upload models (3 splits and 2 languages)

s1/README.md

@@ -0,0 +1,9 @@
+---
+library_name: peft
+---
+## Training procedure
+
+### Framework versions
+
+
+- PEFT 0.4.0

s1/adapter_config.json

@@ -0,0 +1,26 @@
+{
+  "auto_mapping": null,
+  "base_model_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "lora_alpha": 256,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "up_proj",
+    "down_proj",
+    "gate_proj",
+    "v_proj",
+    "q_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM"
+}

s1/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d3febffb0fd3e2a025ce3cb029bebed8acaa25e1000903988aaeb13108c285e
+size 639786637

s1/config.json

@@ -0,0 +1,44 @@
+{
+  "_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "architectures": [
+    "LlavaLlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "freeze_mm_mlp_adapter": false,
+  "freeze_mm_vision_resampler": false,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_aspect_ratio": "pad",
+  "image_grid_pinpoints": null,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_length": 4096,
+  "max_position_embeddings": 4096,
+  "mm_hidden_size": 1024,
+  "mm_projector_lr": 2e-05,
+  "mm_projector_type": "mlp2x_gelu",
+  "mm_resampler_type": null,
+  "mm_use_im_patch_token": false,
+  "mm_use_im_start_end": false,
+  "mm_vision_select_feature": "patch",
+  "mm_vision_select_layer": -2,
+  "mm_vision_tower": "openai/clip-vit-large-patch14-336",
+  "model_type": "llava",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0",
+  "tune_mm_mlp_adapter": false,
+  "tune_mm_vision_resampler": false,
+  "unfreeze_mm_vision_tower": false,
+  "use_cache": true,
+  "use_mm_proj": true,
+  "vocab_size": 32000
+}

s1/global_step16000/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:d7f8eef59b5a3c0c4ec6e82319eeef44e4323b2c8ac07d75566dff8a351b157f
+size 4089599575

s1/global_step16000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d28eb31954dc91f37189a42f4995323cdb662187f21669e7b07705c7ccedd678
+size 28850200126

s1/latest

@@ -0,0 +1 @@
+global_step16000

s1/non_lora_trainables.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4198193940865ffcf86b34f70405ce8adbc1c0071f1b49d6024360adce75538
+size 41961191

s1/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f2b92e3dbbeb0695186b26a8876ed34fe0191fb75d370352321df9b44c28b69
+size 14575

s1/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<unk>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s1/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

s1/tokenizer_config.json

@@ -0,0 +1,35 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": null,
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s1/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87f1ffd8be10f016d5b0f1117c5a9d55c37250231e131f685a241781c4bad5e5
+size 5819

s1/zero_to_fp32.py

@@ -0,0 +1,578 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # 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 files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        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")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)

s1_en/README.md

@@ -0,0 +1,9 @@
+---
+library_name: peft
+---
+## Training procedure
+
+### Framework versions
+
+
+- PEFT 0.4.0

s1_en/adapter_config.json

@@ -0,0 +1,26 @@
+{
+  "auto_mapping": null,
+  "base_model_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "lora_alpha": 256,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "revision": null,
+  "target_modules": [
+    "o_proj",
+    "gate_proj",
+    "k_proj",
+    "up_proj",
+    "v_proj",
+    "down_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM"
+}

s1_en/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc9db79395c61ca721098c42fa0604509f6486b854b170f74f5f3f094450850d
+size 639787082

s1_en/config.json

@@ -0,0 +1,44 @@
+{
+  "_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "architectures": [
+    "LlavaLlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "freeze_mm_mlp_adapter": false,
+  "freeze_mm_vision_resampler": false,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_aspect_ratio": "pad",
+  "image_grid_pinpoints": null,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_length": 4096,
+  "max_position_embeddings": 4096,
+  "mm_hidden_size": 1024,
+  "mm_projector_lr": 2e-05,
+  "mm_projector_type": "mlp2x_gelu",
+  "mm_resampler_type": null,
+  "mm_use_im_patch_token": false,
+  "mm_use_im_start_end": false,
+  "mm_vision_select_feature": "patch",
+  "mm_vision_select_layer": -2,
+  "mm_vision_tower": "openai/clip-vit-large-patch14-336",
+  "model_type": "llava",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0",
+  "tune_mm_mlp_adapter": false,
+  "tune_mm_vision_resampler": false,
+  "unfreeze_mm_vision_tower": false,
+  "use_cache": true,
+  "use_mm_proj": true,
+  "vocab_size": 32000
+}

s1_en/global_step9000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:834a66d3e6038595959d043acd94e8307f2cc3d99b1bd7c086e3ea2b0c2db7f2
+size 4089600080

s1_en/global_step9000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b2df990e43d817071629b6f3d8595491b9b28bd994e08cff95295e85abed0a5
+size 28850200603

s1_en/latest

@@ -0,0 +1 @@
+global_step9000

s1_en/non_lora_trainables.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32b51a7f52ae8e9f73dcaa6749addaa2061b0e2c36c2d8808d0db708c5ff6571
+size 41961648

s1_en/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de6235bea9da7d32a83efa288efc2ba3fb1faa6f936d34421a99a0ab1b4f0e31
+size 14244

s1_en/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<unk>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s1_en/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

s1_en/tokenizer_config.json

@@ -0,0 +1,35 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": null,
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s1_en/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4786c5345d302e6b458af37541702ba130c2950bc2a2c1d2cdf2001843001b92
+size 6264

s1_en/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)

s2/README.md

@@ -0,0 +1,9 @@
+---
+library_name: peft
+---
+## Training procedure
+
+### Framework versions
+
+
+- PEFT 0.4.0

s2/adapter_config.json

@@ -0,0 +1,26 @@
+{
+  "auto_mapping": null,
+  "base_model_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "lora_alpha": 256,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "up_proj",
+    "down_proj",
+    "v_proj",
+    "k_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM"
+}

s2/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:657313672dfd64ea5545a29eeda7f79040e21b724057b163e61b4f9bde5254c1
+size 639786637

s2/config.json

@@ -0,0 +1,44 @@
+{
+  "_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "architectures": [
+    "LlavaLlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "freeze_mm_mlp_adapter": false,
+  "freeze_mm_vision_resampler": false,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_aspect_ratio": "pad",
+  "image_grid_pinpoints": null,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_length": 4096,
+  "max_position_embeddings": 4096,
+  "mm_hidden_size": 1024,
+  "mm_projector_lr": 2e-05,
+  "mm_projector_type": "mlp2x_gelu",
+  "mm_resampler_type": null,
+  "mm_use_im_patch_token": false,
+  "mm_use_im_start_end": false,
+  "mm_vision_select_feature": "patch",
+  "mm_vision_select_layer": -2,
+  "mm_vision_tower": "openai/clip-vit-large-patch14-336",
+  "model_type": "llava",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0",
+  "tune_mm_mlp_adapter": false,
+  "tune_mm_vision_resampler": false,
+  "unfreeze_mm_vision_tower": false,
+  "use_cache": true,
+  "use_mm_proj": true,
+  "vocab_size": 32000
+}

s2/global_step6000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:235c3c2d5ebc5cc4e622b2f6c86c0936e6fbfe1d05cb8e7314f0d6f521c98d02
+size 4089599575

s2/global_step6000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb4898e6378bd9b8e2d3817dfa868077e19aa7156a78daf0eab939a965962028
+size 28850200126

s2/latest

@@ -0,0 +1 @@
+global_step6000

s2/non_lora_trainables.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ca16573ba2935eb1df5756e9b23cc2487e048a9ce1e8a02e76ef2271b47feea
+size 41961191

s2/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b571a0ec7a60a1d36dde4a941ac919260dbdf1599a2822ccc2feaeb81b4eee
+size 14575

s2/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<unk>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s2/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

s2/tokenizer_config.json

@@ -0,0 +1,35 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": null,
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

s2/trainer_state.json

@@ -0,0 +1,3616 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 13.392857142857142,
+  "global_step": 6000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.7174721189591085e-07,
+      "loss": 6.9406,
+      "step": 10
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 7.434944237918217e-07,
+      "loss": 6.9625,
+      "step": 20
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 1.1152416356877324e-06,
+      "loss": 5.4781,
+      "step": 30
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1.4869888475836434e-06,
+      "loss": 4.2969,
+      "step": 40
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1.858736059479554e-06,
+      "loss": 3.7406,
+      "step": 50
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 2.2304832713754648e-06,
+      "loss": 3.3547,
+      "step": 60
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 2.6022304832713758e-06,
+      "loss": 2.9609,
+      "step": 70
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 2.973977695167287e-06,
+      "loss": 2.9625,
+      "step": 80
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 3.3457249070631974e-06,
+      "loss": 2.5938,
+      "step": 90
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 3.717472118959108e-06,
+      "loss": 2.2594,
+      "step": 100
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 4.089219330855019e-06,
+      "loss": 1.9625,
+      "step": 110
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 4.4609665427509296e-06,
+      "loss": 1.8102,
+      "step": 120
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 4.83271375464684e-06,
+      "loss": 1.625,
+      "step": 130
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 5.2044609665427516e-06,
+      "loss": 1.4773,
+      "step": 140
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 5.576208178438662e-06,
+      "loss": 1.3578,
+      "step": 150
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 5.947955390334574e-06,
+      "loss": 1.2027,
+      "step": 160
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 6.319702602230484e-06,
+      "loss": 1.0359,
+      "step": 170
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 6.691449814126395e-06,
+      "loss": 0.917,
+      "step": 180
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 7.063197026022306e-06,
+      "loss": 0.9805,
+      "step": 190
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 7.434944237918216e-06,
+      "loss": 0.7047,
+      "step": 200
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 7.806691449814127e-06,
+      "loss": 0.8602,
+      "step": 210
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.178438661710038e-06,
+      "loss": 0.9711,
+      "step": 220
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.550185873605949e-06,
+      "loss": 0.4922,
+      "step": 230
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 8.921933085501859e-06,
+      "loss": 0.6744,
+      "step": 240
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 9.29368029739777e-06,
+      "loss": 1.0547,
+      "step": 250
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 9.66542750929368e-06,
+      "loss": 0.6121,
+      "step": 260
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1.0037174721189591e-05,
+      "loss": 0.5311,
+      "step": 270
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1.0408921933085503e-05,
+      "loss": 0.4962,
+      "step": 280
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1.0780669144981412e-05,
+      "loss": 0.4607,
+      "step": 290
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1.1152416356877324e-05,
+      "loss": 0.6136,
+      "step": 300
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1.1524163568773235e-05,
+      "loss": 0.6199,
+      "step": 310
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1.1895910780669147e-05,
+      "loss": 0.5402,
+      "step": 320
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1.2267657992565056e-05,
+      "loss": 0.5375,
+      "step": 330
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1.2639405204460968e-05,
+      "loss": 0.4029,
+      "step": 340
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1.3011152416356879e-05,
+      "loss": 0.4354,
+      "step": 350
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1.338289962825279e-05,
+      "loss": 0.7154,
+      "step": 360
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1.37546468401487e-05,
+      "loss": 0.4764,
+      "step": 370
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1.4126394052044612e-05,
+      "loss": 0.4047,
+      "step": 380
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1.4498141263940521e-05,
+      "loss": 0.4208,
+      "step": 390
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1.4869888475836432e-05,
+      "loss": 0.3448,
+      "step": 400
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1.5241635687732344e-05,
+      "loss": 0.4429,
+      "step": 410
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1.5613382899628255e-05,
+      "loss": 0.7079,
+      "step": 420
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1.5985130111524165e-05,
+      "loss": 0.5989,
+      "step": 430
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1.6356877323420076e-05,
+      "loss": 0.4419,
+      "step": 440
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 1.6728624535315986e-05,
+      "loss": 0.3883,
+      "step": 450
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 1.7100371747211897e-05,
+      "loss": 0.3784,
+      "step": 460
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 1.7472118959107808e-05,
+      "loss": 0.3872,
+      "step": 470
+    },
+    {
+      "epoch": 1.07,
+      "learning_rate": 1.7843866171003718e-05,
+      "loss": 0.4399,
+      "step": 480
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 1.8215613382899632e-05,
+      "loss": 0.346,
+      "step": 490
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 1.858736059479554e-05,
+      "loss": 0.3869,
+      "step": 500
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 1.8959107806691453e-05,
+      "loss": 0.3469,
+      "step": 510
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 1.933085501858736e-05,
+      "loss": 0.3942,
+      "step": 520
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 1.970260223048327e-05,
+      "loss": 0.497,
+      "step": 530
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 1.9999999346673545e-05,
+      "loss": 0.3417,
+      "step": 540
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 1.9999976480256544e-05,
+      "loss": 0.41,
+      "step": 550
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 1.9999920947602093e-05,
+      "loss": 0.5033,
+      "step": 560
+    },
+    {
+      "epoch": 1.27,
+      "learning_rate": 1.9999832748891602e-05,
+      "loss": 0.4127,
+      "step": 570
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 1.999971188441319e-05,
+      "loss": 0.4914,
+      "step": 580
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 1.999955835456167e-05,
+      "loss": 0.2726,
+      "step": 590
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 1.9999372159838563e-05,
+      "loss": 0.3555,
+      "step": 600
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 1.9999153300852108e-05,
+      "loss": 0.4463,
+      "step": 610
+    },
+    {
+      "epoch": 1.38,
+      "learning_rate": 1.999890177831723e-05,
+      "loss": 0.3625,
+      "step": 620
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 1.9998617593055565e-05,
+      "loss": 0.559,
+      "step": 630
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 1.9998300745995437e-05,
+      "loss": 0.2857,
+      "step": 640
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 1.9997951238171875e-05,
+      "loss": 0.339,
+      "step": 650
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 1.9997569070726585e-05,
+      "loss": 0.4425,
+      "step": 660
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 1.9997154244907972e-05,
+      "loss": 0.386,
+      "step": 670
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 1.999670676207112e-05,
+      "loss": 0.2583,
+      "step": 680
+    },
+    {
+      "epoch": 1.54,
+      "learning_rate": 1.999622662367778e-05,
+      "loss": 0.2287,
+      "step": 690
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 1.99957138312964e-05,
+      "loss": 0.4197,
+      "step": 700
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 1.999516838660208e-05,
+      "loss": 0.5672,
+      "step": 710
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 1.9994590291376585e-05,
+      "loss": 0.4251,
+      "step": 720
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 1.999397954750834e-05,
+      "loss": 0.2164,
+      "step": 730
+    },
+    {
+      "epoch": 1.65,
+      "learning_rate": 1.999333615699242e-05,
+      "loss": 0.4329,
+      "step": 740
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 1.999266012193054e-05,
+      "loss": 0.3667,
+      "step": 750
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 1.9991951444531067e-05,
+      "loss": 0.5154,
+      "step": 760
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 1.999121012710898e-05,
+      "loss": 0.4191,
+      "step": 770
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 1.9990436172085896e-05,
+      "loss": 0.4372,
+      "step": 780
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 1.9989629581990038e-05,
+      "loss": 0.3543,
+      "step": 790
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 1.9988790359456236e-05,
+      "loss": 0.2781,
+      "step": 800
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 1.998791850722593e-05,
+      "loss": 0.3889,
+      "step": 810
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 1.9987014028147124e-05,
+      "loss": 0.3639,
+      "step": 820
+    },
+    {
+      "epoch": 1.85,
+      "learning_rate": 1.9986076925174433e-05,
+      "loss": 0.2839,
+      "step": 830
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 1.9985107201369024e-05,
+      "loss": 0.3239,
+      "step": 840
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 1.998410485989862e-05,
+      "loss": 0.2576,
+      "step": 850
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 1.9983069904037506e-05,
+      "loss": 0.3892,
+      "step": 860
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 1.99820023371665e-05,
+      "loss": 0.2446,
+      "step": 870
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 1.9980902162772952e-05,
+      "loss": 0.3531,
+      "step": 880
+    },
+    {
+      "epoch": 1.99,
+      "learning_rate": 1.9979769384450728e-05,
+      "loss": 0.4144,
+      "step": 890
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 1.9978604005900187e-05,
+      "loss": 0.3633,
+      "step": 900
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 1.9977406030928205e-05,
+      "loss": 0.3056,
+      "step": 910
+    },
+    {
+      "epoch": 2.05,
+      "learning_rate": 1.9976175463448113e-05,
+      "loss": 0.2005,
+      "step": 920
+    },
+    {
+      "epoch": 2.08,
+      "learning_rate": 1.997491230747973e-05,
+      "loss": 0.3426,
+      "step": 930
+    },
+    {
+      "epoch": 2.1,
+      "learning_rate": 1.9973616567149316e-05,
+      "loss": 0.2796,
+      "step": 940
+    },
+    {
+      "epoch": 2.12,
+      "learning_rate": 1.9972288246689576e-05,
+      "loss": 0.37,
+      "step": 950
+    },
+    {
+      "epoch": 2.14,
+      "learning_rate": 1.9970927350439647e-05,
+      "loss": 0.2958,
+      "step": 960
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 1.9969533882845076e-05,
+      "loss": 0.2895,
+      "step": 970
+    },
+    {
+      "epoch": 2.19,
+      "learning_rate": 1.9968107848457804e-05,
+      "loss": 0.2656,
+      "step": 980
+    },
+    {
+      "epoch": 2.21,
+      "learning_rate": 1.9966649251936164e-05,
+      "loss": 0.2885,
+      "step": 990
+    },
+    {
+      "epoch": 2.23,
+      "learning_rate": 1.996515809804485e-05,
+      "loss": 0.3175,
+      "step": 1000
+    },
+    {
+      "epoch": 2.25,
+      "learning_rate": 1.9963634391654912e-05,
+      "loss": 0.2314,
+      "step": 1010
+    },
+    {
+      "epoch": 2.28,
+      "learning_rate": 1.996207813774374e-05,
+      "loss": 0.3122,
+      "step": 1020
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 1.9960489341395043e-05,
+      "loss": 0.1354,
+      "step": 1030
+    },
+    {
+      "epoch": 2.32,
+      "learning_rate": 1.9958868007798828e-05,
+      "loss": 0.2356,
+      "step": 1040
+    },
+    {
+      "epoch": 2.34,
+      "learning_rate": 1.9957214142251392e-05,
+      "loss": 0.2,
+      "step": 1050
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 1.9955527750155315e-05,
+      "loss": 0.2901,
+      "step": 1060
+    },
+    {
+      "epoch": 2.39,
+      "learning_rate": 1.995380883701941e-05,
+      "loss": 0.2357,
+      "step": 1070
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 1.995205740845874e-05,
+      "loss": 0.3321,
+      "step": 1080
+    },
+    {
+      "epoch": 2.43,
+      "learning_rate": 1.9950273470194566e-05,
+      "loss": 0.227,
+      "step": 1090
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 1.9948457028054364e-05,
+      "loss": 0.2477,
+      "step": 1100
+    },
+    {
+      "epoch": 2.48,
+      "learning_rate": 1.994660808797178e-05,
+      "loss": 0.2848,
+      "step": 1110
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 1.9944726655986618e-05,
+      "loss": 0.2755,
+      "step": 1120
+    },
+    {
+      "epoch": 2.52,
+      "learning_rate": 1.9942812738244827e-05,
+      "loss": 0.2539,
+      "step": 1130
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 1.9940866340998464e-05,
+      "loss": 0.2153,
+      "step": 1140
+    },
+    {
+      "epoch": 2.57,
+      "learning_rate": 1.99388874706057e-05,
+      "loss": 0.3135,
+      "step": 1150
+    },
+    {
+      "epoch": 2.59,
+      "learning_rate": 1.9936876133530766e-05,
+      "loss": 0.2413,
+      "step": 1160
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 1.9934832336343968e-05,
+      "loss": 0.2698,
+      "step": 1170
+    },
+    {
+      "epoch": 2.63,
+      "learning_rate": 1.993275608572163e-05,
+      "loss": 0.3131,
+      "step": 1180
+    },
+    {
+      "epoch": 2.66,
+      "learning_rate": 1.99306473884461e-05,
+      "loss": 0.2826,
+      "step": 1190
+    },
+    {
+      "epoch": 2.68,
+      "learning_rate": 1.992850625140572e-05,
+      "loss": 0.2425,
+      "step": 1200
+    },
+    {
+      "epoch": 2.7,
+      "learning_rate": 1.992633268159479e-05,
+      "loss": 0.2397,
+      "step": 1210
+    },
+    {
+      "epoch": 2.72,
+      "learning_rate": 1.992412668611356e-05,
+      "loss": 0.2258,
+      "step": 1220
+    },
+    {
+      "epoch": 2.75,
+      "learning_rate": 1.992188827216821e-05,
+      "loss": 0.2616,
+      "step": 1230
+    },
+    {
+      "epoch": 2.77,
+      "learning_rate": 1.9919617447070808e-05,
+      "loss": 0.2745,
+      "step": 1240
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 1.991731421823931e-05,
+      "loss": 0.2235,
+      "step": 1250
+    },
+    {
+      "epoch": 2.81,
+      "learning_rate": 1.9914978593197507e-05,
+      "loss": 0.281,
+      "step": 1260
+    },
+    {
+      "epoch": 2.83,
+      "learning_rate": 1.991261057957503e-05,
+      "loss": 0.2768,
+      "step": 1270
+    },
+    {
+      "epoch": 2.86,
+      "learning_rate": 1.9910210185107306e-05,
+      "loss": 0.1956,
+      "step": 1280
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 1.9907777417635537e-05,
+      "loss": 0.4028,
+      "step": 1290
+    },
+    {
+      "epoch": 2.9,
+      "learning_rate": 1.9905312285106682e-05,
+      "loss": 0.2349,
+      "step": 1300
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 1.9902814795573416e-05,
+      "loss": 0.3584,
+      "step": 1310
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 1.9900284957194115e-05,
+      "loss": 0.2127,
+      "step": 1320
+    },
+    {
+      "epoch": 2.97,
+      "learning_rate": 1.9897722778232836e-05,
+      "loss": 0.2892,
+      "step": 1330
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 1.9895128267059268e-05,
+      "loss": 0.2294,
+      "step": 1340
+    },
+    {
+      "epoch": 3.01,
+      "learning_rate": 1.9892501432148728e-05,
+      "loss": 0.1796,
+      "step": 1350
+    },
+    {
+      "epoch": 3.04,
+      "learning_rate": 1.988984228208211e-05,
+      "loss": 0.1547,
+      "step": 1360
+    },
+    {
+      "epoch": 3.06,
+      "learning_rate": 1.988715082554588e-05,
+      "loss": 0.2344,
+      "step": 1370
+    },
+    {
+      "epoch": 3.08,
+      "learning_rate": 1.988442707133204e-05,
+      "loss": 0.1938,
+      "step": 1380
+    },
+    {
+      "epoch": 3.1,
+      "learning_rate": 1.988167102833808e-05,
+      "loss": 0.1642,
+      "step": 1390
+    },
+    {
+      "epoch": 3.12,
+      "learning_rate": 1.9878882705566988e-05,
+      "loss": 0.2037,
+      "step": 1400
+    },
+    {
+      "epoch": 3.15,
+      "learning_rate": 1.9876062112127176e-05,
+      "loss": 0.1806,
+      "step": 1410
+    },
+    {
+      "epoch": 3.17,
+      "learning_rate": 1.9873209257232494e-05,
+      "loss": 0.1858,
+      "step": 1420
+    },
+    {
+      "epoch": 3.19,
+      "learning_rate": 1.987032415020216e-05,
+      "loss": 0.2068,
+      "step": 1430
+    },
+    {
+      "epoch": 3.21,
+      "learning_rate": 1.986740680046076e-05,
+      "loss": 0.1787,
+      "step": 1440
+    },
+    {
+      "epoch": 3.24,
+      "learning_rate": 1.9864457217538198e-05,
+      "loss": 0.1947,
+      "step": 1450
+    },
+    {
+      "epoch": 3.26,
+      "learning_rate": 1.986147541106967e-05,
+      "loss": 0.1804,
+      "step": 1460
+    },
+    {
+      "epoch": 3.28,
+      "learning_rate": 1.9858461390795648e-05,
+      "loss": 0.2003,
+      "step": 1470
+    },
+    {
+      "epoch": 3.3,
+      "learning_rate": 1.985541516656182e-05,
+      "loss": 0.141,
+      "step": 1480
+    },
+    {
+      "epoch": 3.33,
+      "learning_rate": 1.985233674831908e-05,
+      "loss": 0.1854,
+      "step": 1490
+    },
+    {
+      "epoch": 3.35,
+      "learning_rate": 1.984922614612348e-05,
+      "loss": 0.1827,
+      "step": 1500
+    },
+    {
+      "epoch": 3.37,
+      "learning_rate": 1.9846083370136217e-05,
+      "loss": 0.2692,
+      "step": 1510
+    },
+    {
+      "epoch": 3.39,
+      "learning_rate": 1.984290843062358e-05,
+      "loss": 0.1865,
+      "step": 1520
+    },
+    {
+      "epoch": 3.42,
+      "learning_rate": 1.9839701337956922e-05,
+      "loss": 0.1885,
+      "step": 1530
+    },
+    {
+      "epoch": 3.44,
+      "learning_rate": 1.9836462102612643e-05,
+      "loss": 0.2657,
+      "step": 1540
+    },
+    {
+      "epoch": 3.46,
+      "learning_rate": 1.9833190735172117e-05,
+      "loss": 0.1626,
+      "step": 1550
+    },
+    {
+      "epoch": 3.48,
+      "learning_rate": 1.982988724632171e-05,
+      "loss": 0.237,
+      "step": 1560
+    },
+    {
+      "epoch": 3.5,
+      "learning_rate": 1.9826551646852696e-05,
+      "loss": 0.2337,
+      "step": 1570
+    },
+    {
+      "epoch": 3.53,
+      "learning_rate": 1.982318394766124e-05,
+      "loss": 0.2103,
+      "step": 1580
+    },
+    {
+      "epoch": 3.55,
+      "learning_rate": 1.9819784159748394e-05,
+      "loss": 0.217,
+      "step": 1590
+    },
+    {
+      "epoch": 3.57,
+      "learning_rate": 1.9816352294219995e-05,
+      "loss": 0.1827,
+      "step": 1600
+    },
+    {
+      "epoch": 3.59,
+      "learning_rate": 1.981288836228669e-05,
+      "loss": 0.2204,
+      "step": 1610
+    },
+    {
+      "epoch": 3.62,
+      "learning_rate": 1.9809392375263865e-05,
+      "loss": 0.2232,
+      "step": 1620
+    },
+    {
+      "epoch": 3.64,
+      "learning_rate": 1.9805864344571625e-05,
+      "loss": 0.1535,
+      "step": 1630
+    },
+    {
+      "epoch": 3.66,
+      "learning_rate": 1.980230428173474e-05,
+      "loss": 0.1905,
+      "step": 1640
+    },
+    {
+      "epoch": 3.68,
+      "learning_rate": 1.979871219838263e-05,
+      "loss": 0.1729,
+      "step": 1650
+    },
+    {
+      "epoch": 3.71,
+      "learning_rate": 1.9795088106249306e-05,
+      "loss": 0.2792,
+      "step": 1660
+    },
+    {
+      "epoch": 3.73,
+      "learning_rate": 1.9791432017173338e-05,
+      "loss": 0.2183,
+      "step": 1670
+    },
+    {
+      "epoch": 3.75,
+      "learning_rate": 1.978774394309782e-05,
+      "loss": 0.1739,
+      "step": 1680
+    },
+    {
+      "epoch": 3.77,
+      "learning_rate": 1.9784023896070336e-05,
+      "loss": 0.1829,
+      "step": 1690
+    },
+    {
+      "epoch": 3.79,
+      "learning_rate": 1.9780271888242904e-05,
+      "loss": 0.1702,
+      "step": 1700
+    },
+    {
+      "epoch": 3.82,
+      "learning_rate": 1.9776487931871958e-05,
+      "loss": 0.2107,
+      "step": 1710
+    },
+    {
+      "epoch": 3.84,
+      "learning_rate": 1.9772672039318278e-05,
+      "loss": 0.2757,
+      "step": 1720
+    },
+    {
+      "epoch": 3.86,
+      "learning_rate": 1.976882422304699e-05,
+      "loss": 0.2264,
+      "step": 1730
+    },
+    {
+      "epoch": 3.88,
+      "learning_rate": 1.976494449562748e-05,
+      "loss": 0.1797,
+      "step": 1740
+    },
+    {
+      "epoch": 3.91,
+      "learning_rate": 1.9761032869733397e-05,
+      "loss": 0.1772,
+      "step": 1750
+    },
+    {
+      "epoch": 3.93,
+      "learning_rate": 1.9757089358142573e-05,
+      "loss": 0.294,
+      "step": 1760
+    },
+    {
+      "epoch": 3.95,
+      "learning_rate": 1.9753113973737016e-05,
+      "loss": 0.2024,
+      "step": 1770
+    },
+    {
+      "epoch": 3.97,
+      "learning_rate": 1.9749106729502833e-05,
+      "loss": 0.2291,
+      "step": 1780
+    },
+    {
+      "epoch": 4.0,
+      "learning_rate": 1.974506763853022e-05,
+      "loss": 0.1425,
+      "step": 1790
+    },
+    {
+      "epoch": 4.02,
+      "learning_rate": 1.9740996714013394e-05,
+      "loss": 0.1358,
+      "step": 1800
+    },
+    {
+      "epoch": 4.04,
+      "learning_rate": 1.9736893969250567e-05,
+      "loss": 0.143,
+      "step": 1810
+    },
+    {
+      "epoch": 4.06,
+      "learning_rate": 1.97327594176439e-05,
+      "loss": 0.1534,
+      "step": 1820
+    },
+    {
+      "epoch": 4.08,
+      "learning_rate": 1.972859307269944e-05,
+      "loss": 0.1792,
+      "step": 1830
+    },
+    {
+      "epoch": 4.11,
+      "learning_rate": 1.9724394948027102e-05,
+      "loss": 0.1508,
+      "step": 1840
+    },
+    {
+      "epoch": 4.13,
+      "learning_rate": 1.9720165057340616e-05,
+      "loss": 0.1316,
+      "step": 1850
+    },
+    {
+      "epoch": 4.15,
+      "learning_rate": 1.971590341445747e-05,
+      "loss": 0.1316,
+      "step": 1860
+    },
+    {
+      "epoch": 4.17,
+      "learning_rate": 1.9711610033298888e-05,
+      "loss": 0.1558,
+      "step": 1870
+    },
+    {
+      "epoch": 4.2,
+      "learning_rate": 1.9707284927889757e-05,
+      "loss": 0.136,
+      "step": 1880
+    },
+    {
+      "epoch": 4.22,
+      "learning_rate": 1.970292811235861e-05,
+      "loss": 0.1113,
+      "step": 1890
+    },
+    {
+      "epoch": 4.24,
+      "learning_rate": 1.969853960093755e-05,
+      "loss": 0.1663,
+      "step": 1900
+    },
+    {
+      "epoch": 4.26,
+      "learning_rate": 1.969411940796223e-05,
+      "loss": 0.1571,
+      "step": 1910
+    },
+    {
+      "epoch": 4.29,
+      "learning_rate": 1.9689667547871788e-05,
+      "loss": 0.1738,
+      "step": 1920
+    },
+    {
+      "epoch": 4.31,
+      "learning_rate": 1.9685184035208814e-05,
+      "loss": 0.1491,
+      "step": 1930
+    },
+    {
+      "epoch": 4.33,
+      "learning_rate": 1.968066888461929e-05,
+      "loss": 0.206,
+      "step": 1940
+    },
+    {
+      "epoch": 4.35,
+      "learning_rate": 1.9676122110852547e-05,
+      "loss": 0.1792,
+      "step": 1950
+    },
+    {
+      "epoch": 4.38,
+      "learning_rate": 1.9671543728761226e-05,
+      "loss": 0.1409,
+      "step": 1960
+    },
+    {
+      "epoch": 4.4,
+      "learning_rate": 1.9666933753301203e-05,
+      "loss": 0.1634,
+      "step": 1970
+    },
+    {
+      "epoch": 4.42,
+      "learning_rate": 1.9662292199531575e-05,
+      "loss": 0.1211,
+      "step": 1980
+    },
+    {
+      "epoch": 4.44,
+      "learning_rate": 1.9657619082614588e-05,
+      "loss": 0.2405,
+      "step": 1990
+    },
+    {
+      "epoch": 4.46,
+      "learning_rate": 1.965291441781559e-05,
+      "loss": 0.1525,
+      "step": 2000
+    },
+    {
+      "epoch": 4.49,
+      "learning_rate": 1.964817822050299e-05,
+      "loss": 0.1682,
+      "step": 2010
+    },
+    {
+      "epoch": 4.51,
+      "learning_rate": 1.9643410506148196e-05,
+      "loss": 0.1604,
+      "step": 2020
+    },
+    {
+      "epoch": 4.53,
+      "learning_rate": 1.9638611290325576e-05,
+      "loss": 0.1401,
+      "step": 2030
+    },
+    {
+      "epoch": 4.55,
+      "learning_rate": 1.96337805887124e-05,
+      "loss": 0.1205,
+      "step": 2040
+    },
+    {
+      "epoch": 4.58,
+      "learning_rate": 1.9628918417088785e-05,
+      "loss": 0.1611,
+      "step": 2050
+    },
+    {
+      "epoch": 4.6,
+      "learning_rate": 1.9624024791337657e-05,
+      "loss": 0.1259,
+      "step": 2060
+    },
+    {
+      "epoch": 4.62,
+      "learning_rate": 1.9619099727444686e-05,
+      "loss": 0.1404,
+      "step": 2070
+    },
+    {
+      "epoch": 4.64,
+      "learning_rate": 1.961414324149824e-05,
+      "loss": 0.1372,
+      "step": 2080
+    },
+    {
+      "epoch": 4.67,
+      "learning_rate": 1.9609155349689338e-05,
+      "loss": 0.1626,
+      "step": 2090
+    },
+    {
+      "epoch": 4.69,
+      "learning_rate": 1.9604136068311577e-05,
+      "loss": 0.1819,
+      "step": 2100
+    },
+    {
+      "epoch": 4.71,
+      "learning_rate": 1.95990854137611e-05,
+      "loss": 0.1552,
+      "step": 2110
+    },
+    {
+      "epoch": 4.73,
+      "learning_rate": 1.959400340253654e-05,
+      "loss": 0.1783,
+      "step": 2120
+    },
+    {
+      "epoch": 4.75,
+      "learning_rate": 1.9588890051238945e-05,
+      "loss": 0.147,
+      "step": 2130
+    },
+    {
+      "epoch": 4.78,
+      "learning_rate": 1.9583745376571756e-05,
+      "loss": 0.1625,
+      "step": 2140
+    },
+    {
+      "epoch": 4.8,
+      "learning_rate": 1.9578569395340727e-05,
+      "loss": 0.1435,
+      "step": 2150
+    },
+    {
+      "epoch": 4.82,
+      "learning_rate": 1.9573362124453884e-05,
+      "loss": 0.1255,
+      "step": 2160
+    },
+    {
+      "epoch": 4.84,
+      "learning_rate": 1.9568123580921453e-05,
+      "loss": 0.0939,
+      "step": 2170
+    },
+    {
+      "epoch": 4.87,
+      "learning_rate": 1.9562853781855836e-05,
+      "loss": 0.1851,
+      "step": 2180
+    },
+    {
+      "epoch": 4.89,
+      "learning_rate": 1.9557552744471517e-05,
+      "loss": 0.1553,
+      "step": 2190
+    },
+    {
+      "epoch": 4.91,
+      "learning_rate": 1.9552220486085033e-05,
+      "loss": 0.1919,
+      "step": 2200
+    },
+    {
+      "epoch": 4.93,
+      "learning_rate": 1.954685702411491e-05,
+      "loss": 0.1703,
+      "step": 2210
+    },
+    {
+      "epoch": 4.96,
+      "learning_rate": 1.9541462376081594e-05,
+      "loss": 0.1302,
+      "step": 2220
+    },
+    {
+      "epoch": 4.98,
+      "learning_rate": 1.953603655960742e-05,
+      "loss": 0.1564,
+      "step": 2230
+    },
+    {
+      "epoch": 5.0,
+      "learning_rate": 1.9530579592416522e-05,
+      "loss": 0.152,
+      "step": 2240
+    },
+    {
+      "epoch": 5.02,
+      "learning_rate": 1.9525091492334813e-05,
+      "loss": 0.1249,
+      "step": 2250
+    },
+    {
+      "epoch": 5.04,
+      "learning_rate": 1.951957227728988e-05,
+      "loss": 0.1187,
+      "step": 2260
+    },
+    {
+      "epoch": 5.07,
+      "learning_rate": 1.9514021965310972e-05,
+      "loss": 0.091,
+      "step": 2270
+    },
+    {
+      "epoch": 5.09,
+      "learning_rate": 1.950844057452891e-05,
+      "loss": 0.127,
+      "step": 2280
+    },
+    {
+      "epoch": 5.11,
+      "learning_rate": 1.9502828123176042e-05,
+      "loss": 0.11,
+      "step": 2290
+    },
+    {
+      "epoch": 5.13,
+      "learning_rate": 1.9497184629586176e-05,
+      "loss": 0.1098,
+      "step": 2300
+    },
+    {
+      "epoch": 5.16,
+      "learning_rate": 1.949151011219453e-05,
+      "loss": 0.1008,
+      "step": 2310
+    },
+    {
+      "epoch": 5.18,
+      "learning_rate": 1.9485804589537655e-05,
+      "loss": 0.1226,
+      "step": 2320
+    },
+    {
+      "epoch": 5.2,
+      "learning_rate": 1.9480068080253393e-05,
+      "loss": 0.1098,
+      "step": 2330
+    },
+    {
+      "epoch": 5.22,
+      "learning_rate": 1.9474300603080805e-05,
+      "loss": 0.1296,
+      "step": 2340
+    },
+    {
+      "epoch": 5.25,
+      "learning_rate": 1.9468502176860117e-05,
+      "loss": 0.0975,
+      "step": 2350
+    },
+    {
+      "epoch": 5.27,
+      "learning_rate": 1.9462672820532643e-05,
+      "loss": 0.1068,
+      "step": 2360
+    },
+    {
+      "epoch": 5.29,
+      "learning_rate": 1.9456812553140744e-05,
+      "loss": 0.1191,
+      "step": 2370
+    },
+    {
+      "epoch": 5.31,
+      "learning_rate": 1.945092139382776e-05,
+      "loss": 0.102,
+      "step": 2380
+    },
+    {
+      "epoch": 5.33,
+      "learning_rate": 1.944499936183793e-05,
+      "loss": 0.123,
+      "step": 2390
+    },
+    {
+      "epoch": 5.36,
+      "learning_rate": 1.9439046476516356e-05,
+      "loss": 0.1195,
+      "step": 2400
+    },
+    {
+      "epoch": 5.38,
+      "learning_rate": 1.9433062757308914e-05,
+      "loss": 0.1326,
+      "step": 2410
+    },
+    {
+      "epoch": 5.4,
+      "learning_rate": 1.9427048223762212e-05,
+      "loss": 0.1262,
+      "step": 2420
+    },
+    {
+      "epoch": 5.42,
+      "learning_rate": 1.9421002895523515e-05,
+      "loss": 0.0987,
+      "step": 2430
+    },
+    {
+      "epoch": 5.45,
+      "learning_rate": 1.941492679234068e-05,
+      "loss": 0.1205,
+      "step": 2440
+    },
+    {
+      "epoch": 5.47,
+      "learning_rate": 1.9408819934062098e-05,
+      "loss": 0.1237,
+      "step": 2450
+    },
+    {
+      "epoch": 5.49,
+      "learning_rate": 1.9402682340636625e-05,
+      "loss": 0.1588,
+      "step": 2460
+    },
+    {
+      "epoch": 5.51,
+      "learning_rate": 1.9396514032113514e-05,
+      "loss": 0.1303,
+      "step": 2470
+    },
+    {
+      "epoch": 5.54,
+      "learning_rate": 1.9390315028642355e-05,
+      "loss": 0.1057,
+      "step": 2480
+    },
+    {
+      "epoch": 5.56,
+      "learning_rate": 1.9384085350473016e-05,
+      "loss": 0.1137,
+      "step": 2490
+    },
+    {
+      "epoch": 5.58,
+      "learning_rate": 1.9377825017955548e-05,
+      "loss": 0.1101,
+      "step": 2500
+    },
+    {
+      "epoch": 5.6,
+      "learning_rate": 1.9371534051540158e-05,
+      "loss": 0.1159,
+      "step": 2510
+    },
+    {
+      "epoch": 5.62,
+      "learning_rate": 1.9365212471777113e-05,
+      "loss": 0.129,
+      "step": 2520
+    },
+    {
+      "epoch": 5.65,
+      "learning_rate": 1.935886029931668e-05,
+      "loss": 0.1154,
+      "step": 2530
+    },
+    {
+      "epoch": 5.67,
+      "learning_rate": 1.9352477554909067e-05,
+      "loss": 0.1108,
+      "step": 2540
+    },
+    {
+      "epoch": 5.69,
+      "learning_rate": 1.934606425940435e-05,
+      "loss": 0.1231,
+      "step": 2550
+    },
+    {
+      "epoch": 5.71,
+      "learning_rate": 1.93396204337524e-05,
+      "loss": 0.1221,
+      "step": 2560
+    },
+    {
+      "epoch": 5.74,
+      "learning_rate": 1.9333146099002826e-05,
+      "loss": 0.1374,
+      "step": 2570
+    },
+    {
+      "epoch": 5.76,
+      "learning_rate": 1.932664127630488e-05,
+      "loss": 0.1277,
+      "step": 2580
+    },
+    {
+      "epoch": 5.78,
+      "learning_rate": 1.9320105986907433e-05,
+      "loss": 0.1126,
+      "step": 2590
+    },
+    {
+      "epoch": 5.8,
+      "learning_rate": 1.931354025215886e-05,
+      "loss": 0.1074,
+      "step": 2600
+    },
+    {
+      "epoch": 5.83,
+      "learning_rate": 1.9306944093507e-05,
+      "loss": 0.1211,
+      "step": 2610
+    },
+    {
+      "epoch": 5.85,
+      "learning_rate": 1.930031753249907e-05,
+      "loss": 0.124,
+      "step": 2620
+    },
+    {
+      "epoch": 5.87,
+      "learning_rate": 1.9293660590781603e-05,
+      "loss": 0.122,
+      "step": 2630
+    },
+    {
+      "epoch": 5.89,
+      "learning_rate": 1.928697329010037e-05,
+      "loss": 0.1246,
+      "step": 2640
+    },
+    {
+      "epoch": 5.92,
+      "learning_rate": 1.9280255652300326e-05,
+      "loss": 0.1281,
+      "step": 2650
+    },
+    {
+      "epoch": 5.94,
+      "learning_rate": 1.9273507699325513e-05,
+      "loss": 0.1117,
+      "step": 2660
+    },
+    {
+      "epoch": 5.96,
+      "learning_rate": 1.9266729453219008e-05,
+      "loss": 0.1462,
+      "step": 2670
+    },
+    {
+      "epoch": 5.98,
+      "learning_rate": 1.925992093612284e-05,
+      "loss": 0.093,
+      "step": 2680
+    },
+    {
+      "epoch": 6.0,
+      "learning_rate": 1.925308217027792e-05,
+      "loss": 0.1543,
+      "step": 2690
+    },
+    {
+      "epoch": 6.03,
+      "learning_rate": 1.924621317802399e-05,
+      "loss": 0.0921,
+      "step": 2700
+    },
+    {
+      "epoch": 6.05,
+      "learning_rate": 1.9239313981799507e-05,
+      "loss": 0.1139,
+      "step": 2710
+    },
+    {
+      "epoch": 6.07,
+      "learning_rate": 1.92323846041416e-05,
+      "loss": 0.1115,
+      "step": 2720
+    },
+    {
+      "epoch": 6.09,
+      "learning_rate": 1.9225425067685995e-05,
+      "loss": 0.0748,
+      "step": 2730
+    },
+    {
+      "epoch": 6.12,
+      "learning_rate": 1.9218435395166933e-05,
+      "loss": 0.1428,
+      "step": 2740
+    },
+    {
+      "epoch": 6.14,
+      "learning_rate": 1.9211415609417097e-05,
+      "loss": 0.1,
+      "step": 2750
+    },
+    {
+      "epoch": 6.16,
+      "learning_rate": 1.920436573336754e-05,
+      "loss": 0.1159,
+      "step": 2760
+    },
+    {
+      "epoch": 6.18,
+      "learning_rate": 1.919728579004761e-05,
+      "loss": 0.1254,
+      "step": 2770
+    },
+    {
+      "epoch": 6.21,
+      "learning_rate": 1.919017580258487e-05,
+      "loss": 0.0951,
+      "step": 2780
+    },
+    {
+      "epoch": 6.23,
+      "learning_rate": 1.918303579420503e-05,
+      "loss": 0.0798,
+      "step": 2790
+    },
+    {
+      "epoch": 6.25,
+      "learning_rate": 1.917586578823186e-05,
+      "loss": 0.0978,
+      "step": 2800
+    },
+    {
+      "epoch": 6.27,
+      "learning_rate": 1.916866580808714e-05,
+      "loss": 0.1121,
+      "step": 2810
+    },
+    {
+      "epoch": 6.29,
+      "learning_rate": 1.9161435877290538e-05,
+      "loss": 0.0811,
+      "step": 2820
+    },
+    {
+      "epoch": 6.32,
+      "learning_rate": 1.915417601945958e-05,
+      "loss": 0.0841,
+      "step": 2830
+    },
+    {
+      "epoch": 6.34,
+      "learning_rate": 1.9146886258309548e-05,
+      "loss": 0.1058,
+      "step": 2840
+    },
+    {
+      "epoch": 6.36,
+      "learning_rate": 1.9139566617653395e-05,
+      "loss": 0.1103,
+      "step": 2850
+    },
+    {
+      "epoch": 6.38,
+      "learning_rate": 1.9132217121401698e-05,
+      "loss": 0.1101,
+      "step": 2860
+    },
+    {
+      "epoch": 6.41,
+      "learning_rate": 1.912483779356255e-05,
+      "loss": 0.0994,
+      "step": 2870
+    },
+    {
+      "epoch": 6.43,
+      "learning_rate": 1.9117428658241498e-05,
+      "loss": 0.1068,
+      "step": 2880
+    },
+    {
+      "epoch": 6.45,
+      "learning_rate": 1.9109989739641446e-05,
+      "loss": 0.0887,
+      "step": 2890
+    },
+    {
+      "epoch": 6.47,
+      "learning_rate": 1.9102521062062615e-05,
+      "loss": 0.0963,
+      "step": 2900
+    },
+    {
+      "epoch": 6.5,
+      "learning_rate": 1.909502264990241e-05,
+      "loss": 0.1173,
+      "step": 2910
+    },
+    {
+      "epoch": 6.52,
+      "learning_rate": 1.9087494527655383e-05,
+      "loss": 0.1129,
+      "step": 2920
+    },
+    {
+      "epoch": 6.54,
+      "learning_rate": 1.9079936719913138e-05,
+      "loss": 0.1068,
+      "step": 2930
+    },
+    {
+      "epoch": 6.56,
+      "learning_rate": 1.9072349251364238e-05,
+      "loss": 0.0932,
+      "step": 2940
+    },
+    {
+      "epoch": 6.58,
+      "learning_rate": 1.906473214679416e-05,
+      "loss": 0.0824,
+      "step": 2950
+    },
+    {
+      "epoch": 6.61,
+      "learning_rate": 1.9057085431085163e-05,
+      "loss": 0.0976,
+      "step": 2960
+    },
+    {
+      "epoch": 6.63,
+      "learning_rate": 1.904940912921626e-05,
+      "loss": 0.1071,
+      "step": 2970
+    },
+    {
+      "epoch": 6.65,
+      "learning_rate": 1.9041703266263095e-05,
+      "loss": 0.1087,
+      "step": 2980
+    },
+    {
+      "epoch": 6.67,
+      "learning_rate": 1.9033967867397883e-05,
+      "loss": 0.1006,
+      "step": 2990
+    },
+    {
+      "epoch": 6.7,
+      "learning_rate": 1.902620295788932e-05,
+      "loss": 0.1139,
+      "step": 3000
+    },
+    {
+      "epoch": 6.72,
+      "learning_rate": 1.9018408563102505e-05,
+      "loss": 0.1065,
+      "step": 3010
+    },
+    {
+      "epoch": 6.74,
+      "learning_rate": 1.901058470849885e-05,
+      "loss": 0.1183,
+      "step": 3020
+    },
+    {
+      "epoch": 6.76,
+      "learning_rate": 1.900273141963601e-05,
+      "loss": 0.1106,
+      "step": 3030
+    },
+    {
+      "epoch": 6.79,
+      "learning_rate": 1.899484872216778e-05,
+      "loss": 0.1263,
+      "step": 3040
+    },
+    {
+      "epoch": 6.81,
+      "learning_rate": 1.8986936641844025e-05,
+      "loss": 0.0902,
+      "step": 3050
+    },
+    {
+      "epoch": 6.83,
+      "learning_rate": 1.8978995204510605e-05,
+      "loss": 0.1197,
+      "step": 3060
+    },
+    {
+      "epoch": 6.85,
+      "learning_rate": 1.897102443610926e-05,
+      "loss": 0.1006,
+      "step": 3070
+    },
+    {
+      "epoch": 6.88,
+      "learning_rate": 1.8963024362677557e-05,
+      "loss": 0.1146,
+      "step": 3080
+    },
+    {
+      "epoch": 6.9,
+      "learning_rate": 1.895499501034878e-05,
+      "loss": 0.1128,
+      "step": 3090
+    },
+    {
+      "epoch": 6.92,
+      "learning_rate": 1.8946936405351877e-05,
+      "loss": 0.0985,
+      "step": 3100
+    },
+    {
+      "epoch": 6.94,
+      "learning_rate": 1.893884857401133e-05,
+      "loss": 0.091,
+      "step": 3110
+    },
+    {
+      "epoch": 6.96,
+      "learning_rate": 1.8930731542747108e-05,
+      "loss": 0.0872,
+      "step": 3120
+    },
+    {
+      "epoch": 6.99,
+      "learning_rate": 1.8922585338074556e-05,
+      "loss": 0.0998,
+      "step": 3130
+    },
+    {
+      "epoch": 7.01,
+      "learning_rate": 1.8914409986604327e-05,
+      "loss": 0.0923,
+      "step": 3140
+    },
+    {
+      "epoch": 7.03,
+      "learning_rate": 1.8906205515042272e-05,
+      "loss": 0.071,
+      "step": 3150
+    },
+    {
+      "epoch": 7.05,
+      "learning_rate": 1.8897971950189385e-05,
+      "loss": 0.0893,
+      "step": 3160
+    },
+    {
+      "epoch": 7.08,
+      "learning_rate": 1.888970931894169e-05,
+      "loss": 0.0929,
+      "step": 3170
+    },
+    {
+      "epoch": 7.1,
+      "learning_rate": 1.888141764829015e-05,
+      "loss": 0.0892,
+      "step": 3180
+    },
+    {
+      "epoch": 7.12,
+      "learning_rate": 1.8873096965320597e-05,
+      "loss": 0.0739,
+      "step": 3190
+    },
+    {
+      "epoch": 7.14,
+      "learning_rate": 1.886474729721364e-05,
+      "loss": 0.0783,
+      "step": 3200
+    },
+    {
+      "epoch": 7.17,
+      "learning_rate": 1.8856368671244565e-05,
+      "loss": 0.0783,
+      "step": 3210
+    },
+    {
+      "epoch": 7.19,
+      "learning_rate": 1.8847961114783254e-05,
+      "loss": 0.0847,
+      "step": 3220
+    },
+    {
+      "epoch": 7.21,
+      "learning_rate": 1.88395246552941e-05,
+      "loss": 0.0886,
+      "step": 3230
+    },
+    {
+      "epoch": 7.23,
+      "learning_rate": 1.8831059320335902e-05,
+      "loss": 0.088,
+      "step": 3240
+    },
+    {
+      "epoch": 7.25,
+      "learning_rate": 1.882256513756179e-05,
+      "loss": 0.082,
+      "step": 3250
+    },
+    {
+      "epoch": 7.28,
+      "learning_rate": 1.881404213471913e-05,
+      "loss": 0.0779,
+      "step": 3260
+    },
+    {
+      "epoch": 7.3,
+      "learning_rate": 1.8805490339649428e-05,
+      "loss": 0.0926,
+      "step": 3270
+    },
+    {
+      "epoch": 7.32,
+      "learning_rate": 1.879690978028825e-05,
+      "loss": 0.1093,
+      "step": 3280
+    },
+    {
+      "epoch": 7.34,
+      "learning_rate": 1.8788300484665118e-05,
+      "loss": 0.1051,
+      "step": 3290
+    },
+    {
+      "epoch": 7.37,
+      "learning_rate": 1.877966248090343e-05,
+      "loss": 0.0914,
+      "step": 3300
+    },
+    {
+      "epoch": 7.39,
+      "learning_rate": 1.8770995797220356e-05,
+      "loss": 0.0837,
+      "step": 3310
+    },
+    {
+      "epoch": 7.41,
+      "learning_rate": 1.8762300461926766e-05,
+      "loss": 0.0982,
+      "step": 3320
+    },
+    {
+      "epoch": 7.43,
+      "learning_rate": 1.8753576503427107e-05,
+      "loss": 0.0883,
+      "step": 3330
+    },
+    {
+      "epoch": 7.46,
+      "learning_rate": 1.874482395021934e-05,
+      "loss": 0.0949,
+      "step": 3340
+    },
+    {
+      "epoch": 7.48,
+      "learning_rate": 1.8736042830894828e-05,
+      "loss": 0.0872,
+      "step": 3350
+    },
+    {
+      "epoch": 7.5,
+      "learning_rate": 1.8727233174138254e-05,
+      "loss": 0.0851,
+      "step": 3360
+    },
+    {
+      "epoch": 7.52,
+      "learning_rate": 1.871839500872752e-05,
+      "loss": 0.1099,
+      "step": 3370
+    },
+    {
+      "epoch": 7.54,
+      "learning_rate": 1.8709528363533653e-05,
+      "loss": 0.0928,
+      "step": 3380
+    },
+    {
+      "epoch": 7.57,
+      "learning_rate": 1.8700633267520715e-05,
+      "loss": 0.0986,
+      "step": 3390
+    },
+    {
+      "epoch": 7.59,
+      "learning_rate": 1.8691709749745705e-05,
+      "loss": 0.067,
+      "step": 3400
+    },
+    {
+      "epoch": 7.61,
+      "learning_rate": 1.8682757839358472e-05,
+      "loss": 0.0957,
+      "step": 3410
+    },
+    {
+      "epoch": 7.63,
+      "learning_rate": 1.86737775656016e-05,
+      "loss": 0.0998,
+      "step": 3420
+    },
+    {
+      "epoch": 7.66,
+      "learning_rate": 1.866476895781034e-05,
+      "loss": 0.0705,
+      "step": 3430
+    },
+    {
+      "epoch": 7.68,
+      "learning_rate": 1.8655732045412488e-05,
+      "loss": 0.1001,
+      "step": 3440
+    },
+    {
+      "epoch": 7.7,
+      "learning_rate": 1.8646666857928314e-05,
+      "loss": 0.0987,
+      "step": 3450
+    },
+    {
+      "epoch": 7.72,
+      "learning_rate": 1.8637573424970435e-05,
+      "loss": 0.0999,
+      "step": 3460
+    },
+    {
+      "epoch": 7.75,
+      "learning_rate": 1.862845177624375e-05,
+      "loss": 0.0924,
+      "step": 3470
+    },
+    {
+      "epoch": 7.77,
+      "learning_rate": 1.8619301941545323e-05,
+      "loss": 0.0821,
+      "step": 3480
+    },
+    {
+      "epoch": 7.79,
+      "learning_rate": 1.8610123950764288e-05,
+      "loss": 0.0928,
+      "step": 3490
+    },
+    {
+      "epoch": 7.81,
+      "learning_rate": 1.8600917833881765e-05,
+      "loss": 0.0997,
+      "step": 3500
+    },
+    {
+      "epoch": 7.83,
+      "learning_rate": 1.8591683620970737e-05,
+      "loss": 0.0995,
+      "step": 3510
+    },
+    {
+      "epoch": 7.86,
+      "learning_rate": 1.858242134219598e-05,
+      "loss": 0.1074,
+      "step": 3520
+    },
+    {
+      "epoch": 7.88,
+      "learning_rate": 1.8573131027813945e-05,
+      "loss": 0.0789,
+      "step": 3530
+    },
+    {
+      "epoch": 7.9,
+      "learning_rate": 1.856381270817266e-05,
+      "loss": 0.1031,
+      "step": 3540
+    },
+    {
+      "epoch": 7.92,
+      "learning_rate": 1.8554466413711644e-05,
+      "loss": 0.0861,
+      "step": 3550
+    },
+    {
+      "epoch": 7.95,
+      "learning_rate": 1.8545092174961795e-05,
+      "loss": 0.1113,
+      "step": 3560
+    },
+    {
+      "epoch": 7.97,
+      "learning_rate": 1.85356900225453e-05,
+      "loss": 0.0801,
+      "step": 3570
+    },
+    {
+      "epoch": 7.99,
+      "learning_rate": 1.852625998717552e-05,
+      "loss": 0.1016,
+      "step": 3580
+    },
+    {
+      "epoch": 8.01,
+      "learning_rate": 1.8516802099656907e-05,
+      "loss": 0.0901,
+      "step": 3590
+    },
+    {
+      "epoch": 8.04,
+      "learning_rate": 1.8507316390884894e-05,
+      "loss": 0.0762,
+      "step": 3600
+    },
+    {
+      "epoch": 8.06,
+      "learning_rate": 1.84978028918458e-05,
+      "loss": 0.0881,
+      "step": 3610
+    },
+    {
+      "epoch": 8.08,
+      "learning_rate": 1.848826163361671e-05,
+      "loss": 0.0787,
+      "step": 3620
+    },
+    {
+      "epoch": 8.1,
+      "learning_rate": 1.8478692647365402e-05,
+      "loss": 0.0877,
+      "step": 3630
+    },
+    {
+      "epoch": 8.12,
+      "learning_rate": 1.846909596435023e-05,
+      "loss": 0.0935,
+      "step": 3640
+    },
+    {
+      "epoch": 8.15,
+      "learning_rate": 1.845947161592002e-05,
+      "loss": 0.0776,
+      "step": 3650
+    },
+    {
+      "epoch": 8.17,
+      "learning_rate": 1.844981963351397e-05,
+      "loss": 0.0814,
+      "step": 3660
+    },
+    {
+      "epoch": 8.19,
+      "learning_rate": 1.8440140048661547e-05,
+      "loss": 0.0736,
+      "step": 3670
+    },
+    {
+      "epoch": 8.21,
+      "learning_rate": 1.843043289298239e-05,
+      "loss": 0.0783,
+      "step": 3680
+    },
+    {
+      "epoch": 8.24,
+      "learning_rate": 1.8420698198186197e-05,
+      "loss": 0.0766,
+      "step": 3690
+    },
+    {
+      "epoch": 8.26,
+      "learning_rate": 1.841093599607263e-05,
+      "loss": 0.0738,
+      "step": 3700
+    },
+    {
+      "epoch": 8.28,
+      "learning_rate": 1.8401146318531204e-05,
+      "loss": 0.0875,
+      "step": 3710
+    },
+    {
+      "epoch": 8.3,
+      "learning_rate": 1.8391329197541186e-05,
+      "loss": 0.0804,
+      "step": 3720
+    },
+    {
+      "epoch": 8.33,
+      "learning_rate": 1.838148466517149e-05,
+      "loss": 0.0701,
+      "step": 3730
+    },
+    {
+      "epoch": 8.35,
+      "learning_rate": 1.8371612753580583e-05,
+      "loss": 0.0844,
+      "step": 3740
+    },
+    {
+      "epoch": 8.37,
+      "learning_rate": 1.8361713495016354e-05,
+      "loss": 0.0812,
+      "step": 3750
+    },
+    {
+      "epoch": 8.39,
+      "learning_rate": 1.8351786921816037e-05,
+      "loss": 0.0979,
+      "step": 3760
+    },
+    {
+      "epoch": 8.42,
+      "learning_rate": 1.8341833066406083e-05,
+      "loss": 0.0721,
+      "step": 3770
+    },
+    {
+      "epoch": 8.44,
+      "learning_rate": 1.8331851961302075e-05,
+      "loss": 0.0816,
+      "step": 3780
+    },
+    {
+      "epoch": 8.46,
+      "learning_rate": 1.83218436391086e-05,
+      "loss": 0.0714,
+      "step": 3790
+    },
+    {
+      "epoch": 8.48,
+      "learning_rate": 1.8311808132519157e-05,
+      "loss": 0.0623,
+      "step": 3800
+    },
+    {
+      "epoch": 8.5,
+      "learning_rate": 1.830174547431605e-05,
+      "loss": 0.0618,
+      "step": 3810
+    },
+    {
+      "epoch": 8.53,
+      "learning_rate": 1.8291655697370276e-05,
+      "loss": 0.0842,
+      "step": 3820
+    },
+    {
+      "epoch": 8.55,
+      "learning_rate": 1.8281538834641416e-05,
+      "loss": 0.0792,
+      "step": 3830
+    },
+    {
+      "epoch": 8.57,
+      "learning_rate": 1.8271394919177528e-05,
+      "loss": 0.0794,
+      "step": 3840
+    },
+    {
+      "epoch": 8.59,
+      "learning_rate": 1.8261223984115052e-05,
+      "loss": 0.0866,
+      "step": 3850
+    },
+    {
+      "epoch": 8.62,
+      "learning_rate": 1.8251026062678673e-05,
+      "loss": 0.0755,
+      "step": 3860
+    },
+    {
+      "epoch": 8.64,
+      "learning_rate": 1.8240801188181257e-05,
+      "loss": 0.0825,
+      "step": 3870
+    },
+    {
+      "epoch": 8.66,
+      "learning_rate": 1.823054939402369e-05,
+      "loss": 0.0842,
+      "step": 3880
+    },
+    {
+      "epoch": 8.68,
+      "learning_rate": 1.8220270713694803e-05,
+      "loss": 0.0776,
+      "step": 3890
+    },
+    {
+      "epoch": 8.71,
+      "learning_rate": 1.8209965180771262e-05,
+      "loss": 0.097,
+      "step": 3900
+    },
+    {
+      "epoch": 8.73,
+      "learning_rate": 1.8199632828917445e-05,
+      "loss": 0.094,
+      "step": 3910
+    },
+    {
+      "epoch": 8.75,
+      "learning_rate": 1.8189273691885336e-05,
+      "loss": 0.0982,
+      "step": 3920
+    },
+    {
+      "epoch": 8.77,
+      "learning_rate": 1.8178887803514415e-05,
+      "loss": 0.0863,
+      "step": 3930
+    },
+    {
+      "epoch": 8.79,
+      "learning_rate": 1.8168475197731553e-05,
+      "loss": 0.0702,
+      "step": 3940
+    },
+    {
+      "epoch": 8.82,
+      "learning_rate": 1.81580359085509e-05,
+      "loss": 0.0982,
+      "step": 3950
+    },
+    {
+      "epoch": 8.84,
+      "learning_rate": 1.814756997007376e-05,
+      "loss": 0.0843,
+      "step": 3960
+    },
+    {
+      "epoch": 8.86,
+      "learning_rate": 1.8137077416488496e-05,
+      "loss": 0.0759,
+      "step": 3970
+    },
+    {
+      "epoch": 8.88,
+      "learning_rate": 1.8126558282070417e-05,
+      "loss": 0.0956,
+      "step": 3980
+    },
+    {
+      "epoch": 8.91,
+      "learning_rate": 1.8116012601181655e-05,
+      "loss": 0.0852,
+      "step": 3990
+    },
+    {
+      "epoch": 8.93,
+      "learning_rate": 1.810544040827107e-05,
+      "loss": 0.0903,
+      "step": 4000
+    },
+    {
+      "epoch": 8.95,
+      "learning_rate": 1.8094841737874108e-05,
+      "loss": 0.0797,
+      "step": 4010
+    },
+    {
+      "epoch": 8.97,
+      "learning_rate": 1.8084216624612726e-05,
+      "loss": 0.0898,
+      "step": 4020
+    },
+    {
+      "epoch": 9.0,
+      "learning_rate": 1.8073565103195254e-05,
+      "loss": 0.079,
+      "step": 4030
+    },
+    {
+      "epoch": 9.02,
+      "learning_rate": 1.8062887208416282e-05,
+      "loss": 0.1005,
+      "step": 4040
+    },
+    {
+      "epoch": 9.04,
+      "learning_rate": 1.8052182975156557e-05,
+      "loss": 0.0708,
+      "step": 4050
+    },
+    {
+      "epoch": 9.06,
+      "learning_rate": 1.8041452438382873e-05,
+      "loss": 0.0667,
+      "step": 4060
+    },
+    {
+      "epoch": 9.08,
+      "learning_rate": 1.8030695633147926e-05,
+      "loss": 0.06,
+      "step": 4070
+    },
+    {
+      "epoch": 9.11,
+      "learning_rate": 1.801991259459024e-05,
+      "loss": 0.074,
+      "step": 4080
+    },
+    {
+      "epoch": 9.13,
+      "learning_rate": 1.8009103357934024e-05,
+      "loss": 0.0723,
+      "step": 4090
+    },
+    {
+      "epoch": 9.15,
+      "learning_rate": 1.7998267958489076e-05,
+      "loss": 0.0722,
+      "step": 4100
+    },
+    {
+      "epoch": 9.17,
+      "learning_rate": 1.7987406431650653e-05,
+      "loss": 0.0655,
+      "step": 4110
+    },
+    {
+      "epoch": 9.2,
+      "learning_rate": 1.797651881289935e-05,
+      "loss": 0.0819,
+      "step": 4120
+    },
+    {
+      "epoch": 9.22,
+      "learning_rate": 1.7965605137801015e-05,
+      "loss": 0.0928,
+      "step": 4130
+    },
+    {
+      "epoch": 9.24,
+      "learning_rate": 1.79546654420066e-05,
+      "loss": 0.0822,
+      "step": 4140
+    },
+    {
+      "epoch": 9.26,
+      "learning_rate": 1.7943699761252057e-05,
+      "loss": 0.0908,
+      "step": 4150
+    },
+    {
+      "epoch": 9.29,
+      "learning_rate": 1.7932708131358222e-05,
+      "loss": 0.0813,
+      "step": 4160
+    },
+    {
+      "epoch": 9.31,
+      "learning_rate": 1.7921690588230698e-05,
+      "loss": 0.0859,
+      "step": 4170
+    },
+    {
+      "epoch": 9.33,
+      "learning_rate": 1.7910647167859744e-05,
+      "loss": 0.0794,
+      "step": 4180
+    },
+    {
+      "epoch": 9.35,
+      "learning_rate": 1.7899577906320135e-05,
+      "loss": 0.0855,
+      "step": 4190
+    },
+    {
+      "epoch": 9.38,
+      "learning_rate": 1.7888482839771074e-05,
+      "loss": 0.0984,
+      "step": 4200
+    },
+    {
+      "epoch": 9.4,
+      "learning_rate": 1.787736200445606e-05,
+      "loss": 0.0797,
+      "step": 4210
+    },
+    {
+      "epoch": 9.42,
+      "learning_rate": 1.786621543670275e-05,
+      "loss": 0.0772,
+      "step": 4220
+    },
+    {
+      "epoch": 9.44,
+      "learning_rate": 1.7855043172922883e-05,
+      "loss": 0.0666,
+      "step": 4230
+    },
+    {
+      "epoch": 9.46,
+      "learning_rate": 1.7843845249612122e-05,
+      "loss": 0.0728,
+      "step": 4240
+    },
+    {
+      "epoch": 9.49,
+      "learning_rate": 1.7832621703349956e-05,
+      "loss": 0.0825,
+      "step": 4250
+    },
+    {
+      "epoch": 9.51,
+      "learning_rate": 1.7821372570799574e-05,
+      "loss": 0.0698,
+      "step": 4260
+    },
+    {
+      "epoch": 9.53,
+      "learning_rate": 1.781009788870775e-05,
+      "loss": 0.074,
+      "step": 4270
+    },
+    {
+      "epoch": 9.55,
+      "learning_rate": 1.779879769390471e-05,
+      "loss": 0.0862,
+      "step": 4280
+    },
+    {
+      "epoch": 9.58,
+      "learning_rate": 1.7787472023304023e-05,
+      "loss": 0.0768,
+      "step": 4290
+    },
+    {
+      "epoch": 9.6,
+      "learning_rate": 1.7776120913902487e-05,
+      "loss": 0.0814,
+      "step": 4300
+    },
+    {
+      "epoch": 9.62,
+      "learning_rate": 1.7764744402779992e-05,
+      "loss": 0.0829,
+      "step": 4310
+    },
+    {
+      "epoch": 9.64,
+      "learning_rate": 1.77533425270994e-05,
+      "loss": 0.0844,
+      "step": 4320
+    },
+    {
+      "epoch": 9.67,
+      "learning_rate": 1.7741915324106445e-05,
+      "loss": 0.0803,
+      "step": 4330
+    },
+    {
+      "epoch": 9.69,
+      "learning_rate": 1.7730462831129584e-05,
+      "loss": 0.0819,
+      "step": 4340
+    },
+    {
+      "epoch": 9.71,
+      "learning_rate": 1.771898508557989e-05,
+      "loss": 0.0692,
+      "step": 4350
+    },
+    {
+      "epoch": 9.73,
+      "learning_rate": 1.7707482124950923e-05,
+      "loss": 0.0777,
+      "step": 4360
+    },
+    {
+      "epoch": 9.75,
+      "learning_rate": 1.7695953986818625e-05,
+      "loss": 0.089,
+      "step": 4370
+    },
+    {
+      "epoch": 9.78,
+      "learning_rate": 1.7684400708841165e-05,
+      "loss": 0.092,
+      "step": 4380
+    },
+    {
+      "epoch": 9.8,
+      "learning_rate": 1.7672822328758852e-05,
+      "loss": 0.1097,
+      "step": 4390
+    },
+    {
+      "epoch": 9.82,
+      "learning_rate": 1.7661218884393977e-05,
+      "loss": 0.0755,
+      "step": 4400
+    },
+    {
+      "epoch": 9.84,
+      "learning_rate": 1.764959041365073e-05,
+      "loss": 0.0787,
+      "step": 4410
+    },
+    {
+      "epoch": 9.87,
+      "learning_rate": 1.7637936954515026e-05,
+      "loss": 0.0708,
+      "step": 4420
+    },
+    {
+      "epoch": 9.89,
+      "learning_rate": 1.7626258545054425e-05,
+      "loss": 0.0778,
+      "step": 4430
+    },
+    {
+      "epoch": 9.91,
+      "learning_rate": 1.7614555223417987e-05,
+      "loss": 0.0725,
+      "step": 4440
+    },
+    {
+      "epoch": 9.93,
+      "learning_rate": 1.7602827027836153e-05,
+      "loss": 0.077,
+      "step": 4450
+    },
+    {
+      "epoch": 9.96,
+      "learning_rate": 1.7591073996620607e-05,
+      "loss": 0.0809,
+      "step": 4460
+    },
+    {
+      "epoch": 9.98,
+      "learning_rate": 1.757929616816418e-05,
+      "loss": 0.0978,
+      "step": 4470
+    },
+    {
+      "epoch": 10.0,
+      "learning_rate": 1.756749358094069e-05,
+      "loss": 0.0759,
+      "step": 4480
+    },
+    {
+      "epoch": 10.02,
+      "learning_rate": 1.755566627350484e-05,
+      "loss": 0.0678,
+      "step": 4490
+    },
+    {
+      "epoch": 10.04,
+      "learning_rate": 1.754381428449209e-05,
+      "loss": 0.0706,
+      "step": 4500
+    },
+    {
+      "epoch": 10.07,
+      "learning_rate": 1.7531937652618515e-05,
+      "loss": 0.0665,
+      "step": 4510
+    },
+    {
+      "epoch": 10.09,
+      "learning_rate": 1.7520036416680687e-05,
+      "loss": 0.0575,
+      "step": 4520
+    },
+    {
+      "epoch": 10.11,
+      "learning_rate": 1.7508110615555573e-05,
+      "loss": 0.0684,
+      "step": 4530
+    },
+    {
+      "epoch": 10.13,
+      "learning_rate": 1.749616028820036e-05,
+      "loss": 0.0869,
+      "step": 4540
+    },
+    {
+      "epoch": 10.16,
+      "learning_rate": 1.748418547365236e-05,
+      "loss": 0.0648,
+      "step": 4550
+    },
+    {
+      "epoch": 10.18,
+      "learning_rate": 1.7472186211028884e-05,
+      "loss": 0.0673,
+      "step": 4560
+    },
+    {
+      "epoch": 10.2,
+      "learning_rate": 1.7460162539527104e-05,
+      "loss": 0.0677,
+      "step": 4570
+    },
+    {
+      "epoch": 10.22,
+      "learning_rate": 1.7448114498423915e-05,
+      "loss": 0.0687,
+      "step": 4580
+    },
+    {
+      "epoch": 10.25,
+      "learning_rate": 1.743604212707583e-05,
+      "loss": 0.0701,
+      "step": 4590
+    },
+    {
+      "epoch": 10.27,
+      "learning_rate": 1.7423945464918835e-05,
+      "loss": 0.0645,
+      "step": 4600
+    },
+    {
+      "epoch": 10.29,
+      "learning_rate": 1.741182455146827e-05,
+      "loss": 0.084,
+      "step": 4610
+    },
+    {
+      "epoch": 10.31,
+      "learning_rate": 1.739967942631869e-05,
+      "loss": 0.0649,
+      "step": 4620
+    },
+    {
+      "epoch": 10.33,
+      "learning_rate": 1.738751012914375e-05,
+      "loss": 0.0762,
+      "step": 4630
+    },
+    {
+      "epoch": 10.36,
+      "learning_rate": 1.7375316699696042e-05,
+      "loss": 0.0827,
+      "step": 4640
+    },
+    {
+      "epoch": 10.38,
+      "learning_rate": 1.736309917780702e-05,
+      "loss": 0.0691,
+      "step": 4650
+    },
+    {
+      "epoch": 10.4,
+      "learning_rate": 1.7350857603386816e-05,
+      "loss": 0.0816,
+      "step": 4660
+    },
+    {
+      "epoch": 10.42,
+      "learning_rate": 1.733859201642415e-05,
+      "loss": 0.0832,
+      "step": 4670
+    },
+    {
+      "epoch": 10.45,
+      "learning_rate": 1.732630245698617e-05,
+      "loss": 0.0778,
+      "step": 4680
+    },
+    {
+      "epoch": 10.47,
+      "learning_rate": 1.7313988965218337e-05,
+      "loss": 0.0765,
+      "step": 4690
+    },
+    {
+      "epoch": 10.49,
+      "learning_rate": 1.730165158134429e-05,
+      "loss": 0.0632,
+      "step": 4700
+    },
+    {
+      "epoch": 10.51,
+      "learning_rate": 1.7289290345665713e-05,
+      "loss": 0.0822,
+      "step": 4710
+    },
+    {
+      "epoch": 10.54,
+      "learning_rate": 1.7276905298562208e-05,
+      "loss": 0.0615,
+      "step": 4720
+    },
+    {
+      "epoch": 10.56,
+      "learning_rate": 1.7264496480491165e-05,
+      "loss": 0.0701,
+      "step": 4730
+    },
+    {
+      "epoch": 10.58,
+      "learning_rate": 1.7252063931987607e-05,
+      "loss": 0.0682,
+      "step": 4740
+    },
+    {
+      "epoch": 10.6,
+      "learning_rate": 1.7239607693664103e-05,
+      "loss": 0.0991,
+      "step": 4750
+    },
+    {
+      "epoch": 10.62,
+      "learning_rate": 1.7227127806210578e-05,
+      "loss": 0.0764,
+      "step": 4760
+    },
+    {
+      "epoch": 10.65,
+      "learning_rate": 1.7214624310394236e-05,
+      "loss": 0.0706,
+      "step": 4770
+    },
+    {
+      "epoch": 10.67,
+      "learning_rate": 1.7202097247059383e-05,
+      "loss": 0.0863,
+      "step": 4780
+    },
+    {
+      "epoch": 10.69,
+      "learning_rate": 1.7189546657127315e-05,
+      "loss": 0.0813,
+      "step": 4790
+    },
+    {
+      "epoch": 10.71,
+      "learning_rate": 1.717697258159619e-05,
+      "loss": 0.0847,
+      "step": 4800
+    },
+    {
+      "epoch": 10.74,
+      "learning_rate": 1.7164375061540877e-05,
+      "loss": 0.0805,
+      "step": 4810
+    },
+    {
+      "epoch": 10.76,
+      "learning_rate": 1.715175413811283e-05,
+      "loss": 0.0877,
+      "step": 4820
+    },
+    {
+      "epoch": 10.78,
+      "learning_rate": 1.7139109852539954e-05,
+      "loss": 0.0713,
+      "step": 4830
+    },
+    {
+      "epoch": 10.8,
+      "learning_rate": 1.712644224612647e-05,
+      "loss": 0.0776,
+      "step": 4840
+    },
+    {
+      "epoch": 10.83,
+      "learning_rate": 1.7113751360252777e-05,
+      "loss": 0.0685,
+      "step": 4850
+    },
+    {
+      "epoch": 10.85,
+      "learning_rate": 1.7101037236375324e-05,
+      "loss": 0.0825,
+      "step": 4860
+    },
+    {
+      "epoch": 10.87,
+      "learning_rate": 1.708829991602647e-05,
+      "loss": 0.0747,
+      "step": 4870
+    },
+    {
+      "epoch": 10.89,
+      "learning_rate": 1.707553944081434e-05,
+      "loss": 0.086,
+      "step": 4880
+    },
+    {
+      "epoch": 10.92,
+      "learning_rate": 1.7062755852422705e-05,
+      "loss": 0.0729,
+      "step": 4890
+    },
+    {
+      "epoch": 10.94,
+      "learning_rate": 1.7049949192610845e-05,
+      "loss": 0.0688,
+      "step": 4900
+    },
+    {
+      "epoch": 10.96,
+      "learning_rate": 1.7037119503213385e-05,
+      "loss": 0.0755,
+      "step": 4910
+    },
+    {
+      "epoch": 10.98,
+      "learning_rate": 1.7024266826140194e-05,
+      "loss": 0.0877,
+      "step": 4920
+    },
+    {
+      "epoch": 11.0,
+      "learning_rate": 1.701139120337624e-05,
+      "loss": 0.0785,
+      "step": 4930
+    },
+    {
+      "epoch": 11.03,
+      "learning_rate": 1.699849267698143e-05,
+      "loss": 0.0587,
+      "step": 4940
+    },
+    {
+      "epoch": 11.05,
+      "learning_rate": 1.698557128909049e-05,
+      "loss": 0.0584,
+      "step": 4950
+    },
+    {
+      "epoch": 11.07,
+      "learning_rate": 1.6972627081912848e-05,
+      "loss": 0.0641,
+      "step": 4960
+    },
+    {
+      "epoch": 11.09,
+      "learning_rate": 1.695966009773244e-05,
+      "loss": 0.0678,
+      "step": 4970
+    },
+    {
+      "epoch": 11.12,
+      "learning_rate": 1.6946670378907635e-05,
+      "loss": 0.0597,
+      "step": 4980
+    },
+    {
+      "epoch": 11.14,
+      "learning_rate": 1.6933657967871056e-05,
+      "loss": 0.0667,
+      "step": 4990
+    },
+    {
+      "epoch": 11.16,
+      "learning_rate": 1.6920622907129452e-05,
+      "loss": 0.0651,
+      "step": 5000
+    },
+    {
+      "epoch": 11.18,
+      "learning_rate": 1.690756523926356e-05,
+      "loss": 0.0648,
+      "step": 5010
+    },
+    {
+      "epoch": 11.21,
+      "learning_rate": 1.6894485006927972e-05,
+      "loss": 0.0671,
+      "step": 5020
+    },
+    {
+      "epoch": 11.23,
+      "learning_rate": 1.688138225285098e-05,
+      "loss": 0.0708,
+      "step": 5030
+    },
+    {
+      "epoch": 11.25,
+      "learning_rate": 1.6868257019834464e-05,
+      "loss": 0.084,
+      "step": 5040
+    },
+    {
+      "epoch": 11.27,
+      "learning_rate": 1.685510935075371e-05,
+      "loss": 0.0592,
+      "step": 5050
+    },
+    {
+      "epoch": 11.29,
+      "learning_rate": 1.684193928855731e-05,
+      "loss": 0.0681,
+      "step": 5060
+    },
+    {
+      "epoch": 11.32,
+      "learning_rate": 1.682874687626701e-05,
+      "loss": 0.0765,
+      "step": 5070
+    },
+    {
+      "epoch": 11.34,
+      "learning_rate": 1.6815532156977553e-05,
+      "loss": 0.0853,
+      "step": 5080
+    },
+    {
+      "epoch": 11.36,
+      "learning_rate": 1.6802295173856558e-05,
+      "loss": 0.0756,
+      "step": 5090
+    },
+    {
+      "epoch": 11.38,
+      "learning_rate": 1.678903597014437e-05,
+      "loss": 0.0922,
+      "step": 5100
+    },
+    {
+      "epoch": 11.41,
+      "learning_rate": 1.6775754589153913e-05,
+      "loss": 0.0704,
+      "step": 5110
+    },
+    {
+      "epoch": 11.43,
+      "learning_rate": 1.676245107427058e-05,
+      "loss": 0.0697,
+      "step": 5120
+    },
+    {
+      "epoch": 11.45,
+      "learning_rate": 1.6749125468952033e-05,
+      "loss": 0.0698,
+      "step": 5130
+    },
+    {
+      "epoch": 11.47,
+      "learning_rate": 1.673577781672812e-05,
+      "loss": 0.0672,
+      "step": 5140
+    },
+    {
+      "epoch": 11.5,
+      "learning_rate": 1.672240816120071e-05,
+      "loss": 0.0586,
+      "step": 5150
+    },
+    {
+      "epoch": 11.52,
+      "learning_rate": 1.670901654604353e-05,
+      "loss": 0.073,
+      "step": 5160
+    },
+    {
+      "epoch": 11.54,
+      "learning_rate": 1.669560301500205e-05,
+      "loss": 0.0759,
+      "step": 5170
+    },
+    {
+      "epoch": 11.56,
+      "learning_rate": 1.668216761189334e-05,
+      "loss": 0.0817,
+      "step": 5180
+    },
+    {
+      "epoch": 11.58,
+      "learning_rate": 1.6668710380605902e-05,
+      "loss": 0.0642,
+      "step": 5190
+    },
+    {
+      "epoch": 11.61,
+      "learning_rate": 1.6655231365099556e-05,
+      "loss": 0.0669,
+      "step": 5200
+    },
+    {
+      "epoch": 11.63,
+      "learning_rate": 1.6641730609405276e-05,
+      "loss": 0.0647,
+      "step": 5210
+    },
+    {
+      "epoch": 11.65,
+      "learning_rate": 1.6628208157625055e-05,
+      "loss": 0.0711,
+      "step": 5220
+    },
+    {
+      "epoch": 11.67,
+      "learning_rate": 1.6614664053931757e-05,
+      "loss": 0.0764,
+      "step": 5230
+    },
+    {
+      "epoch": 11.7,
+      "learning_rate": 1.6601098342568978e-05,
+      "loss": 0.069,
+      "step": 5240
+    },
+    {
+      "epoch": 11.72,
+      "learning_rate": 1.65875110678509e-05,
+      "loss": 0.0733,
+      "step": 5250
+    },
+    {
+      "epoch": 11.74,
+      "learning_rate": 1.6573902274162135e-05,
+      "loss": 0.0714,
+      "step": 5260
+    },
+    {
+      "epoch": 11.76,
+      "learning_rate": 1.6560272005957604e-05,
+      "loss": 0.0848,
+      "step": 5270
+    },
+    {
+      "epoch": 11.79,
+      "learning_rate": 1.6546620307762364e-05,
+      "loss": 0.0675,
+      "step": 5280
+    },
+    {
+      "epoch": 11.81,
+      "learning_rate": 1.6532947224171482e-05,
+      "loss": 0.0739,
+      "step": 5290
+    },
+    {
+      "epoch": 11.83,
+      "learning_rate": 1.6519252799849887e-05,
+      "loss": 0.085,
+      "step": 5300
+    },
+    {
+      "epoch": 11.85,
+      "learning_rate": 1.650553707953221e-05,
+      "loss": 0.0685,
+      "step": 5310
+    },
+    {
+      "epoch": 11.88,
+      "learning_rate": 1.6491800108022657e-05,
+      "loss": 0.0852,
+      "step": 5320
+    },
+    {
+      "epoch": 11.9,
+      "learning_rate": 1.6478041930194848e-05,
+      "loss": 0.0655,
+      "step": 5330
+    },
+    {
+      "epoch": 11.92,
+      "learning_rate": 1.6464262590991683e-05,
+      "loss": 0.0669,
+      "step": 5340
+    },
+    {
+      "epoch": 11.94,
+      "learning_rate": 1.6450462135425187e-05,
+      "loss": 0.0915,
+      "step": 5350
+    },
+    {
+      "epoch": 11.96,
+      "learning_rate": 1.6436640608576354e-05,
+      "loss": 0.072,
+      "step": 5360
+    },
+    {
+      "epoch": 11.99,
+      "learning_rate": 1.642279805559502e-05,
+      "loss": 0.0792,
+      "step": 5370
+    },
+    {
+      "epoch": 12.01,
+      "learning_rate": 1.6408934521699706e-05,
+      "loss": 0.0632,
+      "step": 5380
+    },
+    {
+      "epoch": 12.03,
+      "learning_rate": 1.639505005217747e-05,
+      "loss": 0.0591,
+      "step": 5390
+    },
+    {
+      "epoch": 12.05,
+      "learning_rate": 1.6381144692383754e-05,
+      "loss": 0.0628,
+      "step": 5400
+    },
+    {
+      "epoch": 12.08,
+      "learning_rate": 1.636721848774224e-05,
+      "loss": 0.0641,
+      "step": 5410
+    },
+    {
+      "epoch": 12.1,
+      "learning_rate": 1.635327148374471e-05,
+      "loss": 0.0625,
+      "step": 5420
+    },
+    {
+      "epoch": 12.12,
+      "learning_rate": 1.633930372595088e-05,
+      "loss": 0.0632,
+      "step": 5430
+    },
+    {
+      "epoch": 12.14,
+      "learning_rate": 1.6325315259988275e-05,
+      "loss": 0.0757,
+      "step": 5440
+    },
+    {
+      "epoch": 12.17,
+      "learning_rate": 1.631130613155205e-05,
+      "loss": 0.0576,
+      "step": 5450
+    },
+    {
+      "epoch": 12.19,
+      "learning_rate": 1.6297276386404872e-05,
+      "loss": 0.0825,
+      "step": 5460
+    },
+    {
+      "epoch": 12.21,
+      "learning_rate": 1.628322607037674e-05,
+      "loss": 0.0584,
+      "step": 5470
+    },
+    {
+      "epoch": 12.23,
+      "learning_rate": 1.626915522936486e-05,
+      "loss": 0.0748,
+      "step": 5480
+    },
+    {
+      "epoch": 12.25,
+      "learning_rate": 1.6255063909333486e-05,
+      "loss": 0.0755,
+      "step": 5490
+    },
+    {
+      "epoch": 12.28,
+      "learning_rate": 1.6240952156313762e-05,
+      "loss": 0.068,
+      "step": 5500
+    },
+    {
+      "epoch": 12.3,
+      "learning_rate": 1.622682001640359e-05,
+      "loss": 0.0621,
+      "step": 5510
+    },
+    {
+      "epoch": 12.32,
+      "learning_rate": 1.6212667535767456e-05,
+      "loss": 0.06,
+      "step": 5520
+    },
+    {
+      "epoch": 12.34,
+      "learning_rate": 1.6198494760636303e-05,
+      "loss": 0.0608,
+      "step": 5530
+    },
+    {
+      "epoch": 12.37,
+      "learning_rate": 1.618430173730736e-05,
+      "loss": 0.074,
+      "step": 5540
+    },
+    {
+      "epoch": 12.39,
+      "learning_rate": 1.617008851214401e-05,
+      "loss": 0.0655,
+      "step": 5550
+    },
+    {
+      "epoch": 12.41,
+      "learning_rate": 1.6155855131575614e-05,
+      "loss": 0.0627,
+      "step": 5560
+    },
+    {
+      "epoch": 12.43,
+      "learning_rate": 1.6141601642097382e-05,
+      "loss": 0.0769,
+      "step": 5570
+    },
+    {
+      "epoch": 12.46,
+      "learning_rate": 1.6127328090270213e-05,
+      "loss": 0.0869,
+      "step": 5580
+    },
+    {
+      "epoch": 12.48,
+      "learning_rate": 1.611303452272053e-05,
+      "loss": 0.0603,
+      "step": 5590
+    },
+    {
+      "epoch": 12.5,
+      "learning_rate": 1.609872098614017e-05,
+      "loss": 0.0733,
+      "step": 5600
+    },
+    {
+      "epoch": 12.52,
+      "learning_rate": 1.608438752728616e-05,
+      "loss": 0.062,
+      "step": 5610
+    },
+    {
+      "epoch": 12.54,
+      "learning_rate": 1.6070034192980638e-05,
+      "loss": 0.0846,
+      "step": 5620
+    },
+    {
+      "epoch": 12.57,
+      "learning_rate": 1.6055661030110655e-05,
+      "loss": 0.068,
+      "step": 5630
+    },
+    {
+      "epoch": 12.59,
+      "learning_rate": 1.6041268085628042e-05,
+      "loss": 0.0653,
+      "step": 5640
+    },
+    {
+      "epoch": 12.61,
+      "learning_rate": 1.602685540654924e-05,
+      "loss": 0.0623,
+      "step": 5650
+    },
+    {
+      "epoch": 12.63,
+      "learning_rate": 1.6012423039955153e-05,
+      "loss": 0.0627,
+      "step": 5660
+    },
+    {
+      "epoch": 12.66,
+      "learning_rate": 1.5997971032991007e-05,
+      "loss": 0.0644,
+      "step": 5670
+    },
+    {
+      "epoch": 12.68,
+      "learning_rate": 1.5983499432866187e-05,
+      "loss": 0.0594,
+      "step": 5680
+    },
+    {
+      "epoch": 12.7,
+      "learning_rate": 1.596900828685407e-05,
+      "loss": 0.0666,
+      "step": 5690
+    },
+    {
+      "epoch": 12.72,
+      "learning_rate": 1.5954497642291897e-05,
+      "loss": 0.0626,
+      "step": 5700
+    },
+    {
+      "epoch": 12.75,
+      "learning_rate": 1.593996754658059e-05,
+      "loss": 0.0587,
+      "step": 5710
+    },
+    {
+      "epoch": 12.77,
+      "learning_rate": 1.5925418047184615e-05,
+      "loss": 0.0645,
+      "step": 5720
+    },
+    {
+      "epoch": 12.79,
+      "learning_rate": 1.591084919163183e-05,
+      "loss": 0.0907,
+      "step": 5730
+    },
+    {
+      "epoch": 12.81,
+      "learning_rate": 1.589626102751331e-05,
+      "loss": 0.059,
+      "step": 5740
+    },
+    {
+      "epoch": 12.83,
+      "learning_rate": 1.588165360248321e-05,
+      "loss": 0.0696,
+      "step": 5750
+    },
+    {
+      "epoch": 12.86,
+      "learning_rate": 1.5867026964258614e-05,
+      "loss": 0.083,
+      "step": 5760
+    },
+    {
+      "epoch": 12.88,
+      "learning_rate": 1.5852381160619343e-05,
+      "loss": 0.0683,
+      "step": 5770
+    },
+    {
+      "epoch": 12.9,
+      "learning_rate": 1.5837716239407855e-05,
+      "loss": 0.0665,
+      "step": 5780
+    },
+    {
+      "epoch": 12.92,
+      "learning_rate": 1.582303224852903e-05,
+      "loss": 0.0703,
+      "step": 5790
+    },
+    {
+      "epoch": 12.95,
+      "learning_rate": 1.580832923595006e-05,
+      "loss": 0.0663,
+      "step": 5800
+    },
+    {
+      "epoch": 12.97,
+      "learning_rate": 1.5793607249700268e-05,
+      "loss": 0.0764,
+      "step": 5810
+    },
+    {
+      "epoch": 12.99,
+      "learning_rate": 1.5778866337870952e-05,
+      "loss": 0.0664,
+      "step": 5820
+    },
+    {
+      "epoch": 13.01,
+      "learning_rate": 1.5764106548615244e-05,
+      "loss": 0.065,
+      "step": 5830
+    },
+    {
+      "epoch": 13.04,
+      "learning_rate": 1.5749327930147932e-05,
+      "loss": 0.0622,
+      "step": 5840
+    },
+    {
+      "epoch": 13.06,
+      "learning_rate": 1.573453053074532e-05,
+      "loss": 0.059,
+      "step": 5850
+    },
+    {
+      "epoch": 13.08,
+      "learning_rate": 1.571971439874505e-05,
+      "loss": 0.0547,
+      "step": 5860
+    },
+    {
+      "epoch": 13.1,
+      "learning_rate": 1.570487958254597e-05,
+      "loss": 0.0512,
+      "step": 5870
+    },
+    {
+      "epoch": 13.12,
+      "learning_rate": 1.569002613060796e-05,
+      "loss": 0.0548,
+      "step": 5880
+    },
+    {
+      "epoch": 13.15,
+      "learning_rate": 1.5675154091451765e-05,
+      "loss": 0.0645,
+      "step": 5890
+    },
+    {
+      "epoch": 13.17,
+      "learning_rate": 1.566026351365886e-05,
+      "loss": 0.0667,
+      "step": 5900
+    },
+    {
+      "epoch": 13.19,
+      "learning_rate": 1.5645354445871274e-05,
+      "loss": 0.0612,
+      "step": 5910
+    },
+    {
+      "epoch": 13.21,
+      "learning_rate": 1.5630426936791433e-05,
+      "loss": 0.0558,
+      "step": 5920
+    },
+    {
+      "epoch": 13.24,
+      "learning_rate": 1.5615481035182013e-05,
+      "loss": 0.0833,
+      "step": 5930
+    },
+    {
+      "epoch": 13.26,
+      "learning_rate": 1.5600516789865767e-05,
+      "loss": 0.0526,
+      "step": 5940
+    },
+    {
+      "epoch": 13.28,
+      "learning_rate": 1.5585534249725362e-05,
+      "loss": 0.0816,
+      "step": 5950
+    },
+    {
+      "epoch": 13.3,
+      "learning_rate": 1.5570533463703233e-05,
+      "loss": 0.0648,
+      "step": 5960
+    },
+    {
+      "epoch": 13.33,
+      "learning_rate": 1.555551448080143e-05,
+      "loss": 0.0613,
+      "step": 5970
+    },
+    {
+      "epoch": 13.35,
+      "learning_rate": 1.5540477350081423e-05,
+      "loss": 0.0567,
+      "step": 5980
+    },
+    {
+      "epoch": 13.37,
+      "learning_rate": 1.5525422120663986e-05,
+      "loss": 0.0538,
+      "step": 5990
+    },
+    {
+      "epoch": 13.39,
+      "learning_rate": 1.5510348841728997e-05,
+      "loss": 0.0582,
+      "step": 6000
+    }
+  ],
+  "max_steps": 17920,
+  "num_train_epochs": 40,
+  "total_flos": 5.287881617912627e+16,
+  "trial_name": null,
+  "trial_params": null
+}

s2/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e58f3087a2c566412b2d6ccff502b7548a8643944ffa0bb008f5e780246067ae
+size 5819

s2/zero_to_fp32.py

@@ -0,0 +1,578 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # 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 files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        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")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)

s2_en/README.md

@@ -0,0 +1,9 @@
+---
+library_name: peft
+---
+## Training procedure
+
+### Framework versions
+
+
+- PEFT 0.4.0

s2_en/adapter_config.json

@@ -0,0 +1,26 @@
+{
+  "auto_mapping": null,
+  "base_model_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "lora_alpha": 256,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "up_proj",
+    "down_proj",
+    "v_proj",
+    "o_proj",
+    "q_proj",
+    "gate_proj"
+  ],
+  "task_type": "CAUSAL_LM"
+}

s2_en/adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:422c0b3bfbe12adcec40ae8e0a603d2bea66f01127e811eb210d95159de1bf73
+size 639787082

s2_en/config.json

@@ -0,0 +1,44 @@
+{
+  "_name_or_path": "liuhaotian/llava-v1.5-7b",
+  "architectures": [
+    "LlavaLlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "freeze_mm_mlp_adapter": false,
+  "freeze_mm_vision_resampler": false,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_aspect_ratio": "pad",
+  "image_grid_pinpoints": null,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_length": 4096,
+  "max_position_embeddings": 4096,
+  "mm_hidden_size": 1024,
+  "mm_projector_lr": 2e-05,
+  "mm_projector_type": "mlp2x_gelu",
+  "mm_resampler_type": null,
+  "mm_use_im_patch_token": false,
+  "mm_use_im_start_end": false,
+  "mm_vision_select_feature": "patch",
+  "mm_vision_select_layer": -2,
+  "mm_vision_tower": "openai/clip-vit-large-patch14-336",
+  "model_type": "llava",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0",
+  "tune_mm_mlp_adapter": false,
+  "tune_mm_vision_resampler": false,
+  "unfreeze_mm_vision_tower": false,
+  "use_cache": true,
+  "use_mm_proj": true,
+  "vocab_size": 32000
+}

s2_en/global_step11000/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:ed9fd169999c88b798a023cbc5a6bff9c471a528683ee25ba2eb508c9009582e
+size 4089600080