Adds support for MQA/GQA and attention mask during training / fine-tuning.

README.md

@@ -94,7 +94,7 @@ with torch.autocast(model.device.type, dtype=torch.float16, enabled=True):
 ```
 
 **Remark.** In the generation function, our model currently does not support beam search (`num_beams` >1).
-Furthermore, in the forward pass of the model, we currently do not support attention mask during training, outputting hidden states or attention values, or using custom input embeddings (instead of the model's).
+Furthermore, in the forward pass of the model, we currently do not support outputting hidden states or attention values, or using custom input embeddings (instead of the model's).
 
 
 ### Citation

configuration_mixformer_sequential.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT license.
 
 import math
-from typing import Any, Dict, List, Optional, Union
+from typing import Optional
 
 from transformers import PretrainedConfig
 
@@ -27,6 +27,7 @@ class MixFormerSequentialConfig(PretrainedConfig):
         n_layer: Optional[int] = 20,
         n_inner: Optional[int] = None,
         n_head: Optional[int] = 16,
+        n_head_kv: Optional[int] = None,
         rotary_dim: Optional[int] = 32,
         activation_function: Optional[str] = "gelu_new",
         embd_pdrop: Optional[float] = 0.0,
@@ -43,6 +44,7 @@ class MixFormerSequentialConfig(PretrainedConfig):
         self.n_layer = n_layer
         self.n_inner = n_inner
         self.n_head = n_head
+        self.n_head_kv = n_head_kv
         self.rotary_dim = min(rotary_dim, n_embd // n_head)
         self.activation_function = activation_function
         self.embd_pdrop = embd_pdrop

modeling_mixformer_sequential.py

@@ -34,20 +34,20 @@
 from __future__ import annotations
 
 import math
-import copy
 from typing import Any, Dict, Optional, Tuple, Union
 from dataclasses import dataclass, field
 
 import torch
 import torch.nn as nn
 
-from einops import rearrange
+from einops import rearrange, repeat
 from transformers.activations import ACT2FN
 from transformers import PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
 from .configuration_mixformer_sequential import MixFormerSequentialConfig
 
+
 @dataclass
 class InferenceParams:
     """Inference parameters passed to model to efficiently calculate
@@ -57,21 +57,20 @@ class InferenceParams:
         https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
 
     Args:
-        max_sequence_len: Maximum sequence length.
+        max_seqlen: Maximum sequence length.
         max_batch_size: Maximum batch size.
-        sequence_len_offset: Sequence length offset.
+        seqlen_offset: Sequence length offset.
         batch_size_offset: Batch size offset.
         key_value_memory_dict: Key value memory dictionary.
-        fused_ft_kernel: Whether to use fused kernel for fast inference.
         lengths_per_sample: Lengths per sample.
 
     """
 
-    max_sequence_len: int = field(metadata={"help": "Maximum sequence length."})
+    max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
 
     max_batch_size: int = field(metadata={"help": "Maximum batch size."})
 
-    sequence_len_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
+    seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
 
     batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
 
@@ -79,8 +78,6 @@ class InferenceParams:
         default_factory=dict, metadata={"help": "Key value memory dictionary."}
     )
 
-    fused_ft_kernel: bool = field(default=False, metadata={"help": "Whether to use fused kernel for fast inference."})
-
     lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
 
 
@@ -103,12 +100,112 @@ class Embedding(nn.Module):
         return hidden_states
 
 
+def _apply_rotary_emb(
+    x: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+) -> torch.FloatTensor:
+    _, seqlen, _, head_dim = x.shape
+    rotary_seqlen, rotary_dim = cos.shape
+    rotary_dim *= 2
+
+    assert rotary_dim <= head_dim
+    assert seqlen <= rotary_seqlen
+    assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
+
+    x_rot = x[:, :, :, :rotary_dim]
+    x_pass = x[:, :, :, rotary_dim:]
+
+    x1, x2 = x_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]
+
+    x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)
+
+    return torch.cat([x_rot, x_pass], axis=-1)
+
+
+def _apply_rotary_emb_kv(
+    kv: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+    cos_k: Optional[torch.FloatTensor] = None,
+    sin_k: Optional[torch.FloatTensor] = None,
+) -> torch.FloatTensor:
+    _, seqlen, two, _, head_dim = kv.shape
+    assert two == 2
+
+    rotary_seqlen, rotary_dim = cos.shape
+    rotary_dim *= 2
+    assert rotary_dim <= head_dim
+    assert seqlen <= rotary_seqlen
+    assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
+
+    k_rot = kv[:, :, 0, :, :rotary_dim]
+    k_pass = kv[:, :, 0, :, rotary_dim:]
+
+    k1, k2 = k_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]
+
+    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)
+
+    return torch.cat(
+        [
+            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+            kv[:, :, 1:2, :, :],
+        ],
+        axis=2,
+    )
+
+
+def _apply_rotary_emb_qkv(
+    qkv: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+    cos_k: Optional[torch.FloatTensor] = None,
+    sin_k: Optional[torch.FloatTensor] = None,
+) -> torch.FloatTensor:
+    _, seqlen, three, _, head_dim = qkv.shape
+    assert three == 3
+
+    rotary_seqlen, rotary_dim = cos.shape
+    rotary_dim *= 2
+    assert rotary_dim <= head_dim
+    assert seqlen <= rotary_seqlen
+    assert cos.shape == sin.shape == (rotary_seqlen, rotary_dim // 2)
+
+    q_rot = qkv[:, :, 0, :, :rotary_dim]
+    q_pass = qkv[:, :, 0, :, rotary_dim:]
+
+    k_rot = qkv[:, :, 1, :, :rotary_dim]
+    k_pass = qkv[:, :, 1, :, rotary_dim:]
+
+    q1, q2 = q_rot.chunk(2, dim=-1)
+    k1, k2 = k_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
+
+    q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
+    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
+
+    return torch.cat(
+        [
+            torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
+            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+            qkv[:, :, 2:3, :, :],
+        ],
+        axis=2,
+    )
+
+
 class RotaryEmbedding(nn.Module):
-    """Rotary embeddings.
+    """Rotary positional embedding (RoPE).
 
     Reference:
-        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/layers/rotary.py.
-    
+        RoFormer: Enhanced Transformer with Rotary Position Embedding.
+        https://arxiv.org/pdf/2104.09864.pdf.
+
     """
 
     def __init__(
@@ -131,14 +228,14 @@ class RotaryEmbedding(nn.Module):
         self.device = device
 
         inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
-        self.register_buffer("inv_freq", inv_freq)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
 
         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)
+        self.register_buffer("scale", scale, persistent=False)
 
         self._seq_len_cached = 0
         self._cos_cached = None
@@ -146,28 +243,26 @@ class RotaryEmbedding(nn.Module):
         self._cos_k_cached = None
         self._sin_k_cached = None
 
-    def _update_cos_sin_cache(self, x: torch.FloatTensor, seqlen_offset: int = 0) -> None:
-        # Reset the tables if the sequence length has changed,
-        # or if we're on a new device (possibly due to tracing for instance)
-        seqlen = x.shape[1] + seqlen_offset
-
+    def _update_cos_sin_cache(
+        self, seqlen: int, device: Optional[str] = None, dtype: Optional[torch.dtype] = None
+    ) -> None:
         # Re-generate the inverse frequency buffer if it's not fp32
         # (for instance if model.half() was called)
         if self.inv_freq.dtype != "torch.float32":
             self.inv_freq = 1.0 / (
-                self.base ** (torch.arange(0, self.dim, 2, device=self.device, dtype=torch.float32) / self.dim)
+                self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim)
             )
 
-        if seqlen > self._seq_len_cached or self._cos_cached.device != x.device or self._cos_cached.dtype != x.dtype:
+        if seqlen > self._seq_len_cached or self._cos_cached.device != device or self._cos_cached.dtype != dtype:
             self._seq_len_cached = seqlen
-            t = torch.arange(seqlen, device=x.device, dtype=torch.float32)
+            t = torch.arange(seqlen, device=device, dtype=torch.float32)
 
             # Don't do einsum, it converts fp32 to fp16
             # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
             freqs = torch.outer(t, self.inv_freq.to(device=t.device, dtype=torch.float32))
             if self.scale is None:
-                self._cos_cached = torch.cos(freqs).to(x.dtype)
-                self._sin_cached = torch.sin(freqs).to(x.dtype)
+                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
@@ -175,62 +270,32 @@ class RotaryEmbedding(nn.Module):
                 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(x.dtype)
-                self._sin_cached = (torch.sin(freqs) * scale).to(x.dtype)
-                self._cos_k_cached = (torch.cos(freqs) / scale).to(x.dtype)
-                self._sin_k_cached = (torch.sin(freqs) / scale).to(x.dtype)
+                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 _apply_rotary_emb_qkv(
+    def forward(
         self,
-        qkv: torch.FloatTensor,
-        sin: torch.FloatTensor,
-        cos: torch.FloatTensor,
-        sin_k: Optional[torch.FloatTensor] = None,
-        cos_k: Optional[torch.FloatTensor] = None,
-    ) -> torch.FloatTensor:
-        _, seqlen, three, _, headdim = qkv.shape
-        assert three == 3
-
-        rotary_seqlen, rotary_dim = cos.shape
-        rotary_dim *= 2
-        assert rotary_dim <= headdim
-        assert seqlen <= rotary_seqlen
-
-        cos_k = cos if cos_k is None else cos_k
-        sin_k = sin if sin_k is None else sin_k
-        assert sin.shape == cos_k.shape == sin_k.shape == (rotary_seqlen, rotary_dim // 2)
-
-        q_rot = qkv[:, :, 0, :, :rotary_dim]
-        q_pass = qkv[:, :, 0, :, rotary_dim:]
-
-        k_rot = qkv[:, :, 1, :, :rotary_dim]
-        k_pass = qkv[:, :, 1, :, rotary_dim:]
-
-        # Splits the queries and keys in half
-        q1, q2 = q_rot.chunk(2, dim=-1)
-        k1, k2 = k_rot.chunk(2, dim=-1)
-        c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
-
-        # Casts to fp32 are necessary to prevent fp16 overflow issues
-        q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
-
-        # Computes the new keys and queries, recasting to original dtype
-        q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
-        k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
-
-        return torch.cat(
-            [
-                torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
-                torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
-                qkv[:, :, 2:3, :, :],
-            ],
-            axis=2,
-        )
+        qkv: torch.Tensor,
+        kv: Optional[torch.Tensor] = None,
+        seqlen_offset: int = 0,
+        max_seqlen: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        seqlen = qkv.shape[1]
+
+        if max_seqlen is not None:
+            self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
+        else:
+            self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
+
+        if kv is None:
+            return _apply_rotary_emb_qkv(qkv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
+        else:
+            q = _apply_rotary_emb(qkv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
+            kv = _apply_rotary_emb_kv(kv, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])
 
-    def forward(self, qkv: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
-        # `qkv` is of shape (batch, seqlen, 3, nheads, headdim)
-        self._update_cos_sin_cache(qkv, seqlen_offset)
-        return self._apply_rotary_emb_qkv(qkv, self._sin_cached[seqlen_offset:], self._cos_cached[seqlen_offset:])
+            return q, kv
 
 
 class MLP(nn.Module):
@@ -290,21 +355,22 @@ class SelfAttention(nn.Module):
         attention_mask: Optional[torch.BoolTensor] = None,
         **kwargs,
     ) -> torch.FloatTensor:
-        causal = self.causal if causal is None else causal
-        batch_size, seq_len = qkv.shape[0], qkv.shape[1]
+        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
         q, k, v = qkv.unbind(dim=2)
 
+        causal = self.causal if causal is None else causal
         softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+
         scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
 
         if attention_mask is not None:
-            padding_mask = torch.full((batch_size, seq_len), -10000.0, dtype=scores.dtype, device=scores.device)
+            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
             padding_mask.masked_fill_(attention_mask, 0.0)
 
             scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
 
         if causal:
-            causal_mask = torch.triu(torch.full((seq_len, seq_len), -10000.0, device=scores.device), 1)
+            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
             scores = scores + causal_mask.to(dtype=scores.dtype)
 
         attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
@@ -343,25 +409,31 @@ class CrossAttention(nn.Module):
         attention_mask: Optional[torch.BoolTensor] = None,
         **kwargs,
     ) -> torch.FloatTensor:
-        causal = self.causal if causal is None else causal
-        batch_size, seq_len_q = q.shape[0], q.shape[1]
-        assert kv.shape[0] == batch_size and kv.shape[3] == q.shape[2] and kv.shape[4] == q.shape[3]
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = kv.shape[1]
+        assert kv.shape[0] == batch_size and kv.shape[4] == q.shape[3]
 
-        seq_len_k = kv.shape[1]
+        if kv.shape[3] != q.shape[2]:
+            kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
         k, v = kv.unbind(dim=2)
 
+        causal = self.causal if causal is None else causal
         softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+
         scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
 
         if attention_mask is not None:
-            padding_mask = torch.full((batch_size, seq_len_k), -10000.0, dtype=scores.dtype, device=scores.device)
+            padding_mask = torch.full((batch_size, seqlen_k), -10000.0, dtype=scores.dtype, device=scores.device)
             padding_mask.masked_fill_(attention_mask, 0.0)
 
             scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
 
         if causal:
-            causal_mask = torch.triu(torch.full((seq_len_q, seq_len_k), -10000.0, device=scores.device), 1)
-            scores = scores + causal_mask.to(dtype=scores.dtype)
+            rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
+            cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
+            causal_mask = cols > rows + seqlen_k - seqlen_q
+
+            scores = scores.masked_fill(causal_mask, -10000.0)
 
         attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
         attention = self.drop(attention)
@@ -371,21 +443,12 @@ class CrossAttention(nn.Module):
         return output
 
 
-def find_mha_dims(
-    config: PretrainedConfig, n_head: Optional[int] = None, head_dim: Optional[int] = None
+def _find_mha_dims(
+    config: PretrainedConfig,
+    n_head: Optional[int] = None,
+    n_head_kv: Optional[int] = None,
+    head_dim: Optional[int] = None,
 ) -> Tuple[int, int]:
-    """Validate and return the number of heads and head dimension for multi-head attention.
-
-    Args:
-        config: Model configuration.
-        n_head: Number of heads.
-        head_dim: Head dimension.
-
-    Returns:
-        Number of heads and head dimension.
-
-    """
-
     assert all(
         hasattr(config, attr) for attr in ["n_embd", "n_head"]
     ), "`config` must have `n_embd` and `n_head` attributes."
@@ -401,31 +464,20 @@ def find_mha_dims(
     elif n_head is None or head_dim is None:
         raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
 
-    return n_head, head_dim
-
-
-def update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
-    """Update the key-value cache for inference.
-
-    Reference:
-        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
-
-    Args:
-        kv: Key-value tensor.
-        inference_params: Inference parameters.
-        layer_idx: Layer index.
+    if n_head_kv is None:
+        n_head_kv = getattr(config, "n_head_kv", None) or n_head
+    assert n_head % n_head_kv == 0, "`n_head` must be divisible by `n_head_kv`."
 
-    Returns:
-        Updated key-value tensor.
+    return n_head, n_head_kv, head_dim
 
-    """
 
+def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
     num_heads, head_dim = kv.shape[-2:]
 
     if layer_idx not in inference_params.key_value_memory_dict:
         kv_cache = torch.empty(
             inference_params.max_batch_size,
-            inference_params.max_sequence_len,
+            inference_params.max_seqlen,
             2,
             num_heads,
             head_dim,
@@ -434,43 +486,19 @@ def update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, la
         )
         inference_params.key_value_memory_dict[layer_idx] = kv_cache
     else:
-        if not inference_params.fused_ft_kernel:
-            kv_cache = inference_params.key_value_memory_dict[layer_idx]
-        else:
-            k_cache, v_cache = inference_params.key_value_memory_dict[layer_idx]
-            kv_cache = None
+        kv_cache = inference_params.key_value_memory_dict[layer_idx]
 
     batch_start = inference_params.batch_size_offset
     batch_end = batch_start + kv.shape[0]
-    assert batch_end <= (kv_cache.shape[0] if kv_cache is not None else v_cache.shape[0])
+    assert batch_end <= kv_cache.shape[0]
 
-    sequence_start = inference_params.sequence_len_offset
+    sequence_start = inference_params.seqlen_offset
     sequence_end = sequence_start + kv.shape[1]
-    assert sequence_end <= (kv_cache.shape[1] if kv_cache is not None else v_cache.shape[2])
-
-    if not inference_params.fused_ft_kernel:
-        assert kv_cache is not None
-
-        kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
-        kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
-
-        return kv
+    assert sequence_end <= kv_cache.shape[1]
 
-    assert inference_params.sequence_len_offset == 0
-    assert kv.dtype in [torch.float16, torch.bfloat16, torch.float32]
-
-    packsize = 4 if kv.dtype == torch.float32 else 8
-
-    if kv_cache is not None:
-        kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
-        k_cache = rearrange(kv_cache[:, :, 0], "b s h (d packsize) -> b h d s packsize", packsize=packsize).contiguous()
-        v_cache = rearrange(kv_cache[:, :, 1], "b s h d -> b h s d").contiguous()
-        inference_params.key_value_memory_dict[layer_idx] = (k_cache, v_cache)
-    else:
-        k_cache[batch_start:batch_end, :, :, :sequence_end, :] = rearrange(
-            kv[:, :, 0], "b s h (d packsize) -> b h d s packsize", packsize=packsize
-        )
-        v_cache[batch_start:batch_end, :, :sequence_end, :] = rearrange(kv[:, :, 1], "b s h d -> b h s d")
+    assert kv_cache is not None
+    kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+    kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
 
     return kv
 
@@ -486,6 +514,7 @@ class MHA(nn.Module):
         rotary_dim: Optional[int] = None,
         rotary_emb_scale_base: Optional[float] = None,
         n_head: Optional[int] = None,
+        n_head_kv: Optional[int] = None,
         head_dim: Optional[int] = None,
         bias: bool = True,
         causal: bool = True,
@@ -506,12 +535,12 @@ class MHA(nn.Module):
             self.rotary_emb = RotaryEmbedding(self.rotary_emb_dim, **rotary_kwargs)
         
         # MLP
-        self.n_head, self.head_dim = find_mha_dims(config, n_head, head_dim)
-        op_size = self.n_head * self.head_dim
+        self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim)
+        op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
         hidden_size = config.n_embd
 
-        self.Wqkv = nn.Linear(hidden_size, 3 * op_size, bias=bias, device=device, dtype=dtype)
-        self.out_proj = nn.Linear(op_size, hidden_size, bias=bias, device=device, dtype=dtype)
+        self.Wqkv = nn.Linear(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
+        self.out_proj = nn.Linear(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)
 
         # Attention
         self.inner_attn = SelfAttention(causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout)
@@ -521,40 +550,75 @@ class MHA(nn.Module):
         self.return_residual = return_residual
         self.checkpointing = checkpointing
 
-    def forward(
+    def _forward_self_attn(
+        self, x: torch.FloatTensor, attention_mask: Optional[torch.BoolTensor]
+    ) -> torch.FloatTensor:
+        qkv = self.Wqkv(x)
+        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+
+        if self.rotary_emb_dim > 0:
+            qkv = self.rotary_emb(qkv)
+
+        if self.checkpointing:
+            return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, attention_mask=attention_mask)
+
+        return self.inner_attn(qkv, attention_mask=attention_mask)
+
+    def _forward_cross_attn(
         self,
         x: torch.FloatTensor,
-        past_key_values: Optional[InferenceParams] = None,
-        attention_mask: Optional[torch.BoolTensor] = None,
-        cu_seqlens: Optional[torch.LongTensor] = None,
-        max_seqlen: Optional[int] = None,
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        past_key_values: Optional[InferenceParams],
+        attention_mask: Optional[torch.BoolTensor],
+    ) -> torch.FloatTensor:
         qkv = self.Wqkv(x)
-        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
 
-        seqlen_offset = past_key_values.sequence_len_offset if past_key_values is not None else 0
+        q = qkv[..., : self.n_head * self.head_dim]
+        q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
+
+        kv = qkv[..., self.n_head * self.head_dim :]
+        kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
+
+        seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
+        causal = None if seqlen_offset == 0 else False
         if self.rotary_emb_dim > 0:
-            qkv = self.rotary_emb(qkv, seqlen_offset=seqlen_offset)
+            q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)
 
         if past_key_values is not None:
-            kv = update_kv_cache(qkv[:, :, 1:], past_key_values, self.layer_idx)
+            kv = _update_kv_cache(kv, past_key_values, self.layer_idx)
 
-        if attention_mask is not None:
-            attention_mask = attention_mask[0] if isinstance(attention_mask, tuple) else attention_mask
-            attention_mask = attention_mask.bool().to(qkv.device)
+        if self.checkpointing:
+            return torch.utils.checkpoint.checkpoint(
+                self.inner_cross_attn, q, kv, attention_mask=attention_mask, causal=causal
+            )
+
+        return self.inner_cross_attn(q, kv, attention_mask=attention_mask, causal=causal)
 
-        attention_kwargs = {"attention_mask": attention_mask}
+    def forward(
+        self,
+        x: torch.FloatTensor,
+        past_key_values: Optional[InferenceParams] = None,
+        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        if attention_mask is not None and torch.any(~attention_mask.bool()):
+            attention_mask = attention_mask.bool()
+        else:
+            attention_mask = None
 
-        if past_key_values is None or seqlen_offset == 0:
-            if self.checkpointing:
-                attn_output = torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, **attention_kwargs)
+        # MHA
+        if self.n_head == self.n_head_kv:
+            if past_key_values is None:
+                # If `past_key_values` are not supplied, we run self-attention
+                attn_output = self._forward_self_attn(x, attention_mask)
             else:
-                attn_output = self.inner_attn(qkv, **attention_kwargs)
+                # If `past_key_values` are supplied, it means that we might have cached values and
+                # could take advantage of cross-attention
+                attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
+        # MQA / GQA
         else:
-            q = qkv[:, :, 0]
-            causal = None if past_key_values.sequence_len_offset == 0 else False
-            attn_output = self.inner_cross_attn(q, kv, causal=causal, **attention_kwargs)
+            # Regardless of `past_key_values` being supplied or not, it always use cross-attention
+            # because `q` and `kv` lengths might be different
+            attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
 
         output = rearrange(attn_output, "... h d -> ... (h d)")
         output = self.out_proj(output)
@@ -672,38 +736,29 @@ class MixFormerSequentialPreTrainedModel(PreTrainedModel):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
         elif isinstance(module, nn.LayerNorm):
-            module.bias.data.zero_()
+            if module.bias is not None:
+                module.bias.data.zero_()
             module.weight.data.fill_(1.0)
 
     def prepare_inputs_for_generation(
         self,
         input_ids: torch.LongTensor,
         past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
-        attention_mask: Optional[torch.BoolTensor] = None,
+        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
         **kwargs,
     ) -> Dict[str, Any]:
-        if attention_mask is not None and torch.any(~attention_mask.bool()):
-            total_seq_len = torch.sum(attention_mask, dim=1)
-            max_seq_len = torch.max(total_seq_len)
-
-            total_seq_len = torch.cat((torch.tensor([0], device=attention_mask.device), total_seq_len)).unsqueeze(1)
-            cumulative_seq_len = torch.cumsum(total_seq_len, dim=0).squeeze(1).to(torch.int32)
-            attention_mask = (attention_mask.bool(), cumulative_seq_len, max_seq_len.item())
-        else:
-            attention_mask = None
-
         if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
             past_key_values = InferenceParams(
+                max_seqlen=self.config.n_positions,
                 max_batch_size=input_ids.shape[0],
-                max_sequence_len=self.config.n_positions,
-                sequence_len_offset=0,
+                seqlen_offset=0,
                 batch_size_offset=0,
-                fused_ft_kernel=False,
                 key_value_memory_dict={},
+                lengths_per_sample=None,
             )
         else:
             # Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
-            past_key_values.sequence_len_offset = len(input_ids[0]) - 1
+            past_key_values.seqlen_offset = len(input_ids[0]) - 1
             input_ids = input_ids[:, -1].unsqueeze(-1)
 
         return {
@@ -711,6 +766,10 @@ class MixFormerSequentialPreTrainedModel(PreTrainedModel):
             "past_key_values": past_key_values,
             "attention_mask": attention_mask,
         }
+    
+    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False) -> None:
+        if isinstance(module, MixFormerSequentialPreTrainedModel):
+            module.gradient_checkpointing = value
 
 
 class MixFormerSequentialForCausalLM(MixFormerSequentialPreTrainedModel):
@@ -752,16 +811,10 @@ class MixFormerSequentialForCausalLM(MixFormerSequentialPreTrainedModel):
         labels: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> CausalLMOutputWithPast:
-        if attention_mask is not None and self.training:
-            print("`attention_mask` is not supported during training. Using it might lead to unexpected results.")
-
-        if past_key_values is None and attention_mask is None:
-            lm_logits = self.layers(input_ids)
-        else:
-            hidden_layer = self.layers[0](input_ids)
-            for module in self.layers[1:-1]:
-                hidden_layer = module(hidden_layer, past_key_values=past_key_values, attention_mask=attention_mask)
-            lm_logits = self.layers[-1](hidden_layer)
+        hidden_layer = self.layers[0](input_ids)
+        for module in self.layers[1:-1]:
+            hidden_layer = module(hidden_layer, past_key_values=past_key_values, attention_mask=attention_mask)
+        lm_logits = self.layers[-1](hidden_layer)
 
         loss = None
         if labels is not None: