in toolkits/model_checkpoints_convertor/qwen/hf2megablocks_qwen1.5.py [0:0]
def save_mgmodel(args, mgmodel, load_path, save_path):
# 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 full_model[k] is None or "_extra_state" in k:
full_model.pop(k)
pattern = r'local_experts\.(\d+)\.'
num_local_experts = args.moe_num_experts // args.target_expert_model_parallel_size if args.moe_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_names(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.moe_num_experts
and args.moe_num_experts % args.target_expert_model_parallel_size == 0
):
checkpoint_name = get_checkpoint_names(save_path, 0, False, True)[0]
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.moe_num_experts is None
):
for tp_rank in range(args.target_tensor_model_parallel_size):
model_split = {}
checkpoint_name = get_checkpoint_names(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.moe_num_experts
and args.moe_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_names(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}')