import os import re import json import torch import torch.nn as nn from collections import defaultdict from transformers import ( AutoConfig, AutoModelForCausalLM, AutoTokenizer, ) from megatron.training.initialize import initialize_megatron from megatron.training import get_args from megatron.training.checkpointing import get_checkpoint_name, get_checkpoint_tracker_filename, read_metadata import sys path_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))) sys.path.append(os.path.join(path_dir, "examples")) from mistral.pretrain_mcore_mistral import model_provider from megatron_patch.arguments import get_patch_args from toolkits.model_checkpoints_convertor.utils import ( save_hfmodel, save_state_dict ) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False torch.backends.cuda.matmul.allow_tf32 = False torch.backends.cudnn.allow_tf32 = False def add_checkpointing_args(parser): parser.add_argument('--megatron-path', type=str, default=None, help='Base directory of Megatron repository') parser.add_argument( '--convert_checkpoint_from_megatron_to_transformers', action='store_true', help= ('If True, convert a Megatron checkpoint to a Transformers checkpoint. ' 'If False, convert a Transformers checkpoint to a Megatron checkpoint.' ), ) parser.add_argument( '--load_path', type=str, required=True, help='Path to the checkpoint to convert.', ) parser.add_argument( '--save_path', type=str, required=True, help='Path to the converted checkpoint.', ) parser.add_argument( '--huggingface_model_path', type=str, required=True, ) return parser def add_megatron_checkpoint_args(parser): parser.add_argument( "--target_tensor_model_parallel_size", type=int, default=1, help=( "The tensor model parallel size of the converted checkpoint. " "Only used when converting a Transformers checkpoint to a Megatron checkpoint." ), ) parser.add_argument( "--target_pipeline_model_parallel_size", type=int, default=1, help=( "The pipeline model parallel size of the converted checkpoint. " "Only used when converting a Transformers checkpoint to a Megatron checkpoint." ), ) parser.add_argument( "--target_expert_model_parallel_size", type=int, default=1, help=( "The data parallel size of the converted checkpoint. " "Only used when converting a Transformers checkpoint to a Megatron checkpoint." ), ) return parser def add_transformers_checkpoint_args(parser): parser.add_argument( "--max_shard_size", type=str, default="10GB", 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`). " "Only used when converting a Megatron checkpoint to a Transformers checkpoint." ), ) return parser def build_huggingface_model(model_to_load, compute_dtype, random_init=False): config = AutoConfig.from_pretrained( model_to_load, trust_remote_code=True, ) if random_init: model = AutoModelForCausalLM.from_config( config=config, torch_dtype=compute_dtype, trust_remote_code=True ) else: model = AutoModelForCausalLM.from_pretrained( model_to_load, torch_dtype=compute_dtype, trust_remote_code=True ) return config, model.eval() def create_huggingface_model(args): if args.num_experts is None: copy_huggingface_tokenizer(args.huggingface_model_path, args.save_path) config, model = build_huggingface_model(args.huggingface_model_path, args.params_dtype) else: copy_huggingface_tokenizer(args.huggingface_model_path, args.save_path, with_code=True) config, model = build_huggingface_model(args.save_path, args.params_dtype, random_init=True) return config, model.eval() def create_megatron_model(args, hf_config): args.hidden_size = hf_config.hidden_size args.num_layers = hf_config.num_hidden_layers args.num_attention_heads = hf_config.num_attention_heads args.kv_channels = args.hidden_size // args.num_attention_heads args.ffn_hidden_size = hf_config.intermediate_size args.num_query_groups = hf_config.num_key_value_heads model = model_provider() return model.eval() def copy_huggingface_tokenizer(src_path, dst_path, with_code=False): assert os.path.exists(src_path) os.makedirs(dst_path, exist_ok=True) os.system("cp -rf " + src_path + "/config*.json " + dst_path) os.system("cp -rf " + src_path + "/tokenizer* " + dst_path) if with_code: cur_dir = os.path.dirname(os.path.abspath(__file__)) code_path = os.path.join(cur_dir, 'hf_mistral_moe') os.system("cp -rf " + code_path + "/*.json " + dst_path) def name_to_expert_rank(key): pattern = r'local_experts\.(\d+)\.' expert_rank = int(re.findall(pattern, key)[0]) return expert_rank def load_megatron_model(args, model): model_path = args.load_path tracker_filename = get_checkpoint_tracker_filename(model_path) iteration, release = read_metadata(tracker_filename) head_dim = args.hidden_size // args.num_attention_heads group_per_split = args.num_query_groups // args.target_tensor_model_parallel_size num_local_experts = args.num_experts // args.target_expert_model_parallel_size if args.num_experts else 0 state_dict = {} mid_state = defaultdict(list) if ( args.target_tensor_model_parallel_size == 1 and args.target_pipeline_model_parallel_size == 1 and args.target_expert_model_parallel_size == 1 ): checkpoint_name = get_checkpoint_name(model_path, iteration, release, None, None, None, None, None) state_dict = torch.load(checkpoint_name, weights_only=False)['model'] elif ( args.target_tensor_model_parallel_size == 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts and args.num_experts % args.target_expert_model_parallel_size == 0 ): for ep_rank in range(args.target_expert_model_parallel_size): checkpoint_name = get_checkpoint_name(model_path, iteration, release, None, None, None, True, ep_rank) print(f'load {checkpoint_name}') split_state = torch.load(checkpoint_name, map_location="cpu", weights_only=False)['model'] for k, v in split_state.items(): if 'local_experts' in k: expert_local_rank = name_to_expert_rank(k) expert_rank = expert_local_rank + num_local_experts * ep_rank k = k.replace(f'local_experts.{expert_local_rank}', f'local_experts.{expert_rank}') state_dict[k] = v elif ( args.target_tensor_model_parallel_size > 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts is None ): for tp_rank in range(args.target_tensor_model_parallel_size): checkpoint_name = get_checkpoint_name(model_path, iteration, release, None, tp_rank, None, None, None) print(f'load {checkpoint_name}') split_state = torch.load(checkpoint_name, map_location="cpu", weights_only=False)['model'] for k, v in split_state.items(): mid_state[k].append(v) for k, v in mid_state.items(): if not isinstance(v[0], torch.Tensor) or 'norm' in k: target_v = v[0] elif 'extra_state' in k: target_v = None elif 'embedding' in k or 'output_layer' in k: target_v = torch.cat(v, dim=0) elif 'linear_proj' in k or 'linear_fc2' in k: target_v = torch.cat(v, dim=1) elif 'linear_qkv.weight' in k: viewed = [x.view(group_per_split, -1, head_dim, args.hidden_size) for x in v] target_v = torch.cat(viewed, dim=0).view(-1, args.hidden_size) elif 'linear_qkv.bias' in k: viewed = [x.view(group_per_split, -1) for x in v] target_v = torch.cat(viewed, dim=0).view(-1) elif 'linear_fc1' in k: viewed = [x.view(2, -1, args.hidden_size) for x in v] target_v = torch.cat(viewed, dim=1).view(-1, args.hidden_size) else: raise ValueError state_dict[k] = target_v elif ( args.target_tensor_model_parallel_size > 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts and args.num_experts % args.target_expert_model_parallel_size == 0 ): for tp_rank in range(args.target_tensor_model_parallel_size): for ep_rank in range(args.target_expert_model_parallel_size): checkpoint_name = get_checkpoint_name(model_path, iteration, release, None, tp_rank, None, True, ep_rank) print(f'load {checkpoint_name}') split_state = torch.load(checkpoint_name, map_location="cpu", weights_only=False)['model'] for k, v in split_state.items(): if 'local_experts' in k and 'norm' not in k: local_expert_rank = name_to_expert_rank(k) expert_rank = local_expert_rank + num_local_experts * ep_rank k = k.replace(f'local_experts.{local_expert_rank}', f'local_experts.{expert_rank}') mid_state[k].append(v) elif ep_rank == 0: mid_state[k].append(v) for k, v in mid_state.items(): if not isinstance(v[0], torch.Tensor) or 'norm' in k or 'router' in k: target_v = v[0] elif 'extra_state' in k: target_v = None elif 'embedding' in k or 'output_layer' in k: target_v = torch.cat(v, dim=0) elif 'linear_proj' in k or 'linear_fc2' in k: target_v = torch.cat(v, dim=1) elif 'linear_qkv.weight' in k: viewed = [x.view(group_per_split, -1, head_dim, args.hidden_size) for x in v] target_v = torch.cat(viewed, dim=0).view(-1, args.hidden_size) elif 'linear_qkv.bias' in k: viewed = [x.view(group_per_split, -1) for x in v] target_v = torch.cat(viewed, dim=0).view(-1) elif 'linear_fc1' in k: viewed = [x.view(2, -1, args.hidden_size) for x in v] target_v = torch.cat(viewed, dim=1).view(-1, args.hidden_size) else: print('passed', k) state_dict[k] = target_v else: raise ValueError('not support yet') model.load_state_dict(state_dict) return model def convert_checkpoint_from_megatron_to_transformers(mgmodel, hgmodel, args): query_group = args.num_query_groups hidden_size = args.hidden_size head_dim = hidden_size // args.num_attention_heads num_experts = args.num_experts value_num_per_group = args.num_attention_heads // query_group with torch.no_grad(): hgmodel.model.embed_tokens.weight.copy_(mgmodel.embedding.word_embeddings.weight) for mglayer, hglayer in zip(mgmodel.decoder.layers, hgmodel.model.layers): hglayer.input_layernorm.weight.copy_(mglayer.self_attention.linear_qkv.layer_norm_weight) qkv_weight = mglayer.self_attention.linear_qkv.weight.view(query_group, -1, head_dim, hidden_size) q_weight, k_weight, v_weight = torch.split(qkv_weight, split_size_or_sections=[value_num_per_group, 1, 1], dim=1) hglayer.self_attn.q_proj.weight.copy_(q_weight.reshape(-1, hidden_size)) hglayer.self_attn.k_proj.weight.copy_(k_weight.reshape(-1, hidden_size)) hglayer.self_attn.v_proj.weight.copy_(v_weight.reshape(-1, hidden_size)) hglayer.self_attn.o_proj.weight.copy_(mglayer.self_attention.linear_proj.weight) if num_experts is None: gate_weight, fc1_weight = torch.split(mglayer.mlp.linear_fc1.weight, split_size_or_sections=args.ffn_hidden_size) hglayer.mlp.gate_proj.weight.copy_(gate_weight) hglayer.mlp.up_proj.weight.copy_(fc1_weight) hglayer.mlp.down_proj.weight.copy_(mglayer.mlp.linear_fc2.weight) hglayer.post_attention_layernorm.weight.copy_(mglayer.mlp.linear_fc1.layer_norm_weight) else: hglayer.post_attention_layernorm.weight.copy_(mglayer.pre_mlp_layernorm.weight) hglayer.mlp.gate.weight.copy_(mglayer.mlp.router.weight) for mgexpert, hgexpert in zip(mglayer.mlp.experts.local_experts, hglayer.mlp.experts): gate_weight, fc1_weight = torch.split(mgexpert.linear_fc1.weight, split_size_or_sections=args.ffn_hidden_size) hgexpert.w1.weight.copy_(gate_weight) hgexpert.w3.weight.copy_(fc1_weight) hgexpert.w2.weight.copy_(mgexpert.linear_fc2.weight) hgmodel.model.norm.weight.copy_(mgmodel.decoder.final_layernorm.weight) hgmodel.lm_head.weight.copy_(mgmodel.output_layer.weight) def convert_checkpoint_from_transformers_to_megatron(mgmodel, hgmodel, args, hf_config): num_query_groups = hf_config.num_key_value_heads hidden_dim = hf_config.hidden_size head_dim = hidden_dim // hf_config.num_attention_heads num_experts = args.num_experts with torch.no_grad(): mgmodel.embedding.word_embeddings.weight.copy_(hgmodel.model.embed_tokens.weight) for mglayer, hglayer in zip(mgmodel.decoder.layers, hgmodel.model.layers): mglayer.self_attention.linear_qkv.layer_norm_weight.copy_(hglayer.input_layernorm.weight) q = hglayer.self_attn.q_proj.weight.view([num_query_groups, -1, head_dim, hidden_dim]) k = hglayer.self_attn.k_proj.weight.view([num_query_groups, -1, head_dim, hidden_dim]) v = hglayer.self_attn.v_proj.weight.view([num_query_groups, -1, head_dim, hidden_dim]) qkv = torch.cat([q, k, v], dim=1).view(-1, hidden_dim).contiguous() mglayer.self_attention.linear_qkv.weight.copy_(qkv) mglayer.self_attention.linear_proj.weight.copy_(hglayer.self_attn.o_proj.weight) fc1_weight = torch.cat([hglayer.mlp.gate_proj.weight, hglayer.mlp.up_proj.weight]) if num_experts is None: mglayer.mlp.linear_fc1.weight.copy_(fc1_weight) mglayer.mlp.linear_fc2.weight.copy_(hglayer.mlp.down_proj.weight) mglayer.mlp.linear_fc1.layer_norm_weight.copy_(hglayer.post_attention_layernorm.weight) else: mglayer.pre_mlp_layernorm.weight.copy_(hglayer.post_attention_layernorm.weight) nn.init.normal_(mglayer.mlp.router.weight, mean=0, std=0.02) for expert in mglayer.mlp.experts.local_experts: expert.linear_fc1.weight.copy_(fc1_weight) expert.linear_fc2.weight.copy_(hglayer.mlp.down_proj.weight) mgmodel.decoder.final_layernorm.weight.copy_(hgmodel.model.norm.weight) mgmodel.output_layer.weight.copy_(hgmodel.lm_head.weight) def save_mgmodel(args, mgmodel, load_path, save_path): args.tensor_model_parallel_size = args.target_tensor_model_parallel_size args.pipeline_model_parallel_size = args.target_pipeline_model_parallel_size if args.num_experts is not None: args.expert_model_parallel_size = args.target_expert_model_parallel_size # Saving config and tokenzier files copy_huggingface_tokenizer(load_path, save_path) tracker_filepath = os.path.join(save_path, 'latest_checkpointed_iteration.txt') with open(tracker_filepath, "w") as f: f.write("release") head_dim = args.hidden_size // args.num_attention_heads group_per_split = args.num_query_groups // args.target_tensor_model_parallel_size full_model = mgmodel.state_dict_for_save_checkpoint() for k in list(full_model.keys()): if 'extra_state' in k: # NOTE: since TE 1.14, fp8 metadata will be saved as tensor. # Always drop these values in the MG ckpt to avoid potential issue. # This should work fine because fp8 metadata is not supported by HF ckpt. full_model[k] = None elif full_model[k] is None: full_model.pop(k) pattern = r'local_experts\.(\d+)\.' num_local_experts = args.num_experts // args.target_expert_model_parallel_size if args.num_experts else 0 if ( args.target_tensor_model_parallel_size == 1 and args.target_pipeline_model_parallel_size == 1 and args.target_expert_model_parallel_size == 1 ): checkpoint_name = get_checkpoint_name(save_path, 0, True) save_state_dict(args, [full_model], checkpoint_name) elif ( args.target_tensor_model_parallel_size == 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts and args.num_experts % args.target_expert_model_parallel_size == 0 ): for ep_rank in range(args.target_expert_model_parallel_size): model_split = {} checkpoint_name = get_checkpoint_name(save_path, 0, True, None, None, None, True, ep_rank) print(f'save ep_rank {ep_rank} model to {checkpoint_name}') for k, v in full_model.items(): if 'local_experts' in k: expert_rank = int(re.findall(pattern, k)[0]) if expert_rank // num_local_experts != ep_rank: continue expert_local_rank = expert_rank % args.target_expert_model_parallel_size k = k.replace(f'local_experts.{expert_rank}', f'local_experts.{expert_local_rank}') model_split[k] = v save_state_dict(args, [model_split], checkpoint_name) elif ( args.target_tensor_model_parallel_size > 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts is None ): for tp_rank in range(args.target_tensor_model_parallel_size): model_split = {} checkpoint_name = get_checkpoint_name(save_path, 0, True, None, tp_rank) print(f'tensor_parallel, save model to {checkpoint_name}') for k, v in full_model.items(): if not isinstance(v, torch.Tensor): target_v = v elif 'linear_qkv.weight' in k and 'norm' not in k: viewed = v.view(args.num_query_groups, -1, head_dim, args.hidden_size) viewed = viewed[group_per_split * tp_rank: group_per_split * (tp_rank + 1)] target_v = viewed.view(-1, args.hidden_size) elif 'linear_qkv.bias' in k and 'norm' not in k: viewed = v.view(args.num_query_groups, -1, head_dim) viewed = viewed[group_per_split * tp_rank: group_per_split * (tp_rank + 1)] target_v = viewed.view(-1) elif 'linear_proj' in k or 'linear_fc2' in k: seg = v.shape[1] // args.target_tensor_model_parallel_size target_v = v[:, seg * tp_rank: seg * (tp_rank + 1)] elif 'embedding' in k or 'output_layer' in k: seg = v.shape[0] // args.target_tensor_model_parallel_size target_v = v[seg * tp_rank: seg * (tp_rank + 1)] elif 'linear_fc1' in k and 'norm' not in k: viewed = v.view(-1, args.ffn_hidden_size, args.hidden_size) seg = args.ffn_hidden_size // args.target_tensor_model_parallel_size target_v = viewed[:, seg * tp_rank: seg * (tp_rank + 1), :].reshape(-1, args.hidden_size) else: target_v = v model_split[k] = target_v save_state_dict(args, [model_split], checkpoint_name) elif ( args.target_tensor_model_parallel_size > 1 and args.target_pipeline_model_parallel_size == 1 and args.num_experts and args.num_experts % args.target_expert_model_parallel_size == 0 ): for tp_rank in range(args.target_tensor_model_parallel_size): for ep_rank in range(args.target_expert_model_parallel_size): model_split = {} checkpoint_name = get_checkpoint_name(save_path, 0, True, None, tp_rank, None, True, ep_rank) for k, v in full_model.items(): if not isinstance(v, torch.Tensor): target_v = v elif 'linear_qkv.weight' in k and 'norm' not in k: viewed = v.view(args.num_query_groups, -1, head_dim, args.hidden_size) viewed = viewed[group_per_split * tp_rank: group_per_split * (tp_rank + 1)] target_v = viewed.view(-1, args.hidden_size) elif 'linear_qkv.bias' in k and 'norm' not in k: viewed = v.view(args.num_query_groups, -1, head_dim) viewed = viewed[group_per_split * tp_rank: group_per_split * (tp_rank + 1)] target_v = viewed.view(-1) elif 'linear_proj' in k: seg = v.shape[1] // args.target_tensor_model_parallel_size target_v = v[:, seg * tp_rank: seg * (tp_rank + 1)] elif 'embedding' in k or 'output_layer' in k: seg = v.shape[0] // args.target_tensor_model_parallel_size target_v = v[seg * tp_rank: seg * (tp_rank + 1)] elif 'local_experts' in k: expert_rank = int(re.findall(pattern, k)[0]) if expert_rank // num_local_experts != ep_rank: continue expert_local_rank = expert_rank % num_local_experts if 'linear_fc1' in k and 'norm' not in k: viewed = v.view(-1, args.ffn_hidden_size, args.hidden_size) seg = args.ffn_hidden_size // args.target_tensor_model_parallel_size target_v = viewed[:, seg * tp_rank: seg * (tp_rank + 1), :].reshape(-1, args.hidden_size) elif 'linear_fc2' in k: seg = v.shape[1] // args.target_tensor_model_parallel_size target_v = v[:, seg * tp_rank: seg * (tp_rank + 1)] k = k.replace(f'local_experts.{expert_rank}', f'local_experts.{expert_local_rank}') else: target_v = v model_split[k] = target_v save_state_dict(args, [model_split], checkpoint_name) else: raise ValueError('not support pp convert') print(f'megatron model is save to {save_path}') def add_ckpt_args(parser): parser = get_patch_args(parser) parser = add_checkpointing_args(parser) parser = add_megatron_checkpoint_args(parser) parser = add_transformers_checkpoint_args(parser) return parser def main(): initialize_megatron(extra_args_provider=add_ckpt_args) args = get_args() hf_config, hf_model = create_huggingface_model(args) mg_model = create_megatron_model(args, hf_config) if args.convert_checkpoint_from_megatron_to_transformers: load_megatron_model(args, mg_model) convert_checkpoint_from_megatron_to_transformers(mg_model, hf_model, args) args.save = args.save_path args.save_safetensors = False save_hfmodel(args, hf_model) else: hf_model.from_pretrained(args.load_path) convert_checkpoint_from_transformers_to_megatron(mg_model, hf_model, args, hf_config) save_mgmodel(args, mg_model, args.load_path, args.save_path) if __name__ == "__main__": main()