feat: support rope

Signed-off-by: jupyterjazz <saba.sturua@jina.ai>

mha.py

@@ -1,6 +1,3 @@
-# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py
-# Commit id: 6bbc532388e61185a92e2a563126739967b4c8c5
-
 # Copyright (c) 2023, Tri Dao.
 
 import math
@@ -10,6 +7,8 @@ import torch
 import torch.nn as nn
 from einops import rearrange, repeat
 
+from flash_attn.utils.distributed import get_dim_for_local_rank
+
 try:
     from flash_attn import (
         flash_attn_kvpacked_func,
@@ -28,10 +27,7 @@ try:
 except ImportError:
     FusedDense, ColumnParallelLinear, RowParallelLinear = None, None, None
 
-try:
-    from flash_attn.layers.rotary import RotaryEmbedding
-except ImportError:
-    RotaryEmbedding = None
+from .rotary import RotaryEmbedding
 
 
 # From https://github.com/ofirpress/attention_with_linear_biases/blob/4b92f28a005ead2567abe2359f633e73e08f3833/fairseq/models/transformer.py#L742
@@ -62,15 +58,7 @@ class FlashSelfAttention(nn.Module):
                            (default: 0.0)
     """
 
-    def __init__(
-        self,
-        causal=False,
-        softmax_scale=None,
-        attention_dropout=0.0,
-        window_size=(-1, -1),
-        alibi_slopes=None,
-        deterministic=False,
-    ):
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0, alibi_slopes=None, deterministic=False):
         super().__init__()
         assert flash_attn_varlen_qkvpacked_func is not None, "FlashAttention is not installed"
         assert flash_attn_qkvpacked_func is not None, "FlashAttention is not installed"
@@ -78,7 +66,6 @@ class FlashSelfAttention(nn.Module):
         self.softmax_scale = softmax_scale
         self.drop = nn.Dropout(attention_dropout)
         self.register_buffer("alibi_slopes", alibi_slopes, persistent=False)
-        self.window_size = window_size
         self.deterministic = deterministic
 
     def forward(self, qkv, causal=None, cu_seqlens=None, max_seqlen=None):
@@ -102,8 +89,6 @@ class FlashSelfAttention(nn.Module):
         assert qkv.is_cuda
         causal = self.causal if causal is None else causal
         unpadded = cu_seqlens is not None
-        if self.alibi_slopes is not None:
-            self.alibi_slopes = self.alibi_slopes.to(torch.float32)
         if unpadded:
             assert cu_seqlens.dtype == torch.int32
             assert max_seqlen is not None
@@ -116,7 +101,6 @@ class FlashSelfAttention(nn.Module):
                 softmax_scale=self.softmax_scale,
                 causal=causal,
                 alibi_slopes=self.alibi_slopes,
-                window_size=self.window_size,
                 deterministic=self.deterministic,
             )
         else:
@@ -126,7 +110,6 @@ class FlashSelfAttention(nn.Module):
                 softmax_scale=self.softmax_scale,
                 causal=causal,
                 alibi_slopes=self.alibi_slopes,
-                window_size=self.window_size,
                 deterministic=self.deterministic,
             )
 
@@ -142,15 +125,7 @@ class FlashCrossAttention(nn.Module):
                            (default: 0.0)
     """
 
-    def __init__(
-        self,
-        causal=False,
-        softmax_scale=None,
-        attention_dropout=0.0,
-        alibi_slopes=None,
-        window_size=(-1, -1),
-        deterministic=False,
-    ):
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0, alibi_slopes=None, deterministic=False):
         super().__init__()
         assert flash_attn_varlen_kvpacked_func is not None, "FlashAttention is not installed"
         assert flash_attn_kvpacked_func is not None, "FlashAttention is not installed"
@@ -158,7 +133,6 @@ class FlashCrossAttention(nn.Module):
         self.softmax_scale = softmax_scale
         self.drop = nn.Dropout(attention_dropout)
         self.register_buffer("alibi_slopes", alibi_slopes, persistent=False)
-        self.window_size = window_size
         self.deterministic = deterministic
 
     def forward(
@@ -188,8 +162,6 @@ class FlashCrossAttention(nn.Module):
         assert q.is_cuda and kv.is_cuda
         causal = self.causal if causal is None else causal
         unpadded = cu_seqlens is not None
-        if self.alibi_slopes is not None:
-            self.alibi_slopes = self.alibi_slopes.to(torch.float32)
         if unpadded:
             assert cu_seqlens.dtype == torch.int32
             assert max_seqlen is not None
@@ -209,7 +181,6 @@ class FlashCrossAttention(nn.Module):
                 softmax_scale=self.softmax_scale,
                 causal=causal,
                 alibi_slopes=self.alibi_slopes,
-                window_size=self.window_size,
                 deterministic=self.deterministic,
             )
         else:
@@ -223,7 +194,6 @@ class FlashCrossAttention(nn.Module):
                 causal=causal,
                 softmax_scale=self.softmax_scale,
                 alibi_slopes=self.alibi_slopes,
-                window_size=self.window_size,
                 deterministic=self.deterministic,
             )
 
@@ -399,7 +369,6 @@ class MHA(nn.Module):
         rotary_emb_scale_base=None,
         rotary_emb_interleaved=False,
         use_alibi=False,
-        window_size=(-1, -1),
         fused_bias_fc=False,
         use_flash_attn=False,
         return_residual=False,
@@ -429,8 +398,6 @@ class MHA(nn.Module):
             alibi_slopes = torch.tensor(get_alibi_slopes(num_heads), device=device)
         else:
             alibi_slopes = None
-        if window_size != (-1, -1):
-            assert use_flash_attn, "Local (sliding window) attention code path requires flash_attn"
 
         self.num_heads = num_heads
         self.num_heads_kv = num_heads_kv if num_heads_kv is not None else num_heads
@@ -461,12 +428,12 @@ class MHA(nn.Module):
         )
         wqkv_cls = linear_cls if not self.return_residual else linear_resid_cls
         inner_attn_cls = (
-            partial(FlashSelfAttention, alibi_slopes=alibi_slopes, window_size=window_size)
+            partial(FlashSelfAttention, alibi_slopes=alibi_slopes)
             if use_flash_attn
             else SelfAttention
         )
         inner_cross_attn_cls = (
-            partial(FlashCrossAttention, alibi_slopes=alibi_slopes, window_size=window_size)
+            partial(FlashCrossAttention, alibi_slopes=alibi_slopes)
             if use_flash_attn
             else CrossAttention
         )
@@ -619,7 +586,7 @@ class MHA(nn.Module):
             assert key_padding_mask is None
             assert self.use_flash_attn
             assert not self.dwconv
-            assert self.rotary_emb_dim == 0
+            # assert self.rotary_emb_dim == 0
         if key_padding_mask is not None:
             assert cu_seqlens is None
             assert max_seqlen is None
@@ -643,7 +610,9 @@ class MHA(nn.Module):
                 else inference_params.seqlen_offset
             )
         )
-        rotary_max_seqlen = inference_params.max_seqlen if inference_params is not None else None
+        rotary_max_seqlen = (
+            inference_params.max_sequence_len if inference_params is not None else max_seqlen
+        )
         batch, seqlen = x.shape[:2]
         if not self.cross_attn and self.num_heads_kv == self.num_heads:
             assert x_kv is None and mixer_subset is None
@@ -664,7 +633,10 @@ class MHA(nn.Module):
             ):
                 if self.rotary_emb_dim > 0:
                     qkv = self.rotary_emb(
-                        qkv, seqlen_offset=seqlen_offset, max_seqlen=rotary_max_seqlen
+                        qkv,
+                        seqlen_offset=seqlen_offset,
+                        cu_seqlens=cu_seqlens,
+                        max_seqlen=rotary_max_seqlen,
                     )
                 if inference_params is None:
                     if not self.checkpointing:
@@ -715,7 +687,11 @@ class MHA(nn.Module):
             ):
                 if self.rotary_emb_dim > 0:
                     q, kv = self.rotary_emb(
-                        q, kv, seqlen_offset=seqlen_offset, max_seqlen=rotary_max_seqlen
+                        q,
+                        kv,
+                        seqlen_offset=seqlen_offset,
+                        cu_seqlens=cu_seqlens,
+                        max_seqlen=rotary_max_seqlen,
                     )
                 if inference_params is None:
                     if not self.checkpointing:
@@ -731,3 +707,315 @@ class MHA(nn.Module):
         out = self.out_proj(rearrange(context, "... h d -> ... (h d)"))
         return out if not self.return_residual else (out, x)
 
+
+class ParallelMHA(nn.Module):
+    """Multi-head self-attention and cross-attention"""
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        process_group,
+        num_heads_kv=None,
+        qkv_proj_bias=True,
+        out_proj_bias=True,
+        dropout=0.0,
+        softmax_scale=None,
+        causal=False,
+        layer_idx=None,
+        rotary_emb_dim=0,
+        rotary_emb_base=10000.0,
+        rotary_emb_scale_base=None,
+        rotary_emb_interleaved=False,
+        use_alibi=False,
+        use_flash_attn=False,
+        checkpointing=False,
+        sequence_parallel=True,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.causal = causal
+        self.layer_idx = layer_idx
+        self.rotary_emb_dim = rotary_emb_dim
+        self.use_flash_attn = use_flash_attn
+        self.checkpointing = checkpointing
+        self.process_group = process_group
+        self.world_size = process_group.size()
+        self.local_rank = torch.distributed.get_rank(process_group)
+
+        self.num_heads = num_heads
+        assert self.embed_dim % self.num_heads == 0, "embed_dim must be divisible by num_heads"
+
+        self.num_heads_kv = num_heads_kv if num_heads_kv is not None else num_heads
+        assert (
+            self.num_heads % self.num_heads_kv == 0
+        ), "num_heads must be divisible by num_heads_kv"
+
+        self.num_heads_per_rank = get_dim_for_local_rank(
+            self.num_heads, self.world_size, self.local_rank
+        )
+        self.num_heads_kv_per_rank = get_dim_for_local_rank(
+            self.num_heads_kv, self.world_size, self.local_rank
+        )
+        self.head_dim = self.embed_dim // num_heads
+        qkv_dim = self.head_dim * (self.num_heads + 2 * self.num_heads_kv)
+
+        if use_alibi:
+            assert use_flash_attn, "ALiBi code path requires flash_attn"
+            num_heads_local = math.ceil(self.num_heads / self.world_size)
+            alibi_slopes = torch.tensor(
+                get_alibi_slopes(num_heads)[
+                    self.local_rank * num_heads_local : (self.local_rank + 1) * num_heads_local
+                ],
+                device=device,
+            )
+        else:
+            alibi_slopes = None
+
+        if self.rotary_emb_dim > 0:
+            assert RotaryEmbedding is not None, "rotary_emb is not installed"
+            self.rotary_emb = RotaryEmbedding(
+                self.rotary_emb_dim,
+                base=rotary_emb_base,
+                scale_base=rotary_emb_scale_base,
+                interleaved=rotary_emb_interleaved,
+                device=device,
+            )
+
+        if ColumnParallelLinear is None or RowParallelLinear is None:
+            raise ImportError("fused_dense is not installed")
+        self.Wqkv = ColumnParallelLinear(
+            embed_dim,
+            qkv_dim,
+            process_group,
+            bias=qkv_proj_bias,
+            sequence_parallel=sequence_parallel,
+            multiple_of=self.head_dim * (self.num_heads // self.num_heads_kv + 2),
+            **factory_kwargs,
+        )
+        inner_attn_cls = (
+            partial(FlashSelfAttention, alibi_slopes=alibi_slopes)
+            if use_flash_attn
+            else SelfAttention
+        )
+        inner_cross_attn_cls = (
+            partial(FlashCrossAttention, alibi_slopes=alibi_slopes)
+            if use_flash_attn
+            else CrossAttention
+        )
+        self.inner_attn = inner_attn_cls(
+            causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout
+        )
+        self.inner_cross_attn = inner_cross_attn_cls(
+            causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout
+        )
+        self.out_proj = RowParallelLinear(
+            embed_dim,
+            embed_dim,
+            process_group,
+            bias=out_proj_bias,
+            sequence_parallel=sequence_parallel,
+            multiple_of=self.head_dim,
+            **factory_kwargs,
+        )
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None):
+        dtype = self.out_proj.weight.dtype if dtype is None else dtype
+        device = self.out_proj.weight.device
+        return torch.empty(
+            batch_size,
+            max_seqlen,
+            2,
+            self.num_heads_kv_per_rank,
+            self.head_dim,
+            dtype=dtype,
+            device=device,
+        )
+
+    def _update_kv_cache(self, kv, inference_params):
+        """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)"""
+        assert self.layer_idx is not None, "Generation requires layer_idx in the constructor"
+        return _update_kv_cache(kv, inference_params, self.layer_idx)
+
+    def _apply_rotary_update_kvcache_attention(self, q, kv, inference_params):
+        """
+        Fast path that combine 3 steps: apply rotary to Q and K, update kv cache, and apply attention.
+        q: (batch_size, seqlen_q, nheads, head_dim)
+        kv: (batch_size, seqlen_k, 2, nheads_kv, head_dim)
+        """
+        assert inference_params is not None and inference_params.seqlen_offset > 0
+        assert self.use_flash_attn
+        if self.rotary_emb_dim > 0:
+            assert self.rotary_emb.scale is None, "This code path does not support xPos"
+            self.rotary_emb._update_cos_sin_cache(
+                inference_params.max_seqlen, device=q.device, dtype=q.dtype
+            )
+            rotary_cos, rotary_sin = self.rotary_emb._cos_cached, self.rotary_emb._sin_cached
+        else:
+            rotary_cos, rotary_sin = None, None
+        batch = q.shape[0]
+        kv_cache = inference_params.key_value_memory_dict[self.layer_idx][:batch]
+        cache_seqlens = (
+            inference_params.lengths_per_sample[:batch]
+            if inference_params.lengths_per_sample is not None
+            else inference_params.seqlen_offset
+        )
+        alibi_slopes = getattr(self.inner_cross_attn, "alibi_slopes", None)
+        context = flash_attn_with_kvcache(
+            q,
+            kv_cache[:, :, 0],
+            kv_cache[:, :, 1],
+            kv[:, :, 0],
+            kv[:, :, 1],
+            rotary_cos=rotary_cos,
+            rotary_sin=rotary_sin,
+            cache_seqlens=cache_seqlens,
+            softmax_scale=self.inner_cross_attn.softmax_scale,
+            causal=self.inner_cross_attn.causal,
+            rotary_interleaved=self.rotary_emb.interleaved if self.rotary_emb_dim > 0 else False,
+            alibi_slopes=alibi_slopes,
+        )
+        return context
+
+    def _update_kvcache_attention(self, q, kv, inference_params):
+        """Write kv to inference_params, then do attention"""
+        if inference_params.seqlen_offset == 0 or not self.use_flash_attn:
+            # TODO: this only uses seqlen_offset and not lengths_per_sample.
+            kv = self._update_kv_cache(kv, inference_params)
+            return self.inner_cross_attn(q, kv)
+        else:
+            batch = q.shape[0]
+            kv_cache = inference_params.key_value_memory_dict[self.layer_idx][:batch]
+            cache_seqlens = (
+                inference_params.lengths_per_sample[:batch]
+                if inference_params.lengths_per_sample is not None
+                else inference_params.seqlen_offset
+            )
+            alibi_slopes = getattr(self.inner_cross_attn, "alibi_slopes", None)
+            context = flash_attn_with_kvcache(
+                q,
+                kv_cache[:, :, 0],
+                kv_cache[:, :, 1],
+                kv[:, :, 0],
+                kv[:, :, 1],
+                cache_seqlens=cache_seqlens,
+                softmax_scale=self.inner_cross_attn.softmax_scale,
+                causal=self.inner_cross_attn.causal,
+                alibi_slopes=alibi_slopes,
+            )
+            return context
+
+    def forward(
+        self, x, seqlen=None, inference_params=None, cu_seqlens=None, max_seqlen=None, **kwargs
+    ):
+        """
+        Arguments:
+            x: (batch, seqlen, hidden_dim) (where hidden_dim = num heads * head dim) if seqlen=None and cu_seqlens=None.
+               (seqlen, hidden_dim) if cu_seqlens not None, seqlen equal cu_seqlens[-1].
+                If seqlen is not None and cu_seqlens=None, x is (batch * seqlen, hidden_dim). This is so that when we
+                split x during sequence parallel, we split the batch * seqlen dimension
+                (in case batch is small).
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into x. Only applicable when using
+                FlashAttention.
+            max_seqlen: int. Maximum sequence length in the batch.
+        """
+        if cu_seqlens is not None:
+            assert max_seqlen is not None
+            assert seqlen is None
+            assert self.use_flash_attn
+        if inference_params is not None:
+            assert cu_seqlens is None and max_seqlen is None
+        qkv = self.Wqkv(x)
+        if seqlen is not None:
+            qkv = rearrange(qkv, "(b s) ... -> b s ...", s=seqlen)
+        kwargs = (
+            {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen, **kwargs}
+            if self.use_flash_attn
+            else kwargs
+        )
+        seqlen_offset = (
+            0
+            if inference_params is None
+            else (
+                inference_params.lengths_per_sample
+                if inference_params.lengths_per_sample is not None
+                else inference_params.seqlen_offset
+            )
+        )
+        rotary_max_seqlen = (
+            inference_params.max_sequence_len if inference_params is not None else max_seqlen
+        )
+        if self.num_heads_kv == self.num_heads:
+            qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+            if (
+                inference_params is None
+                or inference_params.seqlen_offset == 0
+                or (self.rotary_emb_dim == 0 or self.rotary_emb_dim % 16 != 0)
+                or not self.use_flash_attn
+            ):
+                if self.rotary_emb_dim > 0:
+                    qkv = self.rotary_emb(
+                        qkv,
+                        seqlen_offset=seqlen_offset,
+                        cu_seqlens=cu_seqlens,
+                        max_seqlen=rotary_max_seqlen,
+                    )
+                if inference_params is None:
+                    if not self.checkpointing:
+                        context = self.inner_attn(qkv, **kwargs)
+                    else:
+                        context = torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, **kwargs)
+                else:
+                    context = self._update_kvcache_attention(
+                        qkv[:, :, 0], qkv[:, :, 1:], inference_params
+                    )
+            else:
+                context = self._apply_rotary_update_kvcache_attention(
+                    qkv[:, :, 0], qkv[:, :, 1:], inference_params
+                )
+        else:
+            q = rearrange(
+                qkv[..., : self.num_heads_per_rank * self.head_dim],
+                "... (h d) -> ... h d",
+                d=self.head_dim,
+            )
+            kv = rearrange(
+                qkv[..., self.num_heads_per_rank * self.head_dim :],
+                "... (two hkv d) -> ... two hkv d",
+                two=2,
+                d=self.head_dim,
+            )
+            if (
+                inference_params is None
+                or inference_params.seqlen_offset == 0
+                or (self.rotary_emb_dim == 0 or self.rotary_emb_dim % 16 != 0)
+                or not self.use_flash_attn
+            ):
+                if self.rotary_emb_dim > 0:
+                    q, kv = self.rotary_emb(
+                        q,
+                        kv,
+                        seqlen_offset=seqlen_offset,
+                        cu_seqlens=cu_seqlens,
+                        max_seqlen=rotary_max_seqlen,
+                    )
+                if inference_params is None:
+                    if not self.checkpointing:
+                        context = self.inner_cross_attn(q, kv, **kwargs)
+                    else:
+                        context = torch.utils.checkpoint.checkpoint(
+                            self.inner_cross_attn, q, kv, **kwargs
+                        )
+                else:
+                    context = self._update_kvcache_attention(q, kv, inference_params)
+            else:
+                context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
+        context = rearrange(context, "... h d -> ... (h d)")
+        if seqlen is not None:
+            context = rearrange(context, "b s d -> (b s) d")
+        out = self.out_proj(context)
+        return out

modeling_xlm_roberta.py

@@ -45,7 +45,7 @@ from .embedding import XLMRobertaEmbeddings
 from .mha import MHA
 from .mlp import FusedMLP, Mlp
 from .stochastic_depth import StochasticDepth
-
+from .rotary import RotaryEmbedding
 
 try:
     from flash_attn.ops.fused_dense import FusedDense
@@ -91,7 +91,7 @@ def create_mixer_cls(config, cross_attn=False, return_residual=False):
     rotary_kwargs = {}
     if config.position_embedding_type == "rotary":
         rotary_kwargs["rotary_emb_dim"] = getattr(
-            config, "rotary_emb_dim", config.hidden_size
+            config, "rotary_emb_dim", config.hidden_size / 12
         )
         rotary_kwargs["rotary_emb_base"] = getattr(config, "rotary_emb_base", 10000.0)
         rotary_kwargs["rotary_emb_scale_base"] = getattr(

rotary.py

@@ -0,0 +1,570 @@
+# Copyright (c) 2023, Tri Dao.
+
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+from einops import rearrange, repeat
+from flash_attn.ops.triton.rotary import apply_rotary
+
+
+def rotate_half(x, interleaved=False):
+    if not interleaved:
+        x1, x2 = x.chunk(2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+    else:
+        x1, x2 = x[..., ::2], x[..., 1::2]
+        return rearrange(torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2)
+
+
+def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
+    """
+    x: (batch_size, seqlen, nheads, headdim)
+    cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
+    """
+    ro_dim = cos.shape[-1] * 2
+    assert ro_dim <= x.shape[-1]
+    cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
+    sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
+    return torch.cat(
+        [x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]],
+        dim=-1,
+    )
+
+
+class ApplyRotaryEmb(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x,
+        cos,
+        sin,
+        interleaved=False,
+        inplace=False,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+    ):
+        out = apply_rotary(
+            x,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            interleaved=interleaved,
+            inplace=inplace,
+        )
+        if isinstance(seqlen_offsets, int):
+            ctx.save_for_backward(cos, sin, cu_seqlens)  # Can't save int with save_for_backward
+            ctx.seqlen_offsets = seqlen_offsets
+        else:
+            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
+            ctx.seqlen_offsets = None
+        ctx.interleaved = interleaved
+        ctx.inplace = inplace
+        ctx.max_seqlen = max_seqlen
+        return out if not inplace else x
+
+    @staticmethod
+    def backward(ctx, do):
+        seqlen_offsets = ctx.seqlen_offsets
+        if seqlen_offsets is None:
+            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
+        else:
+            cos, sin, cu_seqlens = ctx.saved_tensors
+        # TD [2023-09-02]: For some reason Triton (2.0.0.post1) errors with
+        # "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
+        if not ctx.interleaved and not ctx.inplace:
+            do = do.clone()
+        dx = apply_rotary(
+            do,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=ctx.max_seqlen,
+            interleaved=ctx.interleaved,
+            inplace=ctx.inplace,
+            conjugate=True,
+        )
+        return dx, None, None, None, None, None, None, None
+
+
+def apply_rotary_emb(
+    x,
+    cos,
+    sin,
+    interleaved=False,
+    inplace=False,
+    seqlen_offsets: Union[int, torch.Tensor] = 0,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[int] = None,
+):
+    """
+    Arguments:
+        x: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+        cos, sin: (seqlen_rotary, rotary_dim / 2)
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        inplace: if True, apply rotary embedding in-place.
+        seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Return:
+        out: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+    rotary_dim must be <= headdim
+    Apply rotary embedding to the first rotary_dim of x.
+    """
+    return ApplyRotaryEmb.apply(
+        x, cos, sin, interleaved, inplace, seqlen_offsets, cu_seqlens, max_seqlen
+    )
+
+
+# For backward compatibility
+apply_rotary_emb_func = apply_rotary_emb
+
+
+class ApplyRotaryEmbQKV_(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        qkv,
+        cos,
+        sin,
+        cos_k=None,
+        sin_k=None,
+        interleaved=False,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+    ):
+        # batch, seqlen, three, nheads, headdim = qkv.shape
+        assert qkv.shape[-3] == 3
+        if cos_k is None and sin_k is None and qkv.is_contiguous():
+            # Call 1 kernel instead of 2 kernels
+            # We need qkv to be contiguous so that when we reshape to combine (3, nheads)
+            # dimensions, we get the same tensor
+            qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
+            # qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
+            apply_rotary(
+                qk,
+                cos,
+                sin,
+                seqlen_offsets=seqlen_offsets,
+                interleaved=interleaved,
+                inplace=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+        else:
+            cos_k = cos if cos_k is None else cos_k
+            sin_k = sin if sin_k is None else sin_k
+            q, k = qkv[..., 0, :, :], qkv[..., 1, :, :]
+            apply_rotary(
+                q,
+                cos,
+                sin,
+                seqlen_offsets,
+                interleaved=interleaved,
+                inplace=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+            apply_rotary(
+                k,
+                cos_k,
+                sin_k,
+                seqlen_offsets,
+                interleaved=interleaved,
+                inplace=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+            ctx.save_for_backward(cos, sin, cos_k, sin_k)
+        if isinstance(seqlen_offsets, int):
+            ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens)
+            ctx.seqlen_offsets = seqlen_offsets
+        else:
+            ctx.save_for_backward(cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets)
+            ctx.seqlen_offsets = None
+        ctx.max_seqlen = max_seqlen
+        ctx.interleaved = interleaved
+        return qkv
+
+    @staticmethod
+    def backward(ctx, dqkv):
+        seqlen_offsets = ctx.seqlen_offsets
+        if seqlen_offsets is None:
+            cos, sin, cos_k, sin_k, cu_seqlens, seqlen_offsets = ctx.saved_tensors
+        else:
+            cos, sin, cos_k, sin_k, cu_seqlens = ctx.saved_tensors
+        if cos_k is None and sin_k is None and dqkv.is_contiguous():
+            # Call 1 kernel instead of 2 kernels
+            # We need dqkv to be contiguous so that when we reshape to combine (3, nheads)
+            # dimensions, we get the same tensor
+            dqk = rearrange(dqkv[..., :2, :, :], "... t h d -> ... (t h) d")
+            apply_rotary(
+                dqk,
+                cos,
+                sin,
+                seqlen_offsets=seqlen_offsets,
+                interleaved=ctx.interleaved,
+                inplace=True,
+                conjugate=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=ctx.max_seqlen,
+            )
+        else:
+            cos_k = cos if cos_k is None else cos_k
+            sin_k = sin if sin_k is None else sin_k
+            dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
+            apply_rotary(
+
+                dq,
+                cos,
+                sin,
+                seqlen_offsets,
+                interleaved=ctx.interleaved,
+                inplace=True,
+                conjugate=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=ctx.max_seqlen,
+            )
+            apply_rotary(
+                dk,
+                cos_k,
+                sin_k,
+                seqlen_offsets,
+                interleaved=ctx.interleaved,
+                inplace=True,
+                conjugate=True,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=ctx.max_seqlen,
+            )
+        return dqkv, None, None, None, None, None, None, None, None
+
+
+def apply_rotary_emb_qkv_(
+    qkv,
+    cos,
+    sin,
+    cos_k=None,
+    sin_k=None,
+    interleaved=False,
+    seqlen_offsets: Union[int, torch.Tensor] = 0,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[int] = None,
+):
+    """
+    Arguments:
+        qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, 3, nheads, headdim)
+        cos, sin: (seqlen, rotary_dim / 2)
+        cos_k, sin_k: (seqlen, rotary_dim / 2), optional
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
+            1st half and 2nd half (GPT-NeoX style).
+        seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+            cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Return:
+        qkv: (batch_size, seqlen, 3, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, 3, nheads, headdim)
+    rotary_dim must be <= headdim
+    Apply rotary embedding *inplace* to the first rotary_dim of Q and K.
+    """
+    return ApplyRotaryEmbQKV_.apply(
+        qkv, cos, sin, cos_k, sin_k, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
+    )
+
+
+class ApplyRotaryEmbKV_(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        kv,
+        cos,
+        sin,
+        interleaved=False,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+    ):
+        # batch, seqlen, two, nheads, headdim = kv.shape
+        assert kv.shape[-3] == 2
+        k = kv[..., 0, :, :]
+        apply_rotary(
+            k,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            interleaved=interleaved,
+            inplace=True,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )
+        if isinstance(seqlen_offsets, int):
+            ctx.save_for_backward(cos, sin, cu_seqlens)  # Can't save int with save_for_backward
+            ctx.seqlen_offsets = seqlen_offsets
+        else:
+            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
+            ctx.seqlen_offsets = None
+        ctx.max_seqlen = max_seqlen
+        ctx.interleaved = interleaved
+        return kv
+
+    @staticmethod
+    def backward(ctx, dkv):
+        seqlen_offsets = ctx.seqlen_offsets
+        if seqlen_offsets is None:
+            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
+        else:
+            cos, sin, cu_seqlens = ctx.saved_tensors
+        apply_rotary(
+            dkv[..., 0, :, :],
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            interleaved=ctx.interleaved,
+            inplace=True,
+            conjugate=True,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=ctx.max_seqlen,
+        )
+        return dkv, None, None, None, None, None, None
+
+
+apply_rotary_emb_kv_ = ApplyRotaryEmbKV_.apply
+
+
+def apply_rotary_emb_kv_(
+    kv,
+    cos,
+    sin,
+    interleaved=False,
+    seqlen_offsets: Union[int, torch.Tensor] = 0,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[int] = None,
+):
+    """
+    Arguments:
+        kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, 2, nheads, headdim)
+        cos, sin: (seqlen, rotary_dim / 2)
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead of
+            1st half and 2nd half (GPT-NeoX style).
+        seqlen_offsets: (batch_size,) or int. Each sequence in Q and K is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Return:
+        kv: (batch_size, seqlen, 2, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, 2, nheads, headdim)
+    rotary_dim must be <= headdim
+    Apply rotary embedding *inplace* to the first rotary_dim of K.
+    """
+    return ApplyRotaryEmbKV_.apply(
+        kv, cos, sin, interleaved, seqlen_offsets, cu_seqlens, max_seqlen
+    )
+
+
+class RotaryEmbedding(torch.nn.Module):
+    """
+    The rotary position embeddings from RoFormer_ (Su et. al).
+    A crucial insight from the method is that the query and keys are
+    transformed by rotation matrices which depend on the relative positions.
+
+    Other implementations are available in the Rotary Transformer repo_ and in
+    GPT-NeoX_, GPT-NeoX was an inspiration
+
+    .. _RoFormer: https://arxiv.org/abs/2104.09864
+    .. _repo: https://github.com/ZhuiyiTechnology/roformer
+    .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
+
+    If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
+    A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
+    Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        base=10000.0,
+        interleaved=False,
+        scale_base=None,
+        pos_idx_in_fp32=True,
+        device=None,
+    ):
+        """
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
+            otherwise they might be in lower precision.
+            This option was added because previously (before 2023-07-02), when we construct
+            the position indices, we use the dtype of self.inv_freq. In most cases this would
+            be fp32, but if the model is trained in pure bf16 (not mixed precision), then
+            self.inv_freq would be bf16, and the position indices are also in bf16.
+            Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
+            embeddings for some positions will coincide.
+            To maintain compatibility with models previously trained in pure bf16,
+            we add this option.
+        """
+        super().__init__()
+        self.dim = dim
+        self.base = float(base)
+        self.pos_idx_in_fp32 = pos_idx_in_fp32
+        # Generate and save the inverse frequency buffer (non trainable)
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.interleaved = interleaved
+        self.scale_base = scale_base
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            else None
+        )
+        self.register_buffer("scale", scale, persistent=False)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+
+    def _compute_inv_freq(self, device=None):
+        return 1.0 / (
+            self.base
+            ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim)
+        )
+
+    def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
+        # Reset the tables if the sequence length has changed,
+        # if we're on a new device (possibly due to tracing for instance),
+        # or if we're switching from inference mode to training
+        if (
+            seqlen > self._seq_len_cached
+            or self._cos_cached is None
+            or self._cos_cached.device != device
+            or self._cos_cached.dtype != dtype
+            or (self.training and self._cos_cached.is_inference())
+        ):
+            self._seq_len_cached = seqlen
+            # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
+            # And the output of arange can be quite large, so bf16 would lose a lot of precision.
+            # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
+            if self.pos_idx_in_fp32:
+                t = torch.arange(seqlen, device=device, dtype=torch.float32)
+                # We want fp32 here as well since inv_freq will be multiplied with t, and the output
+                # will be large. Having it in bf16 will lose a lot of precision and cause the
+                # cos & sin output to change significantly.
+                # We want to recompute self.inv_freq if it was not loaded in fp32
+                if self.inv_freq.dtype != torch.float32:
+                    inv_freq = self._compute_inv_freq(device=device)
+                else:
+                    inv_freq = self.inv_freq
+            else:
+                t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+                inv_freq = self.inv_freq
+            # Don't do einsum, it converts fp32 to fp16 under AMP
+            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            freqs = torch.outer(t, inv_freq)
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(dtype)
+                self._sin_cached = torch.sin(freqs).to(dtype)
+            else:
+                power = (
+                    torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
+                    - seqlen // 2
+                ) / self.scale_base
+                scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
+
+    def forward(
+        self,
+        qkv: torch.Tensor,
+        kv: Optional[torch.Tensor] = None,
+        seqlen_offset: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        qkv: (batch, seqlen, 3, nheads, headdim) if kv is none,
+             else it's just q of shape (batch, seqlen, nheads, headdim)
+        kv: (batch, seqlen, 2, nheads, headdim)
+        seqlen_offset: (batch_size,) or int. Each sequence in x is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+            If it's a tensor of shape (batch_size,), then to update the cos / sin cache, one
+            should pass in max_seqlen, which will update the cos / sin cache up to that length.
+        Apply rotary embedding *inplace* to qkv and / or kv.
+        """
+        if cu_seqlens is not None:
+            assert max_seqlen is not None
+        seqlen = qkv.shape[1] if max_seqlen is None else max_seqlen
+        if max_seqlen is not None:
+            self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
+        elif isinstance(seqlen_offset, int):
+            self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
+        if kv is None:
+            if self.scale is None:
+                return apply_rotary_emb_qkv_(
+                    qkv,
+                    self._cos_cached,
+                    self._sin_cached,
+                    interleaved=self.interleaved,
+                    seqlen_offsets=seqlen_offset,
+                    cu_seqlens=cu_seqlens,
+                    max_seqlen=max_seqlen,
+                )
+            else:
+                return apply_rotary_emb_qkv_(
+                    qkv,
+                    self._cos_cached,
+                    self._sin_cached,
+                    self._cos_k_cached,
+                    self._sin_k_cached,
+                    interleaved=self.interleaved,
+                    seqlen_offsets=seqlen_offset,
+                    cu_seqlens=cu_seqlens,
+                    max_seqlen=max_seqlen,
+                )
+        else:
+            q = qkv
+            q = apply_rotary_emb_func(
+                q,
+                self._cos_cached,
+                self._sin_cached,
+                interleaved=self.interleaved,
+                inplace=True,
+                seqlen_offsets=seqlen_offset,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+            if self.scale is None:
+                kv = apply_rotary_emb_kv_(
+                    kv,
+                    self._cos_cached,
+                    self._sin_cached,
+                    interleaved=self.interleaved,
+                    seqlen_offsets=seqlen_offset,
+                    cu_seqlens=cu_seqlens,
+                    max_seqlen=max_seqlen,
+                )
+            else:
+                kv = apply_rotary_emb_kv_(
+                    kv,
+                    self._cos_k_cached,
+                    self._sin_k_cached,
+                    interleaved=self.interleaved,
+                    seqlen_offsets=seqlen_offset,
+                    cu_seqlens=cu_seqlens,
+                    max_seqlen=max_seqlen,
+                )
+            return q, kv