Upload folder using huggingface_hub

checkpoints/step_5200/config.json

@@ -0,0 +1,22 @@
+{
+  "activation_hidden_dim": 1536,
+  "architectures": [
+    "PicoDecoderHF"
+  ],
+  "attention_n_heads": 12,
+  "attention_n_kv_heads": 4,
+  "auto_map": {
+    "AutoConfig": "pico_decoder.PicoDecoderHFConfig",
+    "AutoModelForCausalLM": "pico_decoder.PicoDecoderHF"
+  },
+  "batch_size": 1024,
+  "d_model": 384,
+  "max_seq_len": 512,
+  "model_type": "pico_decoder",
+  "n_layers": 12,
+  "norm_eps": 1e-06,
+  "position_emb_theta": 10000.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.3",
+  "vocab_size": 50281
+}

checkpoints/step_5200/fabric_state/checkpoint/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d7360a5e3f18a0e7c383cf716e33d8f55212c8113c5f35de8a925a86a89a35a
+size 213443205

checkpoints/step_5200/fabric_state/checkpoint/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61356e5e3d0471fe4c0a7c3d1360fb3ae934b1f145f2ebdaed0992088cf4e654
+size 213447621

checkpoints/step_5200/fabric_state/checkpoint/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f24bf89675997f72b6a0694ffb3142698c52c08a1463e63fe7811891f67e094
+size 213445701

checkpoints/step_5200/fabric_state/checkpoint/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ccfd0f779945f7d0fe8f1912d22f3bda6189fa07637022440fab07bc768ac7f
+size 213443589

checkpoints/step_5200/fabric_state/checkpoint/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6ea5f4570d262ffd9ad1eb5bdd4466e854e90fe62ab152d7f03433d1f26c04f
+size 142325529

checkpoints/step_5200/fabric_state/latest

@@ -0,0 +1 @@
+checkpoint

checkpoints/step_5200/fabric_state/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/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 1, 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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+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, weights_only=False)
+
+        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())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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 _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoints/step_5200/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf333c8a1518ae90ee1b99f66d7d2f77597186661b7796b82552c5c3fed0efea
+size 258323520

checkpoints/step_5200/pico_decoder.py

@@ -0,0 +1,623 @@
+"""
+Pico Decoder: A Lightweight Causal Transformer Language Model
+
+Pico Decoder uses a simple LLAMA-style transformer architecture, written for clarity and educational purposes.
+
+Everything is written with a modular design for easy modification and experimentation.
+
+Key features:
+- RMSNorm for layer normalization
+- Rotary Positional Embeddings (RoPE)
+- Multi-head attention with KV-cache support
+- SwiGLU activation function
+- Residual connections throughout
+
+- KV-cache for faster autoregressive generation
+
+References:
+    - RoPE: https://arxiv.org/abs/2104.09864
+    - SwiGLU: https://arxiv.org/abs/2002.05202
+    - LLAMA: https://arxiv.org/abs/2302.13971
+
+Adapted from:
+    - OLMO: https://github.com/allenai/OLMo
+    - LLAMA: https://github.com/meta/llama
+"""
+
+from dataclasses import asdict
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutput, CausalLMOutputWithPast
+
+try:
+    if TYPE_CHECKING:
+        # We need to do this to avoid importing these when creating the HF-compatible models
+        from src.config import ModelConfig
+except ImportError:
+    pass
+
+########################################################
+#
+# Layer Normalization
+#
+########################################################
+
+
+class RMSNorm(torch.nn.Module):
+    """Root Mean Square Layer Normalization.
+
+    A variant of Layer Normalization that uses RMS statistics instead of mean/variance,
+    resulting in improved stability and performance.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Configuration object containing normalization parameters
+            - config.norm_eps: Small constant for numerical stability
+            - config.d_model: Model dimension for the weight parameter
+
+    References:
+        https://arxiv.org/abs/1910.07467
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+        self.eps = config.norm_eps
+        self.weight = nn.Parameter(torch.ones(config.d_model))
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Normalizes the input tensor by its RMS value.
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Applies RMS normalization to the input tensor and scales it by the weight parameter.
+        """
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+########################################################
+#
+# Positional Embedding
+#
+########################################################
+
+
+class RoPE(nn.Module):
+    """Rotary Positional Embeddings (RoPE).
+
+    Implements position-dependent rotation of keys and queries in attention mechanism,
+    allowing better modeling of relative positions in sequences. Uses complex number
+    operations for efficient rotation.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Model configuration containing:
+            - config.position_emb_theta: Base for frequency computation
+            - config.d_model: Model dimension
+            - config.attention_n_heads: Number of attention heads
+            - config.max_seq_len: Maximum sequence length
+
+    References:
+        https://arxiv.org/abs/2104.09864
+    """
+
+    _freqs_cis_tensor: torch.Tensor | None = None
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        self.theta = config.position_emb_theta
+        self.dim = config.d_model // config.attention_n_heads
+
+        max_seq_len = config.max_seq_len
+
+        # only gets set once, and then reused for all RoPE instances
+        if RoPE._freqs_cis_tensor is None:
+            RoPE._freqs_cis_tensor = self._setup_freqs_cis(
+                max_seq_len, self.theta, self.dim
+            )
+
+        # register _freqs_cis buffer
+        # can be easily recomputed so persistent=False
+        self.register_buffer("_freqs_cis", self._freqs_cis_tensor, persistent=False)
+
+    @classmethod
+    def _setup_freqs_cis(cls, seq_len: int, theta: float, dim: int) -> torch.Tensor:
+        """Setup Frequency Tensor for RoPE Embeddings
+
+        Initializes the complex frequency tensor that is used to compute the RoPE embeddings.
+
+        Note other implementations will use cos and sin directly, but using the complex
+        number representation is (probably?) more efficient:
+
+            e^(theta * i * t) = cos(theta * t) + i * sin(theta * t) [Euler's formula]
+        """
+        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+        positions = torch.arange(seq_len)
+        freqs = torch.outer(positions, _freqs)
+        return torch.polar(torch.ones_like(freqs), freqs)  # complex64
+
+    def get_freqs_cis(
+        self, input_shape: torch.Size, start_pos: int, end_pos: int
+    ) -> torch.Tensor:
+        """Reshape Frequency Tensor for RoPE Embeddings
+
+        Makes the frequency tensor broadcastable with the input tensor.
+        """
+        _freqs_cis = self._freqs_cis[start_pos:end_pos]
+        ndim = len(input_shape)
+        assert 0 <= 1 < ndim
+        assert _freqs_cis.shape == (input_shape[1], input_shape[-1])
+
+        # TODO: Check whether this is correct (might be able to remove this)
+        shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)]
+        return _freqs_cis.view(*shape)
+
+    def forward(
+        self,
+        queries: torch.Tensor,
+        keys: torch.Tensor,
+        start_pos: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply RoPE Embeddings to Queries and Keys
+
+        Applies the rotary positional embeddings to the input tensors via complex num multiplication
+
+        NOTE: The start_pos is used if we want to use the kv_cache in the attention mechanism.
+        """
+        queries_ = torch.view_as_complex(
+            queries.float().reshape(*queries.shape[:-1], -1, 2)
+        )
+        keys_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
+
+        input_shape = (
+            queries_.shape
+        )  # same as keys: (batch_size, seq_len, n_heads, head_dim/2)
+        freqs_start_pos = start_pos
+        freqs_end_pos = freqs_start_pos + queries_.shape[1]
+
+        freqs_cis = self.get_freqs_cis(input_shape, freqs_start_pos, freqs_end_pos)
+
+        queries_rotated = torch.view_as_real(queries_ * freqs_cis).flatten(3)
+        keys_rotated = torch.view_as_real(keys_ * freqs_cis).flatten(3)
+        return queries_rotated.type_as(queries), keys_rotated.type_as(keys)
+
+
+########################################################
+#
+# Attention
+#
+########################################################
+
+
+class Attention(nn.Module):
+    """Multi-head Attention with Group Query Attention support.
+
+    Implements scaled dot-product attention and supports:
+    - Grouped Query Attention (GQA)
+    - Key-Value caching for efficient inference
+    - RoPE integration
+
+    Args:
+        config (Union[ModelConfig, PretrainedConfig]): Configuration containing:
+            - config.attention_n_heads: Number of attention heads
+            - config.attention_n_kv_heads: Number of key/value heads
+            - config.d_model: Model dimension
+            - config.batch_size: Maximum batch size
+            - config.max_seq_len: Maximum sequence length
+
+    Shape:
+        - Input: (batch_size, seq_len, d_model)
+        - Output: (batch_size, seq_len, d_model)
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.n_heads = config.attention_n_heads
+        self.n_kv_heads = config.attention_n_kv_heads
+
+        self.batch_size = config.batch_size
+        self.max_seq_len = config.max_seq_len
+
+        d_model = config.d_model
+        self.head_dim = d_model // self.n_heads
+
+        self.n_rep = self.n_heads // self.n_kv_heads
+
+        self.q_proj = nn.Linear(d_model, self.n_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.n_heads * self.head_dim, d_model, bias=False)
+
+        self.rope = RoPE(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        """Forward pass for the attention mechanism.
+
+        Computes queries, keys, and values for the attention mechanism. Applies rotary positional
+        embeddings to the queries and keys, and then computes attention scores and outputs.
+
+        For an introduction to the attention mechanism, see:
+        https://arxiv.org/abs/1706.03762
+
+        A few things to note:
+        - The past_key_values is used to implement the KV cache, which is used to speed up
+          generation by caching the KV pairs from previous forward passes. This is useful when doing
+          tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+          modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+          its own KV cache - this KV cache is implemented as a tuple.
+        """
+        bsz, seq_len, _ = input.shape
+        _queries, _keys, _values = (
+            self.q_proj(input),
+            self.k_proj(input),
+            self.v_proj(input),
+        )
+
+        # Reshaping for multi-head attention
+        queries = _queries.view(bsz, seq_len, self.n_heads, self.head_dim)
+        keys = _keys.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        values = _values.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+
+        # The start position is used to apply the RoPE embeddings to only the new tokens
+        # when using the kv_cache in the attention mechanism.
+        # We want to start from the last position in the cache.
+        start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
+
+        # apply rotary positional embeddings
+        queries, keys = self.rope(queries, keys, start_pos)
+
+        if past_key_values is not None:
+            keys = torch.cat([past_key_values[0], keys], dim=1)
+            values = torch.cat([past_key_values[1], values], dim=1)
+
+        if use_cache:
+            cached_keys = keys
+            cached_values = values
+        else:
+            cached_keys = None
+            cached_values = None
+
+        queries = queries.transpose(1, 2)
+        keys = keys.transpose(1, 2)
+        values = values.transpose(1, 2)
+
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and queries.device.type == "mps":
+            # NOTE: MPS does not support GQA in the SDPA kernel, but we can repeat the keys and values
+            # outside of the kernel to get the same effect.
+            # See: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
+            keys = keys.repeat_interleave(self.n_rep, dim=-3)
+            values = values.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+
+        backends = [SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]
+
+        with sdpa_kernel(backends=backends):
+            attn_output = F.scaled_dot_product_attention(
+                queries.contiguous(),
+                keys.contiguous(),
+                values.contiguous(),
+                attn_mask=mask.to(queries.dtype),
+                enable_gqa=apply_gqa,
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        output = self.o_proj(attn_output)
+
+        return output, (cached_keys, cached_values)
+
+
+########################################################
+#
+# SwiGLU (Combines MLP and Activation)
+#
+########################################################
+
+
+class SwiGLU(nn.Module):
+    """SwiGLU Activation Function with Linear Projections.
+
+    Implements the SwiGLU activation function combined with linear transformations,
+    serving as the feed-forward network in transformer blocks.
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Configuration containing:
+            - config.d_model: Model dimension
+            - config.activation_hidden_dim: Hidden dimension (typically 4 * d_model)
+
+    References:
+        https://arxiv.org/abs/2002.05202
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        model_dim = config.d_model
+        act_hidden_dim = config.activation_hidden_dim  # usually 4 * d_model
+
+        self.w_0 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_1 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_2 = nn.Linear(act_hidden_dim, model_dim, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
+
+
+########################################################
+#
+# PicoDecoderBlock
+#
+########################################################
+
+
+class PicoDecoderBlock(nn.Module):
+    """Single Transformer Block with Attention and Feed-forward layers.
+
+    Implements a standard transformer block with:
+    - Multi-head attention with normalization and residual connection
+    - SwiGLU feed-forward network with normalization and residual connection
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Model configuration; either a dataclass or
+            a HuggingFace PicoDecoderHFConfig
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.attention = Attention(config)
+        self.swiglu = SwiGLU(config)
+        self.attention_norm = RMSNorm(config)
+        self.swiglu_norm = RMSNorm(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        attention_output, cached_key_values = self.attention(
+            self.attention_norm(input),
+            mask=mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        # NOTE: cached_key_values is None if use_cache is False
+
+        h = input + attention_output
+        out = h + self.swiglu(self.swiglu_norm(h))
+        return out, cached_key_values
+
+
+########################################################
+#
+# Pico Decoder (Causal Transformer Model)
+#
+########################################################
+
+
+class PicoDecoder(nn.Module):
+    """
+    Pico Decoder: combines the embedding, causal decoder blocks, and output projection into a
+    single autoregressive model.
+
+    For more information on the model, see the classes for the modules that make up the model.
+    """
+
+    def __init__(
+        self,
+        model_config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+        self.config = model_config
+
+        self.embedding_proj = nn.Embedding(self.config.vocab_size, self.config.d_model)
+        self.layers = nn.ModuleList(
+            [PicoDecoderBlock(self.config) for _ in range(self.config.n_layers)]
+        )
+        self.output_norm = RMSNorm(self.config)
+        self.de_embedding_proj = nn.Linear(
+            self.config.d_model, self.config.vocab_size, bias=False
+        )
+
+    def convert_to_hf_model(self) -> "PicoDecoderHF":
+        """Convert the Lightning model to a HuggingFace model."""
+        # Build HF config
+        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
+
+        # Instantiate the HF-wrapped model
+        hf_model = PicoDecoderHF(hf_config)
+
+        # Grab our full state dict, prefixing module names
+        raw_state = self.state_dict(prefix="pico_decoder.")
+
+        # Only keep keys that exist in the HF model (drops classifier_head, etc.)
+        hf_keys = set(hf_model.state_dict().keys())
+        filtered_state = {k: v for k, v in raw_state.items() if k in hf_keys}
+
+        # Load into HF model, ignore any missing keys
+        hf_model.load_state_dict(filtered_state, strict=False)
+
+        return hf_model
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        return_hidden: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]]]:
+        """
+        This is the forward pass for the entire Pico model. It boils down to:
+        - Embedding the input ids
+        - Creating a causal mask
+        - Processing through the pico layers
+        - Projecting the output to logits
+
+        NOTE: One feature that might be confusing is the KV cache. The KV cache is used to speed up
+        generation by caching the KV pairs from previous forward passes. This is useful when doing
+        tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+        modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+        its own KV cache which is stored as a tuple. The whole model then stores a tuple of these
+        KV caches (so a tuple of tuples).
+        """
+
+        seq_len = input_ids.shape[-1]
+        h = self.embedding_proj(input_ids)
+
+        # Calculate start position from past cached KV pairs. Remember that each layer has its
+        # own KV Cache. So when we index past_key_values, we need to index into the KV pairs for the
+        # correct layer and then for either the keys or values.
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
+        # Create causal mask for current sequence
+        mask = None
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
+
+            # If using KV cache, extend mask to cover cached sequence length
+            if past_key_values is not None:
+                # Add zeros for cached tokens (we can attend to all of them)
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+
+            mask = mask.to(h.device)
+
+        # NOTE: If we are using the cache, we need to store the cached KV pairs for each layer
+        #       in a tuple. Each layer will have its own cached KV pair which we aggregate in a tuple.
+        cached_key_values = () if use_cache else None
+
+        # Process through transformer blocks
+        for idx, layer in enumerate(self.layers):
+            layer_past_key_values = (
+                past_key_values[idx] if past_key_values is not None else None
+            )
+
+            h, layer_cached_key_values = layer(
+                h, mask=mask, past_key_values=layer_past_key_values, use_cache=use_cache
+            )
+
+            if use_cache:
+                cached_key_values += (layer_cached_key_values,)
+
+        # Final norm and projection
+        h = self.output_norm(h)
+
+        if return_hidden:
+            return h, cached_key_values
+
+        logits = self.de_embedding_proj(h).float()
+
+        return logits, cached_key_values
+
+
+########################################################
+#
+# HuggingFace Wrapper for the Pico Decoder model.
+#
+########################################################
+
+
+class PicoDecoderHFConfig(PretrainedConfig):
+    """Config class for the Pico Decoder HuggingFace wrapper."""
+
+    model_type = "pico_decoder"
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
+        # NOTE The typical from_dict method doesn't actually set the attributes unless they are
+        # defined in the constructor.
+
+        pico_config = cls(**kwargs)
+
+        # Because this class is just a wrapper around the ModelConfig dataclass, we need to do
+        # a little extra work to ensure that the attributes are actually set.
+        for key, value in config_dict.items():
+            setattr(pico_config, key, value)
+
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+        unused_kwargs = {
+            key: value for key, value in kwargs.items() if not hasattr(pico_config, key)
+        }
+
+        if return_unused_kwargs:
+            return pico_config, unused_kwargs
+        return pico_config
+
+    @classmethod
+    def from_dataclass(cls, model_config: "ModelConfig"):
+        """Initialise from our custom config dataclass."""
+        return cls.from_dict(asdict(model_config))
+
+
+class PicoDecoderHF(PreTrainedModel):
+    """
+    HuggingFace wrapper for the Pico model.
+
+    Many evaluation frameworks require a model be setup as a HuggingFace model, so we provide a simple
+    wrapper that does just that. When we save checkpoints of the Pico model, we save both the normal
+    Pico model as well as the model wrapped in this HuggingFace class.
+
+    This also lets you do cool things like:
+
+    `model = AutoModelForCausalLM.from_pretrained("path/to/checkpoint")`
+    """
+
+    config_class = PicoDecoderHFConfig
+    _no_split_modules = ["PicoBlock", "Attention", "SwiGLU", "RMSNorm"]
+
+    def __init__(self, config: PicoDecoderHFConfig):
+        super().__init__(config)
+        self.pico_decoder = PicoDecoder(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Union[CausalLMOutput, CausalLMOutputWithPast]:
+        """HuggingFace forward pass wrapper.
+
+        Forwards pass for the HuggingFace version of the Pico Model. Basic wrapper around the
+        Pico model's forward pass, and returns the output as a HuggingFace CausalLMOutput.
+        """
+        logits, past_key_values = self.pico_decoder(
+            input_ids, past_key_values, use_cache
+        )
+        if use_cache:
+            return CausalLMOutputWithPast(
+                logits=logits,
+                past_key_values=past_key_values,
+            )
+        else:
+            return CausalLMOutput(
+                logits=logits,
+            )
+
+
+# Register for auto classes
+PicoDecoderHFConfig.register_for_auto_class()
+PicoDecoderHF.register_for_auto_class("AutoModel")
+PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")

checkpoints/step_5200/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|padding|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoints/step_5200/tokenizer_config.json

@@ -0,0 +1,248 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": null,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<|padding|>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

checkpoints/step_5300/config.json

@@ -0,0 +1,22 @@
+{
+  "activation_hidden_dim": 1536,
+  "architectures": [
+    "PicoDecoderHF"
+  ],
+  "attention_n_heads": 12,
+  "attention_n_kv_heads": 4,
+  "auto_map": {
+    "AutoConfig": "pico_decoder.PicoDecoderHFConfig",
+    "AutoModelForCausalLM": "pico_decoder.PicoDecoderHF"
+  },
+  "batch_size": 1024,
+  "d_model": 384,
+  "max_seq_len": 512,
+  "model_type": "pico_decoder",
+  "n_layers": 12,
+  "norm_eps": 1e-06,
+  "position_emb_theta": 10000.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.3",
+  "vocab_size": 50281
+}

checkpoints/step_5300/fabric_state/checkpoint/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4f186a3d24e46bc7eb2320e1abc93fe4c2c61ae04225fe7e7a232fccbe4b90a
+size 213443205

checkpoints/step_5300/fabric_state/checkpoint/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:519ed47009fa2de676b8c554fba58d109c801e21bc741272d3b7adba0784a5c8
+size 213447621

checkpoints/step_5300/fabric_state/checkpoint/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d45e4d5a6a067656dcea1048de78666348140c64a519b4b54226c06703700bf
+size 213445701

checkpoints/step_5300/fabric_state/checkpoint/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5ca182c75d3603ccc982676373a8594197d90c5eb271f727d124b1878904f9f
+size 213443589

checkpoints/step_5300/fabric_state/checkpoint/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d598c440f11a42ce85595a6c1fd78d3496041202bdbc30339389da15f5bc6f3
+size 142325529

checkpoints/step_5300/fabric_state/latest

@@ -0,0 +1 @@
+checkpoint

checkpoints/step_5300/fabric_state/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/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 1, 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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+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, weights_only=False)
+
+        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())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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 _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoints/step_5300/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6af91360b36720cb427673d3d611531378c8698713d6565b8bde714726b5e08f
+size 258323520

checkpoints/step_5300/pico_decoder.py

@@ -0,0 +1,623 @@
+"""
+Pico Decoder: A Lightweight Causal Transformer Language Model
+
+Pico Decoder uses a simple LLAMA-style transformer architecture, written for clarity and educational purposes.
+
+Everything is written with a modular design for easy modification and experimentation.
+
+Key features:
+- RMSNorm for layer normalization
+- Rotary Positional Embeddings (RoPE)
+- Multi-head attention with KV-cache support
+- SwiGLU activation function
+- Residual connections throughout
+
+- KV-cache for faster autoregressive generation
+
+References:
+    - RoPE: https://arxiv.org/abs/2104.09864
+    - SwiGLU: https://arxiv.org/abs/2002.05202
+    - LLAMA: https://arxiv.org/abs/2302.13971
+
+Adapted from:
+    - OLMO: https://github.com/allenai/OLMo
+    - LLAMA: https://github.com/meta/llama
+"""
+
+from dataclasses import asdict
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutput, CausalLMOutputWithPast
+
+try:
+    if TYPE_CHECKING:
+        # We need to do this to avoid importing these when creating the HF-compatible models
+        from src.config import ModelConfig
+except ImportError:
+    pass
+
+########################################################
+#
+# Layer Normalization
+#
+########################################################
+
+
+class RMSNorm(torch.nn.Module):
+    """Root Mean Square Layer Normalization.
+
+    A variant of Layer Normalization that uses RMS statistics instead of mean/variance,
+    resulting in improved stability and performance.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Configuration object containing normalization parameters
+            - config.norm_eps: Small constant for numerical stability
+            - config.d_model: Model dimension for the weight parameter
+
+    References:
+        https://arxiv.org/abs/1910.07467
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+        self.eps = config.norm_eps
+        self.weight = nn.Parameter(torch.ones(config.d_model))
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Normalizes the input tensor by its RMS value.
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Applies RMS normalization to the input tensor and scales it by the weight parameter.
+        """
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+########################################################
+#
+# Positional Embedding
+#
+########################################################
+
+
+class RoPE(nn.Module):
+    """Rotary Positional Embeddings (RoPE).
+
+    Implements position-dependent rotation of keys and queries in attention mechanism,
+    allowing better modeling of relative positions in sequences. Uses complex number
+    operations for efficient rotation.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Model configuration containing:
+            - config.position_emb_theta: Base for frequency computation
+            - config.d_model: Model dimension
+            - config.attention_n_heads: Number of attention heads
+            - config.max_seq_len: Maximum sequence length
+
+    References:
+        https://arxiv.org/abs/2104.09864
+    """
+
+    _freqs_cis_tensor: torch.Tensor | None = None
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        self.theta = config.position_emb_theta
+        self.dim = config.d_model // config.attention_n_heads
+
+        max_seq_len = config.max_seq_len
+
+        # only gets set once, and then reused for all RoPE instances
+        if RoPE._freqs_cis_tensor is None:
+            RoPE._freqs_cis_tensor = self._setup_freqs_cis(
+                max_seq_len, self.theta, self.dim
+            )
+
+        # register _freqs_cis buffer
+        # can be easily recomputed so persistent=False
+        self.register_buffer("_freqs_cis", self._freqs_cis_tensor, persistent=False)
+
+    @classmethod
+    def _setup_freqs_cis(cls, seq_len: int, theta: float, dim: int) -> torch.Tensor:
+        """Setup Frequency Tensor for RoPE Embeddings
+
+        Initializes the complex frequency tensor that is used to compute the RoPE embeddings.
+
+        Note other implementations will use cos and sin directly, but using the complex
+        number representation is (probably?) more efficient:
+
+            e^(theta * i * t) = cos(theta * t) + i * sin(theta * t) [Euler's formula]
+        """
+        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+        positions = torch.arange(seq_len)
+        freqs = torch.outer(positions, _freqs)
+        return torch.polar(torch.ones_like(freqs), freqs)  # complex64
+
+    def get_freqs_cis(
+        self, input_shape: torch.Size, start_pos: int, end_pos: int
+    ) -> torch.Tensor:
+        """Reshape Frequency Tensor for RoPE Embeddings
+
+        Makes the frequency tensor broadcastable with the input tensor.
+        """
+        _freqs_cis = self._freqs_cis[start_pos:end_pos]
+        ndim = len(input_shape)
+        assert 0 <= 1 < ndim
+        assert _freqs_cis.shape == (input_shape[1], input_shape[-1])
+
+        # TODO: Check whether this is correct (might be able to remove this)
+        shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)]
+        return _freqs_cis.view(*shape)
+
+    def forward(
+        self,
+        queries: torch.Tensor,
+        keys: torch.Tensor,
+        start_pos: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply RoPE Embeddings to Queries and Keys
+
+        Applies the rotary positional embeddings to the input tensors via complex num multiplication
+
+        NOTE: The start_pos is used if we want to use the kv_cache in the attention mechanism.
+        """
+        queries_ = torch.view_as_complex(
+            queries.float().reshape(*queries.shape[:-1], -1, 2)
+        )
+        keys_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
+
+        input_shape = (
+            queries_.shape
+        )  # same as keys: (batch_size, seq_len, n_heads, head_dim/2)
+        freqs_start_pos = start_pos
+        freqs_end_pos = freqs_start_pos + queries_.shape[1]
+
+        freqs_cis = self.get_freqs_cis(input_shape, freqs_start_pos, freqs_end_pos)
+
+        queries_rotated = torch.view_as_real(queries_ * freqs_cis).flatten(3)
+        keys_rotated = torch.view_as_real(keys_ * freqs_cis).flatten(3)
+        return queries_rotated.type_as(queries), keys_rotated.type_as(keys)
+
+
+########################################################
+#
+# Attention
+#
+########################################################
+
+
+class Attention(nn.Module):
+    """Multi-head Attention with Group Query Attention support.
+
+    Implements scaled dot-product attention and supports:
+    - Grouped Query Attention (GQA)
+    - Key-Value caching for efficient inference
+    - RoPE integration
+
+    Args:
+        config (Union[ModelConfig, PretrainedConfig]): Configuration containing:
+            - config.attention_n_heads: Number of attention heads
+            - config.attention_n_kv_heads: Number of key/value heads
+            - config.d_model: Model dimension
+            - config.batch_size: Maximum batch size
+            - config.max_seq_len: Maximum sequence length
+
+    Shape:
+        - Input: (batch_size, seq_len, d_model)
+        - Output: (batch_size, seq_len, d_model)
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.n_heads = config.attention_n_heads
+        self.n_kv_heads = config.attention_n_kv_heads
+
+        self.batch_size = config.batch_size
+        self.max_seq_len = config.max_seq_len
+
+        d_model = config.d_model
+        self.head_dim = d_model // self.n_heads
+
+        self.n_rep = self.n_heads // self.n_kv_heads
+
+        self.q_proj = nn.Linear(d_model, self.n_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.n_heads * self.head_dim, d_model, bias=False)
+
+        self.rope = RoPE(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        """Forward pass for the attention mechanism.
+
+        Computes queries, keys, and values for the attention mechanism. Applies rotary positional
+        embeddings to the queries and keys, and then computes attention scores and outputs.
+
+        For an introduction to the attention mechanism, see:
+        https://arxiv.org/abs/1706.03762
+
+        A few things to note:
+        - The past_key_values is used to implement the KV cache, which is used to speed up
+          generation by caching the KV pairs from previous forward passes. This is useful when doing
+          tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+          modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+          its own KV cache - this KV cache is implemented as a tuple.
+        """
+        bsz, seq_len, _ = input.shape
+        _queries, _keys, _values = (
+            self.q_proj(input),
+            self.k_proj(input),
+            self.v_proj(input),
+        )
+
+        # Reshaping for multi-head attention
+        queries = _queries.view(bsz, seq_len, self.n_heads, self.head_dim)
+        keys = _keys.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        values = _values.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+
+        # The start position is used to apply the RoPE embeddings to only the new tokens
+        # when using the kv_cache in the attention mechanism.
+        # We want to start from the last position in the cache.
+        start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
+
+        # apply rotary positional embeddings
+        queries, keys = self.rope(queries, keys, start_pos)
+
+        if past_key_values is not None:
+            keys = torch.cat([past_key_values[0], keys], dim=1)
+            values = torch.cat([past_key_values[1], values], dim=1)
+
+        if use_cache:
+            cached_keys = keys
+            cached_values = values
+        else:
+            cached_keys = None
+            cached_values = None
+
+        queries = queries.transpose(1, 2)
+        keys = keys.transpose(1, 2)
+        values = values.transpose(1, 2)
+
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and queries.device.type == "mps":
+            # NOTE: MPS does not support GQA in the SDPA kernel, but we can repeat the keys and values
+            # outside of the kernel to get the same effect.
+            # See: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
+            keys = keys.repeat_interleave(self.n_rep, dim=-3)
+            values = values.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+
+        backends = [SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]
+
+        with sdpa_kernel(backends=backends):
+            attn_output = F.scaled_dot_product_attention(
+                queries.contiguous(),
+                keys.contiguous(),
+                values.contiguous(),
+                attn_mask=mask.to(queries.dtype),
+                enable_gqa=apply_gqa,
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        output = self.o_proj(attn_output)
+
+        return output, (cached_keys, cached_values)
+
+
+########################################################
+#
+# SwiGLU (Combines MLP and Activation)
+#
+########################################################
+
+
+class SwiGLU(nn.Module):
+    """SwiGLU Activation Function with Linear Projections.
+
+    Implements the SwiGLU activation function combined with linear transformations,
+    serving as the feed-forward network in transformer blocks.
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Configuration containing:
+            - config.d_model: Model dimension
+            - config.activation_hidden_dim: Hidden dimension (typically 4 * d_model)
+
+    References:
+        https://arxiv.org/abs/2002.05202
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        model_dim = config.d_model
+        act_hidden_dim = config.activation_hidden_dim  # usually 4 * d_model
+
+        self.w_0 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_1 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_2 = nn.Linear(act_hidden_dim, model_dim, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
+
+
+########################################################
+#
+# PicoDecoderBlock
+#
+########################################################
+
+
+class PicoDecoderBlock(nn.Module):
+    """Single Transformer Block with Attention and Feed-forward layers.
+
+    Implements a standard transformer block with:
+    - Multi-head attention with normalization and residual connection
+    - SwiGLU feed-forward network with normalization and residual connection
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Model configuration; either a dataclass or
+            a HuggingFace PicoDecoderHFConfig
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.attention = Attention(config)
+        self.swiglu = SwiGLU(config)
+        self.attention_norm = RMSNorm(config)
+        self.swiglu_norm = RMSNorm(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        attention_output, cached_key_values = self.attention(
+            self.attention_norm(input),
+            mask=mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        # NOTE: cached_key_values is None if use_cache is False
+
+        h = input + attention_output
+        out = h + self.swiglu(self.swiglu_norm(h))
+        return out, cached_key_values
+
+
+########################################################
+#
+# Pico Decoder (Causal Transformer Model)
+#
+########################################################
+
+
+class PicoDecoder(nn.Module):
+    """
+    Pico Decoder: combines the embedding, causal decoder blocks, and output projection into a
+    single autoregressive model.
+
+    For more information on the model, see the classes for the modules that make up the model.
+    """
+
+    def __init__(
+        self,
+        model_config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+        self.config = model_config
+
+        self.embedding_proj = nn.Embedding(self.config.vocab_size, self.config.d_model)
+        self.layers = nn.ModuleList(
+            [PicoDecoderBlock(self.config) for _ in range(self.config.n_layers)]
+        )
+        self.output_norm = RMSNorm(self.config)
+        self.de_embedding_proj = nn.Linear(
+            self.config.d_model, self.config.vocab_size, bias=False
+        )
+
+    def convert_to_hf_model(self) -> "PicoDecoderHF":
+        """Convert the Lightning model to a HuggingFace model."""
+        # Build HF config
+        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
+
+        # Instantiate the HF-wrapped model
+        hf_model = PicoDecoderHF(hf_config)
+
+        # Grab our full state dict, prefixing module names
+        raw_state = self.state_dict(prefix="pico_decoder.")
+
+        # Only keep keys that exist in the HF model (drops classifier_head, etc.)
+        hf_keys = set(hf_model.state_dict().keys())
+        filtered_state = {k: v for k, v in raw_state.items() if k in hf_keys}
+
+        # Load into HF model, ignore any missing keys
+        hf_model.load_state_dict(filtered_state, strict=False)
+
+        return hf_model
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        return_hidden: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]]]:
+        """
+        This is the forward pass for the entire Pico model. It boils down to:
+        - Embedding the input ids
+        - Creating a causal mask
+        - Processing through the pico layers
+        - Projecting the output to logits
+
+        NOTE: One feature that might be confusing is the KV cache. The KV cache is used to speed up
+        generation by caching the KV pairs from previous forward passes. This is useful when doing
+        tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+        modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+        its own KV cache which is stored as a tuple. The whole model then stores a tuple of these
+        KV caches (so a tuple of tuples).
+        """
+
+        seq_len = input_ids.shape[-1]
+        h = self.embedding_proj(input_ids)
+
+        # Calculate start position from past cached KV pairs. Remember that each layer has its
+        # own KV Cache. So when we index past_key_values, we need to index into the KV pairs for the
+        # correct layer and then for either the keys or values.
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
+        # Create causal mask for current sequence
+        mask = None
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
+
+            # If using KV cache, extend mask to cover cached sequence length
+            if past_key_values is not None:
+                # Add zeros for cached tokens (we can attend to all of them)
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+
+            mask = mask.to(h.device)
+
+        # NOTE: If we are using the cache, we need to store the cached KV pairs for each layer
+        #       in a tuple. Each layer will have its own cached KV pair which we aggregate in a tuple.
+        cached_key_values = () if use_cache else None
+
+        # Process through transformer blocks
+        for idx, layer in enumerate(self.layers):
+            layer_past_key_values = (
+                past_key_values[idx] if past_key_values is not None else None
+            )
+
+            h, layer_cached_key_values = layer(
+                h, mask=mask, past_key_values=layer_past_key_values, use_cache=use_cache
+            )
+
+            if use_cache:
+                cached_key_values += (layer_cached_key_values,)
+
+        # Final norm and projection
+        h = self.output_norm(h)
+
+        if return_hidden:
+            return h, cached_key_values
+
+        logits = self.de_embedding_proj(h).float()
+
+        return logits, cached_key_values
+
+
+########################################################
+#
+# HuggingFace Wrapper for the Pico Decoder model.
+#
+########################################################
+
+
+class PicoDecoderHFConfig(PretrainedConfig):
+    """Config class for the Pico Decoder HuggingFace wrapper."""
+
+    model_type = "pico_decoder"
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
+        # NOTE The typical from_dict method doesn't actually set the attributes unless they are
+        # defined in the constructor.
+
+        pico_config = cls(**kwargs)
+
+        # Because this class is just a wrapper around the ModelConfig dataclass, we need to do
+        # a little extra work to ensure that the attributes are actually set.
+        for key, value in config_dict.items():
+            setattr(pico_config, key, value)
+
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+        unused_kwargs = {
+            key: value for key, value in kwargs.items() if not hasattr(pico_config, key)
+        }
+
+        if return_unused_kwargs:
+            return pico_config, unused_kwargs
+        return pico_config
+
+    @classmethod
+    def from_dataclass(cls, model_config: "ModelConfig"):
+        """Initialise from our custom config dataclass."""
+        return cls.from_dict(asdict(model_config))
+
+
+class PicoDecoderHF(PreTrainedModel):
+    """
+    HuggingFace wrapper for the Pico model.
+
+    Many evaluation frameworks require a model be setup as a HuggingFace model, so we provide a simple
+    wrapper that does just that. When we save checkpoints of the Pico model, we save both the normal
+    Pico model as well as the model wrapped in this HuggingFace class.
+
+    This also lets you do cool things like:
+
+    `model = AutoModelForCausalLM.from_pretrained("path/to/checkpoint")`
+    """
+
+    config_class = PicoDecoderHFConfig
+    _no_split_modules = ["PicoBlock", "Attention", "SwiGLU", "RMSNorm"]
+
+    def __init__(self, config: PicoDecoderHFConfig):
+        super().__init__(config)
+        self.pico_decoder = PicoDecoder(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Union[CausalLMOutput, CausalLMOutputWithPast]:
+        """HuggingFace forward pass wrapper.
+
+        Forwards pass for the HuggingFace version of the Pico Model. Basic wrapper around the
+        Pico model's forward pass, and returns the output as a HuggingFace CausalLMOutput.
+        """
+        logits, past_key_values = self.pico_decoder(
+            input_ids, past_key_values, use_cache
+        )
+        if use_cache:
+            return CausalLMOutputWithPast(
+                logits=logits,
+                past_key_values=past_key_values,
+            )
+        else:
+            return CausalLMOutput(
+                logits=logits,
+            )
+
+
+# Register for auto classes
+PicoDecoderHFConfig.register_for_auto_class()
+PicoDecoderHF.register_for_auto_class("AutoModel")
+PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")

checkpoints/step_5300/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|padding|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoints/step_5300/tokenizer_config.json

@@ -0,0 +1,248 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": null,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<|padding|>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

checkpoints/step_5400/config.json

@@ -0,0 +1,22 @@
+{
+  "activation_hidden_dim": 1536,
+  "architectures": [
+    "PicoDecoderHF"
+  ],
+  "attention_n_heads": 12,
+  "attention_n_kv_heads": 4,
+  "auto_map": {
+    "AutoConfig": "pico_decoder.PicoDecoderHFConfig",
+    "AutoModelForCausalLM": "pico_decoder.PicoDecoderHF"
+  },
+  "batch_size": 1024,
+  "d_model": 384,
+  "max_seq_len": 512,
+  "model_type": "pico_decoder",
+  "n_layers": 12,
+  "norm_eps": 1e-06,
+  "position_emb_theta": 10000.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.3",
+  "vocab_size": 50281
+}

checkpoints/step_5400/fabric_state/checkpoint/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bb2974d5d7a309dfb70d05621411367069d349ecfa2b54da2a361d6c51c5e4c
+size 213443205

checkpoints/step_5400/fabric_state/checkpoint/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c24fcb923e02569ec6d17cc371b12a18c79e63d4b64e37d9abb70c543de1bb6
+size 213447621

checkpoints/step_5400/fabric_state/checkpoint/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1164e4daa4174c3a50e41da74a9e8ee6ff1d861c29946a5fd28fa9880cee66e0
+size 213445701

checkpoints/step_5400/fabric_state/checkpoint/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:799887faf69fa4db010a991ec75fbe3b0eded5dea3d7f7b910173c9f612fffed
+size 213443589

checkpoints/step_5400/fabric_state/checkpoint/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f643728d9146644cb97689aff1a6cd27be9b20f64fbf4a437ebd49148fed335
+size 142325529

checkpoints/step_5400/fabric_state/latest

@@ -0,0 +1 @@
+checkpoint

checkpoints/step_5400/fabric_state/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/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 1, 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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+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, weights_only=False)
+
+        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())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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 _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoints/step_5400/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d3d8534f6b4788c4ce6ca6ccc61354a31368ce12866d7ef6ca23b3591e1acee
+size 258323520

checkpoints/step_5400/pico_decoder.py

@@ -0,0 +1,623 @@
+"""
+Pico Decoder: A Lightweight Causal Transformer Language Model
+
+Pico Decoder uses a simple LLAMA-style transformer architecture, written for clarity and educational purposes.
+
+Everything is written with a modular design for easy modification and experimentation.
+
+Key features:
+- RMSNorm for layer normalization
+- Rotary Positional Embeddings (RoPE)
+- Multi-head attention with KV-cache support
+- SwiGLU activation function
+- Residual connections throughout
+
+- KV-cache for faster autoregressive generation
+
+References:
+    - RoPE: https://arxiv.org/abs/2104.09864
+    - SwiGLU: https://arxiv.org/abs/2002.05202
+    - LLAMA: https://arxiv.org/abs/2302.13971
+
+Adapted from:
+    - OLMO: https://github.com/allenai/OLMo
+    - LLAMA: https://github.com/meta/llama
+"""
+
+from dataclasses import asdict
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutput, CausalLMOutputWithPast
+
+try:
+    if TYPE_CHECKING:
+        # We need to do this to avoid importing these when creating the HF-compatible models
+        from src.config import ModelConfig
+except ImportError:
+    pass
+
+########################################################
+#
+# Layer Normalization
+#
+########################################################
+
+
+class RMSNorm(torch.nn.Module):
+    """Root Mean Square Layer Normalization.
+
+    A variant of Layer Normalization that uses RMS statistics instead of mean/variance,
+    resulting in improved stability and performance.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Configuration object containing normalization parameters
+            - config.norm_eps: Small constant for numerical stability
+            - config.d_model: Model dimension for the weight parameter
+
+    References:
+        https://arxiv.org/abs/1910.07467
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+        self.eps = config.norm_eps
+        self.weight = nn.Parameter(torch.ones(config.d_model))
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Normalizes the input tensor by its RMS value.
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Applies RMS normalization to the input tensor and scales it by the weight parameter.
+        """
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+########################################################
+#
+# Positional Embedding
+#
+########################################################
+
+
+class RoPE(nn.Module):
+    """Rotary Positional Embeddings (RoPE).
+
+    Implements position-dependent rotation of keys and queries in attention mechanism,
+    allowing better modeling of relative positions in sequences. Uses complex number
+    operations for efficient rotation.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Model configuration containing:
+            - config.position_emb_theta: Base for frequency computation
+            - config.d_model: Model dimension
+            - config.attention_n_heads: Number of attention heads
+            - config.max_seq_len: Maximum sequence length
+
+    References:
+        https://arxiv.org/abs/2104.09864
+    """
+
+    _freqs_cis_tensor: torch.Tensor | None = None
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        self.theta = config.position_emb_theta
+        self.dim = config.d_model // config.attention_n_heads
+
+        max_seq_len = config.max_seq_len
+
+        # only gets set once, and then reused for all RoPE instances
+        if RoPE._freqs_cis_tensor is None:
+            RoPE._freqs_cis_tensor = self._setup_freqs_cis(
+                max_seq_len, self.theta, self.dim
+            )
+
+        # register _freqs_cis buffer
+        # can be easily recomputed so persistent=False
+        self.register_buffer("_freqs_cis", self._freqs_cis_tensor, persistent=False)
+
+    @classmethod
+    def _setup_freqs_cis(cls, seq_len: int, theta: float, dim: int) -> torch.Tensor:
+        """Setup Frequency Tensor for RoPE Embeddings
+
+        Initializes the complex frequency tensor that is used to compute the RoPE embeddings.
+
+        Note other implementations will use cos and sin directly, but using the complex
+        number representation is (probably?) more efficient:
+
+            e^(theta * i * t) = cos(theta * t) + i * sin(theta * t) [Euler's formula]
+        """
+        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+        positions = torch.arange(seq_len)
+        freqs = torch.outer(positions, _freqs)
+        return torch.polar(torch.ones_like(freqs), freqs)  # complex64
+
+    def get_freqs_cis(
+        self, input_shape: torch.Size, start_pos: int, end_pos: int
+    ) -> torch.Tensor:
+        """Reshape Frequency Tensor for RoPE Embeddings
+
+        Makes the frequency tensor broadcastable with the input tensor.
+        """
+        _freqs_cis = self._freqs_cis[start_pos:end_pos]
+        ndim = len(input_shape)
+        assert 0 <= 1 < ndim
+        assert _freqs_cis.shape == (input_shape[1], input_shape[-1])
+
+        # TODO: Check whether this is correct (might be able to remove this)
+        shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)]
+        return _freqs_cis.view(*shape)
+
+    def forward(
+        self,
+        queries: torch.Tensor,
+        keys: torch.Tensor,
+        start_pos: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply RoPE Embeddings to Queries and Keys
+
+        Applies the rotary positional embeddings to the input tensors via complex num multiplication
+
+        NOTE: The start_pos is used if we want to use the kv_cache in the attention mechanism.
+        """
+        queries_ = torch.view_as_complex(
+            queries.float().reshape(*queries.shape[:-1], -1, 2)
+        )
+        keys_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
+
+        input_shape = (
+            queries_.shape
+        )  # same as keys: (batch_size, seq_len, n_heads, head_dim/2)
+        freqs_start_pos = start_pos
+        freqs_end_pos = freqs_start_pos + queries_.shape[1]
+
+        freqs_cis = self.get_freqs_cis(input_shape, freqs_start_pos, freqs_end_pos)
+
+        queries_rotated = torch.view_as_real(queries_ * freqs_cis).flatten(3)
+        keys_rotated = torch.view_as_real(keys_ * freqs_cis).flatten(3)
+        return queries_rotated.type_as(queries), keys_rotated.type_as(keys)
+
+
+########################################################
+#
+# Attention
+#
+########################################################
+
+
+class Attention(nn.Module):
+    """Multi-head Attention with Group Query Attention support.
+
+    Implements scaled dot-product attention and supports:
+    - Grouped Query Attention (GQA)
+    - Key-Value caching for efficient inference
+    - RoPE integration
+
+    Args:
+        config (Union[ModelConfig, PretrainedConfig]): Configuration containing:
+            - config.attention_n_heads: Number of attention heads
+            - config.attention_n_kv_heads: Number of key/value heads
+            - config.d_model: Model dimension
+            - config.batch_size: Maximum batch size
+            - config.max_seq_len: Maximum sequence length
+
+    Shape:
+        - Input: (batch_size, seq_len, d_model)
+        - Output: (batch_size, seq_len, d_model)
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.n_heads = config.attention_n_heads
+        self.n_kv_heads = config.attention_n_kv_heads
+
+        self.batch_size = config.batch_size
+        self.max_seq_len = config.max_seq_len
+
+        d_model = config.d_model
+        self.head_dim = d_model // self.n_heads
+
+        self.n_rep = self.n_heads // self.n_kv_heads
+
+        self.q_proj = nn.Linear(d_model, self.n_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.n_heads * self.head_dim, d_model, bias=False)
+
+        self.rope = RoPE(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        """Forward pass for the attention mechanism.
+
+        Computes queries, keys, and values for the attention mechanism. Applies rotary positional
+        embeddings to the queries and keys, and then computes attention scores and outputs.
+
+        For an introduction to the attention mechanism, see:
+        https://arxiv.org/abs/1706.03762
+
+        A few things to note:
+        - The past_key_values is used to implement the KV cache, which is used to speed up
+          generation by caching the KV pairs from previous forward passes. This is useful when doing
+          tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+          modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+          its own KV cache - this KV cache is implemented as a tuple.
+        """
+        bsz, seq_len, _ = input.shape
+        _queries, _keys, _values = (
+            self.q_proj(input),
+            self.k_proj(input),
+            self.v_proj(input),
+        )
+
+        # Reshaping for multi-head attention
+        queries = _queries.view(bsz, seq_len, self.n_heads, self.head_dim)
+        keys = _keys.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        values = _values.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+
+        # The start position is used to apply the RoPE embeddings to only the new tokens
+        # when using the kv_cache in the attention mechanism.
+        # We want to start from the last position in the cache.
+        start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
+
+        # apply rotary positional embeddings
+        queries, keys = self.rope(queries, keys, start_pos)
+
+        if past_key_values is not None:
+            keys = torch.cat([past_key_values[0], keys], dim=1)
+            values = torch.cat([past_key_values[1], values], dim=1)
+
+        if use_cache:
+            cached_keys = keys
+            cached_values = values
+        else:
+            cached_keys = None
+            cached_values = None
+
+        queries = queries.transpose(1, 2)
+        keys = keys.transpose(1, 2)
+        values = values.transpose(1, 2)
+
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and queries.device.type == "mps":
+            # NOTE: MPS does not support GQA in the SDPA kernel, but we can repeat the keys and values
+            # outside of the kernel to get the same effect.
+            # See: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
+            keys = keys.repeat_interleave(self.n_rep, dim=-3)
+            values = values.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+
+        backends = [SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]
+
+        with sdpa_kernel(backends=backends):
+            attn_output = F.scaled_dot_product_attention(
+                queries.contiguous(),
+                keys.contiguous(),
+                values.contiguous(),
+                attn_mask=mask.to(queries.dtype),
+                enable_gqa=apply_gqa,
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        output = self.o_proj(attn_output)
+
+        return output, (cached_keys, cached_values)
+
+
+########################################################
+#
+# SwiGLU (Combines MLP and Activation)
+#
+########################################################
+
+
+class SwiGLU(nn.Module):
+    """SwiGLU Activation Function with Linear Projections.
+
+    Implements the SwiGLU activation function combined with linear transformations,
+    serving as the feed-forward network in transformer blocks.
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Configuration containing:
+            - config.d_model: Model dimension
+            - config.activation_hidden_dim: Hidden dimension (typically 4 * d_model)
+
+    References:
+        https://arxiv.org/abs/2002.05202
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        model_dim = config.d_model
+        act_hidden_dim = config.activation_hidden_dim  # usually 4 * d_model
+
+        self.w_0 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_1 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_2 = nn.Linear(act_hidden_dim, model_dim, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
+
+
+########################################################
+#
+# PicoDecoderBlock
+#
+########################################################
+
+
+class PicoDecoderBlock(nn.Module):
+    """Single Transformer Block with Attention and Feed-forward layers.
+
+    Implements a standard transformer block with:
+    - Multi-head attention with normalization and residual connection
+    - SwiGLU feed-forward network with normalization and residual connection
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Model configuration; either a dataclass or
+            a HuggingFace PicoDecoderHFConfig
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.attention = Attention(config)
+        self.swiglu = SwiGLU(config)
+        self.attention_norm = RMSNorm(config)
+        self.swiglu_norm = RMSNorm(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        attention_output, cached_key_values = self.attention(
+            self.attention_norm(input),
+            mask=mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        # NOTE: cached_key_values is None if use_cache is False
+
+        h = input + attention_output
+        out = h + self.swiglu(self.swiglu_norm(h))
+        return out, cached_key_values
+
+
+########################################################
+#
+# Pico Decoder (Causal Transformer Model)
+#
+########################################################
+
+
+class PicoDecoder(nn.Module):
+    """
+    Pico Decoder: combines the embedding, causal decoder blocks, and output projection into a
+    single autoregressive model.
+
+    For more information on the model, see the classes for the modules that make up the model.
+    """
+
+    def __init__(
+        self,
+        model_config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+        self.config = model_config
+
+        self.embedding_proj = nn.Embedding(self.config.vocab_size, self.config.d_model)
+        self.layers = nn.ModuleList(
+            [PicoDecoderBlock(self.config) for _ in range(self.config.n_layers)]
+        )
+        self.output_norm = RMSNorm(self.config)
+        self.de_embedding_proj = nn.Linear(
+            self.config.d_model, self.config.vocab_size, bias=False
+        )
+
+    def convert_to_hf_model(self) -> "PicoDecoderHF":
+        """Convert the Lightning model to a HuggingFace model."""
+        # Build HF config
+        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
+
+        # Instantiate the HF-wrapped model
+        hf_model = PicoDecoderHF(hf_config)
+
+        # Grab our full state dict, prefixing module names
+        raw_state = self.state_dict(prefix="pico_decoder.")
+
+        # Only keep keys that exist in the HF model (drops classifier_head, etc.)
+        hf_keys = set(hf_model.state_dict().keys())
+        filtered_state = {k: v for k, v in raw_state.items() if k in hf_keys}
+
+        # Load into HF model, ignore any missing keys
+        hf_model.load_state_dict(filtered_state, strict=False)
+
+        return hf_model
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        return_hidden: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]]]:
+        """
+        This is the forward pass for the entire Pico model. It boils down to:
+        - Embedding the input ids
+        - Creating a causal mask
+        - Processing through the pico layers
+        - Projecting the output to logits
+
+        NOTE: One feature that might be confusing is the KV cache. The KV cache is used to speed up
+        generation by caching the KV pairs from previous forward passes. This is useful when doing
+        tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+        modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+        its own KV cache which is stored as a tuple. The whole model then stores a tuple of these
+        KV caches (so a tuple of tuples).
+        """
+
+        seq_len = input_ids.shape[-1]
+        h = self.embedding_proj(input_ids)
+
+        # Calculate start position from past cached KV pairs. Remember that each layer has its
+        # own KV Cache. So when we index past_key_values, we need to index into the KV pairs for the
+        # correct layer and then for either the keys or values.
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
+        # Create causal mask for current sequence
+        mask = None
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
+
+            # If using KV cache, extend mask to cover cached sequence length
+            if past_key_values is not None:
+                # Add zeros for cached tokens (we can attend to all of them)
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+
+            mask = mask.to(h.device)
+
+        # NOTE: If we are using the cache, we need to store the cached KV pairs for each layer
+        #       in a tuple. Each layer will have its own cached KV pair which we aggregate in a tuple.
+        cached_key_values = () if use_cache else None
+
+        # Process through transformer blocks
+        for idx, layer in enumerate(self.layers):
+            layer_past_key_values = (
+                past_key_values[idx] if past_key_values is not None else None
+            )
+
+            h, layer_cached_key_values = layer(
+                h, mask=mask, past_key_values=layer_past_key_values, use_cache=use_cache
+            )
+
+            if use_cache:
+                cached_key_values += (layer_cached_key_values,)
+
+        # Final norm and projection
+        h = self.output_norm(h)
+
+        if return_hidden:
+            return h, cached_key_values
+
+        logits = self.de_embedding_proj(h).float()
+
+        return logits, cached_key_values
+
+
+########################################################
+#
+# HuggingFace Wrapper for the Pico Decoder model.
+#
+########################################################
+
+
+class PicoDecoderHFConfig(PretrainedConfig):
+    """Config class for the Pico Decoder HuggingFace wrapper."""
+
+    model_type = "pico_decoder"
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
+        # NOTE The typical from_dict method doesn't actually set the attributes unless they are
+        # defined in the constructor.
+
+        pico_config = cls(**kwargs)
+
+        # Because this class is just a wrapper around the ModelConfig dataclass, we need to do
+        # a little extra work to ensure that the attributes are actually set.
+        for key, value in config_dict.items():
+            setattr(pico_config, key, value)
+
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+        unused_kwargs = {
+            key: value for key, value in kwargs.items() if not hasattr(pico_config, key)
+        }
+
+        if return_unused_kwargs:
+            return pico_config, unused_kwargs
+        return pico_config
+
+    @classmethod
+    def from_dataclass(cls, model_config: "ModelConfig"):
+        """Initialise from our custom config dataclass."""
+        return cls.from_dict(asdict(model_config))
+
+
+class PicoDecoderHF(PreTrainedModel):
+    """
+    HuggingFace wrapper for the Pico model.
+
+    Many evaluation frameworks require a model be setup as a HuggingFace model, so we provide a simple
+    wrapper that does just that. When we save checkpoints of the Pico model, we save both the normal
+    Pico model as well as the model wrapped in this HuggingFace class.
+
+    This also lets you do cool things like:
+
+    `model = AutoModelForCausalLM.from_pretrained("path/to/checkpoint")`
+    """
+
+    config_class = PicoDecoderHFConfig
+    _no_split_modules = ["PicoBlock", "Attention", "SwiGLU", "RMSNorm"]
+
+    def __init__(self, config: PicoDecoderHFConfig):
+        super().__init__(config)
+        self.pico_decoder = PicoDecoder(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Union[CausalLMOutput, CausalLMOutputWithPast]:
+        """HuggingFace forward pass wrapper.
+
+        Forwards pass for the HuggingFace version of the Pico Model. Basic wrapper around the
+        Pico model's forward pass, and returns the output as a HuggingFace CausalLMOutput.
+        """
+        logits, past_key_values = self.pico_decoder(
+            input_ids, past_key_values, use_cache
+        )
+        if use_cache:
+            return CausalLMOutputWithPast(
+                logits=logits,
+                past_key_values=past_key_values,
+            )
+        else:
+            return CausalLMOutput(
+                logits=logits,
+            )
+
+
+# Register for auto classes
+PicoDecoderHFConfig.register_for_auto_class()
+PicoDecoderHF.register_for_auto_class("AutoModel")
+PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")

checkpoints/step_5400/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|padding|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoints/step_5400/tokenizer_config.json

@@ -0,0 +1,248 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": null,
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<|padding|>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

checkpoints/step_5500/config.json

@@ -0,0 +1,22 @@
+{
+  "activation_hidden_dim": 1536,
+  "architectures": [
+    "PicoDecoderHF"
+  ],
+  "attention_n_heads": 12,
+  "attention_n_kv_heads": 4,
+  "auto_map": {
+    "AutoConfig": "pico_decoder.PicoDecoderHFConfig",
+    "AutoModelForCausalLM": "pico_decoder.PicoDecoderHF"
+  },
+  "batch_size": 1024,
+  "d_model": 384,
+  "max_seq_len": 512,
+  "model_type": "pico_decoder",
+  "n_layers": 12,
+  "norm_eps": 1e-06,
+  "position_emb_theta": 10000.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.3",
+  "vocab_size": 50281
+}

checkpoints/step_5500/fabric_state/checkpoint/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42cfb84cd4b9670005de66448b51f322fe2ee3383f4e2af0f577b203cb3be1ad
+size 213443205

checkpoints/step_5500/fabric_state/checkpoint/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:302a06c18d5912123fd9c802a060dfdeb46390adb8900be59c8279203167b24b
+size 213447621

checkpoints/step_5500/fabric_state/checkpoint/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6aa53d8c64a21998f9c22a5173ef2bd2aa51b47ca9c768f6c6e9f4dc3f9a62db
+size 213445701

checkpoints/step_5500/fabric_state/checkpoint/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c573374b708dccaf59983870aa51ddcea316cc4bf0b3f1a6497698dcae10613
+size 213443589

checkpoints/step_5500/fabric_state/checkpoint/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb1aed7f056a9d362c605072a22a32c1995fead54473ef4553ae1c2697fa3b2b
+size 142325529

checkpoints/step_5500/fabric_state/latest

@@ -0,0 +1 @@
+checkpoint

checkpoints/step_5500/fabric_state/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/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 1, 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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+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, weights_only=False)
+
+        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())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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 _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoints/step_5500/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9244cd4b73321a3763989ecad43fc2222f503768f9eb53a7091d1967d4ea9310
+size 258323520

checkpoints/step_5500/pico_decoder.py

@@ -0,0 +1,623 @@
+"""
+Pico Decoder: A Lightweight Causal Transformer Language Model
+
+Pico Decoder uses a simple LLAMA-style transformer architecture, written for clarity and educational purposes.
+
+Everything is written with a modular design for easy modification and experimentation.
+
+Key features:
+- RMSNorm for layer normalization
+- Rotary Positional Embeddings (RoPE)
+- Multi-head attention with KV-cache support
+- SwiGLU activation function
+- Residual connections throughout
+
+- KV-cache for faster autoregressive generation
+
+References:
+    - RoPE: https://arxiv.org/abs/2104.09864
+    - SwiGLU: https://arxiv.org/abs/2002.05202
+    - LLAMA: https://arxiv.org/abs/2302.13971
+
+Adapted from:
+    - OLMO: https://github.com/allenai/OLMo
+    - LLAMA: https://github.com/meta/llama
+"""
+
+from dataclasses import asdict
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutput, CausalLMOutputWithPast
+
+try:
+    if TYPE_CHECKING:
+        # We need to do this to avoid importing these when creating the HF-compatible models
+        from src.config import ModelConfig
+except ImportError:
+    pass
+
+########################################################
+#
+# Layer Normalization
+#
+########################################################
+
+
+class RMSNorm(torch.nn.Module):
+    """Root Mean Square Layer Normalization.
+
+    A variant of Layer Normalization that uses RMS statistics instead of mean/variance,
+    resulting in improved stability and performance.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Configuration object containing normalization parameters
+            - config.norm_eps: Small constant for numerical stability
+            - config.d_model: Model dimension for the weight parameter
+
+    References:
+        https://arxiv.org/abs/1910.07467
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+        self.eps = config.norm_eps
+        self.weight = nn.Parameter(torch.ones(config.d_model))
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Normalizes the input tensor by its RMS value.
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Applies RMS normalization to the input tensor and scales it by the weight parameter.
+        """
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+########################################################
+#
+# Positional Embedding
+#
+########################################################
+
+
+class RoPE(nn.Module):
+    """Rotary Positional Embeddings (RoPE).
+
+    Implements position-dependent rotation of keys and queries in attention mechanism,
+    allowing better modeling of relative positions in sequences. Uses complex number
+    operations for efficient rotation.
+
+    Args:
+        config (Union[ModelConfig, PicoHFConfig]): Model configuration containing:
+            - config.position_emb_theta: Base for frequency computation
+            - config.d_model: Model dimension
+            - config.attention_n_heads: Number of attention heads
+            - config.max_seq_len: Maximum sequence length
+
+    References:
+        https://arxiv.org/abs/2104.09864
+    """
+
+    _freqs_cis_tensor: torch.Tensor | None = None
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        self.theta = config.position_emb_theta
+        self.dim = config.d_model // config.attention_n_heads
+
+        max_seq_len = config.max_seq_len
+
+        # only gets set once, and then reused for all RoPE instances
+        if RoPE._freqs_cis_tensor is None:
+            RoPE._freqs_cis_tensor = self._setup_freqs_cis(
+                max_seq_len, self.theta, self.dim
+            )
+
+        # register _freqs_cis buffer
+        # can be easily recomputed so persistent=False
+        self.register_buffer("_freqs_cis", self._freqs_cis_tensor, persistent=False)
+
+    @classmethod
+    def _setup_freqs_cis(cls, seq_len: int, theta: float, dim: int) -> torch.Tensor:
+        """Setup Frequency Tensor for RoPE Embeddings
+
+        Initializes the complex frequency tensor that is used to compute the RoPE embeddings.
+
+        Note other implementations will use cos and sin directly, but using the complex
+        number representation is (probably?) more efficient:
+
+            e^(theta * i * t) = cos(theta * t) + i * sin(theta * t) [Euler's formula]
+        """
+        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+        positions = torch.arange(seq_len)
+        freqs = torch.outer(positions, _freqs)
+        return torch.polar(torch.ones_like(freqs), freqs)  # complex64
+
+    def get_freqs_cis(
+        self, input_shape: torch.Size, start_pos: int, end_pos: int
+    ) -> torch.Tensor:
+        """Reshape Frequency Tensor for RoPE Embeddings
+
+        Makes the frequency tensor broadcastable with the input tensor.
+        """
+        _freqs_cis = self._freqs_cis[start_pos:end_pos]
+        ndim = len(input_shape)
+        assert 0 <= 1 < ndim
+        assert _freqs_cis.shape == (input_shape[1], input_shape[-1])
+
+        # TODO: Check whether this is correct (might be able to remove this)
+        shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)]
+        return _freqs_cis.view(*shape)
+
+    def forward(
+        self,
+        queries: torch.Tensor,
+        keys: torch.Tensor,
+        start_pos: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply RoPE Embeddings to Queries and Keys
+
+        Applies the rotary positional embeddings to the input tensors via complex num multiplication
+
+        NOTE: The start_pos is used if we want to use the kv_cache in the attention mechanism.
+        """
+        queries_ = torch.view_as_complex(
+            queries.float().reshape(*queries.shape[:-1], -1, 2)
+        )
+        keys_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
+
+        input_shape = (
+            queries_.shape
+        )  # same as keys: (batch_size, seq_len, n_heads, head_dim/2)
+        freqs_start_pos = start_pos
+        freqs_end_pos = freqs_start_pos + queries_.shape[1]
+
+        freqs_cis = self.get_freqs_cis(input_shape, freqs_start_pos, freqs_end_pos)
+
+        queries_rotated = torch.view_as_real(queries_ * freqs_cis).flatten(3)
+        keys_rotated = torch.view_as_real(keys_ * freqs_cis).flatten(3)
+        return queries_rotated.type_as(queries), keys_rotated.type_as(keys)
+
+
+########################################################
+#
+# Attention
+#
+########################################################
+
+
+class Attention(nn.Module):
+    """Multi-head Attention with Group Query Attention support.
+
+    Implements scaled dot-product attention and supports:
+    - Grouped Query Attention (GQA)
+    - Key-Value caching for efficient inference
+    - RoPE integration
+
+    Args:
+        config (Union[ModelConfig, PretrainedConfig]): Configuration containing:
+            - config.attention_n_heads: Number of attention heads
+            - config.attention_n_kv_heads: Number of key/value heads
+            - config.d_model: Model dimension
+            - config.batch_size: Maximum batch size
+            - config.max_seq_len: Maximum sequence length
+
+    Shape:
+        - Input: (batch_size, seq_len, d_model)
+        - Output: (batch_size, seq_len, d_model)
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.n_heads = config.attention_n_heads
+        self.n_kv_heads = config.attention_n_kv_heads
+
+        self.batch_size = config.batch_size
+        self.max_seq_len = config.max_seq_len
+
+        d_model = config.d_model
+        self.head_dim = d_model // self.n_heads
+
+        self.n_rep = self.n_heads // self.n_kv_heads
+
+        self.q_proj = nn.Linear(d_model, self.n_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.n_heads * self.head_dim, d_model, bias=False)
+
+        self.rope = RoPE(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        """Forward pass for the attention mechanism.
+
+        Computes queries, keys, and values for the attention mechanism. Applies rotary positional
+        embeddings to the queries and keys, and then computes attention scores and outputs.
+
+        For an introduction to the attention mechanism, see:
+        https://arxiv.org/abs/1706.03762
+
+        A few things to note:
+        - The past_key_values is used to implement the KV cache, which is used to speed up
+          generation by caching the KV pairs from previous forward passes. This is useful when doing
+          tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+          modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+          its own KV cache - this KV cache is implemented as a tuple.
+        """
+        bsz, seq_len, _ = input.shape
+        _queries, _keys, _values = (
+            self.q_proj(input),
+            self.k_proj(input),
+            self.v_proj(input),
+        )
+
+        # Reshaping for multi-head attention
+        queries = _queries.view(bsz, seq_len, self.n_heads, self.head_dim)
+        keys = _keys.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        values = _values.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+
+        # The start position is used to apply the RoPE embeddings to only the new tokens
+        # when using the kv_cache in the attention mechanism.
+        # We want to start from the last position in the cache.
+        start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
+
+        # apply rotary positional embeddings
+        queries, keys = self.rope(queries, keys, start_pos)
+
+        if past_key_values is not None:
+            keys = torch.cat([past_key_values[0], keys], dim=1)
+            values = torch.cat([past_key_values[1], values], dim=1)
+
+        if use_cache:
+            cached_keys = keys
+            cached_values = values
+        else:
+            cached_keys = None
+            cached_values = None
+
+        queries = queries.transpose(1, 2)
+        keys = keys.transpose(1, 2)
+        values = values.transpose(1, 2)
+
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and queries.device.type == "mps":
+            # NOTE: MPS does not support GQA in the SDPA kernel, but we can repeat the keys and values
+            # outside of the kernel to get the same effect.
+            # See: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
+            keys = keys.repeat_interleave(self.n_rep, dim=-3)
+            values = values.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+
+        backends = [SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]
+
+        with sdpa_kernel(backends=backends):
+            attn_output = F.scaled_dot_product_attention(
+                queries.contiguous(),
+                keys.contiguous(),
+                values.contiguous(),
+                attn_mask=mask.to(queries.dtype),
+                enable_gqa=apply_gqa,
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        output = self.o_proj(attn_output)
+
+        return output, (cached_keys, cached_values)
+
+
+########################################################
+#
+# SwiGLU (Combines MLP and Activation)
+#
+########################################################
+
+
+class SwiGLU(nn.Module):
+    """SwiGLU Activation Function with Linear Projections.
+
+    Implements the SwiGLU activation function combined with linear transformations,
+    serving as the feed-forward network in transformer blocks.
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Configuration containing:
+            - config.d_model: Model dimension
+            - config.activation_hidden_dim: Hidden dimension (typically 4 * d_model)
+
+    References:
+        https://arxiv.org/abs/2002.05202
+    """
+
+    def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
+        super().__init__()
+
+        model_dim = config.d_model
+        act_hidden_dim = config.activation_hidden_dim  # usually 4 * d_model
+
+        self.w_0 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_1 = nn.Linear(model_dim, act_hidden_dim, bias=False)
+        self.w_2 = nn.Linear(act_hidden_dim, model_dim, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
+
+
+########################################################
+#
+# PicoDecoderBlock
+#
+########################################################
+
+
+class PicoDecoderBlock(nn.Module):
+    """Single Transformer Block with Attention and Feed-forward layers.
+
+    Implements a standard transformer block with:
+    - Multi-head attention with normalization and residual connection
+    - SwiGLU feed-forward network with normalization and residual connection
+
+    Args:
+        config (Union[ModelConfig, PicoDecoderHFConfig]): Model configuration; either a dataclass or
+            a HuggingFace PicoDecoderHFConfig
+    """
+
+    def __init__(
+        self,
+        config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+
+        self.attention = Attention(config)
+        self.swiglu = SwiGLU(config)
+        self.attention_norm = RMSNorm(config)
+        self.swiglu_norm = RMSNorm(config)
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+        attention_output, cached_key_values = self.attention(
+            self.attention_norm(input),
+            mask=mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        # NOTE: cached_key_values is None if use_cache is False
+
+        h = input + attention_output
+        out = h + self.swiglu(self.swiglu_norm(h))
+        return out, cached_key_values
+
+
+########################################################
+#
+# Pico Decoder (Causal Transformer Model)
+#
+########################################################
+
+
+class PicoDecoder(nn.Module):
+    """
+    Pico Decoder: combines the embedding, causal decoder blocks, and output projection into a
+    single autoregressive model.
+
+    For more information on the model, see the classes for the modules that make up the model.
+    """
+
+    def __init__(
+        self,
+        model_config: Union["ModelConfig", "PicoDecoderHFConfig"],
+    ):
+        super().__init__()
+        self.config = model_config
+
+        self.embedding_proj = nn.Embedding(self.config.vocab_size, self.config.d_model)
+        self.layers = nn.ModuleList(
+            [PicoDecoderBlock(self.config) for _ in range(self.config.n_layers)]
+        )
+        self.output_norm = RMSNorm(self.config)
+        self.de_embedding_proj = nn.Linear(
+            self.config.d_model, self.config.vocab_size, bias=False
+        )
+
+    def convert_to_hf_model(self) -> "PicoDecoderHF":
+        """Convert the Lightning model to a HuggingFace model."""
+        # Build HF config
+        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
+
+        # Instantiate the HF-wrapped model
+        hf_model = PicoDecoderHF(hf_config)
+
+        # Grab our full state dict, prefixing module names
+        raw_state = self.state_dict(prefix="pico_decoder.")
+
+        # Only keep keys that exist in the HF model (drops classifier_head, etc.)
+        hf_keys = set(hf_model.state_dict().keys())
+        filtered_state = {k: v for k, v in raw_state.items() if k in hf_keys}
+
+        # Load into HF model, ignore any missing keys
+        hf_model.load_state_dict(filtered_state, strict=False)
+
+        return hf_model
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        return_hidden: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]]]:
+        """
+        This is the forward pass for the entire Pico model. It boils down to:
+        - Embedding the input ids
+        - Creating a causal mask
+        - Processing through the pico layers
+        - Projecting the output to logits
+
+        NOTE: One feature that might be confusing is the KV cache. The KV cache is used to speed up
+        generation by caching the KV pairs from previous forward passes. This is useful when doing
+        tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
+        modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
+        its own KV cache which is stored as a tuple. The whole model then stores a tuple of these
+        KV caches (so a tuple of tuples).
+        """
+
+        seq_len = input_ids.shape[-1]
+        h = self.embedding_proj(input_ids)
+
+        # Calculate start position from past cached KV pairs. Remember that each layer has its
+        # own KV Cache. So when we index past_key_values, we need to index into the KV pairs for the
+        # correct layer and then for either the keys or values.
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
+        # Create causal mask for current sequence
+        mask = None
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
+
+            # If using KV cache, extend mask to cover cached sequence length
+            if past_key_values is not None:
+                # Add zeros for cached tokens (we can attend to all of them)
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+
+            mask = mask.to(h.device)
+
+        # NOTE: If we are using the cache, we need to store the cached KV pairs for each layer
+        #       in a tuple. Each layer will have its own cached KV pair which we aggregate in a tuple.
+        cached_key_values = () if use_cache else None
+
+        # Process through transformer blocks
+        for idx, layer in enumerate(self.layers):
+            layer_past_key_values = (
+                past_key_values[idx] if past_key_values is not None else None
+            )
+
+            h, layer_cached_key_values = layer(
+                h, mask=mask, past_key_values=layer_past_key_values, use_cache=use_cache
+            )
+
+            if use_cache:
+                cached_key_values += (layer_cached_key_values,)
+
+        # Final norm and projection
+        h = self.output_norm(h)
+
+        if return_hidden:
+            return h, cached_key_values
+
+        logits = self.de_embedding_proj(h).float()
+
+        return logits, cached_key_values
+
+
+########################################################
+#
+# HuggingFace Wrapper for the Pico Decoder model.
+#
+########################################################
+
+
+class PicoDecoderHFConfig(PretrainedConfig):
+    """Config class for the Pico Decoder HuggingFace wrapper."""
+
+    model_type = "pico_decoder"
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
+        # NOTE The typical from_dict method doesn't actually set the attributes unless they are
+        # defined in the constructor.
+
+        pico_config = cls(**kwargs)
+
+        # Because this class is just a wrapper around the ModelConfig dataclass, we need to do
+        # a little extra work to ensure that the attributes are actually set.
+        for key, value in config_dict.items():
+            setattr(pico_config, key, value)
+
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+        unused_kwargs = {
+            key: value for key, value in kwargs.items() if not hasattr(pico_config, key)
+        }
+
+        if return_unused_kwargs:
+            return pico_config, unused_kwargs
+        return pico_config
+
+    @classmethod
+    def from_dataclass(cls, model_config: "ModelConfig"):
+        """Initialise from our custom config dataclass."""
+        return cls.from_dict(asdict(model_config))
+
+
+class PicoDecoderHF(PreTrainedModel):
+    """
+    HuggingFace wrapper for the Pico model.
+
+    Many evaluation frameworks require a model be setup as a HuggingFace model, so we provide a simple
+    wrapper that does just that. When we save checkpoints of the Pico model, we save both the normal
+    Pico model as well as the model wrapped in this HuggingFace class.
+
+    This also lets you do cool things like:
+
+    `model = AutoModelForCausalLM.from_pretrained("path/to/checkpoint")`
+    """
+
+    config_class = PicoDecoderHFConfig
+    _no_split_modules = ["PicoBlock", "Attention", "SwiGLU", "RMSNorm"]
+
+    def __init__(self, config: PicoDecoderHFConfig):
+        super().__init__(config)
+        self.pico_decoder = PicoDecoder(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Union[CausalLMOutput, CausalLMOutputWithPast]:
+        """HuggingFace forward pass wrapper.
+
+        Forwards pass for the HuggingFace version of the Pico Model. Basic wrapper around the
+        Pico model's forward pass, and returns the output as a HuggingFace CausalLMOutput.
+        """
+        logits, past_key_values = self.pico_decoder(
+            input_ids, past_key_values, use_cache
+        )
+        if use_cache:
+            return CausalLMOutputWithPast(
+                logits=logits,
+                past_key_values=past_key_values,
+            )
+        else:
+            return CausalLMOutput(
+                logits=logits,
+            )
+
+
+# Register for auto classes
+PicoDecoderHFConfig.register_for_auto_class()
+PicoDecoderHF.register_for_auto_class("AutoModel")
+PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")

checkpoints/step_5500/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|padding|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}