Upload folder using huggingface_hub

checkpoint-100/adapter_config.json

@@ -0,0 +1,19 @@
+{
+  "base_model_name_or_path": "/tmp/falcon-40b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 64,
+  "target_modules": [
+    "dense",
+    "query_key_value",
+    "dense_h_to_4h",
+    "dense_4h_to_h"
+  ],
+  "task_type": "CAUSAL_LM"
+}

checkpoint-100/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66c10e793fd79eea15522bf32fe6486cfa94ec0c22dcf34b0ade12e48a5aae5d
+size 888747453

checkpoint-100/latest

@@ -0,0 +1 @@
+global_step100

checkpoint-100/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": "<|endoftext|>"
+}

checkpoint-100/tokenizer_config.json

@@ -0,0 +1,7 @@
+{
+  "add_prefix_space": false,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 2048,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

checkpoint-100/training_config.yaml

@@ -0,0 +1,143 @@
+aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
+test_file: null
+model:
+  _target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
+  pad_token_id: 11
+  use_peft: true
+  lora_config:
+    _recursive_: false
+    _target_: models.rw.LoraConfig
+    task_type: CAUSAL_LM
+    inference_mode: false
+    target_modules:
+      _target_: models.rw.find_all_linear_names
+      bits: 4
+    r: 64
+    lora_alpha: 16
+    lora_dropout: 0.05
+  gradient_checkpointing: true
+  torch_dtype:
+    _target_: general_util.training_utils.return_torch_dtype
+    dtype: bfloat16
+  quantization_config:
+    _target_: transformers.utils.quantization_config.BitsAndBytesConfig
+    load_in_4bit: true
+    bnb_4bit_compute_dtype:
+      _target_: general_util.training_utils.return_torch_dtype
+      dtype: bfloat16
+    bnb_4bit_use_double_quant: true
+    bnb_4bit_quant_type: nf4
+  device_map:
+    _target_: models.rw.return_single_device_map
+  load_in_4bit: true
+  max_memory: true
+read_tensor_train:
+  _target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
+  max_neg_num: 3
+  aug_num: 3
+  max_seq_length: 512
+  shuffle_context: true
+  min_rep_num: 5
+  geo_p: 0.4
+  deduct_ratio: 1.0
+  context_ratio: 1.0
+  noise_sent_ratio: 0.0
+  num_workers: 128
+extended_vocab: null
+collator:
+  _target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
+  max_seq_length: 512
+  tokenizer: ${model_name_or_path}
+  causal_lm: true
+  causal_lm_add_eos: false
+  generative_mode: true
+num_workers: 4
+prefetch_factor: 2
+do_preprocess: false
+model_name_or_path: /tmp/falcon-40b
+pretrain: null
+exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+exp_notes: null
+output_dir: /tmp/${exp_name}
+do_train: true
+evaluate_during_training: false
+do_eval: true
+eval_sub_path: checkpoint-*
+per_gpu_train_batch_size: 16
+per_gpu_eval_batch_size: 8
+learning_rate: 0.0005
+gradient_accumulation_steps: 16
+weight_decay: 0.0
+adam_epsilon: 1.0e-06
+adam_betas: (0.9, 0.99)
+max_grad_norm: 0.3
+num_train_epochs: 1
+max_steps: -1
+warmup_proportion: 0
+warmup_steps: 50
+optimizer: null
+use_nvlamb: null
+bit_training: null
+logging_steps: 1
+save_best: false
+save_steps: 100
+eval_steps: -1
+ddp_eval: true
+no_cuda: false
+seed: 42
+local_rank: 0
+fp16: true
+fp16_opt_level: O1
+fp16_bfloat16: true
+prediction_cfg:
+  metric: acc
+  measure: 1
+  best_checkpoint: null
+  best_result: null
+eval_forward_fn:
+  _target_: general_util.evaluator.DiscriminatorForwardFn
+post_process: null
+compile: false
+fairscale_config: null
+fsdp_config: null
+ds_cfg:
+  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
+  gradient_accumulation_steps: ${gradient_accumulation_steps}
+  optimizer:
+    type: AdamW
+    params:
+      lr: ${learning_rate}
+      betas:
+      - 0.9
+      - 0.999
+      eps: ${adam_epsilon}
+      weight_decay: ${weight_decay}
+  scheduler:
+    type: WarmupLR
+    params:
+      warmup_max_lr: ${learning_rate}
+      warmup_num_steps: 50
+      warmup_type: linear
+  gradient_clipping: ${max_grad_norm}
+  bf16:
+    enabled: ${fp16}
+  zero_optimization:
+    stage: 1
+    contiguous_gradients: true
+    overlap_comm: true
+    reduce_scatter: true
+    reduce_bucket_size: 500000000.0
+    allgather_bucket_size: 500000000.0
+  steps_per_print: 1024
+with_lightseq: false
+summary_helper:
+  _target_: general_util.tensorboard_helper.WandbWriter
+  batch_index_or_keys: null
+  outputs_index_or_keys:
+    train/mlm_loss: mlm_loss
+n_gpu: 1
+device: cuda:0
+train_batch_size: 16
+eval_batch_size: null
+world_size: 16

checkpoint-100/zero_to_fp32.py

@@ -0,0 +1,483 @@
+#!/usr/bin/env python
+'''Copyright The Microsoft 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
+
+# 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)
+
+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_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    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]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+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_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"
+    )
+
+    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 _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # 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
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"{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 fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    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)

checkpoint-200/adapter_config.json

@@ -0,0 +1,19 @@
+{
+  "base_model_name_or_path": "/tmp/falcon-40b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 64,
+  "target_modules": [
+    "dense",
+    "query_key_value",
+    "dense_h_to_4h",
+    "dense_4h_to_h"
+  ],
+  "task_type": "CAUSAL_LM"
+}

checkpoint-200/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa1c94a8e5e5bf4b41ac24e627215502d3592c241d511b1106844eda5fef35c1
+size 888747453

checkpoint-200/latest

@@ -0,0 +1 @@
+global_step200

checkpoint-200/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": "<|endoftext|>"
+}

checkpoint-200/tokenizer_config.json

@@ -0,0 +1,7 @@
+{
+  "add_prefix_space": false,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 2048,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

checkpoint-200/training_config.yaml

@@ -0,0 +1,143 @@
+aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
+test_file: null
+model:
+  _target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
+  pad_token_id: 11
+  use_peft: true
+  lora_config:
+    _recursive_: false
+    _target_: models.rw.LoraConfig
+    task_type: CAUSAL_LM
+    inference_mode: false
+    target_modules:
+      _target_: models.rw.find_all_linear_names
+      bits: 4
+    r: 64
+    lora_alpha: 16
+    lora_dropout: 0.05
+  gradient_checkpointing: true
+  torch_dtype:
+    _target_: general_util.training_utils.return_torch_dtype
+    dtype: bfloat16
+  quantization_config:
+    _target_: transformers.utils.quantization_config.BitsAndBytesConfig
+    load_in_4bit: true
+    bnb_4bit_compute_dtype:
+      _target_: general_util.training_utils.return_torch_dtype
+      dtype: bfloat16
+    bnb_4bit_use_double_quant: true
+    bnb_4bit_quant_type: nf4
+  device_map:
+    _target_: models.rw.return_single_device_map
+  load_in_4bit: true
+  max_memory: true
+read_tensor_train:
+  _target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
+  max_neg_num: 3
+  aug_num: 3
+  max_seq_length: 512
+  shuffle_context: true
+  min_rep_num: 5
+  geo_p: 0.4
+  deduct_ratio: 1.0
+  context_ratio: 1.0
+  noise_sent_ratio: 0.0
+  num_workers: 128
+extended_vocab: null
+collator:
+  _target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
+  max_seq_length: 512
+  tokenizer: ${model_name_or_path}
+  causal_lm: true
+  causal_lm_add_eos: false
+  generative_mode: true
+num_workers: 4
+prefetch_factor: 2
+do_preprocess: false
+model_name_or_path: /tmp/falcon-40b
+pretrain: null
+exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+exp_notes: null
+output_dir: /tmp/${exp_name}
+do_train: true
+evaluate_during_training: false
+do_eval: true
+eval_sub_path: checkpoint-*
+per_gpu_train_batch_size: 16
+per_gpu_eval_batch_size: 8
+learning_rate: 0.0005
+gradient_accumulation_steps: 16
+weight_decay: 0.0
+adam_epsilon: 1.0e-06
+adam_betas: (0.9, 0.99)
+max_grad_norm: 0.3
+num_train_epochs: 1
+max_steps: -1
+warmup_proportion: 0
+warmup_steps: 50
+optimizer: null
+use_nvlamb: null
+bit_training: null
+logging_steps: 1
+save_best: false
+save_steps: 100
+eval_steps: -1
+ddp_eval: true
+no_cuda: false
+seed: 42
+local_rank: 0
+fp16: true
+fp16_opt_level: O1
+fp16_bfloat16: true
+prediction_cfg:
+  metric: acc
+  measure: 1
+  best_checkpoint: null
+  best_result: null
+eval_forward_fn:
+  _target_: general_util.evaluator.DiscriminatorForwardFn
+post_process: null
+compile: false
+fairscale_config: null
+fsdp_config: null
+ds_cfg:
+  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
+  gradient_accumulation_steps: ${gradient_accumulation_steps}
+  optimizer:
+    type: AdamW
+    params:
+      lr: ${learning_rate}
+      betas:
+      - 0.9
+      - 0.999
+      eps: ${adam_epsilon}
+      weight_decay: ${weight_decay}
+  scheduler:
+    type: WarmupLR
+    params:
+      warmup_max_lr: ${learning_rate}
+      warmup_num_steps: 50
+      warmup_type: linear
+  gradient_clipping: ${max_grad_norm}
+  bf16:
+    enabled: ${fp16}
+  zero_optimization:
+    stage: 1
+    contiguous_gradients: true
+    overlap_comm: true
+    reduce_scatter: true
+    reduce_bucket_size: 500000000.0
+    allgather_bucket_size: 500000000.0
+  steps_per_print: 1024
+with_lightseq: false
+summary_helper:
+  _target_: general_util.tensorboard_helper.WandbWriter
+  batch_index_or_keys: null
+  outputs_index_or_keys:
+    train/mlm_loss: mlm_loss
+n_gpu: 1
+device: cuda:0
+train_batch_size: 16
+eval_batch_size: null
+world_size: 16

checkpoint-200/zero_to_fp32.py

@@ -0,0 +1,483 @@
+#!/usr/bin/env python
+'''Copyright The Microsoft 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
+
+# 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)
+
+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_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    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]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+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_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"
+    )
+
+    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 _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # 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
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"{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 fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    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)

checkpoint-300/adapter_config.json

@@ -0,0 +1,19 @@
+{
+  "base_model_name_or_path": "/tmp/falcon-40b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 64,
+  "target_modules": [
+    "dense",
+    "query_key_value",
+    "dense_h_to_4h",
+    "dense_4h_to_h"
+  ],
+  "task_type": "CAUSAL_LM"
+}

checkpoint-300/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:128586f354a558aadb064f6441acdc59655fde2f22a204c4ad3d1765b7b455b8
+size 888747453

checkpoint-300/latest

@@ -0,0 +1 @@
+global_step300

checkpoint-300/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": "<|endoftext|>"
+}

checkpoint-300/tokenizer_config.json

@@ -0,0 +1,7 @@
+{
+  "add_prefix_space": false,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 2048,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

checkpoint-300/training_config.yaml

@@ -0,0 +1,143 @@
+aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
+test_file: null
+model:
+  _target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
+  pad_token_id: 11
+  use_peft: true
+  lora_config:
+    _recursive_: false
+    _target_: models.rw.LoraConfig
+    task_type: CAUSAL_LM
+    inference_mode: false
+    target_modules:
+      _target_: models.rw.find_all_linear_names
+      bits: 4
+    r: 64
+    lora_alpha: 16
+    lora_dropout: 0.05
+  gradient_checkpointing: true
+  torch_dtype:
+    _target_: general_util.training_utils.return_torch_dtype
+    dtype: bfloat16
+  quantization_config:
+    _target_: transformers.utils.quantization_config.BitsAndBytesConfig
+    load_in_4bit: true
+    bnb_4bit_compute_dtype:
+      _target_: general_util.training_utils.return_torch_dtype
+      dtype: bfloat16
+    bnb_4bit_use_double_quant: true
+    bnb_4bit_quant_type: nf4
+  device_map:
+    _target_: models.rw.return_single_device_map
+  load_in_4bit: true
+  max_memory: true
+read_tensor_train:
+  _target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
+  max_neg_num: 3
+  aug_num: 3
+  max_seq_length: 512
+  shuffle_context: true
+  min_rep_num: 5
+  geo_p: 0.4
+  deduct_ratio: 1.0
+  context_ratio: 1.0
+  noise_sent_ratio: 0.0
+  num_workers: 128
+extended_vocab: null
+collator:
+  _target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
+  max_seq_length: 512
+  tokenizer: ${model_name_or_path}
+  causal_lm: true
+  causal_lm_add_eos: false
+  generative_mode: true
+num_workers: 4
+prefetch_factor: 2
+do_preprocess: false
+model_name_or_path: /tmp/falcon-40b
+pretrain: null
+exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
+exp_notes: null
+output_dir: /tmp/${exp_name}
+do_train: true
+evaluate_during_training: false
+do_eval: true
+eval_sub_path: checkpoint-*
+per_gpu_train_batch_size: 16
+per_gpu_eval_batch_size: 8
+learning_rate: 0.0005
+gradient_accumulation_steps: 16
+weight_decay: 0.0
+adam_epsilon: 1.0e-06
+adam_betas: (0.9, 0.99)
+max_grad_norm: 0.3
+num_train_epochs: 1
+max_steps: -1
+warmup_proportion: 0
+warmup_steps: 50
+optimizer: null
+use_nvlamb: null
+bit_training: null
+logging_steps: 1
+save_best: false
+save_steps: 100
+eval_steps: -1
+ddp_eval: true
+no_cuda: false
+seed: 42
+local_rank: 0
+fp16: true
+fp16_opt_level: O1
+fp16_bfloat16: true
+prediction_cfg:
+  metric: acc
+  measure: 1
+  best_checkpoint: null
+  best_result: null
+eval_forward_fn:
+  _target_: general_util.evaluator.DiscriminatorForwardFn
+post_process: null
+compile: false
+fairscale_config: null
+fsdp_config: null
+ds_cfg:
+  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
+  gradient_accumulation_steps: ${gradient_accumulation_steps}
+  optimizer:
+    type: AdamW
+    params:
+      lr: ${learning_rate}
+      betas:
+      - 0.9
+      - 0.999
+      eps: ${adam_epsilon}
+      weight_decay: ${weight_decay}
+  scheduler:
+    type: WarmupLR
+    params:
+      warmup_max_lr: ${learning_rate}
+      warmup_num_steps: 50
+      warmup_type: linear
+  gradient_clipping: ${max_grad_norm}
+  bf16:
+    enabled: ${fp16}
+  zero_optimization:
+    stage: 1
+    contiguous_gradients: true
+    overlap_comm: true
+    reduce_scatter: true
+    reduce_bucket_size: 500000000.0
+    allgather_bucket_size: 500000000.0
+  steps_per_print: 1024
+with_lightseq: false
+summary_helper:
+  _target_: general_util.tensorboard_helper.WandbWriter
+  batch_index_or_keys: null
+  outputs_index_or_keys:
+    train/mlm_loss: mlm_loss
+n_gpu: 1
+device: cuda:0
+train_batch_size: 16
+eval_batch_size: null
+world_size: 16

checkpoint-300/zero_to_fp32.py

@@ -0,0 +1,483 @@
+#!/usr/bin/env python
+'''Copyright The Microsoft 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
+
+# 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)
+
+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_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    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]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+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_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"
+    )
+
+    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 _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # 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
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"{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 fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    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)

logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:694b78b68739e7ae2f39c441f3ee284269b7e67cd2a1580373889c7c2ab6cf82
+size 12704

logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3861}

logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:337e73e47317aa6d269c6d88cd5637dbb2ae32505de6cc1fcf56bcfa81a72598
+size 12704

logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3696}

logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fcd86d2b06e468632ac7cb064f9f47febb6ff75f798b7675581b3e4f2c751e4
+size 12704

logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3715}

logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f640f89a04bcaa15d9ee7318aedd944d973dd0a6ff8bc530c6aecdbec53ab222
+size 12680

logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3926}

logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99d2f4dc6e59ca6038743efb507a2bc1451c919e1e5fbc22785d67fbc4b973fc
+size 12680

logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3881}

logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57ed8b5971038f010c9ada982042cfb43409adb836ddb8f015471686dff0ec64
+size 12680

logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.3862}

reclor_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d624748d81e273100278a09e8d72d41a3b5aea0a6e2c6c0986657661cc7818e
+size 8128

reclor_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.0}

reclor_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f85fb5eb848073e0793f0c2f25d362d609868d2fca08041b99aa5b923df7c40
+size 8128

reclor_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json

@@ -0,0 +1 @@
+{"acc": 0.0}

reclor_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84cd970cf400a19de4ffd32874c7b6068c509713eb1f3656d9767c6cae4e9f2b
+size 8128