update convert_scripts

tools/convert_scripts/convert_scripts/Megatron-Deepspeed/convert_megatron_gpt2_checkpoint.py

+####################################################################################################
+
+# Copyright (c) 2021-, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+####################################################################################################
+
+#
+# Note: If when running this conversion script you're getting an exception:
+#     ModuleNotFoundError: No module named 'megatron.model.enums'
+# you need to tell python where to find the clone of Megatron-LM, e.g.:
+#
+# cd /tmp
+# git clone https://github.com/NVIDIA/Megatron-LM
+# PYTHONPATH=/tmp/Megatron-LM python src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py ...
+#
+# if you already have it cloned elsewhere, simply adjust the path to the existing path
+#
+# If the training was done using a Megatron-LM fork, e.g.,
+# https://github.com/microsoft/Megatron-DeepSpeed/ then chances are that you need to have that one
+# in your path, i.e., /path/to/Megatron-DeepSpeed/
+#
+
+import argparse
+import os
+import re
+import zipfile
+
+import torch
+
+from transformers import AutoTokenizer, GPT2Config
+import pdb
+
+####################################################################################################
+
+
+def recursive_print(name, val, spaces=0):
+    # Format the message.
+    if name is None:
+        msg = None
+    else:
+        fmt = "." * max(0, spaces - 2) + "# {:" + str(50 - spaces) + "s}"
+        msg = fmt.format(name)
+
+    # Print and recurse (if needed).
+    if isinstance(val, dict):
+        if msg is not None:
+            print(msg)
+        for k in val.keys():
+            recursive_print(k, val[k], spaces + 2)
+    elif isinstance(val, torch.Tensor):
+        print(msg, ":", val.size())
+    else:
+        print(msg, ":", val)
+
+
+def fix_query_key_value_ordering(param, checkpoint_version, num_splits, num_heads, hidden_size):
+    # Permutes layout of param tensor to [num_splits * num_heads * hidden_size, :]
+    # for compatibility with later versions of NVIDIA Megatron-LM.
+    # The inverse operation is performed inside Megatron-LM to read checkpoints:
+    # https://github.com/NVIDIA/Megatron-LM/blob/v2.4/megatron/checkpointing.py#L209
+    # If param is the weight tensor of the self-attention block, the returned tensor
+    # will have to be transposed one more time to be read by HuggingFace GPT2.
+    input_shape = param.size()
+    if checkpoint_version == 1.0:
+        # version 1.0 stores [num_heads * hidden_size * num_splits, :]
+        saved_shape = (num_heads, hidden_size, num_splits) + input_shape[1:]
+        param = param.view(*saved_shape)
+        param = param.transpose(0, 2)
+        param = param.transpose(1, 2).contiguous()
+    elif checkpoint_version >= 2.0:
+        # other versions store [num_heads * num_splits * hidden_size, :]
+        saved_shape = (num_heads, num_splits, hidden_size) + input_shape[1:]
+        param = param.view(*saved_shape)
+        param = param.transpose(0, 1).contiguous()
+    param = param.view(*input_shape)
+    return param
+
+
+####################################################################################################
+
+
+def convert_megatron_checkpoint(args, input_state_dict, config, origin_tp_degree=1):
+    # The converted output model.
+    output_state_dict = {}
+
+    # old versions did not store training args
+    ds_args = input_state_dict.get("args", None)
+    if ds_args is not None:
+        # do not make the user write a config file when the exact dimensions/sizes are already in the checkpoint
+        # from pprint import pprint
+        # pprint(vars(ds_args))
+
+        config.vocab_size = ds_args.padded_vocab_size
+        config.n_positions = ds_args.max_position_embeddings
+        config.n_embd = ds_args.hidden_size
+        config.n_layer = ds_args.num_layers
+        config.n_head = ds_args.num_attention_heads
+        config.n_inner = ds_args.ffn_hidden_size
+        # pprint(config)
+
+    # The number of heads.
+    heads = config.n_head
+    # The hidden_size per head.
+    hidden_size_per_head = config.n_embd // config.n_head
+    # Megatron-LM checkpoint version
+    if "checkpoint_version" in input_state_dict.keys():
+        checkpoint_version = input_state_dict["checkpoint_version"]
+    else:
+        checkpoint_version = 0.0
+
+    # The model.
+    model = input_state_dict["model"] if "model" in input_state_dict else input_state_dict["module"]
+    for key in model.keys():
+        print(f">> {key} in model: {model[key].keys()}")
+        for sub_key in model[key].keys():
+            print(f"\t>> {sub_key} in {key} in model: {model[key][sub_key].keys()}")
+    # The language model.
+    lm = model["language_model"]
+    # The embeddings.
+    embeddings = lm["embedding"]
+
+    # The word embeddings.
+    word_embeddings = embeddings["word_embeddings"]["weight"]
+    # Truncate the embedding table to vocab_size rows.
+    word_embeddings = word_embeddings[: config.vocab_size, :]
+    #output_state_dict["transformer.wte.weight"] = word_embeddings
+    output_state_dict["model.embed_tokens.weight"] = word_embeddings
+
+    # for LLAMA2
+    lm_head = lm['output_layer']['weight'] if 'output_layer' in lm else word_embeddings
+
+    # The position embeddings.
+    #pos_embeddings = embeddings["position_embeddings"]["weight"]
+    # Read the causal mask dimension (seqlen). [max_sequence_length, hidden_size]
+    #n_positions = pos_embeddings.size(0)
+    n_positions = config.n_positions
+    if n_positions != config.n_positions:
+        raise ValueError(
+            f"pos_embeddings.max_sequence_length={n_positions} and config.n_positions={config.n_positions} don't match"
+        )
+    # Store the position embeddings.
+    #output_state_dict["transformer.wpe.weight"] = pos_embeddings
+
+    # The transformer.
+    transformer = lm["transformer"] if "transformer" in lm.keys() else lm["encoder"]
+
+    # The regex to extract layer names.
+    layer_re = re.compile(r"layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+
+    # The simple map of names for "automated" rules.
+    megatron_to_transformers = {
+        #"attention.dense": ".attn.c_proj.",
+        #"self_attention.dense": ".attn.c_proj.",
+        "attention.dense": ".self_attn.o_proj.",
+        "self_attention.dense": ".self_attn.o_proj.",
+        "mlp.dense_h_to_4h": ".mlp.c_fc.",
+        #"mlp.dense_4h_to_h": ".mlp.c_proj.",
+        "mlp.dense_4h_to_h": ".mlp.down_proj.",
+    }
+
+    # Extract the layers.
+    for key, val in transformer.items():
+        # Match the name.
+        m = layer_re.match(key)
+
+        # Stop if that's not a layer
+        if m is None:
+            break
+
+        # The index of the layer.
+        layer_idx = int(m.group(1))
+        # The name of the operation.
+        op_name = m.group(2)
+        # Is it a weight or a bias?
+        weight_or_bias = m.group(3)
+
+        # The name of the layer.
+        layer_name = f"transformer.h.{layer_idx}"
+        layer_name = f"model.layers.{layer_idx}"
+
+        # For layernorm(s), simply store the layer norm.
+        if op_name.endswith("layernorm"):
+            #ln_name = "ln_1" if op_name.startswith("input") else "ln_2"
+            ln_name = op_name
+            output_state_dict[layer_name + "." + ln_name + "." + weight_or_bias] = val
+
+        # Transpose the QKV matrix.
+        elif (
+            op_name == "attention.query_key_value" or op_name == "self_attention.query_key_value"
+        ) and weight_or_bias == "weight":
+            # Insert a tensor of 1x1xDxD bias.
+            #causal_mask = torch.tril(torch.ones((n_positions, n_positions), dtype=torch.float16)).view(
+            #    1, 1, n_positions, n_positions
+            #)
+            #output_state_dict[layer_name + ".attn.bias"] = causal_mask
+
+            # Insert a "dummy" tensor for masked_bias.
+            #masked_bias = torch.tensor(-1e4, dtype=torch.float16)
+            #output_state_dict[layer_name + ".attn.masked_bias"] = masked_bias
+
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 3, heads, hidden_size_per_head)
+            # Megatron stores (3*D) x D but transformers-GPT2 expects D x 3*D.
+            #out_val = out_val.transpose(0, 1).contiguous()
+            out_val = out_val.contiguous()
+            # Store.
+            #output_state_dict[layer_name + ".attn.c_attn.weight"] = out_val
+            output_state_dict[layer_name + ".self_attn.q_proj.weight"] = out_val[:config.n_embd, :]
+            output_state_dict[layer_name + ".self_attn.k_proj.weight"] = out_val[config.n_embd:config.n_embd * 2, :]
+            output_state_dict[layer_name + ".self_attn.v_proj.weight"] = out_val[config.n_embd * 2 :, :]
+        elif (
+            op_name == "self_attention.query"
+        ) and weight_or_bias == "weight":
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 1, heads, hidden_size_per_head)
+            #out_val = out_val.transpose(0, 1).contiguous()
+            out_val = out_val.contiguous()
+            output_state_dict[layer_name + ".self_attn.q_proj.weight"] = out_val
+        elif (
+            op_name == "self_attention.key_value"
+        )   and weight_or_bias == "weight":
+            #print(f">> key_value origin size: {val.size()}")
+            size_per_weight = val.size(0) // 2
+            #please set the NUM_KV_HEADS used to replace number "4" in fix_query_key_value_ordering function
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 2, 4, hidden_size_per_head)
+            #print(f">> key_value output size: {out_val.size()}")
+            out_val = out_val.contiguous()
+            output_state_dict[layer_name + ".self_attn.k_proj.weight"] = out_val[:size_per_weight, :]
+            output_state_dict[layer_name + ".self_attn.v_proj.weight"] = out_val[size_per_weight:, :]
+        # Transpose the bias.
+        elif (
+            op_name == "attention.query_key_value" or op_name == "self_attention.query_key_value"
+        ) and weight_or_bias == "bias":
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 3, heads, hidden_size_per_head)
+            # Store. No change of shape.
+            output_state_dict[layer_name + ".attn.c_attn.bias"] = out_val
+        elif op_name == "mlp.dense_h_to_4h":
+            # this 2 lines for TP=1 (swiglu)
+            if origin_tp_degree == 1:
+                output_state_dict[layer_name + ".mlp.gate_proj.weight"] = val[:config.n_inner, :]
+                output_state_dict[layer_name + ".mlp.up_proj.weight"] = val[config.n_inner:, :]
+            elif origin_tp_degree == 2:
+            # this 2 lines for TP=2 (swiglu)
+                output_state_dict[layer_name + ".mlp.gate_proj.weight"] = torch.cat([val[:config.n_inner//2, :], val[config.n_inner:config.n_inner + config.n_inner // 2, :]])
+                output_state_dict[layer_name + ".mlp.up_proj.weight"] = torch.cat([val[config.n_inner//2:config.n_inner, :], val[config.n_inner + config.n_inner // 2:, :]])
+            elif origin_tp_degree == 4:
+                output_state_dict[layer_name + ".mlp.gate_proj.weight"] = torch.cat([val[:config.n_inner//4, :],val[config.n_inner//2:config.n_inner//2+config.n_inner//4, :],val[config.n_inner:config.n_inner+config.n_inner//4, :],val[config.n_inner+config.n_inner//2:config.n_inner+config.n_inner//4*3, :] ])
+                output_state_dict[layer_name + ".mlp.up_proj.weight"] = torch.cat([val[config.n_inner//4:config.n_inner//2,:], val[config.n_inner//2+config.n_inner//4:config.n_inner,:],val[config.n_inner+config.n_inner//4:config.n_inner+config.n_inner//2,:],val[config.n_inner+config.n_inner//4*3:config.n_inner*2, :] ])
+            else:
+                raise ValueError("Not Implemented Yet for TP /= 1 && 2 && 4.")
+        # Transpose the weights.
+        elif weight_or_bias == "weight":
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "weight"] = val#.transpose(0, 1)
+
+        # Copy the bias.
+        elif weight_or_bias == "bias":
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "bias"] = val
+
+    # DEBUG.
+    assert config.n_layer == layer_idx + 1
+
+    # The final layernorm.
+    #output_state_dict["transformer.ln_f.weight"] = transformer["final_layernorm.weight"]
+    #pdb.set_trace()
+    output_state_dict["model.norm.weight"] = transformer["final_layernorm.weight"]
+    #output_state_dict["transformer.ln_f.bias"] = transformer["final_layernorm.bias"]
+
+    # For LM head, transformers' wants the matrix to weight embeddings.
+    output_state_dict["lm_head.weight"] = lm_head
+
+    # transform the key for LLAMA2
+    transform_dict = {
+        "transformer.h": "model.layers",
+
+    }
+    # It should be done!
+    return output_state_dict
+
+
+####################################################################################################
+
+
+def main():
+    # Create the argument parser.
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--print-checkpoint-structure", action="store_true")
+    parser.add_argument(
+        "path_to_checkpoint",
+        type=str,
+        help="Path to the checkpoint file (.zip archive or direct .pt file)",
+    )
+    parser.add_argument(
+        "--config_file",
+        default="",
+        type=str,
+        help="An optional config json file describing the pre-trained model.",
+    )
+    args = parser.parse_args()
+
+    # Extract the basename.
+    basename = os.path.dirname(args.path_to_checkpoint)
+
+    # Load the model.
+    # the .zip is very optional, let's keep it for backward compatibility
+    print(f"Extracting PyTorch state dictionary from {args.path_to_checkpoint}")
+    if args.path_to_checkpoint.endswith(".zip"):
+        with zipfile.ZipFile(args.path_to_checkpoint, "r") as checkpoint:
+            with checkpoint.open("release/mp_rank_00/model_optim_rng.pt") as pytorch_dict:
+                input_state_dict = torch.load(pytorch_dict, map_location="cpu")
+    else:
+        input_state_dict = torch.load(args.path_to_checkpoint, map_location="cpu")
+    print(f">> keys: {input_state_dict.keys()}")
+    ds_args = input_state_dict.get("args", None)
+    #print(f">> ds_args: {ds_args}")
+    # Read the config, or default to the model released by NVIDIA.
+    if args.config_file == "":
+        if ds_args is not None:
+            if ds_args.bias_gelu_fusion:
+                activation_function = "gelu_fast"
+            elif ds_args.openai_gelu:
+                activation_function = "gelu_new"
+            else:
+                activation_function = "gelu"
+        else:
+            # in the very early days this used to be "gelu_new"
+            activation_function = "gelu_new"
+
+        # Spell out all parameters in case the defaults change.
+        config = GPT2Config(
+            vocab_size=50257,
+            n_positions=1024,
+            n_embd=1024,
+            n_layer=24,
+            n_head=16,
+            n_inner=4096,
+            activation_function=activation_function,
+            resid_pdrop=0.1,
+            embd_pdrop=0.1,
+            attn_pdrop=0.1,
+            layer_norm_epsilon=1e-5,
+            initializer_range=0.02,
+            summary_type="cls_index",
+            summary_use_proj=True,
+            summary_activation=None,
+            summary_proj_to_labels=True,
+            summary_first_dropout=0.1,
+            scale_attn_weights=True,
+            use_cache=True,
+            bos_token_id=50256,
+            eos_token_id=50256,
+        )
+    else:
+        config = GPT2Config.from_json_file(args.config_file)
+
+    config.architectures = ["GPT2LMHeadModel"]
+
+    # Convert.
+    print("Converting")
+    output_state_dict = convert_megatron_checkpoint(args, input_state_dict, config)
+
+    # Print the structure of converted state dict.
+    if args.print_checkpoint_structure:
+        recursive_print(None, output_state_dict)
+
+    tokenizer_model_name = ""
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_model_name)
+    tokenizer_class = type(tokenizer).__name__
+    config.tokenizer_class = tokenizer_class
+
+    # Store the config to file.
+    print("Saving config")
+    #config.save_pretrained(basename)
+    print(f">> here is the local converter")
+    # Save tokenizer based on args
+    print(f"Adding {tokenizer_class} tokenizer files")
+    #tokenizer.save_pretrained(basename)
+
+    # Store the state_dict to file.
+    output_checkpoint_file = os.path.join(basename, "pytorch_model.bin")
+    print(f'Saving checkpoint to "{output_checkpoint_file}"')
+    torch.save(output_state_dict, output_checkpoint_file)
+
+
+####################################################################################################
+
+if __name__ == "__main__":
+    main()
+
+####################################################################################################
+

tools/convert_scripts/convert_scripts/Megatron-Deepspeed/tools/covert_checkpoint/deepspeed_checkpoint.py

+import os
+from typing import Dict
+import torch 
+
+ZERO_FILE_PREFIX = 'zero_pp_rank_'
+LAYER_FILE_PREFIX = 'layer_'
+MP_RANK_FILE_PREFIX = 'mp_rank_'
+EMBEDDING_LAYER_INDEX = 0
+LM_HEAD_LAYER_INDEX = -1
+FINAL_LAYER_NORM_INDEX = -2 # original -1 but -1 is the output_layer
+ARGS_KEY = 'args'
+ITERATION_KEY = 'iteration'
+SEQUENTIAL_LAYERS = [
+    'input_layernorm.weight', 'input_layernorm.bias',
+    'self_attention.dense.bias',
+    'post_attention_layernorm.weight', 'post_attention_layernorm.bias',
+    'mlp.dense_4h_to_h.bias',
+    'position_embeddings.weight'
+]
+
+LAYER_CONCAT_DIM = {
+    'self_attention.dense.weight': 1,
+    'mlp.dense_4h_to_h.weight': 1
+}
+
+class DeepSpeedCheckpoint(object):
+    def __init__(self, dir, tp_degree=None, pp_degree=None, no_pp=False, share_emb=False):
+        self.dir = dir
+        self.no_pp = no_pp
+        self.file_list = self._get_files(dir)
+        self.zero_files = self._get_files_with_context(self.file_list, ZERO_FILE_PREFIX)
+        self.layer_files = self._get_files_with_prefix(self.file_list, LAYER_FILE_PREFIX)
+        self.mp_rank_files = self._get_files_with_prefix(self.file_list, MP_RANK_FILE_PREFIX)
+        self.layer_keys = self._get_layer_keys()
+        self.layer_count = len(self.layer_keys)
+        if not self.no_pp:
+            self.original_tp_degree = len(self._get_files_with_prefix(self.layer_files, f'{LAYER_FILE_PREFIX}01'))
+            self.original_pp_degree = len(self.mp_rank_files) // self.original_tp_degree
+        else:
+            self.original_tp_degree = len(self.mp_rank_files)
+            self.original_pp_degree = 1
+        self.dp_degree = len(self.zero_files) // (self.original_pp_degree * self.original_tp_degree)
+        self.tp_degree = self.original_tp_degree if tp_degree is None else tp_degree
+        self.pp_degree = self.original_pp_degree if pp_degree is None else pp_degree
+        print(f">> layer_files:")
+        for item in self.layer_files[:2]:
+            print(f"> {item}")
+        print("")
+        print(f">> mp_rank_files:")
+        for item in self.mp_rank_files:
+            print(f"> {item}")
+        print("")
+        print(f">> layer_keys: {self.layer_keys}")
+        print(f">> dp_degree: {self.dp_degree}, tp_degree: {self.tp_degree}, pp_degree: {self.pp_degree}")
+        print("")
+        self.global_state = {}
+    
+        self._sanity_check()
+        self.pp_to_transformer_map = self._build_pp_transformer_map()
+        if self.original_pp_degree == 1 and self.original_tp_degree == 1:
+            self.transformer_file_map = self._build_transformer_file_map_nopp()
+        else:
+            self.transformer_file_map = self._build_transformer_file_map()
+        #self.show_transformer_file_map()
+        if not self.no_pp:
+            self.tp_to_embedding_map = self._build_tp_other_layer_map(EMBEDDING_LAYER_INDEX)
+            if share_emb:
+                FINAL_LAYER_NORM_INDEX = -1
+                self.tp_to_final_norm_map = self._build_tp_other_layer_map(FINAL_LAYER_NORM_INDEX)
+                self.tp_to_lm_head_map = {}
+            else:
+                FINAL_LAYER_NORM_INDEX = -2
+                self.tp_to_lm_head_map = self._build_tp_other_layer_map(LM_HEAD_LAYER_INDEX)
+                self.tp_to_final_norm_map = self._build_tp_other_layer_map(FINAL_LAYER_NORM_INDEX)
+        self._build_global_state()
+
+
+
+    def show_tp_embedding_map(self):
+        self._dump_mapping(self.tp_to_embedding_map, 'tp_to_embedding_layers')
+
+    def show_tp_final_norm_map(self):
+        self._dump_mapping(self.tp_to_final_norm_map, 'tp_to_final_norm_layers')
+
+    def show_pp_tranformer_map(self):
+        self._dump_mapping(self.pp_to_transformer_map, 'pp_to_tranformer_layers')
+
+    def show_transformer_file_map(self):
+        self._dump_mapping(self.transformer_file_map, 'rank_to_tranformer_files')
+
+    def _build_global_state(self):
+        sd = torch.load(self.mp_rank_files[0], map_location=torch.device('cpu'))
+        self.global_state[ITERATION_KEY] = sd.get(ITERATION_KEY, 0)
+        self.global_state[ARGS_KEY] = sd.get(ARGS_KEY, None)
+
+    def get_iteration(self):
+        if not ITERATION_KEY in self.global_state:
+            sd = torch.load(self.mp_rank_files[0], map_location=torch.device('cpu'))
+            self.global_state[ITERATION_KEY] = sd.get(ITERATION_KEY, 0)
+
+        return self.global_state[ITERATION_KEY]
+
+    def get_embedding_state(self, tp_index: int) -> Dict:
+        print(f">> tp_index: {tp_index} searched keys: {self.tp_to_embedding_map.keys()}")
+        assert tp_index in self.tp_to_embedding_map.keys()
+        for name in self.tp_to_embedding_map[tp_index]:
+            print(f">> deepspeed checkpoint.py, name in tp_to_embedding_map: {name}")
+        sd_list = [torch.load(fname, map_location=torch.device('cpu')) for fname in self.tp_to_embedding_map[tp_index]]
+        sd = self._merge_state_dicts(sd_list)
+        return sd
+    
+    def get_lm_head_state(self, tp_index: int) -> Dict:
+        assert tp_index in self.tp_to_embedding_map.keys()
+        for name in self.tp_to_lm_head_map[tp_index]:
+            print(f">> deepspeed checkpoint.py, name in tp_to_lm_head_map: {name}")
+        sd_list = [torch.load(fname, map_location=torch.device('cpu')) for fname in self.tp_to_lm_head_map[tp_index]]
+        print(f">> lm head sd_list: {sd_list}")
+        sd = self._merge_state_dicts(sd_list)
+        return sd
+
+    def get_args(self):
+        if not ARGS_KEY in self.global_state:
+            sd = torch.load(self.mp_rank_files[0], map_location=torch.device('cpu'))
+            self.global_state[ARGS_KEY] = sd.get(ARGS_KEY, None)
+
+        return self.global_state[ARGS_KEY]
+    
+
+    def get_transformer_state(self, tp_index: int, pp_index: int) -> list:
+        assert tp_index < self.tp_degree
+        assert pp_index < self.pp_degree
+        t_list = []
+        for fname_list in self.transformer_file_map[(tp_index, pp_index)]:
+            sd_list = [torch.load(fname, map_location=torch.device('cpu')) for fname in fname_list]
+            sd = self._merge_state_dicts(sd_list)
+            t_list.append(sd)
+        return t_list   
+
+    def get_final_norm_state(self, tp_index:int) -> Dict:
+        assert tp_index in self.tp_to_final_norm_map.keys()
+        sd = torch.load(self.tp_to_final_norm_map[tp_index][0], map_location=torch.device('cpu'))
+        return sd
+
+    def _build_tp_other_layer_map(self, layer_index:int):
+        assert layer_index < len(self.layer_files)
+        layer_files = self._get_files_with_prefix(self.layer_files, self.layer_keys[layer_index])
+        layer_file_partitions = self._partition_data(layer_files, self.tp_degree)
+        data_map = {i:flist for i, flist in enumerate(layer_file_partitions)}
+        return data_map
+
+    def _build_pp_transformer_map(self):
+        data_map = {}
+        transformer_layers = self.layer_keys[1:-1]
+        layers_per_pp = len(transformer_layers) // self.pp_degree
+        data_map = {i:transformer_layers[i*layers_per_pp:(i+1)*layers_per_pp] for i in range(0, self.pp_degree)}
+        return data_map
+
+    def _dump_mapping(self, data_map, map_tag = None):
+        if map_tag is not None:
+            print(f'Dump mapping: {map_tag}')
+        for k, v in data_map.items():
+            print(f'{k} = {v}')
+    
+    # xianfzeng
+    def _build_transformer_file_map_nopp(self):
+        file_map = {(0, 0): [self.mp_rank_files]}
+        return file_map
+
+    def _build_transformer_file_map(self):
+        transformer_layer_keys = self.layer_keys[1:-1]
+        file_map = {}
+        layers_per_pp = len(transformer_layer_keys) // self.pp_degree
+        for key_index, layer_key in enumerate(transformer_layer_keys):
+            pp_index = key_index // layers_per_pp
+            layer_files = self._get_files_with_prefix(self.layer_files, layer_key)
+            layer_file_partitions = self._partition_data(layer_files, self.tp_degree)
+            for tp_index in range(self.tp_degree):
+                map_key = (tp_index, pp_index)
+                if not map_key in file_map.keys():
+                    file_map[map_key] = []
+                file_map[map_key].append(layer_file_partitions[tp_index])
+        
+        return file_map
+        
+    def _sanity_check(self):
+        assert len(self.mp_rank_files) % self.tp_degree == 0
+        assert len(self.zero_files) % (self.pp_degree * self.tp_degree) == 0
+        if not self.no_pp:
+            assert len(self.layer_keys) > 2
+            assert (len(self.layer_keys) - 2) % self.pp_degree == 0
+     
+    def _get_files_with_prefix(self, all_files, prefix):
+        file_list = []
+        for file_path in all_files:
+            _, fname = os.path.split(file_path)
+            if fname.startswith(prefix):
+                file_list.append(file_path)
+        
+        return sorted(file_list)
+
+    # @ xianfzeng
+    def _get_files_with_context(self, all_files, context):
+        file_list = []
+        for file_path in all_files:
+            _, fname = os.path.split(file_path)
+            if context in fname:
+                file_list.append(file_path)
+        
+        return sorted(file_list)
+
+    def validate_files(self):
+        for file in self.file_list:
+            if not os.path.isfile(file):
+                print(f'Error: {file} is not existent')
+        
+    def _get_files(self, dir):
+        file_list = []
+        for root, dirs, files in os.walk(dir):
+            for file in files:
+                file_list.append(os.path.join(root, file))
+        return file_list
+
+    def _get_layer_keys(self):
+        key_set = set()
+        key_len = len(LAYER_FILE_PREFIX) + 2 
+        for file_path in self.layer_files:
+            _, fname = os.path.split(file_path)
+            key_set.add(fname[:key_len])
+        return sorted(list(key_set))
+
+    def _partition_data(self, data_list, num_partitions):
+        num_elems = len(data_list)
+        assert num_elems % num_partitions == 0
+        partition_size = num_elems // num_partitions
+        partitions_list = [data_list[i:i+partition_size] for i in range(0, num_elems, partition_size)]
+        return partitions_list
+
+    def _merge_state_dicts(self, sd_list):
+        print(f">> sd_List: {sd_list[0].keys()}")
+        #print(f">> sd_list: {sd_list[0]['module']['language_model'].keys()}")
+        merged_sd = {}
+        for key in sd_list[0].keys():
+            if not key in SEQUENTIAL_LAYERS:
+                cat_dim = LAYER_CONCAT_DIM.get(key, 0)
+                merged_sd[key] = torch.cat([sd[key] for sd in sd_list], dim=cat_dim)
+            else:
+                merged_sd[key] = sd_list[0][key]
+        return merged_sd

tools/convert_scripts/convert_scripts/Megatron-Deepspeed/tools/covert_checkpoint/deepspeed_to_megatron.py

+#!/usr/bin/env python
+
+import argparse
+import os
+import torch
+from collections import OrderedDict
+from deepspeed_checkpoint import ARGS_KEY, DeepSpeedCheckpoint
+
+MODEL_KEY = 'model'
+ARGS_KEY = 'args'
+LANGUAGE_MODEL_KEY = 'language_model'
+EMBEDDING_KEY = 'embedding'
+ENCODER_KEY = 'encoder'
+WORD_EMBEDDINGS_FOR_HEAD_KEY = 'word_embeddings_for_head'
+WORD_EMBEDDINGS_FOR_HEAD_KEY = 'output_layer'
+WORD_EMBEDDINGS_FOR_HEAD_LLAMA_KEY = 'lm_head'
+WORD_EMBEDDINGS_KEY = 'word_embeddings'
+FINAL_LAYER_NORM_KEY ='final_layernorm'
+CHECKPOINT_VERSION_KEY = 'checkpoint_version'
+CHECKPOINT_VERSION_VALUE = 3.0
+ITERATION_KEY = 'iteration'
+
+def parse_arguments():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input_folder', default=None, type=str, help='Input DeepSpeed Checkpoint folder')
+    parser.add_argument('--output_folder', default=None, type=str, help='Output Megatron checkpoint folder')
+    parser.add_argument('--target_tp', default=1, type=int, help='Target TP degree')
+    parser.add_argument('--target_pp', default=1, type=int, help='Target PP degree')
+    parser.add_argument('--no_pp', default=False, action='store_true', help='Target PP degree')
+    parser.add_argument('--share_emb', default=False, action='store_true', help='Share embedding and lm_head')
+
+    parser.add_argument('--for_release', action='store_true', help='Convert for release purpose, reset some (progress) counters.')
+    args = parser.parse_args()
+    print(f'args = {args}')
+    return args
+
+
+def _convert_ds_transformer_state(sd_list):
+    new_sd = OrderedDict()
+    for i, sd in enumerate(sd_list):
+        for key, value in sd.items():
+            if key == "weight":
+                continue
+            new_key = f'layers.{i}.{key}'
+            new_sd[new_key] = value
+
+    return new_sd
+
+def _create_checkpoint_paths(base_folder, iteration, tp_degree, pp_degree):
+    path_list = []
+    iter_folder = f'iter_{iteration:07d}'
+    for i in range(0, tp_degree):
+        path_list.append([])
+        for j in range(0, pp_degree):
+            rank_folder = f'mp_rank_{i:02d}' if pp_degree == 1 else f'mp_rank_{i:02d}_{j:03d}'
+            ckpt_path = os.path.join(rank_folder, 'model_optim_rng.pt')
+            path_list[i].append(os.path.join(base_folder, iter_folder, ckpt_path))
+
+    return path_list
+
+
+def _create_megatron_dict():
+    language_model_dict = {
+        EMBEDDING_KEY: {},
+        ENCODER_KEY: {}
+    }
+    megatron_dict = {
+        MODEL_KEY: {LANGUAGE_MODEL_KEY: language_model_dict},
+        CHECKPOINT_VERSION_KEY: CHECKPOINT_VERSION_VALUE
+    }
+    return megatron_dict
+
+
+def _save_checkpoint(file_path, chkpt_sd):
+    dir, _ = os.path.split(file_path)
+    os.makedirs(dir, exist_ok=True)
+    torch.save(chkpt_sd, file_path)
+
+
+def _renest_sd(sd):
+    new_sd = OrderedDict()
+    for key, value in sd.items():
+        a, b = key.split('.')
+        new_sd[a] = {b: value}
+    return new_sd
+
+
+def _create_rank_checkpoint(ds_checkpoint, tp_index, pp_index, for_release=False, share_emb=False):
+    meg_encoder_sd = OrderedDict()
+    meg_embedding_sd = OrderedDict()
+    meg_embedding_for_head_sd = OrderedDict()
+
+    transformer_sd = ds_checkpoint.get_transformer_state(tp_index, pp_index)
+    meg_encoder_sd.update(_convert_ds_transformer_state(transformer_sd))
+
+    if pp_index in [0, ds_checkpoint.pp_degree - 1]:
+        embedding_sd = ds_checkpoint.get_embedding_state(tp_index)
+        nested_embedding_sd = _renest_sd(embedding_sd)
+        if pp_index == 0:
+            meg_embedding_sd.update(nested_embedding_sd)
+
+        if pp_index == ds_checkpoint.pp_degree -1:
+            for key, value in embedding_sd.items():
+                print(f">> deepspeed to megatron.py, key: {key}, value: {value}")
+                if key.startswith(WORD_EMBEDDINGS_KEY):
+                    fields = key.split('.')
+                    new_fields = fields[1:]
+                    new_key = '.'.join(new_fields)
+                    #new_key = "weight"
+                    meg_embedding_sd[new_key] = value
+                    print(f">> embedding sd: {new_key} {key}")
+
+            if not share_emb:
+                for key, value in ds_checkpoint.get_lm_head_state(tp_index).items():
+                    if WORD_EMBEDDINGS_FOR_HEAD_LLAMA_KEY in key:
+                        fields = key.split('.')
+                        new_fields = fields[1:]
+                        new_key = '.'.join(new_fields)
+                        #new_key = "weight"
+                        meg_embedding_for_head_sd[new_key] = value
+                        print(f">> embedding for head sd: {new_key} {key}")
+
+            final_norm_sd = ds_checkpoint.get_final_norm_state(tp_index)
+            new_final_norm_sd = {}
+            for key, value in final_norm_sd.items():
+                print(f">> final_norm_sd, key: {key}, value: {value.size()}")
+                new_final_norm_sd[f'{FINAL_LAYER_NORM_KEY}.{key}'] = value
+            #new_final_norm_sd = {f'{FINAL_LAYER_NORM_KEY}.{key}': value for key, value in final_norm_sd.items()}
+            meg_encoder_sd.update(new_final_norm_sd)
+
+    checkpoint_sd = _create_megatron_dict()
+
+    iteration = ds_checkpoint.get_iteration()
+    checkpoint_sd[ITERATION_KEY] = iteration
+    if pp_index == 0:
+        checkpoint_sd[MODEL_KEY][LANGUAGE_MODEL_KEY][EMBEDDING_KEY] = meg_embedding_sd
+        print(f">> deepspeed to megatron.py, meg_embedding_sd: {meg_embedding_sd}")
+    checkpoint_sd[MODEL_KEY][LANGUAGE_MODEL_KEY][ENCODER_KEY] = meg_encoder_sd
+    print(f">> pp_index: {pp_index} pp_degree: {ds_checkpoint.pp_degree}")
+    if pp_index == ds_checkpoint.pp_degree -1 and not share_emb:
+        #checkpoint_sd[MODEL_KEY][WORD_EMBEDDINGS_FOR_HEAD_KEY] = meg_embedding_for_head_sd
+        checkpoint_sd[MODEL_KEY][LANGUAGE_MODEL_KEY][WORD_EMBEDDINGS_FOR_HEAD_KEY] = meg_embedding_for_head_sd
+        print(f">> deepspeed to megatron.py, meg_embedding_for_head_sd: {meg_embedding_for_head_sd}")
+    #print(f">> checkpoint_sd[model][output_layer]: {checkpoint_sd[MODEL_KEY][LANGUAGE_MODEL_KEY][WORD_EMBEDDINGS_FOR_HEAD_KEY]}")
+    checkpoint_sd[ARGS_KEY] = ds_checkpoint.get_args()
+    # Adjust specific fields
+    checkpoint_sd[ARGS_KEY].tensor_model_parallel_size = ds_checkpoint.tp_degree
+    checkpoint_sd[ARGS_KEY].pipeline_model_parallel_size = ds_checkpoint.pp_degree
+    if for_release:
+        checkpoint_sd[ARGS_KEY].consumed_train_samples = 0
+        checkpoint_sd[ARGS_KEY].consumed_valid_samples = 0
+
+    return checkpoint_sd
+
+
+def _create_latest_file(base_folder, iteration):
+    file_path = os.path.join(base_folder, 'latest_checkpointed_iteration.txt')
+    os.makedirs(base_folder, exist_ok=True)
+    with open(file_path, 'w') as f:
+        f.write(str(iteration))
+
+def main():
+    print(f'Convert DeepSpeed Checkpoint to Megatron Checkpoint')
+
+    args = parse_arguments()
+    print(f'Converting DeepSpeed checkpoint in {args.input_folder} to Megatron checkpoint in {args.output_folder}')
+
+    ds_checkpoint = DeepSpeedCheckpoint(args.input_folder, args.target_tp, args.target_pp)
+    iteration = ds_checkpoint.get_iteration()
+    _create_latest_file(args.output_folder, iteration)
+    checkpoint_paths = _create_checkpoint_paths(args.output_folder, iteration, ds_checkpoint.tp_degree, ds_checkpoint.pp_degree)
+    for i in range(0, ds_checkpoint.tp_degree):
+        for j in range(0, ds_checkpoint.pp_degree):
+            sd = _create_rank_checkpoint(ds_checkpoint, i, j, args.for_release)
+            _save_checkpoint(checkpoint_paths[i][j], sd)
+
+if __name__ == "__main__":
+    main()

tools/convert_scripts/convert_scripts/Megatron-Deepspeed/tools/covert_checkpoint/deepspeed_to_transformers.py

+#!/usr/bin/env python
+
+import os
+import torch
+import json
+
+from deepspeed_checkpoint import DeepSpeedCheckpoint
+from deepspeed_to_megatron import _create_rank_checkpoint, parse_arguments
+
+# the import was tested to work with this version
+# https://github.com/huggingface/transformers/commit/0af901e83 if it diverges we may consider
+# copying that version here instead
+#from transformers.models.megatron_gpt2.convert_megatron_gpt2_checkpoint import convert_megatron_checkpoint
+from convert_megatron_gpt2_checkpoint import convert_megatron_checkpoint
+from transformers import GPT2Config, AutoTokenizer, LlamaTokenizer
+
+def main():
+
+    # this first part comes mainly from deepspeed_to_megatron.main
+    args = parse_arguments()
+    print(f'Converting DeepSpeed checkpoint in {args.input_folder} to HF Transformers checkpoint in {args.output_folder}')
+
+    ds_checkpoint = DeepSpeedCheckpoint(args.input_folder, args.target_tp, args.target_pp, args.no_pp)
+
+    iteration = ds_checkpoint.get_iteration()
+    input_state_dict = _create_rank_checkpoint(ds_checkpoint, 0, 0, args.for_release)
+
+    # the 2nd part comes from transformers.models.megatron_gpt2.convert_megatron_gpt2_checkpoint.main
+    # Spell out all parameters in case the defaults change.
+    config = None
+    config = GPT2Config(
+        vocab_size=50257,
+        n_positions=1024,
+        n_ctx=1024,
+        n_embd=1024,
+        n_layer=24,
+        n_head=16,
+        n_inner=4096,
+        activation_function="gelu",  # used to be "gelu_new" in earlier versions
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        summary_type="cls_index",
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        scale_attn_weights=True,
+        gradient_checkpointing=False,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+    )
+
+    # Convert.
+    print("Converting to HF Checkpoint")
+    print(f"input_state_dict: {input_state_dict['model'].keys()}")
+    print(f"language model: {input_state_dict['model']['language_model'].keys()}")
+    print(f"language model: {input_state_dict['model']['language_model']['encoder']['layers.23.mlp.dense_h_to_4h.weight'].size()}")
+    output_state_dict = convert_megatron_checkpoint(args, input_state_dict, config, ds_checkpoint.original_tp_degree)
+
+    basename = args.output_folder
+    os.makedirs(basename, exist_ok=True)
+
+    # Print the structure of converted state dict.
+    #if args.print_checkpoint_structure:
+    #    recursive_print(None, output_state_dict)
+
+    # Store the config to file.
+    #please prepare your config.json according to your model
+
+    # Store the state_dict to file.
+    output_checkpoint_file = os.path.join(basename, "pytorch_model.bin")
+    print(f'Saving checkpoint to "{output_checkpoint_file}"')
+    torch.save(output_state_dict, output_checkpoint_file)
+ 
+
+if __name__ == "__main__":
+    main()

tools/convert_scripts/convert_scripts/convert.sh

+
+# please change the relating path according to your env
+WORK_DIR=/path_to_convert_scripts/Megatron-Deepspeed
+CODE_DIR=$WORK_DIR
+export PYTHONPATH=${WORK_DIR}:$PYTHONPATH
+export PYTHONPATH=/path_to_Megatron-DeepSpeed-main:$PYTHONPATH
+
+checkpoint_dir=/path_to_checkpoint/global_step1xxxx
+
+python3 $WORK_DIR/tools/covert_checkpoint/deepspeed_to_transformers.py \
+    --input_folder ${checkpoint_dir} \
+    --output_folder ${checkpoint_dir}_hf \