add falcon landmark code (incomplete)

code/configuration_RW.py

@@ -0,0 +1,86 @@
+# coding=utf-8
+# Copyright 2022 the Big Science Workshop and HuggingFace Inc. team.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Bloom configuration"""
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class RWConfig(PretrainedConfig):
+    model_type = "RefinedWebModel"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_hidden_layers": "n_layer",
+        "num_attention_heads": "n_head",
+    }
+
+    def __init__(
+        self,
+        vocab_size=250880,
+        hidden_size=64,
+        n_layer=2,
+        n_head=8,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        apply_residual_connection_post_layernorm=False,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        multi_query=False,
+        alibi=False,
+        bias=False,
+        parallel_attn=False,
+        mem_id=None,
+        mem_freq=50,
+        train_context_length=512,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        # Backward compatibility with n_embed kwarg
+        n_embed = kwargs.pop("n_embed", None)
+        self.hidden_size = hidden_size if n_embed is None else n_embed
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.multi_query = multi_query
+        self.alibi = alibi
+        self.bias = bias
+        self.parallel_attn = parallel_attn
+
+        self.mem_id = mem_id
+        self.mem_freq = mem_freq
+        self.train_context_length = train_context_length
+
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+    @property
+    def head_dim(self):
+        return self.hidden_size // self.n_head
+
+    @property
+    def rotary(self):
+        return not self.alibi

code/install_deps.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+pip install -r requirements.txt
+pip install "git+https://github.com/openai/triton.git#subdirectory=python"

code/llama_landmark_config.py

@@ -0,0 +1,16 @@
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+class LlamaLandmarkConfig(LlamaConfig):
+    model_type = "llama_with_landmark"
+
+    def __init__(
+        self,
+        mem_id=32001,
+        mem_freq=50,
+        train_context_length=512,
+        **kwargs,
+    ):
+        self.mem_id = mem_id
+        self.mem_freq = mem_freq
+        self.train_context_length = train_context_length
+        super().__init__(**kwargs)

code/llama_mem.py

@@ -0,0 +1,1295 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch LLaMA model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from llama_landmark_config import LlamaLandmarkConfig
+from ltriton.flash_landmark_attention import fused_landmark_attention
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaLandmarkConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class LlamaRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+            self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
+    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
+    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
+    if position_ids.ndim == 2:
+        cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+        sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    else:
+        cos = cos[position_ids]
+        sin = sin[position_ids]
+    if q is None:
+        q_embed = None
+    else:
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+class LandmarkGroupedSoftmaxFunction(torch.autograd.Function):
+
+    # Note that forward, setup_context, and backward are @staticmethods
+    @staticmethod
+    def forward(ctx, x, dim, mem_cnt, resp_mem_idx):
+        new_shape = list(x.shape)
+        new_shape[dim] = mem_cnt # max_mem_cnt.item()
+        max_by_group = x.new_zeros((*new_shape,))
+        max_by_group.scatter_reduce_(src=x, index=resp_mem_idx, dim=dim, reduce="amax", include_self=False)
+            
+        maxes = torch.gather(max_by_group, dim, resp_mem_idx)
+        #x_exp = torch.exp(x - torch.where(torch.isinf(maxes), 0, maxes))
+        x_exp = torch.exp((x - maxes).to(torch.float32))
+
+        cumsum_by_group = torch.zeros_like(max_by_group, dtype=x_exp.dtype)
+
+        cumsum_by_group.scatter_add_(dim, resp_mem_idx, x_exp, )
+        denom = torch.gather(cumsum_by_group, dim, resp_mem_idx)
+
+        #probs = torch.where(denom < 0.5, 0, x_exp / denom)
+        probs = x_exp / denom
+        
+        
+        ctx.mem_cnt = mem_cnt
+        ctx.dim = dim
+        ctx.save_for_backward(resp_mem_idx, probs)
+
+        return probs
+
+    @staticmethod
+    def backward(ctx, grad_probs):
+        mem_cnt = ctx.mem_cnt
+        dim = ctx.dim
+        resp_mem_idx, probs = ctx.saved_tensors
+        grad_x = grad_dim = grad_mem_cnt = grad_resp_mem_idx = None
+
+        if ctx.needs_input_grad[0] or ctx.needs_input_grad[4]:
+            grad_pair = grad_probs * probs
+
+            new_shape = list(probs.shape)
+            new_shape[dim] = mem_cnt # max_mem_cnt.item()
+            cumsum_by_group = grad_pair.new_zeros((*new_shape,))
+            cumsum_by_group.scatter_add_(dim, resp_mem_idx, grad_pair)
+
+
+        if ctx.needs_input_grad[0]:
+            grad_sum = torch.gather(cumsum_by_group, dim, resp_mem_idx)
+            grad_x = grad_pair - probs * grad_sum
+        assert not ctx.needs_input_grad[1]
+        assert not ctx.needs_input_grad[2]
+        assert not ctx.needs_input_grad[3]
+        
+        return grad_x, grad_dim, grad_mem_cnt, grad_resp_mem_idx
+
+def landmark_grouped_softmax(x, dim, is_mem, last_section_mask):
+    
+    last_and_rest_mask = last_section_mask # | mask
+
+    full_access_mask =  is_mem | last_and_rest_mask
+
+    max_mem_cnt = 64
+    mem_group_idx = torch.cumsum(is_mem, dim=dim)
+    mem_bucket_id = max_mem_cnt - 1
+    resp_mem_idx = torch.where(last_and_rest_mask, 
+                                max_mem_cnt - 1,
+                                torch.where(is_mem, mem_bucket_id, mem_group_idx))
+    probs = LandmarkGroupedSoftmaxFunction.apply(x, dim, max_mem_cnt, resp_mem_idx)
+
+    new_shape = list(x.shape)
+    new_shape[dim] = max_mem_cnt
+    group_prob = probs.new_zeros((*new_shape, ))
+    group_prob.scatter_(dim, torch.where(is_mem, mem_group_idx - 1, max_mem_cnt - 1), probs)
+    probs = probs.mul(torch.where(full_access_mask, last_section_mask, torch.gather(group_prob, dim, resp_mem_idx)))
+
+
+    return probs
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaLandmarkConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        is_mem: Optional[torch.Tensor] = None,
+        last_section_mask: Optional[torch.Tensor] = None,
+        offload_cache_to_cpu: bool = False,
+        use_flash: bool = False,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+            if len(past_key_value) > 2:
+                kv_seq_len += past_key_value[3].shape[2] * past_key_value[3].shape[3]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        key_states_before_pos = key_states
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        attn_prefix = None
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            if mem_freq is None:
+                cache_len = past_key_value[0].shape[2]
+                if is_mem is not None:
+                    if use_flash:
+                        is_mem = torch.cat((is_mem.new_zeros((1, cache_len)), is_mem), dim=-1)
+                    else:
+                        is_mem = torch.cat((is_mem.new_zeros((1, 1, q_len, cache_len)), is_mem), dim=-1)
+                        last_section_mask = torch.cat((last_section_mask.new_ones((1, 1, q_len, cache_len)), last_section_mask), dim=-1)
+
+                past_key_states = torch.cat([past_key_value[0], key_states], dim=2)
+                past_value_states = torch.cat([past_key_value[1], value_states], dim=2)
+                key_states = past_key_states[:, :, -(q_len + cache_len):]
+                value_states = past_value_states[:, :, -(q_len + cache_len):]
+                expected_att_size = (bsz, self.num_heads, q_len, cache_len + q_len)
+            else:
+                orig_value_states = value_states
+
+                incomplete_len = past_key_value[0].shape[2] % (mem_freq + 1)
+                full_len = past_key_value[0].shape[2] - incomplete_len
+                past_key_mem, past_key_incomplete = torch.split(past_key_value[0], (full_len, incomplete_len), dim=2)
+                past_value_mem, past_value_incomplete = torch.split(past_key_value[1], (full_len, incomplete_len), dim=2)
+
+                if offload_cache_to_cpu:
+                    past_key_value = (past_key_incomplete, past_value_incomplete, *past_key_value[2:])
+                
+                if incomplete_len > 0:
+                    assert q_len + incomplete_len <= (mem_freq + 1)
+                if use_flash:
+                    is_mem = torch.cat((is_mem.new_zeros((1, incomplete_len)), is_mem), dim=-1)
+                else:
+                    is_mem = torch.cat((is_mem.new_zeros((1, 1, q_len, incomplete_len)), is_mem), dim=-1)
+                    last_section_mask = torch.cat((last_section_mask.new_ones((1, 1, q_len, incomplete_len)), last_section_mask), dim=-1)
+
+                if len(past_key_value) > 2:
+                    full_len += past_key_value[3].shape[2] * past_key_value[3].shape[3]
+                past_key_incomplete_pos = torch.arange(full_len, full_len + incomplete_len, dtype=torch.long, device=position_ids.device).unsqueeze(0)
+                _, past_key_incomplete = apply_rotary_pos_emb(None, past_key_incomplete, cos, sin, past_key_incomplete_pos)
+                key_states = torch.cat((past_key_incomplete, key_states), dim=2)
+                value_states = torch.cat((past_value_incomplete, value_states), dim=2)
+
+                past_key_mem = past_key_mem.view(bsz, self.num_heads, -1, mem_freq + 1, self.head_dim)
+                past_value_mem = past_value_mem.view(bsz, self.num_heads, -1, mem_freq + 1, self.head_dim)
+
+                if len(past_key_value) > 2:
+                    mem_key_nopos = torch.cat((
+                        past_key_value[2], 
+                        past_key_mem.select(dim=3, index=mem_freq)), dim=2)
+                    past_key_mem_offload = past_key_value[3]
+                    past_key_mem = torch.cat((
+                        past_key_mem_offload, 
+                        past_key_mem.to(past_key_mem_offload.device)), dim=2)
+                    past_value_mem = torch.cat((past_key_value[4], past_value_mem.to(past_key_mem_offload.device)), dim=2)
+                else:
+                    mem_key_nopos = past_key_mem.select(dim=3, index=mem_freq)    
+                
+                num_mems = past_key_mem.shape[2]
+                top_k = min(cache_top_k, num_mems)
+                prefix_len = full_len - (top_k + 1) * (mem_freq + 1)
+                mem_indices = torch.cat(
+                    (position_ids.new_zeros((max(0, num_mems - top_k), )),
+                    torch.arange(1, top_k + 1, device=query_states.device, dtype=position_ids.dtype)), dim=0)
+                mem_pos = (mem_indices * (mem_freq + 1) + mem_freq).unsqueeze(0).expand(bsz, -1) + prefix_len
+                _, mem_key = apply_rotary_pos_emb(None, mem_key_nopos, cos, sin, mem_pos)
+                mem_attn_weights = torch.matmul(query_states, mem_key.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+                if offload_cache_to_cpu:
+                    aggregate = "max_over_tokens"
+                else:
+                    aggregate = None
+                if aggregate == "max_over_tokens":
+                    token_retrievers = 1
+                    head_retrievers = self.num_heads
+                    mem_attn_weights = torch.nn.functional.softmax(mem_attn_weights, dim=-1,dtype=torch.float32).to(query_states.dtype)
+                    mem_attn_weights = mem_attn_weights.amax(dim=2, keepdim=True)
+                elif aggregate is None:
+                    token_retrievers = q_len
+                    head_retrievers = self.num_heads
+                else:
+                    raise NotImplementedError()
+
+                mem_selected_idx = mem_attn_weights.topk(dim=-1,k=top_k)[1].sort(dim=-1)[0].view(bsz, head_retrievers, token_retrievers, top_k)
+                
+                selected_indices = torch.arange(0, top_k * (mem_freq + 1), device=query_states.device, dtype=position_ids.dtype)
+                selected_indices = torch.where(mem_selected_idx >= num_mems - top_k, mem_freq + 1, 0).unsqueeze(-1) + selected_indices.view(1, 1, 1, top_k, mem_freq + 1)
+                selected_indices = selected_indices.view(bsz, head_retrievers, token_retrievers, -1) + prefix_len
+
+                
+                
+                
+                mem_selected_idx = mem_selected_idx.to(past_key_mem.device)
+                
+                mem_selected_idx = mem_selected_idx.view(bsz, self.num_heads, token_retrievers, top_k, 1, 1).expand(bsz, self.num_heads, token_retrievers, top_k, mem_freq + 1, self.head_dim) 
+                selected_keys = past_key_mem.unsqueeze(2).expand(bsz, self.num_heads, token_retrievers, -1, mem_freq + 1, self.head_dim)
+                selected_keys = selected_keys.take_along_dim(mem_selected_idx, dim=3).to(query_states.device)
+                selected_values = past_value_mem.unsqueeze(2).expand(bsz, self.num_heads, token_retrievers, -1, mem_freq + 1, self.head_dim).take_along_dim(mem_selected_idx, dim=3).to(query_states.device)
+
+                if aggregate == "max_over_tokens":
+                    selected_indices = selected_indices.squeeze(2)
+                    selected_keys = selected_keys.view(bsz, self.num_heads, -1, self.head_dim)
+                    selected_keys = apply_rotary_pos_emb(None, selected_keys, cos, sin, selected_indices)[1] 
+                    key_states = torch.cat((selected_keys, key_states), dim=2)
+                    value_states = torch.cat((selected_values.view(bsz, self.num_heads, -1, self.head_dim), value_states), dim=2)
+                    expected_att_size = (bsz, self.num_heads, q_len, key_states.shape[2])
+                else:
+                    selected_indices = selected_indices.expand(bsz, self.num_heads, q_len, -1)
+                    selected_keys = selected_keys.view(bsz, self.num_heads, token_retrievers, -1, self.head_dim).expand(bsz, self.num_heads, q_len, -1, self.head_dim)
+                    selected_keys = apply_rotary_pos_emb(None, selected_keys, cos, sin, selected_indices)[1]
+                    selected_values = selected_values.view(bsz, self.num_heads, token_retrievers, -1, self.head_dim).expand(bsz, self.num_heads, q_len, -1, self.head_dim)
+                    attn_prefix = torch.matmul(query_states.unsqueeze(3), selected_keys.transpose(3, 4)).squeeze(3) / math.sqrt(self.head_dim)
+                    expected_att_size = (bsz, self.num_heads, q_len, q_len + incomplete_len)
+                
+                is_mem_prefix = torch.cat((is_mem.new_zeros((mem_freq, )), is_mem.new_ones((1, )))).unsqueeze(0).repeat((top_k, 1))
+                if use_flash:
+                    is_mem_prefix = is_mem_prefix.view(1, -1)
+                else:
+                    is_mem_prefix = is_mem_prefix.view(1, 1, 1, -1).expand(1, 1, q_len, -1)
+                    last_section_mask = torch.cat((last_section_mask.new_zeros((1, 1, q_len, top_k * (mem_freq + 1))), last_section_mask), dim=-1)
+                is_mem = torch.cat((is_mem_prefix, is_mem), dim=-1)
+
+
+                past_key_states = torch.cat([past_key_value[0], key_states_before_pos], dim=2)
+                past_value_states = torch.cat([past_key_value[1], orig_value_states], dim=2)
+
+                if offload_cache_to_cpu:
+                    past_key_value = (past_key_states, past_value_states, mem_key_nopos, past_key_mem.to("cpu"), past_value_mem.to("cpu"), *past_key_value[5:]) if use_cache else None
+                else:
+                    past_key_value = (past_key_states, past_value_states) if use_cache else None
+
+        else:
+            if mem_freq is None:
+                past_key_states = key_states
+            else:
+                past_key_states = key_states_before_pos
+            past_value_states = value_states
+            expected_att_size = (bsz, self.num_heads, q_len, kv_seq_len)
+            past_key_value = (past_key_states, past_value_states) if use_cache else None
+
+        if use_flash:
+            assert attn_prefix is None
+            assert not output_attentions
+            assert mem_freq is not None
+            attn_output = fused_landmark_attention(query_states, key_states, value_states, is_mem, block_size=mem_freq+1)
+            attn_weights = None
+        else:
+            attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+            if attn_weights.size() != expected_att_size:
+                raise ValueError(
+                    f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                    f" {attn_weights.size()}"
+                )
+
+            if attention_mask is not None:
+                if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                    raise ValueError(
+                        f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                    )
+                attn_weights = attn_weights + attention_mask[...,-attn_weights.shape[-1]:]
+                attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+            if attn_prefix is not None:
+                attn_weights = torch.cat((attn_prefix, attn_weights), dim=-1)
+            # upcast attention to fp32
+            if is_mem is None:
+                raise ValueError("Don't use this without landmarks")
+                attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+            else:
+                attn_weights = landmark_grouped_softmax(attn_weights, dim=-1, is_mem=is_mem.expand(-1, self.num_heads, -1, -1), last_section_mask=last_section_mask).to(query_states.dtype)
+            if attn_prefix is not None:
+                attn_prefix, attn_weights = torch.split(attn_weights, (attn_prefix.shape[-1], attn_weights.shape[-1] - attn_prefix.shape[-1]), dim=-1)
+            attn_output = torch.matmul(attn_weights, value_states)
+            if attn_prefix is not None:
+                attn_output += torch.matmul(attn_prefix.unsqueeze(3), selected_values).squeeze(3)
+
+            if not output_attentions:
+                attn_weights = None
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaLandmarkConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+    
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        is_mem: Optional[torch.Tensor] = None,
+        last_section_mask: Optional[torch.Tensor] = None,
+        offload_cache_to_cpu: bool = False,
+        use_flash: bool = False,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            is_mem=is_mem,
+            last_section_mask=last_section_mask,
+            offload_cache_to_cpu=offload_cache_to_cpu,
+            use_flash=use_flash,
+            cache_top_k=cache_top_k,
+            mem_freq=mem_freq
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlamaLandmarkConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaLandmarkConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaLandmarkConfig
+    """
+
+    def __init__(self, config: LlamaLandmarkConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        offload_cache_to_cpu: Optional[bool] = None,
+        use_flash: Optional[bool] = None,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        is_mem = None
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+            if self.config.mem_id is not None:
+                with torch.no_grad():
+                    is_mem = input_ids == self.config.mem_id
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            if self.config.mem_id is not None:
+                raise NotImplementedError
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            if is_mem is not None:
+                pass
+                #raise NotImplementedError
+            past_key_values_length = past_key_values[0][0].shape[2]
+            if len(past_key_values[0]) > 2:
+                past_key_values_length += past_key_values[0][3].shape[2] * past_key_values[0][3].shape[3]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+        
+        last_section_mask = None
+        if is_mem is not None and not use_flash:
+            is_mem = is_mem.unsqueeze(1).unsqueeze(2)
+            current_len = input_ids.shape[1]
+            mem_ids = torch.where(attention_mask[..., -current_len:] < -1, 0, torch.cumsum(is_mem, -1) - is_mem.int())
+            last_section_mask = torch.amax(mem_ids, -1, keepdim=True) == mem_ids
+            attention_mask[..., -current_len:].masked_fill_(last_section_mask & is_mem, torch.tensor(torch.finfo(inputs_embeds.dtype).min, device=inputs_embeds.device))
+            last_section_mask.logical_and_(attention_mask[..., -current_len:] > -1)
+            is_mem = is_mem.logical_and(attention_mask[..., -current_len:] > -1)
+            
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                    is_mem,
+                    last_section_mask,
+                    offload_cache_to_cpu,
+                    use_flash,
+                    cache_top_k,
+                    mem_freq
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    is_mem=is_mem,
+                    last_section_mask=last_section_mask,
+                    offload_cache_to_cpu=offload_cache_to_cpu,
+                    use_flash=use_flash,
+                    cache_top_k=cache_top_k,
+                    mem_freq=mem_freq,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.auto_insert_landmarks = False
+        self.always_use_flash = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        offload_cache_to_cpu: Optional[bool] = None,
+        use_flash: Optional[bool] = None,
+        cache_top_k: Optional[int] = None,
+        max_chunk_length: Optional[int] = 0,
+        mem_freq: Optional[int] = None,
+        drop_last_logit_if_mem: Optional[bool] = False,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        ```"""
+
+        use_flash = use_flash if use_flash is not None else self.always_use_flash
+
+        if self.auto_insert_landmarks:
+            mem_freq = self.config.mem_freq
+            assert self.config.mem_freq is not None
+            block_size = self.config.mem_freq + 1
+            input_ids = input_ids.view(input_ids.shape[0], -1, block_size - 1)
+            input_ids = torch.cat((input_ids, input_ids.new_full((input_ids.shape[0], input_ids.shape[1], 1), self.config.mem_id)), dim=-1)
+            input_ids = input_ids.view(input_ids.shape[0], -1)
+            if attention_mask is not None:
+                attention_mask = attention_mask.view(attention_mask.shape[0], -1, block_size - 1)
+                attention_mask = torch.cat((attention_mask, attention_mask.new_ones((attention_mask.shape[0], attention_mask.shape[1], 1))), dim=-1)
+                attention_mask = attention_mask.view(attention_mask.shape[0], -1)
+
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if max_chunk_length == 0:
+            if cache_top_k is not None:
+                max_chunk_length = self.config.train_context_length - self.config.train_context_length % (mem_freq + 1) - (cache_top_k + 1) * (mem_freq + 1)
+                if max_chunk_length <= 0:
+                    raise ValueError("K is too large for this model.")
+            else:
+                max_chunk_length = None
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        window_len = max_chunk_length or input_ids.shape[1]
+        if use_flash:
+            assert window_len % (mem_freq + 1) == 0
+        last_logits = None
+        for step, idx in enumerate(range(0, input_ids.shape[1], window_len)):
+            if idx >= 1:
+                if output_attentions or output_hidden_states:
+                    raise NotImplementedError
+                if not use_cache:
+                    raise NotImplementedError
+            outputs = self.model(
+                input_ids=input_ids[:, idx:idx + window_len],
+                attention_mask=attention_mask[:, :idx + window_len + attention_mask.shape[1] - input_ids.shape[1]] if attention_mask is not None else None,
+                position_ids=position_ids[:, idx:idx + window_len] if position_ids is not None else None,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds[:, idx:idx + window_len] if inputs_embeds is not None else None,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                offload_cache_to_cpu=offload_cache_to_cpu,
+                use_flash=(use_flash or self.auto_insert_landmarks),
+                cache_top_k=cache_top_k,
+                mem_freq=mem_freq,
+            )
+            past_key_values = outputs[1]
+            if last_logits is not None:
+                last_logits = torch.cat((last_logits, outputs[0]), dim=-2)
+            last_logits = outputs[0]
+
+        hidden_states = last_logits
+        if self.auto_insert_landmarks:
+            block_size = self.config.mem_freq + 1
+            hidden_states = hidden_states.reshape(hidden_states.shape[0], hidden_states.shape[1] // block_size, block_size, hidden_states.shape[2])
+            hidden_states = hidden_states[:, :, :block_size - 1]
+            hidden_states = hidden_states.reshape(hidden_states.shape[0], -1, hidden_states.shape[3])
+        if drop_last_logit_if_mem:
+            is_any_mem = (input_ids[:, -1] == self.config.mem_id).any()
+            are_all_mem = (input_ids[:, -1] == self.config.mem_id).all()
+            assert is_any_mem == are_all_mem
+            if is_any_mem:
+                hidden_states = hidden_states[:, :-1]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def set_mem_id(self, mem_id):
+        if self.config.mem_id is not None:
+            assert mem_id == self.config.mem_id, "Chanigng mem_id can break the model. If you really intend to do this, manually disable this check"
+        self.config.mem_id = mem_id
+
+    def enable_landmark_insertion(self):
+        self.auto_insert_landmarks = True
+
+    def enable_flash(self):
+        self.always_use_flash = True
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        total_len = input_ids.shape[1]
+        if past_key_values:
+            prev_len = input_ids.shape[1] - 1
+            use_flash = False if kwargs.get("use_flash") is not None else None
+        else:
+            prev_len = 0
+            use_flash = kwargs.get("use_flash")
+
+        position_ids = kwargs.get("position_ids", None)
+
+        mem_freq = kwargs.get("mem_freq") or self.config.mem_freq
+
+        if mem_freq is not None:
+            if position_ids is not None:
+                raise NotImplementedError
+            T = input_ids.shape[1]
+
+            prev_incomplete_len = prev_len % mem_freq
+            prev_complete_len = prev_len - prev_incomplete_len
+            incomplete_len = total_len % mem_freq
+            new_full_len = total_len - prev_complete_len - incomplete_len
+
+            prev_input, input_ids_with_mem, input_ids_without_mem = torch.split(input_ids, (prev_complete_len, new_full_len, incomplete_len), dim=-1)
+            
+            bsz, q_len = input_ids.size()
+            input_ids_with_mem = input_ids_with_mem.view(bsz, -1, mem_freq)            
+            input_ids_with_mem = torch.cat(
+                (
+                    input_ids_with_mem, 
+                    input_ids_with_mem.new_full((bsz, input_ids_with_mem.shape[1], 1), self.config.mem_id)
+                ), 
+                dim=-1
+            ).view(bsz, -1)
+            input_ids = torch.cat((prev_input, input_ids_with_mem, input_ids_without_mem), dim=-1)
+            if attention_mask is not None:
+                attention_mask_with_mem, attention_mask_without_mem = torch.split(attention_mask, (prev_complete_len + new_full_len, incomplete_len), dim=-1)
+                attention_mask_with_mem = attention_mask_with_mem.view(bsz, -1, mem_freq)
+                attention_mask_with_mem = torch.cat(
+                    (
+                        attention_mask_with_mem, 
+                        attention_mask_with_mem.new_ones((bsz, attention_mask_with_mem.shape[1], 1))
+                    ), 
+                    dim=-1
+                ).view(bsz, -1)
+                attention_mask = torch.cat((attention_mask_with_mem, attention_mask_without_mem), dim=-1)
+            
+            
+        input_ids = input_ids[:, prev_len:]
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            position_ids = position_ids[:, -input_ids.shape[1]:].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None and mem_freq is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "offload_cache_to_cpu": kwargs.get("offload_cache_to_cpu"),
+                "use_flash": use_flash,
+                "cache_top_k": kwargs.get("cache_top_k"),
+                "max_chunk_length": kwargs.get("max_chunk_length", 0),
+                "mem_freq": mem_freq,
+                "drop_last_logit_if_mem": True,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The LLaMa Model transformer with a sequence classification head on top (linear layer).
+
+    [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    LLAMA_START_DOCSTRING,
+)
+class LlamaForSequenceClassification(LlamaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = LlamaModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

code/llama_orig.py

@@ -0,0 +1,888 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch LLaMA model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from ...modeling_utils import PreTrainedModel
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_llama import LlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        return (self.weight * hidden_states).to(input_dtype)
+
+
+class LlamaRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        dtype = torch.get_default_dtype()
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(x.dtype), persistent=False)
+            self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(x.dtype), persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
+    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
+    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
+    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(
+                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min, device=attn_weights.device)
+            )
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlamaConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The LLaMa Model transformer with a sequence classification head on top (linear layer).
+
+    [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    LLAMA_START_DOCSTRING,
+)
+class LlamaForSequenceClassification(LlamaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = LlamaModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

code/modelling_RW.py

@@ -0,0 +1,1362 @@
+# port of models described in RW
+# We use the bloom model as a starting point for these model.
+# Please refer to the bloom models for usage instructions.
+
+import math
+import warnings
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+from torch.nn import functional as F
+
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from configuration_RW import RWConfig
+from ltriton.flash_landmark_attention import fused_landmark_attention
+
+logger = logging.get_logger(__name__)
+
+# NOTE(Hesslow): Unfortunately we did not fuse matmul and bias during training, this means that there's one additional quantization to bfloat16 between the operations.
+# In order not to degrade the quality of our HF-port, we keep these characteristics in the final model.
+class Linear(nn.Linear):
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        ret = input @ self.weight.T
+        if self.bias is None:
+            return ret
+        else:
+            return ret + self.bias
+
+
+from einops import rearrange
+
+# rotary pos emb helpers (torch.jit.script does not seem to support staticmethod...)
+def rotate_half(x):
+    x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=x1.ndim - 1)  # dim=-1 triggers a bug in torch < 1.8.0
+
+
+class RotaryEmbedding(torch.nn.Module):
+    """Implementation of RotaryEmbedding from GPT-NeoX.
+    This implementation is design to operate on queries and keys that are compatible with
+    [batch_size, n_heads_per_partition, seq_len, head_dim] (e.g. MinGPTAttention format).
+    """
+
+    def __init__(
+        self,
+        head_dim: int,
+        base=10000,
+    ):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.head_dim = head_dim
+        self.seq_len_cached = None
+        self.batch_size_cached = None
+        self.cos_cached: torch.Tensor | None = None
+        self.sin_cached: torch.Tensor | None = None
+
+    def cos_sin(
+        self,
+        seq_len: int,
+        device="cuda",
+        dtype=torch.bfloat16,
+    ) -> torch.Tensor:
+        if seq_len != self.seq_len_cached:
+            self.seq_len_cached = seq_len
+            t = torch.arange(seq_len, device=device).type_as(self.inv_freq)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            emb = torch.cat((freqs, freqs), dim=-1).to(device)
+
+            if dtype in [torch.float16, torch.bfloat16]:
+                emb = emb.float()
+
+            self.cos_cached = emb.cos()[None, :, :]
+            self.sin_cached = emb.sin()[None, :, :]
+
+            self.cos_cached = self.cos_cached.type(dtype)
+            self.sin_cached = self.sin_cached.type(dtype)
+
+        return self.cos_cached, self.sin_cached
+
+    def forward(self, q, k):
+        batch, seq_len, head_dim = q.shape
+        cos, sin = self.cos_sin(seq_len)
+        return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+
+
+def _make_causal_mask(
+    input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
+) -> torch.BoolTensor:
+    batch_size, target_length = input_ids_shape
+    mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
+    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
+    seq_ids = torch.arange(target_length, device=device)
+    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]
+
+    if past_key_values_length > 0:
+        mask[:, :past_key_values_length] = False
+
+    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
+    return expanded_mask
+
+
+def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
+    batch_size, src_length = mask.shape
+    tgt_length = tgt_length if tgt_length is not None else src_length
+
+    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
+    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)
+
+
+def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
+    batch_size, seq_length = attention_mask.shape
+    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+    base = torch.tensor(
+        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+    )
+    powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != num_heads:
+        extra_base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+        )
+        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
+        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
+    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
+    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
+    # => the query_length dimension will then be broadcasted correctly
+    # This is more or less identical to T5's relative position bias:
+    # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
+    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
+    alibi = slopes[..., None].bfloat16() * arange_tensor
+    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
+
+
+def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
+    out = F.dropout(x, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+class LandmarkGroupedSoftmaxFunction(torch.autograd.Function):
+
+    # Note that forward, setup_context, and backward are @staticmethods
+    @staticmethod
+    def forward(ctx, x, dim, mem_cnt, resp_mem_idx):
+        new_shape = list(x.shape)
+        new_shape[dim] = mem_cnt # max_mem_cnt.item()
+        max_by_group = x.new_zeros((*new_shape,))
+        max_by_group.scatter_reduce_(src=x, index=resp_mem_idx, dim=dim, reduce="amax", include_self=False)
+            
+        maxes = torch.gather(max_by_group, dim, resp_mem_idx)
+        #x_exp = torch.exp(x - torch.where(torch.isinf(maxes), 0, maxes))
+        x_exp = torch.exp((x - maxes).to(torch.float32))
+
+        cumsum_by_group = torch.zeros_like(max_by_group, dtype=x_exp.dtype)
+
+        cumsum_by_group.scatter_add_(dim, resp_mem_idx, x_exp, )
+        denom = torch.gather(cumsum_by_group, dim, resp_mem_idx)
+
+        #probs = torch.where(denom < 0.5, 0, x_exp / denom)
+        probs = x_exp / denom
+        
+        
+        ctx.mem_cnt = mem_cnt
+        ctx.dim = dim
+        ctx.save_for_backward(resp_mem_idx, probs)
+
+        return probs
+
+    @staticmethod
+    def backward(ctx, grad_probs):
+        mem_cnt = ctx.mem_cnt
+        dim = ctx.dim
+        resp_mem_idx, probs = ctx.saved_tensors
+        grad_x = grad_dim = grad_mem_cnt = grad_resp_mem_idx = None
+
+        if ctx.needs_input_grad[0] or ctx.needs_input_grad[4]:
+            grad_pair = grad_probs * probs
+
+            new_shape = list(probs.shape)
+            new_shape[dim] = mem_cnt # max_mem_cnt.item()
+            cumsum_by_group = grad_pair.new_zeros((*new_shape,))
+            cumsum_by_group.scatter_add_(dim, resp_mem_idx, grad_pair)
+
+        if ctx.needs_input_grad[0]:
+            grad_sum = torch.gather(cumsum_by_group, dim, resp_mem_idx)
+            grad_x = grad_pair - probs * grad_sum
+        assert not ctx.needs_input_grad[1]
+        assert not ctx.needs_input_grad[2]
+        assert not ctx.needs_input_grad[3]
+        
+        return grad_x, grad_dim, grad_mem_cnt, grad_resp_mem_idx
+
+
+def landmark_grouped_softmax(x, dim, is_mem, last_section_mask):
+
+    last_and_rest_mask = last_section_mask # | mask
+
+    full_access_mask =  is_mem | last_and_rest_mask
+
+    max_mem_cnt = 64
+    mem_group_idx = torch.cumsum(is_mem, dim=dim)
+    mem_bucket_id = max_mem_cnt - 1
+    resp_mem_idx = torch.where(last_and_rest_mask, 
+                                max_mem_cnt - 1,
+                                torch.where(is_mem, mem_bucket_id, mem_group_idx))
+    probs = LandmarkGroupedSoftmaxFunction.apply(x, dim, max_mem_cnt, resp_mem_idx)
+
+    new_shape = list(x.shape)
+    new_shape[dim] = max_mem_cnt
+    group_prob = probs.new_zeros((*new_shape, ))
+    group_prob.scatter_(dim, torch.where(is_mem, mem_group_idx - 1, max_mem_cnt - 1), probs)
+    probs = probs.mul(torch.where(full_access_mask, last_section_mask, torch.gather(group_prob, dim, resp_mem_idx)))
+
+    return probs
+
+
+class Attention(nn.Module):
+    def __init__(self, config: RWConfig):
+        super().__init__()
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.n_head
+        self.head_dim = self.hidden_size // self.num_heads
+        self.split_size = self.hidden_size
+        self.hidden_dropout = config.hidden_dropout
+
+        if self.head_dim * self.num_heads != self.hidden_size:
+            raise ValueError(
+                f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+
+        self.maybe_rotary = RotaryEmbedding(config.head_dim) if config.rotary else lambda q, k: (q, k)
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+        self.beta = self.inv_norm_factor
+
+        self.query_key_value = Linear(
+            self.hidden_size,
+            3 * self.hidden_size if not config.multi_query else (self.hidden_size + 2 * self.head_dim),
+            bias=config.bias,
+        )
+        self.multi_query = config.multi_query
+        self.dense = Linear(self.hidden_size, self.hidden_size, bias=config.bias)
+        self.attention_dropout = nn.Dropout(config.attention_dropout)
+        self.num_kv = config.n_head if not self.multi_query else 1
+
+    def _split_heads(self, fused_qkv: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Split the last dimension into (num_heads, head_dim) without making any copies, results share same memory
+        storage as `fused_qkv`
+
+        Args:
+            fused_qkv (`torch.tensor`, *required*): [batch_size, seq_length, num_heads * 3 * head_dim]
+
+        Returns:
+            query: [batch_size, seq_length, num_heads, head_dim] key: [batch_size, seq_length, num_heads, head_dim]
+            value: [batch_size, seq_length, num_heads, head_dim]
+        """
+        if not self.multi_query:
+            batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
+            fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads, 3, self.head_dim)
+            return fused_qkv[..., 0, :], fused_qkv[..., 1, :], fused_qkv[..., 2, :]
+        else:
+            batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
+            fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads + 2, self.head_dim)
+            return fused_qkv[..., :-2, :], fused_qkv[..., [-2], :], fused_qkv[..., [-1], :]
+
+    def _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Merge heads together over the last dimenstion
+
+        Args:
+            x: (`torch.tensor`, *required*): [batch_size * num_heads, seq_length, head_dim]
+
+        Returns:
+            torch.tensor: [batch_size, seq_length, num_heads * head_dim]
+        """
+        # What we want to achieve is:
+        # batch_size * num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads * head_dim
+        batch_size_and_num_heads, seq_length, _ = x.shape
+        batch_size = batch_size_and_num_heads // self.num_heads
+
+        # First view to decompose the batch size
+        # batch_size * num_heads, seq_length, head_dim -> batch_size, num_heads, seq_length, head_dim
+        x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)
+
+        # batch_size, num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads, head_dim
+        x = x.permute(0, 2, 1, 3)
+
+        # batch_size, seq_length, num_heads, head_dim -> batch_size, seq_length, num_heads * head_dim
+        return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        alibi: torch.Tensor,
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+        is_mem: Optional[torch.Tensor] = None,
+        last_section_mask: Optional[torch.Tensor] = None,
+        offload_cache_to_cpu: bool = False,
+        use_flash: bool = False,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None
+    ):
+        fused_qkv = self.query_key_value(hidden_states)  # [batch_size, seq_length, 3 x hidden_size]
+
+        # 3 x [batch_size, seq_length, num_heads, head_dim]
+        (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
+
+        batch_size, q_length, _, _ = query_layer.shape
+
+        query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, q_length, self.head_dim)
+        key_layer = key_layer.transpose(1, 2).reshape(
+            batch_size * self.num_kv,
+            q_length,
+            self.head_dim,
+        )
+        value_layer = value_layer.transpose(1, 2).reshape(batch_size * self.num_kv, q_length, self.head_dim)
+
+        query_layer, key_layer = self.maybe_rotary(query_layer, key_layer)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            # concatenate along seq_length dimension:
+            #  - key: [batch_size * self.num_heads, head_dim, kv_length]
+            #  - value: [batch_size * self.num_heads, kv_length, head_dim]
+            key_layer = torch.cat((past_key, key_layer), dim=1)
+            value_layer = torch.cat((past_value, value_layer), dim=1)
+
+        _, kv_length, _ = key_layer.shape
+
+        if use_cache is True:
+            present = (key_layer, value_layer)
+        else:
+            present = None
+
+        if alibi is None:
+            query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
+            key_layer_ = key_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
+            value_layer_ = value_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
+
+            assert self.num_kv == 1
+
+            key_layer_ = key_layer_.expand(query_layer_.shape)
+            value_layer_ = value_layer_.expand(query_layer_.shape)
+
+            assert not output_attentions    # not supported.
+            assert mem_freq is not None
+            attn_output = fused_landmark_attention(query_layer_, key_layer_, value_layer_, is_mem, block_size=mem_freq+1)
+
+            # attn_output = F.scaled_dot_product_attention(
+            #     query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
+            # )
+
+            x = attn_output.view(batch_size, self.num_heads, q_length, self.head_dim)
+            x = x.permute(0, 2, 1, 3)
+            attn_output = x.reshape(batch_size, q_length, self.num_heads * self.head_dim)
+
+            output_tensor = self.dense(attn_output)
+
+            outputs = (output_tensor, present)
+            return outputs
+        else:
+            attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, -1e9).to(torch.bfloat16)
+            matmul_result = query_layer @ key_layer.transpose(-1, -2)
+
+            # change view to [batch_size, num_heads, q_length, kv_length]
+            attention_scores = matmul_result.view(batch_size, self.num_heads, q_length, kv_length)
+
+            # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype - [batch_size, num_heads, q_length, kv_length]
+            input_dtype = attention_scores.dtype
+            # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
+            if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
+                attention_scores = attention_scores.to(torch.float32)
+            # attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
+            attention_probs = F.softmax(
+                (attention_scores + alibi.view(batch_size, self.num_heads, 1, -1)) * self.inv_norm_factor + attention_mask_float,
+                dim=-1,
+                dtype=hidden_states.dtype,
+            )
+            # [batch_size, num_heads, q_length, kv_length]
+            attention_probs = self.attention_dropout(attention_probs)
+
+            if head_mask is not None:
+                attention_probs = attention_probs * head_mask
+
+            # change view [batch_size x num_heads, q_length, kv_length]
+            attention_probs_reshaped = attention_probs.view(batch_size * self.num_heads, q_length, kv_length)
+
+            # matmul: [batch_size * num_heads, q_length, head_dim]
+            context_layer = attention_probs_reshaped @ value_layer
+
+            # change view [batch_size, num_heads, q_length, head_dim]
+            context_layer = self._merge_heads(context_layer)
+
+            output_tensor = self.dense(context_layer)
+
+            outputs = (output_tensor, present)
+            if output_attentions:
+                outputs += (attention_probs,)
+
+            return outputs
+
+
+class MLP(nn.Module):
+    def __init__(self, config: RWConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.dense_h_to_4h = Linear(hidden_size, 4 * hidden_size, bias=config.bias)
+        self.act = nn.GELU()
+        self.dense_4h_to_h = Linear(4 * hidden_size, hidden_size, bias=config.bias)
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.act(self.dense_h_to_4h(x))
+        x = self.dense_4h_to_h(x)
+        return x
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, config: RWConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.num_heads = config.n_head
+        self.self_attention = Attention(config)
+
+        if not config.parallel_attn:
+            # unused if parallel attn
+            self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = MLP(config)
+
+        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
+        self.hidden_dropout = config.hidden_dropout
+
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        alibi: torch.Tensor,
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+        is_mem: Optional[torch.Tensor] = None,
+        last_section_mask: Optional[torch.Tensor] = None,
+        offload_cache_to_cpu: bool = False,
+        use_flash: bool = False,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None,
+    ):
+
+        layernorm_output = self.input_layernorm(hidden_states)
+        residual = hidden_states
+
+        # Self attention.
+        attn_outputs = self.self_attention(
+            layernorm_output,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            alibi=alibi,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            is_mem=is_mem,
+            last_section_mask=last_section_mask,
+            offload_cache_to_cpu=offload_cache_to_cpu,
+            use_flash=use_flash,
+            cache_top_k=cache_top_k,
+            mem_freq=mem_freq
+        )
+
+        attention_output = attn_outputs[0]
+
+        if not self.config.parallel_attn:
+            residual = dropout_add(attention_output, residual, self.config.attention_dropout, training=self.training)
+            layernorm_output = self.post_attention_layernorm(residual)
+
+        outputs = attn_outputs[1:]
+
+        # MLP.
+        mlp_output = self.mlp(layernorm_output)
+
+        if self.config.parallel_attn:
+            mlp_output += attention_output
+
+        output = dropout_add(mlp_output, residual, self.config.hidden_dropout, training=self.training)
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+
+class RWPreTrainedModel(PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RWConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["DecoderLayer"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear) or isinstance(module, Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
+        if isinstance(module, RWModel):
+            module.gradient_checkpointing = value
+
+    @staticmethod
+    def _convert_to_standard_cache(
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Standardizes the format of the cache so as to match most implementations, i.e. to tuple(tuple([batch_size,
+        num_heads, ...]))
+        """
+        batch_size_times_num_heads, head_dim, seq_length = past_key_value[0][0].shape
+        num_heads = batch_size_times_num_heads // batch_size
+        # key: [batch_size * num_heads, head_dim, seq_length] -> [batch_size, num_heads, head_dim, seq_length]
+        # value: [batch_size * num_heads, seq_length, head_dim] -> [batch_size, num_heads, seq_length, head_dim]
+        return tuple(
+            (
+                layer_past[0].view(batch_size, num_heads, head_dim, seq_length),
+                layer_past[1].view(batch_size, num_heads, seq_length, head_dim),
+            )
+            for layer_past in past_key_value
+        )
+
+    @staticmethod
+    def _convert_to_rw_cache(
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]]
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
+        batch_size, num_heads, head_dim, seq_length = past_key_value[0][0].shape
+        batch_size_times_num_heads = batch_size * num_heads
+        # key:  [batch_size, num_heads, head_dim, seq_length] -> [batch_size * num_heads, head_dim, seq_length]
+        # value: [batch_size, num_heads, seq_length, head_dim] -> [batch_size * num_heads, seq_length, head_dim]
+        return tuple(
+            (
+                layer_past[0].view(batch_size_times_num_heads, head_dim, seq_length),
+                layer_past[1].view(batch_size_times_num_heads, seq_length, head_dim),
+            )
+            for layer_past in past_key_value
+        )
+
+
+class RWModel(RWPreTrainedModel):
+    def __init__(self, config: RWConfig):
+        super().__init__(config)
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.n_head
+        self.alibi = config.alibi
+
+        # Embedding + LN Embedding
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)
+
+        # Transformer blocks
+        self.h = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
+
+        # Final Layer Norm
+        self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def _prepare_attn_mask(
+        self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
+    ) -> torch.BoolTensor:
+        # create causal mask
+        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
+        combined_attention_mask = None
+        device = attention_mask.device
+        _, src_length = input_shape
+
+        if src_length > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, device=device, past_key_values_length=past_key_values_length
+            )
+
+        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
+        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
+        combined_attention_mask = (
+            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
+        )
+
+        return combined_attention_mask
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        offload_cache_to_cpu: Optional[bool] = None,
+        use_flash: Optional[bool] = None,
+        cache_top_k: Optional[int] = None,
+        mem_freq: Optional[int] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
+        if deprecated_arguments.pop("position_ids", False) is not False:
+            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
+            warnings.warn(
+                "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
+                " passing `position_ids`.",
+                FutureWarning,
+            )
+        if len(deprecated_arguments) > 0:
+            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        is_mem = None
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+            if self.config.mem_id is not None:
+                with torch.no_grad():
+                    is_mem = input_ids == self.config.mem_id
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            if self.config.mem_id is not None:
+                raise NotImplementedError
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if past_key_values is None:
+            past_key_values = tuple([None] * len(self.h))
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape batch_size x num_heads x N x N
+        # head_mask has shape n_layer x batch x num_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        last_section_mask = None
+        if is_mem is not None and not use_flash:
+            is_mem = is_mem.unsqueeze(1).unsqueeze(2)
+            current_len = input_ids.shape[1]
+            mem_ids = torch.where(attention_mask[..., -current_len:] < -1, 0, torch.cumsum(is_mem, -1) - is_mem.int())
+            last_section_mask = torch.amax(mem_ids, -1, keepdim=True) == mem_ids
+            attention_mask[..., -current_len:].masked_fill_(last_section_mask & is_mem, torch.tensor(torch.finfo(inputs_embeds.dtype).min, device=inputs_embeds.device))
+            last_section_mask.logical_and_(attention_mask[..., -current_len:] > -1)
+            is_mem = is_mem.logical_and(attention_mask[..., -current_len:] > -1)
+
+        hidden_states = inputs_embeds
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        # Compute alibi tensor: check build_alibi_tensor documentation
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+        if past_key_values[0] is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            if len(past_key_values[0]) > 2:
+                past_key_values_length += past_key_values[0][3].shape[2] * past_key_values[0][3].shape[3]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+
+        if self.alibi:
+            alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
+        else:
+            alibi = None
+
+        causal_mask = self._prepare_attn_mask(
+            attention_mask,
+            input_shape=(batch_size, seq_length),
+            past_key_values_length=past_key_values_length,
+        )
+
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs, **kwargs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions, **kwargs)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    alibi,
+                    causal_mask,
+                    head_mask[i],
+                    use_reentrant=False,
+                    is_mem=is_mem,
+                    last_section_mask=last_section_mask,
+                    offload_cache_to_cpu=offload_cache_to_cpu,
+                    use_flash=use_flash,
+                    cache_top_k=cache_top_k,
+                    mem_freq=mem_freq,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=causal_mask,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    alibi=alibi,
+                    is_mem=is_mem,
+                    last_section_mask=last_section_mask,
+                    offload_cache_to_cpu=offload_cache_to_cpu,
+                    use_flash=use_flash,
+                    cache_top_k=cache_top_k,
+                    mem_freq=mem_freq,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        # Add last hidden state
+        hidden_states = self.ln_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class RWForCausalLM(RWPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
+
+    def __init__(self, config: RWConfig):
+        super().__init__(config)
+        self.transformer = RWModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.auto_insert_landmarks = False
+        self.always_use_flash = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: torch.Tensor):
+        self.lm_head = new_embeddings
+
+    def set_mem_id(self, mem_id):
+        if self.config.mem_id is not None:
+            assert mem_id == self.config.mem_id, "Chanigng mem_id can break the model. If you really intend to do this, manually disable this check"
+        self.config.mem_id = mem_id
+
+    def enable_landmark_insertion(self):
+        self.auto_insert_landmarks = True
+
+    def enable_flash(self):
+        self.always_use_flash = True
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> dict:
+        total_len = input_ids.shape[1]
+
+        # only last token for input_ids if past is not None
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+            # the cache may be in the stardard format (e.g. in contrastive search), convert to our's format if needed
+            if past[0][0].shape[0] == input_ids.shape[0]:
+                past = self._convert_to_rw_cache(past)
+            
+            use_flash = False if kwargs.get("use_flash") is not None else None
+        else:
+            prev_len = 0
+            use_flash = kwargs.get("use_flash")
+
+        mem_freq = kwargs.get("mem_freq") or self.config.mem_freq
+
+        if mem_freq is not None:
+            # if position_ids is not None:
+            #     raise NotImplementedError
+            T = input_ids.shape[1]
+
+            prev_incomplete_len = prev_len % mem_freq
+            prev_complete_len = prev_len - prev_incomplete_len
+            incomplete_len = total_len % mem_freq
+            new_full_len = total_len - prev_complete_len - incomplete_len
+
+            prev_input, input_ids_with_mem, input_ids_without_mem = torch.split(input_ids, (prev_complete_len, new_full_len, incomplete_len), dim=-1)
+            
+            bsz, q_len = input_ids.size()
+            input_ids_with_mem = input_ids_with_mem.view(bsz, -1, mem_freq)
+            input_ids_with_mem = torch.cat(
+                (
+                    input_ids_with_mem, 
+                    input_ids_with_mem.new_full((bsz, input_ids_with_mem.shape[1], 1), self.config.mem_id)
+                ), 
+                dim=-1
+            ).view(bsz, -1)
+            input_ids = torch.cat((prev_input, input_ids_with_mem, input_ids_without_mem), dim=-1)
+            if attention_mask is not None:
+                attention_mask_with_mem, attention_mask_without_mem = torch.split(attention_mask, (prev_complete_len + new_full_len, incomplete_len), dim=-1)
+                attention_mask_with_mem = attention_mask_with_mem.view(bsz, -1, mem_freq)
+                attention_mask_with_mem = torch.cat(
+                    (
+                        attention_mask_with_mem, 
+                        attention_mask_with_mem.new_ones((bsz, attention_mask_with_mem.shape[1], 1))
+                    ), 
+                    dim=-1
+                ).view(bsz, -1)
+                attention_mask = torch.cat((attention_mask_with_mem, attention_mask_without_mem), dim=-1)
+
+        input_ids = input_ids[:, prev_len:]
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+            "offload_cache_to_cpu": kwargs.get("offload_cache_to_cpu"),
+            "use_flash": use_flash,
+            "cache_top_k": kwargs.get("cache_top_k"),
+            "max_chunk_length": kwargs.get("max_chunk_length", 0),
+            "mem_freq": mem_freq,
+            "drop_last_logit_if_mem": True,
+        }
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        offload_cache_to_cpu: Optional[bool] = None,
+        use_flash: Optional[bool] = None,
+        cache_top_k: Optional[int] = None,
+        max_chunk_length: Optional[int] = 0,
+        mem_freq: Optional[int] = None,
+        drop_last_logit_if_mem: Optional[bool] = False,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        if deprecated_arguments.pop("position_ids", False) is not False:
+            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
+            warnings.warn(
+                "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
+                " passing `position_ids`.",
+                FutureWarning,
+            )
+        if len(deprecated_arguments) > 0:
+            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        use_flash = use_flash if use_flash is not None else self.always_use_flash
+
+        if self.auto_insert_landmarks:
+            mem_freq = self.config.mem_freq
+            assert self.config.mem_freq is not None
+            block_size = self.config.mem_freq + 1
+            input_ids = input_ids.view(input_ids.shape[0], -1, block_size - 1)
+            input_ids = torch.cat((input_ids, input_ids.new_full((input_ids.shape[0], input_ids.shape[1], 1), self.config.mem_id)), dim=-1)
+            input_ids = input_ids.view(input_ids.shape[0], -1)
+            if attention_mask is not None:
+                attention_mask = attention_mask.view(attention_mask.shape[0], -1, block_size - 1)
+                attention_mask = torch.cat((attention_mask, attention_mask.new_ones((attention_mask.shape[0], attention_mask.shape[1], 1))), dim=-1)
+                attention_mask = attention_mask.view(attention_mask.shape[0], -1)
+
+        if max_chunk_length == 0:
+            if cache_top_k is not None:
+                max_chunk_length = self.config.train_context_length - self.config.train_context_length % (mem_freq + 1) - (cache_top_k + 1) * (mem_freq + 1)
+                if max_chunk_length <= 0:
+                    raise ValueError("K is too large for this model.")
+            else:
+                max_chunk_length = None
+
+        window_len = max_chunk_length or input_ids.shape[1]
+        if use_flash:
+            assert window_len % (mem_freq + 1) == 0
+        last_logits = None
+        for step, idx in enumerate(range(0, input_ids.shape[1], window_len)):
+            if idx >= 1:
+                if output_attentions or output_hidden_states:
+                    raise NotImplementedError
+                if not use_cache:
+                    raise NotImplementedError
+
+            outputs = self.transformer(
+                input_ids[:, idx:idx + window_len],
+                past_key_values=past_key_values,
+                attention_mask=attention_mask[:, :idx + window_len + attention_mask.shape[1] - input_ids.shape[1]] if attention_mask is not None else None,
+                head_mask=head_mask,    ## ??
+                inputs_embeds=inputs_embeds[:, idx:idx + window_len] if inputs_embeds is not None else None,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                offload_cache_to_cpu=offload_cache_to_cpu,
+                use_flash=(use_flash or self.auto_insert_landmarks),
+                cache_top_k=cache_top_k,
+                mem_freq=mem_freq,
+            )
+            past_key_values = outputs[1]
+            if last_logits is not None:
+                last_logits = torch.cat((last_logits, outputs[0]), dim=-2)
+            last_logits = outputs[0]
+
+        hidden_states = last_logits
+        if self.auto_insert_landmarks:
+            block_size = self.config.mem_freq + 1
+            hidden_states = hidden_states.reshape(hidden_states.shape[0], hidden_states.shape[1] // block_size, block_size, hidden_states.shape[2])
+            hidden_states = hidden_states[:, :, :block_size - 1]
+            hidden_states = hidden_states.reshape(hidden_states.shape[0], -1, hidden_states.shape[3])
+        if drop_last_logit_if_mem:
+            is_any_mem = (input_ids[:, -1] == self.config.mem_id).any()
+            are_all_mem = (input_ids[:, -1] == self.config.mem_id).all()
+            assert is_any_mem == are_all_mem
+            if is_any_mem:
+                hidden_states = hidden_states[:, :-1]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            batch_size, seq_length, vocab_size = shift_logits.shape
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _reorder_cache(
+        self, past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+
+        Output shares the same memory storage as `past`.
+        """
+        standardized_past = self._convert_to_standard_cache(past, batch_size=len(beam_idx))
+
+        # Get a copy of `beam_idx` on all the devices where we need those indices.
+        device_to_beam_idx = {
+            past_state.device: beam_idx.to(past_state.device) for layer_past in past for past_state in layer_past
+        }
+        reordered_past = tuple(
+            (
+                layer_past[0].index_select(0, device_to_beam_idx[layer_past[0].device]),
+                layer_past[1].index_select(0, device_to_beam_idx[layer_past[0].device]),
+            )
+            for layer_past in standardized_past
+        )
+        return self._convert_to_rw_cache(reordered_past)
+
+
+class RWForSequenceClassification(RWPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
+
+    def __init__(self, config: RWConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = RWModel(config)
+        self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        if deprecated_arguments.pop("position_ids", False) is not False:
+            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
+            warnings.warn(
+                "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
+                " passing `position_ids`.",
+                FutureWarning,
+            )
+        if len(deprecated_arguments) > 0:
+            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(dim=-1) - 1
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+class RWForTokenClassification(RWPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
+
+    def __init__(self, config: RWConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = RWModel(config)
+        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        if deprecated_arguments.pop("position_ids", False) is not False:
+            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
+            warnings.warn(
+                "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
+                " passing `position_ids`.",
+                FutureWarning,
+            )
+        if len(deprecated_arguments) > 0:
+            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            batch_size, seq_length = labels.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length))
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+class RWForQuestionAnswering(RWPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = RWModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

code/redpajama.py

@@ -0,0 +1,174 @@
+# Copyright 2023 Together Computer
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""RedPajama: An Open-Source, Clean-Room 1.2 Trillion Token Dataset."""
+
+
+import json
+
+import datasets
+import traceback
+import numpy as np
+import math
+
+logger = datasets.logging.get_logger(__name__)
+
+
+_DESCRIPTION = """\
+RedPajama is a clean-room, fully open-source implementation of the LLaMa dataset.
+"""
+
+_URL_LISTS = {
+    "arxiv": "urls/arxiv.txt",
+    "book": "urls/book.txt",
+    "c4": "urls/c4.txt",
+    "common_crawl": "urls/common_crawl.txt",
+    "github": "urls/github.txt",
+    "stackexchange": "urls/stackexchange.txt",
+    "wikipedia": "urls/wikipedia.txt",
+}
+
+
+class RedPajama1TConfig(datasets.BuilderConfig):
+    """BuilderConfig for RedPajama sample."""
+
+    def __init__(self, *args, subsets, p_sample=None, **kwargs):
+        """BuilderConfig for RedPajama.
+        Args:
+          **kwargs: keyword arguments forwarded to super.
+        """
+        super(RedPajama1TConfig, self).__init__(**kwargs)
+
+        self.subsets = subsets
+        self.p_sample = p_sample
+
+
+class RedPajama1T(datasets.GeneratorBasedBuilder):
+    """RedPajama: Reproducing the LLaMA training dataset of over 1.2 trillion tokens. Version 1.0.0."""
+    BUILDER_CONFIG_CLASS = RedPajama1TConfig
+    BUILDER_CONFIGS = [
+        RedPajama1TConfig(
+            subsets = list(_URL_LISTS.keys()),
+            name="plain_text",
+            version=datasets.Version("1.0.0", ""),
+            description="Plain text",
+        ),
+        RedPajama1TConfig(
+            subsets = list(_URL_LISTS.keys()),
+            name="plain_text_tenpercent",
+            version=datasets.Version("1.0.0", ""),
+            description="Plain text",
+            p_sample=0.1
+        ),
+    ]
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(
+                {
+                    "text": datasets.Value("string"),
+                    "meta": datasets.Value("string"),
+                    "red_pajama_subset": datasets.Value("string"),
+                }
+            ),
+            supervised_keys=None,
+        )
+
+    def _split_generators(self, dl_manager):
+        url_lists = dl_manager.download_and_extract({
+            subset: _URL_LISTS[subset] for subset in self.config.subsets
+        })
+
+        urls = {}
+        rng = np.random.default_rng(seed=2)
+
+        for subset, url_list in url_lists.items():
+            with open(url_list, encoding="utf-8") as f:
+                urls[subset] = [line.strip() for line in f]
+            if self.config.p_sample is not None:
+                urls[subset] = rng.choice(
+                    urls[subset], 
+                    size=int(math.ceil(len(urls[subset]) * self.config.p_sample)), replace=False).tolist()
+
+        downloaded_files = dl_manager.download(urls)
+
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs = {
+                    "files": {
+                        subset: downloaded_files[subset]
+                        for subset in self.config.subsets
+                    }
+                }
+            )
+        ]
+
+    def _generate_examples(self, files):
+        """This function returns the examples in the raw (text) form."""
+        key = 0
+        for subset in files:
+            if subset == "common_crawl":
+                import zstandard as zstd
+
+                for path in files[subset]:
+                    with zstd.open(open(path, "rb"), "rt", encoding="utf-8") as f:
+                        for i, row in enumerate(f):
+                            try:
+                                data = json.loads(row)
+                                text = data["text"]
+                                del data["text"]
+                                yield key, {
+                                    "text": text,
+                                    "meta": json.dumps(data),
+                                    "red_pajama_subset": subset,
+                                }
+                                key += 1
+                            except Exception as e:
+                                print(f'Subset: {subset}')
+                                print(f'Path: {path}')
+                                print(f'Row: {row}')
+                                traceback.print_exc()
+
+                                raise e
+            else:
+                for path in files[subset]:
+                    with open(path, encoding="utf-8") as f:
+                        for i, row in enumerate(f):
+                            try:
+                                data = json.loads(row)
+                                if "meta" not in data:
+                                    text = data["text"]
+                                    del data["text"]
+                                    yield key, {
+                                        "text": text,
+                                        "meta": json.dumps(data),
+                                        "red_pajama_subset": subset,
+                                    }
+                                else:
+                                    yield key, {
+                                        "text": data["text"],
+                                        "meta": data["meta"],
+                                        "red_pajama_subset": subset,
+                                    }
+                                key += 1
+                            except Exception as e:
+                                print(f'Subset: {subset}')
+                                print(f'Path: {path}')
+                                print(f'Row: {row}')
+                                traceback.print_exc()
+
+                                raise e

code/run_test.py

@@ -0,0 +1,171 @@
+# Copyright 2023 Amirkeivan Mohtashami, Martin Jaggi
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+import os
+import random
+import re
+import requests
+
+
+llama_weights_7b_base = "/llama_weights/7B_hf/"
+llama_weights_7b_tuned = "/llama-redpajama-mem-15000-with-mem/"
+cache_path = "/hf-cache/"
+use_flash = False # using flash for inference is only implemented for when offloading kv to cpu
+top_k = 5
+dtype = torch.bfloat16
+
+def make_llama_base_pipe():
+
+    from transformers import pipeline
+
+    from transformers.models.llama import LlamaForCausalLM
+
+    llama_base = LlamaForCausalLM.from_pretrained(
+        llama_weights_7b_base,
+        cache_dir=cache_path,
+    )
+
+    llama_base = llama_base.to('cuda:0')
+
+    import transformers
+    
+    tokenizer = transformers.AutoTokenizer.from_pretrained(
+        llama_weights_7b_base,
+        cache_dir=cache_path,
+        model_max_length=2048,
+        padding_side="right",
+        use_fast=False,
+    )
+
+    llama_base_pipe = pipeline("text-generation", model=llama_base, tokenizer=tokenizer, device=llama_base.device)
+    return llama_base_pipe
+
+
+
+llama_base_pipe = make_llama_base_pipe()
+
+def make_llama_mem_pipe():
+    from llama_mem import LlamaForCausalLM
+
+    model = LlamaForCausalLM.from_pretrained(
+        llama_weights_7b_tuned,
+        cache_dir=cache_path,
+        torch_dtype=dtype
+    )
+
+    model.to('cuda:1')
+
+    import transformers
+
+    tokenizer = transformers.AutoTokenizer.from_pretrained(
+            llama_weights_7b_tuned,
+            cache_dir=cache_path,
+            model_max_length=model.config.train_context_length,
+            padding_side="right",
+            use_fast=False,
+        )
+    mem_id = tokenizer.convert_tokens_to_ids("<landmark>")
+    model.set_mem_id(mem_id)
+    from transformers import pipeline
+    llama_mem_pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=model.device,
+                              offload_cache_to_cpu=use_flash, use_flash=use_flash, 
+                              cache_top_k=top_k)
+    return llama_mem_pipe
+
+
+llama_mem_pipe = make_llama_mem_pipe()
+
+
+
+pipes = {"base": llama_base_pipe, "mem": llama_mem_pipe}
+
+
+def generate_prompt(n_garbage):
+    """Generates a text file and inserts an execute line at a random position."""
+    n_garbage_prefix = random.randint(0, n_garbage)
+    n_garbage_suffix = n_garbage - n_garbage_prefix
+    
+    task_description = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there."
+    garbage = "The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again."
+    garbage_inf = " ".join([garbage] * 2000)
+    assert len(garbage_inf) >= n_garbage
+    garbage_prefix = garbage_inf[:n_garbage_prefix]
+    garbage_suffix = garbage_inf[:n_garbage_suffix]
+    pass_key = random.randint(1, 50000)
+    information_line = f"The pass key is {pass_key}. Remember it. {pass_key} is the pass key."
+    final_question = "What is the pass key? The pass key is"
+    lines = [
+        task_description,
+        garbage_prefix,
+        information_line,
+        garbage_suffix,
+        final_question
+    ]
+    return "\n".join(lines), pass_key
+            
+
+
+def test_model(prompt_text, pass_key, model_name):
+    response = pipes[model_name](prompt_text,num_return_sequences=1, max_new_tokens=10)[0]["generated_text"][len(prompt_text):]
+    assert f"The pass key is {pass_key}" in prompt_text
+
+    try:
+        pass_key = int(re.search(r'\d+', response).group())
+    except:
+        pass_key = response[:20]
+    
+    return pass_key
+
+
+n_values = [0, 100, 500, 1000, 5000, 8000, 10000, 12000, 14000, 18000, 20000, 25000, 38000]
+num_tests = 50
+models = ["base", "mem"]
+accuracies = {x: [] for x in models}
+individual_results = {x: [] for x in models}
+
+for n in n_values:
+    
+    correct_count = {x: 0 for x in models}
+    
+    n_results = {x: [] for x in models}
+    for i in range(num_tests):
+        print(f"\nRunning test {i + 1}/{num_tests} for n = {n}...")
+        prompt_text, pass_key = generate_prompt(n)
+        
+        
+        
+        for model_name in models:
+            if pipes[model_name] is None:
+                continue
+            num_tokens = len(pipes[model_name].tokenizer.encode(prompt_text))
+
+            print("Number of tokens in this prompt: ", num_tokens)
+            model_output = test_model(prompt_text, pass_key, model_name)
+            print(f"Expected number in the prompt: {pass_key}, {model_name} output: {model_output}")
+
+            if pass_key == model_output:
+                correct_count[model_name] += 1
+                n_results[model_name].append(1)
+                print("Success!")
+            else:
+                n_results[model_name].append(0)
+                print("Fail.")
+    
+    for model in models:
+        accuracy = (correct_count[model] / num_tests) * 100
+        print(f"Accuracy {model} for n = {n}: {accuracy}%")
+        accuracies[model].append(accuracy)
+        individual_results[model].append(n_results)

code/run_train_1x.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+CUDA_VISIBLE_DEVICES=0 python train.py  \
+    --model_name_or_path tiiuae/falcon-7b \
+    --bf16 True \
+    --output_dir ./out_dir/  \
+    --cache_dir ./hf-cache/ \
+    --num_train_epochs 1  \
+    --per_device_train_batch_size 1     \
+    --per_device_eval_batch_size 1     \
+    --gradient_accumulation_steps 1     \
+    --evaluation_strategy "no"     \
+    --save_strategy "steps"     \
+    --save_steps 2000     \
+    --save_total_limit 2     \
+    --learning_rate 2e-5     \
+    --weight_decay 0.1     \
+    --warmup_ratio 0.03     \
+    --lr_scheduler_type "cosine"     \
+    --logging_steps 1     \
+    --tf32 True \
+    --max_steps 15000 \
+    --model_max_length 1024 \
+    --mem_freq 31

code/run_train_8x.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+torchrun --nproc_per_node=8 train.py  \
+    --model_name_or_path tiiuae/falcon-7b \
+    --bf16 True \
+    --output_dir ./out_dir/  \
+    --cache_dir ./hf-cache/ \
+    --num_train_epochs 1  \
+    --per_device_train_batch_size 2     \
+    --per_device_eval_batch_size 2     \
+    --gradient_accumulation_steps 8     \
+    --evaluation_strategy "no"     \
+    --save_strategy "steps"     \
+    --save_steps 2000     \
+    --save_total_limit 2     \
+    --learning_rate 2e-5     \
+    --weight_decay 0.1     \
+    --warmup_ratio 0.03     \
+    --lr_scheduler_type "cosine"     \
+    --logging_steps 1     \
+    --tf32 True \
+    --max_steps 15000 \
+    --model_max_length 2048 \
+    --mem_freq 31 \
+    --fsdp "full_shard auto_wrap"     \
+    --fsdp_transformer_layer_cls_to_wrap 'DecoderLayer'

code/train.py

@@ -0,0 +1,206 @@
+# Copyright 2023 Amirkeivan Mohtashami, Martin Jaggi
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+#import copy
+#import logging
+from dataclasses import dataclass, field
+from functools import partial
+from typing import Dict, Optional, Sequence
+
+
+import torch
+import transformers
+#from torch.utils.data import Dataset
+from transformers import Trainer, DataCollatorForLanguageModeling, get_cosine_schedule_with_warmup
+
+from modelling_RW import RWForCausalLM
+#from transformers import AutoModelForCausalLM
+
+
+from torch.distributed import barrier
+import os
+
+
+from datasets import load_dataset
+
+IGNORE_INDEX = -100
+DEFAULT_PAD_TOKEN = "[PAD]"
+DEFAULT_EOS_TOKEN = "</s>"
+DEFAULT_BOS_TOKEN = "<s>"
+DEFAULT_UNK_TOKEN = "<unk>"
+
+
+@dataclass
+class ModelArguments:
+    model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    cache_dir: Optional[str] = field(default=None)
+    #optim: str = field(default="adamw_hf")
+    optim: str = field(default="adamw_torch")
+    model_max_length: int = field(
+        default=128,
+        metadata={"help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)."},
+    )
+    use_flash: bool = field(default=False)
+    mem_freq: int = field(default=63)
+    #report_to: str = "none" # disable logging
+
+
+class TrainerCosine(Trainer):
+    def create_scheduler(self, num_training_steps: int, optimizer: torch.optim.Optimizer = None):
+        """
+        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        passed as an argument.
+
+        Args:
+            num_training_steps (int): The number of training steps to do.
+        """
+        if self.args.lr_scheduler_type != "cosine":
+            return super().create_scheduler(num_training_steps, optimizer)
+        if self.lr_scheduler is None:
+            self.lr_scheduler = get_cosine_schedule_with_warmup(
+                optimizer=self.optimizer if optimizer is None else optimizer,
+                num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                num_training_steps=num_training_steps,
+                num_cycles=0.4 # ~10% of the init lr
+            )
+        return self.lr_scheduler
+
+
+def smart_tokenizer_and_embedding_resize(
+    special_tokens_dict: Dict,
+    tokenizer: transformers.PreTrainedTokenizer,
+    model: transformers.PreTrainedModel,
+):
+    """Resize tokenizer and embedding.
+
+    Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
+    """
+    num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
+    model.resize_token_embeddings(len(tokenizer))
+
+    if num_new_tokens > 0:
+        input_embeddings = model.get_input_embeddings().weight.data
+        output_embeddings = model.get_output_embeddings().weight.data
+
+        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+
+        input_embeddings[-num_new_tokens:] = input_embeddings_avg
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+def tokenize_fn(tokenizer, example):
+    context_length = tokenizer.model_max_length
+    outputs = tokenizer(
+        tokenizer.eos_token.join(example["text"]),
+        truncation=False,
+        return_tensors="pt",
+        pad_to_multiple_of=context_length,
+        padding=True,
+    )
+    return {"input_ids": outputs["input_ids"].view(-1, context_length)}
+
+def train():
+    parser = transformers.HfArgumentParser((ModelArguments, TrainingArguments))
+    model_args, training_args = parser.parse_args_into_dataclasses()
+
+    # ensure max length leaves room for landmark tokens
+    model_max_length = training_args.model_max_length - (training_args.model_max_length // training_args.mem_freq)
+    model_max_length = model_max_length // training_args.mem_freq * training_args.mem_freq
+
+    tokenizer = transformers.AutoTokenizer.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=training_args.cache_dir,
+        model_max_length=model_max_length,
+        padding_side="right",
+        use_fast=False,
+    )
+    special_tokens_dict = dict()
+    if tokenizer.pad_token is None:
+        special_tokens_dict["pad_token"] = DEFAULT_PAD_TOKEN
+    if tokenizer.eos_token is None:
+        special_tokens_dict["eos_token"] = DEFAULT_EOS_TOKEN
+    if tokenizer.bos_token is None:
+        special_tokens_dict["bos_token"] = DEFAULT_BOS_TOKEN
+    if tokenizer.unk_token is None:
+        special_tokens_dict["unk_token"] = DEFAULT_UNK_TOKEN
+    mem_token = "<landmark>"
+    special_tokens_dict["additional_special_tokens"] = [mem_token]
+
+    model = RWForCausalLM.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=training_args.cache_dir,
+        mem_freq=training_args.mem_freq,
+        torch_dtype=torch.bfloat16,
+    )
+    # model = AutoModelForCausalLM.from_pretrained(
+    #     model_args.model_name_or_path,
+    #     cache_dir=training_args.cache_dir,
+    #     torch_dtype=torch.bfloat16,
+    #     trust_remote_code=True,
+    # )
+
+    smart_tokenizer_and_embedding_resize(
+        special_tokens_dict=special_tokens_dict,
+        tokenizer=tokenizer,
+        model=model,
+    )
+
+    mem_id = tokenizer.convert_tokens_to_ids(mem_token)
+    model.set_mem_id(mem_id)
+    print(f"Landmark token: {mem_token}: {mem_id}")
+    
+    rank = int(os.environ.get('RANK', -1))
+    if rank > 0:
+        barrier()
+    #dataset = load_dataset("togethercomputer/RedPajama-Data-1T-Sample", cache_dir=training_args.cache_dir, split='train[:100]')
+    dataset = load_dataset("togethercomputer/RedPajama-Data-1T-Sample", cache_dir=training_args.cache_dir, split='train')
+
+    dataset = dataset.map(partial(tokenize_fn, tokenizer), batched=True, num_proc=32, remove_columns=["text", "meta"])
+
+    model.enable_landmark_insertion()
+    model.enable_flash()
+    
+    # if training_args.use_flash:
+    #     model.enable_landmark_insertion()
+    #     model.enable_flash()
+    # else:
+    #     dataset = dataset.map(
+    #         partial(
+    #             add_mem_tokens, 
+    #             mem_freq=training_args.mem_freq, 
+    #             mem_id=mem_id
+    #         ), batched=False, num_proc=32)
+
+    if rank == 0:
+        barrier()
+    print(dataset)
+
+    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+    trainer = TrainerCosine(
+        model=model, tokenizer=tokenizer, args=training_args, 
+        train_dataset=dataset,  #dataset["train"],
+        eval_dataset=None,
+        data_collator=data_collator)
+    trainer.train()
+    trainer.save_state()
+    trainer.save_model(output_dir=training_args.output_dir)
+
+
+if __name__ == "__main__":
+    train()

code/weight_diff.py

@@ -0,0 +1,158 @@
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+# This file has been changed by Amirkeivan Mohtashami
+# to take into account the new token in the embedding layer
+
+import os
+from typing import Optional
+
+import fire
+import torch
+import tqdm
+import transformers
+from train import smart_tokenizer_and_embedding_resize
+import llama_mem
+
+@torch.inference_mode()
+def make_diff(
+    path_raw: str, path_tuned: str, path_diff: str, device="cpu",  # "cuda" or "cpu"
+):
+    """Make the weight diff.
+
+    This function is given to present full transparency of how the weight diff was created.
+
+    Run:
+        python weight_diff.py make_diff --path_raw <your_path_raw> --path_tuned <your_path_tuned> --path_diff <your_path_diff>
+    """
+    model_tuned: transformers.PreTrainedModel = llama_mem.LlamaForCausalLM.from_pretrained(
+        path_tuned,
+        device_map={"": torch.device(device)},
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=True,
+    )
+    model_raw: transformers.PreTrainedModel = transformers.AutoModelForCausalLM.from_pretrained(
+        path_raw,
+        device_map={"": torch.device(device)},
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=True,
+    )
+
+    tokenizer_tuned: transformers.PreTrainedTokenizer = transformers.AutoTokenizer.from_pretrained(
+        path_tuned
+    )
+    tokenizer_raw: transformers.PreTrainedTokenizer = transformers.AutoTokenizer.from_pretrained(
+        path_raw
+    )
+    smart_tokenizer_and_embedding_resize(
+        special_tokens_dict=dict(pad_token="[PAD]", additional_special_tokens=["<landmark>"]),
+        model=model_raw,
+        tokenizer=tokenizer_raw,
+    )
+
+
+
+    state_dict_tuned = model_tuned.state_dict()
+    state_dict_raw = model_raw.state_dict()
+    with open(os.path.join(path_diff, "checksum_psum.txt"), "w") as f:
+        f.write(str(sum(state_dict_tuned[key].sum().item() for key in state_dict_tuned)))
+
+    for key in tqdm.tqdm(state_dict_tuned):
+        state_dict_tuned[key].add_(-state_dict_raw[key])
+
+    model_tuned.save_pretrained(path_diff)
+    tokenizer_tuned.save_pretrained(path_diff)
+
+
+@torch.inference_mode()
+def recover(
+    path_raw,
+    path_diff,
+    path_tuned: Optional[str] = None,
+    device="cpu",
+    test_inference=True,
+    check_integrity_naively=True,
+):
+    """Recover the original weights from the released weight diff.
+
+    This function is given for you to run.
+
+    Things to do before running this:
+        1. Convert Meta's released weights into huggingface format. Follow this guide:
+            https://huggingface.co/docs/transformers/main/model_doc/llama
+        2. Make sure you cloned the released weight diff into your local machine. The weight diff is located at:
+            https://huggingface.co/tatsu-lab/alpaca-7b/tree/main
+        3. Run this function with the correct paths. E.g.,
+            python weight_diff.py recover --path_raw <path_to_step_1_dir> --path_diff <path_to_step_2_dir>
+
+    Additional notes:
+        - If things run too slowly, and you have an 80G GPU lying around, let GPU go brrr by setting `--device "cuda"`.
+        - If you want to save the recovered weights, set `--path_tuned <your_path_tuned>`.
+            Next time you can load the recovered weights directly from `<your_path_tuned>`.
+    """
+    model_raw: transformers.PreTrainedModel = transformers.AutoModelForCausalLM.from_pretrained(
+        path_raw,
+        device_map={"": torch.device(device)},
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=True,
+    )
+    model_recovered: transformers.PreTrainedModel = llama_mem.LlamaForCausalLM.from_pretrained(
+        path_diff,
+        device_map={"": torch.device(device)},
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=True,
+    )
+
+    tokenizer_raw: transformers.PreTrainedTokenizer = transformers.AutoTokenizer.from_pretrained(
+        path_raw
+    )
+    smart_tokenizer_and_embedding_resize(
+        special_tokens_dict=dict(pad_token="[PAD]", additional_special_tokens=["<landmark>"]),
+        model=model_raw,
+        tokenizer=tokenizer_raw,
+    )
+    tokenizer_recovered: transformers.PreTrainedTokenizer = transformers.AutoTokenizer.from_pretrained(
+        path_diff
+    )
+
+    state_dict_recovered = model_recovered.state_dict()
+    state_dict_raw = model_raw.state_dict()
+    for key in tqdm.tqdm(state_dict_recovered):
+        state_dict_recovered[key].add_(state_dict_raw[key])
+
+    if check_integrity_naively:
+        # This is not a rigorous, cryptographically strong integrity check :)
+        allsum = sum(state_dict_recovered[key].sum() for key in state_dict_recovered)
+        if os.path.exists(os.path.join(path_diff, "checksum_psum.txt")):
+            with open(os.path.join(path_diff, "checksum_psum.txt")) as f:
+                expected_sum = float(f.read())
+        else:
+            expected_sum = 49798.7656 # backward compatibility with the first released weights
+        assert torch.allclose(
+            allsum, torch.full_like(allsum, fill_value=expected_sum), atol=1e-2, rtol=0
+        ), "Naive integrity check failed. This could imply that some of the checkpoint files are corrupted."
+
+    if path_tuned is not None:
+        model_recovered.save_pretrained(path_tuned)
+        tokenizer_recovered.save_pretrained(path_tuned)
+
+    return model_recovered, tokenizer_recovered
+
+
+def main(task, **kwargs):
+    globals()[task](**kwargs)
+
+
+if __name__ == "__main__":
+    fire.Fire(main)