Upload folder using huggingface_hub

LICENSE.txt

@@ -0,0 +1,72 @@
+TII Falcon LLM License Version 1.0
+May 2023
+falconllm.tii.ae
+
+INTRODUCTORY NOTE
+This license is, in part, based on the Apache License Version 2.0 (available at http://www.apache.org/licenses/), with a series of modifications. The contribution of the Apache License 2.0 to the framing of this document is acknowledged. Please read this license carefully, as it is different to other ‘open source’ licenses you may have encountered previously. In particular, note that this license contains obligations on those of you who are commercially exploiting Falcon LLM or any Derivative Work to make royalty payments.
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 
+1	Definitions.
+"Commercial Application Address” means Falconllm.sales@tii.ae.
+“Commercial Use” means use where there is, or, in relation to a new use case, a reasonable expectation that there will be revenue directly attributable to the use of the Work for that use case.
+“Commercial User” means someone who has applied to make Commercial Use of the Work and been granted permission by the Licensor to make such Commercial Use in accordance with Section 8. 
+“Contribution” shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, “submitted” means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as “Not a Contribution.” 
+“Contributor” shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
+“Derivative Works” shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall include machine learning models trained using outputs from the Falcon LLM or any other Derivative Work, but shall not otherwise include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
+“Falcon LLM” shall mean only the following releases of TII’s Falcon large language models: (i) Falcon-RW-1B; (ii) Falcon-RW-7B; (iii) Falcon-7B; (iv) Falcon-40B; (v) Falcon-7B-Instruct; or (vi) Falcon-40B-Instruct; each of which is initially made available in Object form only under this license at FalconLLM.tii.ae. No other sizes or versions of the ‘Falcon’ family of large language models is made available under this license. 
+“Legal Entity” shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, “control” means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
+“License” shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 to 11 of this document.
+“Licensor” shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
+“Object” form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, a trained and/or fine-tuned machine learning model or artificial intelligence model, generated documentation, and conversions to other media types.
+“Source” form shall mean the preferred form for making modifications, including but not limited to software source code, training datasets used for training or fine tuning a machine learning model or artificial intelligence model, documentation source, and configuration files.
+“TII” shall mean the Technology Innovation Institute – Sole Proprietorship L.L.C., or any party nominated in writing by Technology Innovation Institute – Sole Proprietorship L.L.C. as its successor for the purposes of this License, or any party nominated in writing to be a successor to any successor for the purposes of this license.
+“Work” shall mean the work of authorship, which in relation to the initial release of Falcon LLM is in Object form only, but in the case of any and all Derivative Works means the work of authorship whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). 
+“You” (or “Your”) shall mean an individual or Legal Entity exercising permissions granted by this License.
+2	Grant of Copyright License. 
+2.1	Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
+2.2	Other than where you are a Commercial User in accordance with Section 8, Your copyright license to use the Work shall be royalty free and without charge. 
+3	Grant of Patent License. 
+3.1	Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
+3.2	Other than where you are a Commercial User in accordance with Section 8, Your patent license to use the Work shall be royalty free and without charge. 
+4	Redistribution. 
+4.1	Irrespective of whether you are a Commercial User, You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: 
+(a)	You must give any other recipients of the Work or Derivative Works a copy of this License; and 
+(b)	You must cause any modified files to carry prominent notices stating that You changed the files; and 
+(c)	You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and 
+(d)	If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. 
+4.2	Subject to Sections 5 and 8 below, where you are a Commercial User , then You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
+4.3	Where You are not a Commercial User, then: 
+(a)	You must apply this license in an unmodified form as the terms under which you distribute Your modifications or any Derivative Work. Making the Work or Derivative Work available to any third party users via any means will constitute distribution of the Work or Derivative Work for these purposes. 
+(b)	Other than in respect of Falcon LLM, which may only be distributed in Object form, where you distribute Your modifications or any Derivative Work then You must provide both Object and Source versions of the modifications or Derivative Work, provided that this obligation may be satisfied by providing users of the Object version of Your modifications or Derivative Work a link to a location where they may freely download a copy of the Source version.
+(c)	You may include a copyright statement recording the nature of your modifications or Derivative Work, but such copyright statement shall be based on the copyright statement provided with the Work and you shall make only such changes as are strictly necessary to record your contribution.  
+5	Publication 
+5.1	You shall include prominently in any public statement regarding a Derivative Work the following statement: 
+“[name of relevant Derivative Work] is built using Falcon LLM technology from the Technology Innovation Institute”. 
+5.2	You may request from TII a reasonably adjusted version of the above statement to suit the publication the relevant statement is being made in. TII will not unreasonably withhold or delay approval of such a request. 
+5.3	You shall comply with any reasonable request from TII regarding the prominence and phrasing of any reference to the ‘Falcon LLM technology’ or the ‘Technology Innovation Institute’.
+6	Submission of Contributions. 
+6.1	Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions.
+6.2	Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
+7	Trademarks. 
+7.1	Except as required for compliance with Section 5 of this License, this License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
+8	Commercial Use
+8.1	Where You wish to make Commercial Use of Falcon LLM or any Derivative Work, You must apply to TII for permission to make Commercial Use of that Work in writing via the means specified from time to time at the Commercial Application Address, providing such information as may be required.
+8.2	Where TII grants permission for You to make Commercial Use of the relevant Work, then for that purpose You shall be considered a Commercial User, and:
+(a)	In its written grant of permission, TII shall set the royalty rate that will apply to you as a Commercial User as a percentage of revenue ( “Relevant Percentage”), where, unless otherwise specified in the grant of permission, the Relevant Percentage shall be 10%; and 
+(b)	Each year on the anniversary of the date upon which you were granted permission by TII to make Commercial Use of the relevant Work (the “Anniversary Date") You shall account to TII in writing in full for all revenue you have received in the previous 12 months which is attributable (whether directly or indirectly) to Your use of the relevant Work (“Attributable Revenue”); and
+(c)	Where, on the Anniversary Date, the Attributable Revenue for the preceding 12 months is greater than $1m or its equivalent in the currency or currencies in which the revenue has been earned (the “Royalty Threshold”) then You shall make a payment of the Relevant Percentage of the relevant Attributable Revenue that exceeds the Royalty Threshold in full in cleared funds to TII into the account specified by TII from time to time in writing for such purpose within 30 days of that Anniversary Date.
+9	Disclaimer of Warranty. 
+9.1	Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
+10	Limitation of Liability. 
+10.1	In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+11	Accepting Warranty or Additional Liability. 
+11.1	While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS 
+
+APPENDIX: How to apply the TII Falcon LLM License to your work.
+To apply the TII Falcon LLM License to your work, attach the following boilerplate notice, with the fields enclosed by brackets "[]" replaced with your own identifying information. (Don't include the brackets!) The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same "printed page" as the copyright notice for easier identification within third-party archives.
+
+Copyright [yyyy] [name of copyright owner] Licensed under the TII Falcon LLM License, Version 1.0 (the "License"); you may not use this file except in compliance with the License.
+You may obtain a copy of the License at FalconLLM.tii.ae. 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.

config.json

@@ -0,0 +1,29 @@
+{
+  "_name_or_path": "/workspace/models/WizardLM-Uncensored-falcon-7b",
+  "alibi": false,
+  "apply_residual_connection_post_layernorm": false,
+  "architectures": [
+    "RWForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_RW.RWConfig",
+    "AutoModelForCausalLM": "modelling_RW.RWForCausalLM"
+  },
+  "bias": false,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_dropout": 0.0,
+  "hidden_size": 4544,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "RefinedWebModel",
+  "multi_query": true,
+  "n_head": 71,
+  "n_layer": 32,
+  "parallel_attn": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.30.0.dev0",
+  "use_cache": true,
+  "vocab_size": 65025
+}

configuration_RW.py

@@ -0,0 +1,79 @@
+# 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,
+        **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
+
+        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

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.30.0.dev0"
+}

latest

@@ -0,0 +1 @@
+global_step1718

modelling_RW.py

@@ -0,0 +1,1100 @@
+# 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
+
+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, q.device)
+        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 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,
+    ):
+        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)
+
+            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)
+            assert not output_attentions  # not supported.
+            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,
+    ):
+
+        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,
+        )
+
+        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,
+        **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
+
+        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
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        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)
+
+        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]
+            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):
+                        # None for past_key_value
+                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    alibi,
+                    causal_mask,
+                    head_mask[i],
+                )
+            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,
+                )
+
+            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)
+
+        # 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 prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> dict:
+        # 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)
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+        }
+
+    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], 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
+
+        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]
+
+        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,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_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,
+        )

pytorch_model-00001-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:408c8f2810bd0d6440d72ce9b94430dd29baac07d2c19c02cb085115d1928509
+size 9962633779

pytorch_model-00002-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3be0f862e96f34058540076f4d43cccce3329b1b5e093003f99fafd2449c0b74
+size 9939388767

pytorch_model-00003-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ad524dae2a47c61c9793ea65dfef593d68ba3f07cdfc1cf41af9d7d695fa542
+size 8966840507

pytorch_model.bin.index.json

@@ -0,0 +1,203 @@
+{
+  "metadata": {
+    "total_size": 28868795392
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.0.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.0.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.0.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.0.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.0.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.0.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.1.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.10.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.10.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.10.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.10.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.10.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.10.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.11.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.11.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.11.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.11.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.11.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.11.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.12.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.13.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.14.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.15.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.16.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.17.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.18.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.19.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.2.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.2.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.2.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.2.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.2.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.2.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.20.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.20.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.20.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.20.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.20.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.20.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.21.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.22.input_layernorm.bias": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.22.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.22.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.22.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.22.self_attention.dense.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.22.self_attention.query_key_value.weight": "pytorch_model-00002-of-00003.bin",
+    "transformer.h.23.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.23.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.23.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.23.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.23.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.23.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.24.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.25.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.26.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.27.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.28.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.29.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.3.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.3.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.3.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.3.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.3.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.3.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.30.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.30.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.30.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.30.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.30.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.30.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.input_layernorm.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.mlp.dense_4h_to_h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.mlp.dense_h_to_4h.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.self_attention.dense.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.31.self_attention.query_key_value.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.h.4.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.4.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.4.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.4.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.4.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.4.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.5.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.6.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.7.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.8.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.input_layernorm.bias": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.self_attention.dense.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.h.9.self_attention.query_key_value.weight": "pytorch_model-00001-of-00003.bin",
+    "transformer.ln_f.bias": "pytorch_model-00003-of-00003.bin",
+    "transformer.ln_f.weight": "pytorch_model-00003-of-00003.bin",
+    "transformer.word_embeddings.weight": "pytorch_model-00001-of-00003.bin"
+  }
+}

special_tokens_map.json

@@ -0,0 +1,17 @@
+{
+  "additional_special_tokens": [
+    ">>TITLE<<",
+    ">>ABSTRACT<<",
+    ">>INTRODUCTION<<",
+    ">>SUMMARY<<",
+    ">>COMMENT<<",
+    ">>ANSWER<<",
+    ">>QUESTION<<",
+    ">>DOMAIN<<",
+    ">>PREFIX<<",
+    ">>SUFFIX<<",
+    ">>MIDDLE<<"
+  ],
+  "eos_token": "<|endoftext|>",
+  "pad_token": "[PAD]"
+}

tokenizer_config.json

@@ -0,0 +1,8 @@
+{
+  "add_prefix_space": false,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 2048,
+  "padding_side": "right",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2032fd61844abb20c2eb49f6eab73e177bbe0f4c4414647499751e223939d488
+size 5115

zero_to_fp32.py

@@ -0,0 +1,584 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # record shared parameters so that they can be recovered based on partners
+        # this is because such parameters holding reference only are not saved by optimizer
+        shared_params = []
+        for param in state_dict["module"]:
+            if param not in [*param_names, *buffer_names]:
+                for share_param in state_dict["module"]:
+                    if (state_dict["module"][share_param].data_ptr() == state_dict["module"][param].data_ptr()
+                            and share_param != param):
+                        shared_params.append([param, share_param])
+                        break
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)