accelerator.gather_for_metrics(completion_mask.sum(1)).float().mean().item()
+        self._metrics["completion_length"].append(completion_length)
+
+        mean_kl = ((per_token_kl * completion_mask).sum(dim=1) / completion_mask.sum(dim=1)).mean()
+        self._metrics["kl"].append(self.accelerator.gather_for_metrics(mean_kl).mean().item())
+
+        return loss
+"""
+
+
+import torch
+import triton
+import triton.language as tl
+
+from fla.ops.utils.op import exp, log
+from fla.utils import input_guard
+
+
+@triton.autotune(
+    [triton.Config({'BLOCK_SIZE': BLOCK_SIZE},  num_warps=NUM_WARPS, num_stages=NUM_STAGES)
+     for BLOCK_SIZE in [1024, 2048, 4096, 8192]
+     for NUM_WARPS in [8, 16, 32]
+     for NUM_STAGES in [1, 2, 4]
+     ], key=['B', 'N']
+)
+@triton.jit
+def grpo_fwd_kernel(
+    logits_ptr,
+    ref_logp_ptr,
+    input_ids_ptr,
+    advantages_ptr,
+    completion_mask_ptr,
+    loss_ptr,
+    lse_ptr,
+    beta,
+    save_kl: tl.constexpr,
+    B,
+    M,
+    N,
+    L,
+    start_idx,
+    BLOCK_SIZE: tl.constexpr
+):
+    row_idx = tl.program_id(0)
+
+    off_b = row_idx // L
+    N = tl.cast(N, tl.int64)
+
+    loss_ptr += row_idx
+
+    completion_mask_ptr += row_idx
+    not_skip = tl.load(completion_mask_ptr).to(tl.int1)
+    if not_skip == 1:
+        ref_logp_ptr += row_idx
+        lse_ptr += row_idx
+        advantages_ptr += off_b
+        logits_ptr += N * (row_idx + off_b)
+        input_ids_ptr += row_idx + (off_b+1) * start_idx
+        base_cols = tl.arange(0, BLOCK_SIZE)
+
+        m_i = -float("inf")
+        l_i = 0.0
+        for start_n in tl.range(0, N, BLOCK_SIZE):
+            cols = start_n + base_cols
+            mask = cols < N
+            logits = tl.load(logits_ptr+cols, mask=mask, other=-float('inf')).to(tl.float32)
+            m_ij = tl.max(logits)
+            new_m_i = tl.maximum(m_i, m_ij)
+            l_i = l_i * exp(m_i - new_m_i) + tl.sum(exp(logits - new_m_i))
+            m_i = new_m_i
+        lse = log(l_i) + m_i
+
+        idx = tl.load(input_ids_ptr)
+        x = tl.load(logits_ptr+idx).to(tl.float32)
+        advantage = tl.load(advantages_ptr).to(tl.float32)
+        ref_logp = tl.load(ref_logp_ptr)
+        logp = x - lse
+        diff = ref_logp - logp
+        kl = exp(diff) - diff - 1
+        loss = kl * beta - advantage
+
+        tl.store(loss_ptr, loss.to(loss_ptr.dtype.element_ty))
+        tl.store(lse_ptr, lse.to(lse_ptr.dtype.element_ty))
+        if save_kl:
+            tl.store(loss_ptr+M, kl.to(loss_ptr.dtype.element_ty))
+    else:
+        # store 0
+        tl.store(loss_ptr, 0.0)
+        if save_kl:
+            tl.store(loss_ptr+M, 0.0)
+
+
+@triton.autotune(
+    [triton.Config({'BLOCK_SIZE': BLOCK_SIZE},  num_warps=NUM_WARPS, num_stages=NUM_STAGES)
+     for BLOCK_SIZE in [1024, 2048, 4096, 8192]
+     for NUM_WARPS in [8, 16, 32]
+     for NUM_STAGES in [1, 2, 4]
+     ], key=['B', 'N']
+)
+@triton.jit
+def grpo_bwd_kernel(
+    dloss_ptr,
+    dlogits_ptr,
+    logits_ptr,
+    ref_logp_ptr,
+    input_ids_ptr,
+    advantages_ptr,
+    completion_mask_ptr,
+    lse_ptr,
+    beta,
+    B,
+    N,
+    L,
+    start_idx,
+    BLOCK_SIZE: tl.constexpr
+):
+
+    row_idx = tl.program_id(0)  # B*L
+    off_b = row_idx // L
+
+    N = tl.cast(N, tl.int64)
+
+    dlogits_ptr += N * (row_idx + off_b)
+    base_cols = tl.arange(0, BLOCK_SIZE)
+    completion_mask_ptr += row_idx
+    not_skip = tl.load(completion_mask_ptr).to(tl.int1)
+
+    if not_skip == 1:
+        lse_ptr += row_idx
+        dloss_ptr += row_idx
+        advantages_ptr += off_b
+        ref_logp_ptr += row_idx
+        logits_ptr += N * (row_idx + off_b)
+        input_ids_ptr += row_idx + (off_b+1) * start_idx
+        dloss = tl.load(dloss_ptr).to(tl.float32)
+        lse = tl.load(lse_ptr).to(tl.float32)
+        idx = tl.load(input_ids_ptr)
+        x = tl.load(logits_ptr+idx).to(tl.float32)
+        advantage = tl.load(advantages_ptr).to(tl.float32)
+        ref_logp = tl.load(ref_logp_ptr)
+        logp = x - lse
+
+        dlogp = (beta * (-1.0 * exp(ref_logp - logp) + 1)
+                 - advantage) * dloss
+
+        for start_n in tl.range(0, N, BLOCK_SIZE):
+            cols = start_n + base_cols
+            mask = cols < N
+            logits = tl.load(logits_ptr+cols, mask=mask, other=-float('inf')).to(tl.float32)
+            probs = exp(logits - lse)
+            dlogits = tl.where(cols == idx, 1-probs, -probs) * dlogp
+
+            tl.store(dlogits_ptr+cols, dlogits.to(dlogits_ptr.dtype.element_ty), mask=mask)
+    else:
+        dlogits = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+        for start_n in tl.range(0, N, BLOCK_SIZE):
+            cols = start_n + base_cols
+            mask = cols < N
+
+            tl.store(dlogits_ptr+cols, dlogits.to(dlogits_ptr.dtype.element_ty), mask=mask)
+
+
+class GrpoLoss(torch.autograd.Function):
+
+    @input_guard
+    @staticmethod
+    def forward(ctx, logits, ref_logp, input_ids, advantages, beta, completion_mask, save_kl):
+        ctx.input_shape = logits.shape
+        B, L_ADD_1, N = ctx.input_shape
+        L = L_ADD_1 - 1
+        M = B * L
+        input_ids_start_index = input_ids.size(1) - L
+
+        if not save_kl:
+            loss = torch.empty(B, L, device=logits.device, dtype=torch.float32)
+        else:
+            loss = torch.empty(B*2, L, device=logits.device, dtype=torch.float32)
+
+        lse = torch.empty(B, L, device=logits.device, dtype=torch.float32)
+
+        if completion_mask is None:
+            completion_mask = torch.ones(B, L, device=logits.device, dtype=torch.int32)
+        else:
+            loss[:B].masked_fill_(completion_mask.logical_not(), 0.0)
+
+        grpo_fwd_kernel[(M,)](
+            logits_ptr=logits,
+            ref_logp_ptr=ref_logp,
+            input_ids_ptr=input_ids,
+            advantages_ptr=advantages,
+            completion_mask_ptr=completion_mask,
+            loss_ptr=loss,
+            lse_ptr=lse,
+            beta=beta,
+            save_kl=save_kl,
+            B=B, M=M, N=N, L=L,
+            start_idx=input_ids_start_index,
+        )
+        ctx.beta = beta
+        ctx.save_for_backward(lse, logits, input_ids, advantages, completion_mask)
+        ctx.ref_logp = ref_logp
+        return loss
+
+    @input_guard
+    @staticmethod
+    def backward(ctx, dloss):
+        # The grad of logits comes from two parts, the reward part and the kl part
+        lse, logits, input_ids, advantages, completion_mask = ctx.saved_tensors
+        B, L_ADD_1, N = ctx.input_shape
+        L = L_ADD_1 - 1
+        M = B * L
+
+        input_ids_start_index = input_ids.size(1) - L
+
+        dlogits = torch.empty_like(logits)  # B, L_ADD_1, N
+
+        grpo_bwd_kernel[(M,)](
+            dloss_ptr=dloss,
+            dlogits_ptr=dlogits,
+            logits_ptr=logits,
+            ref_logp_ptr=ctx.ref_logp,
+            input_ids_ptr=input_ids,
+            advantages_ptr=advantages,
+            completion_mask_ptr=completion_mask,
+            lse_ptr=lse,
+            beta=ctx.beta,
+            B=B, N=N, L=L,
+            start_idx=input_ids_start_index,
+        )
+        # The last token in the completion is not used in the loss computation
+        # and therefore its gradient should be set to 0
+        dlogits[:, -1, :].fill_(0.0)
+        return dlogits.view(*ctx.input_shape), None, None, None, None, None, None
+
+
+def fused_grpo_loss(logits, ref_logp, input_ids, advantages, beta=0.1, completion_mask=None, save_kl=False) -> torch.Tensor:
+    '''
+    compute grpo loss, save memory(no addition usage) and fast speed(6X for A800)
+
+    Args:
+        logtits: Tensor, [B, L+1, vocab_size], the origin output of model, it's not logits[:, :-1]
+        ref_logp: Tensor, [B, L], the origin output of model, it's not ref_logits[:, :-1]
+        input_ids: Tensor, [B, K+L], it's prompt_completion_id, it contains the prompt ids and output ids
+        advantages: Tensor, [B], the advantages of each prompt
+        beta: float, the weight of kl loss
+        completion_mask: Tensor, loss mask
+        save_kl: bool, if true will save kl
+
+    Retutn:
+        loss: Tensor, [B, L], the loss of grpo, it contains the advantage part and kl part
+
+    NOTE: logits(ref_logits) is computed by these steps
+        logits_to_keep = completion_ids.size(1)
+
+        def get_per_token_logits(model, input_ids, attention_mask, logits_to_keep):
+            # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
+            logits = model(
+                input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1
+            ).logits
+            return logits
+
+        logits = get_per_token_logits(model, prompt_completion_ids, attention_mask, logits_to_keep)
+    '''
+    out = GrpoLoss.apply(logits, ref_logp, input_ids, advantages, beta, completion_mask, save_kl)
+    if not save_kl:
+        return out
+    else:
+        return out.chunk(2, axis=0)
+
+
+def grpo_loss_torch(logits, ref_logp, input_ids, advantages, beta=0.1, completion_mask=None, save_kl=False):
+    def get_log_probs(logits, input_ids):
+        per_token_logps = []
+        for logits_row, input_ids_row in zip(logits, input_ids[:, -logits.size(1):]):
+            log_probs = logits_row.log_softmax(dim=-1)
+            token_log_prob = torch.gather(log_probs, dim=1, index=input_ids_row.unsqueeze(1)).squeeze(1)
+            per_token_logps.append(token_log_prob)
+        return torch.stack(per_token_logps)
+
+    logits = logits[:, :-1]
+    per_token_logps = get_log_probs(logits, input_ids)
+    ref_per_token_logps = ref_logp
+    per_token_kl = torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1
+
+    per_token_loss = torch.exp(per_token_logps - per_token_logps.detach()) * advantages.unsqueeze(1)
+    per_token_loss = -(per_token_loss - beta * per_token_kl)
+    if completion_mask is not None:
+        per_token_loss *= completion_mask
+        if save_kl:
+            per_token_kl *= completion_mask
+    return per_token_loss if not save_kl else (per_token_loss, per_token_kl)
+
+
+@torch.compile(fullgraph=True)
+def grpo_loss_with_old_logps(
+    logps: torch.Tensor,
+    ref_logps: torch.Tensor,
+    old_logps: torch.Tensor,
+    pad_mask: torch.Tensor,
+    logits_to_keep: int,
+    rewards: torch.Tensor,
+    beta: float = 0.2,
+    epsilon: float = 0.2
+):
+    """
+    Compute the GRPO (Group Relative Policy Optimization) loss.
+
+    Args:
+        logps (torch.Tensor): [Batch, Token_length] Log probabilities of the current policy.
+        ref_logps (torch.Tensor):[Batch, Token_length]  Log probabilities of the reference policy.
+        old_logps (torch.Tensor): [Batch, Token_length] Log probabilities of the old policy.
+        completion_ids (torch.Tensor): [Batch, Token_length] Completion token IDs (bool).
+        pad_token_id: Pad token ID.
+        logits_to_keep (int): Number of logits to keep for masking.
+        rewards (torch.Tensor): [Batch] Rewards for each generation.
+        beta (float) = 0.2: A hyperparameter for weighting the KL divergence term.
+        epsilon (float) = 0.2: An float hyperparameter for clipping the importance weights.
+
+    Returns:
+        torch.Tensor: The computed GRPO loss.
+    """
+    B = logps.shape[0]
+    assert B > 1, "Batch * Num generations should be greater than 1"
+
+    rewards_shaped = rewards.view(-1, B)  # B,num_generations
+    advantages = (rewards_shaped - rewards_shaped.mean(dim=1, keepdim=True)) / \
+        (rewards_shaped.std(dim=1, keepdim=True) + 1e-8)
+    advantages = advantages.view(-1)  # B*num_generations
+    # Calculate the per - token KL divergence
+    per_token_kl = torch.exp(ref_logps - logps) - (ref_logps - logps) - 1
+
+    # Calculate the ratio of probabilities (importance weights)
+    # Importance weights are calculated as exp(log_pi_theta - log_pi_theta_old)
+    importance_weights = torch.exp(logps - old_logps)
+
+    # Clip the importance weights to the range [1 - epsilon, 1 + epsilon]
+    importance_weights_clipped = torch.clamp(importance_weights, 1 - epsilon, 1 + epsilon)
+
+    # Create a completion mask. It checks which positions are valid based on logits_to_keep
+    completion_mask = torch.arange(logits_to_keep, device=logps.device)[None, :] >= 0
+
+    # Combine the completion mask and padding mask
+    completion_mask = completion_mask & pad_mask  # Ensure matching shape
+
+    # Add an extra dimension to advantages to match the shape for element - wise multiplication
+    advantages = advantages.unsqueeze(1)
+
+    # Calculate the per - token loss. It takes the minimum of the unclipped and clipped importance weights
+    # and subtracts the KL divergence term weighted by beta, then multiplies by the completion mask
+    token_loss = -(torch.min(advantages * importance_weights, advantages *
+                   importance_weights_clipped) - beta * per_token_kl) * completion_mask
+
+    # Calculate the final loss by summing the token losses and normalizing by the number of valid tokens
+    loss = -token_loss.sum() / completion_mask.sum()
+
+    return loss

fla/modules/l2norm.py

@@ -0,0 +1,176 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from fla.utils import input_guard
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+    ],
+    key=['N']
+)
+@triton.jit
+def l2norm_fwd_kernel(
+    X,
+    Y,
+    N,
+    eps,
+    BLOCK_N: tl.constexpr,
+):
+    i_m = tl.program_id(0)
+    X += i_m * N
+    Y += i_m * N
+    # Compute mean and variance
+    cols = tl.arange(0, BLOCK_N)
+    mask = cols < N
+    x = tl.load(X + cols, mask=mask, other=0.0).to(tl.float32)
+    xbar = tl.where(mask, x, 0.0)
+    var = tl.sum(xbar * xbar, axis=0)
+    rstd = 1 / tl.sqrt(var + eps)
+    # tl.store(Rstd + i_m, rstd)
+    # Normalize and apply linear transformation
+    y = x * rstd
+    # Write output
+    tl.store(Y + cols, y, mask=mask)
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+    ],
+    key=['N']
+)
+@triton.jit
+def l2norm_bwd_kernel(
+    X,
+    DY,
+    DX,
+    N,
+    eps,
+    BLOCK_N: tl.constexpr,
+):
+    i_m = tl.program_id(0)
+    X += i_m * N
+    DX += i_m * N
+    DY += i_m * N
+
+    # Y += i_m * stride_y_row
+    cols = tl.arange(0, BLOCK_N)
+    mask = cols < N
+    x = tl.load(X + cols, mask=mask, other=0.0).to(tl.float32)
+    x = tl.where(mask, x, 0.0)
+    var = tl.sum(x * x)
+    rstd = 1 / tl.sqrt(var + eps)
+    # tl.store(Rstd + i_m, rstd)
+    # Normalize and apply linear transformation
+    # y = x * rstd
+    dy = tl.load(DY + cols, mask=mask, other=0.0).to(tl.float32)
+    dy = tl.where(mask, dy, 0.0)
+    dx = dy * rstd - tl.sum(dy * x) * (1 / (var+eps)) * rstd * x
+    tl.store(DX + cols, dx, mask=mask)
+
+
+def l2norm_fwd(
+    x: torch.Tensor,
+    eps: float = 1e-6,
+    output_dtype: Optional[torch.dtype] = None
+):
+    x_shape_og = x.shape
+    x = x.reshape(-1, x.shape[-1])
+    # allocate output
+    if output_dtype is None:
+        y = torch.empty_like(x)
+    else:
+        y = torch.empty_like(x, dtype=output_dtype)
+    assert y.stride(-1) == 1
+    N = x.shape[-1]
+    M = x.shape[0]
+    # rstd = torch.empty((M,), dtype=torch.float32, device=x.device)
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    # heuristics for number of warps
+    l2norm_fwd_kernel[(M,)](
+        x,
+        y,
+        N,
+        eps,
+        BLOCK_N,
+    )
+    return y.reshape(x_shape_og)
+
+
+def l2norm_bwd(
+    x: torch.Tensor,
+    dy: torch.Tensor,
+    eps: float = 1e-5
+):
+    x_shape_og = x.shape
+    x = x.reshape(-1, dy.shape[-1])
+    dy = dy.reshape(-1, dy.shape[-1])
+    if dy.stride(-1) != 1:
+        dy = dy.contiguous()
+    assert dy.shape == x.shape
+    # allocate output
+    dx = torch.empty_like(x)
+    M = x.shape[0]
+    N = x.shape[-1]
+    # rstd = torch.empty((M,), dtype=torch.float32, device=x.device)
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    # heuristics for number of warps
+    l2norm_bwd_kernel[(M,)](
+        x,
+        dy,
+        dx,
+        N,
+        eps,
+        BLOCK_N,
+    )
+    return dx.reshape(x_shape_og)
+
+
+class L2NormFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(
+        ctx,
+        x,
+        eps=1e-6,
+        output_dtype=None
+    ):
+        y = l2norm_fwd(x, eps, output_dtype)
+        ctx.eps = eps
+        ctx.x_dtype = x.dtype
+        ctx.save_for_backward(x)
+        return y
+
+    @staticmethod
+    @input_guard
+    def backward(ctx, dy):
+        x, = ctx.saved_tensors
+        dx = l2norm_bwd(x, dy, ctx.eps)
+        return dx, None, None
+
+
+def l2_norm(
+    x: torch.Tensor,
+    eps: float = 1e-6,
+    output_dtype: Optional[torch.dtype] = None
+) -> torch.Tensor:
+    return L2NormFunction.apply(x, eps, output_dtype)