in toolkits/model_checkpoints_convertor/qwen/hf2megatron_qwen1.0.py [0:0]
def convert_checkpoint_from_transformers_to_megatron(args):
"""
Convert a checkpoint from HuggingFace Transformers to Megatron-LM. This allows converted checkpoints with variable
tensor parallelism and pipeline parallelism sizes. It takes as input a checkpoint from HuggingFace Transformers
which can have multiple shards.
Args:
args (argparse.Namespace): the arguments to the script
"""
os.makedirs(args.save_path, exist_ok=True)
# Search in directory above this
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
if args.megatron_path is not None:
sys.path.insert(0, args.megatron_path)
try:
from megatron.tokenizer.tokenizer import _vocab_size_with_padding
except ModuleNotFoundError:
print("Unable to import Megatron, please specify the path to Megatron using --megatron-path. Exiting.")
exit(1)
# load the transformers model state dict and config
sub_dirs = [x for x in os.listdir(args.load_path) if x.startswith("pytorch_model") or x.endswith("safetensors") ]
if len(sub_dirs) == 1:
checkpoint_name = "pytorch_model.bin"
hf_state_dict = torch.load(os.path.join(args.load_path, checkpoint_name), map_location="cpu")
else:
if args.model_name == "qwen-7b" or args.model_name == "qwen-14b" or args.model_name == "qwen-72b":
hf_state_dict = AutoModelForCausalLM.from_pretrained(args.load_path, trust_remote_code=True).state_dict()
else:
print("Unrecognized model name, choose from qwen-7b, qwen-14b, qwen-72b!")
exit(1)
config = AutoConfig.from_pretrained(args.load_path, trust_remote_code=True)
internal_state_dict = {}
for layer_id in range(config.num_hidden_layers):
internal_state_dict['transformer.layers.'+str(layer_id)+'.self_attn.query_key_value.weight'] =\
hf_state_dict['transformer.h.'+str(layer_id)+'.attn.c_attn.weight']
internal_state_dict['transformer.layers.'+str(layer_id)+'.self_attn.query_key_value.bias'] =\
hf_state_dict['transformer.h.'+str(layer_id)+'.attn.c_attn.bias']
internal_state_dict['transformer.layers.' + str(layer_id) + '.self_attn.dense.weight'] =\
hf_state_dict['transformer.h.' + str(layer_id) + '.attn.c_proj.weight']
dense_h_to_4h_1_weight = hf_state_dict[
'transformer.h.' + str(layer_id) + '.mlp.w1.weight']
dense_h_to_4h_2_weight = hf_state_dict[
'transformer.h.' + str(layer_id) + '.mlp.w2.weight']
internal_state_dict['transformer.layers.' + str(layer_id) + '.mlp.dense_h_to_4h_1.weight'] =\
dense_h_to_4h_1_weight
internal_state_dict['transformer.layers.' + str(layer_id) + '.mlp.dense_h_to_4h_2.weight'] =\
dense_h_to_4h_2_weight
internal_state_dict['transformer.layers.' + str(layer_id) + '.mlp.dense_4h_to_h.weight'] =\
hf_state_dict['transformer.h.' + str(layer_id) + '.mlp.c_proj.weight']
internal_state_dict['transformer.layers.' + str(layer_id) + '.input_layernorm.weight'] = hf_state_dict[
'transformer.h.' + str(layer_id) + '.ln_1.weight']
internal_state_dict['transformer.layers.' + str(layer_id) + '.post_attention_layernorm.weight'] = hf_state_dict[
'transformer.h.' + str(layer_id) + '.ln_2.weight']
internal_state_dict["transformer.word_embeddings.weight"] = hf_state_dict['transformer.wte.weight']
internal_state_dict["transformer.final_layernorm.weight"] = hf_state_dict['transformer.ln_f.weight']
internal_state_dict["transformer.lm_head.weight"] = hf_state_dict['lm_head.weight']
state_dict = internal_state_dict
# Saving config and tokenzier files
os.system("cp -rf "+args.load_path+"/*.json "+args.save_path)
os.system("cp -rf " + args.load_path + "/qwen.tiktoken " + args.save_path)
os.system("cp -rf " + args.load_path + "/tokeni* " + args.save_path)
os.system("cp -rf " + args.load_path + "/*.py " + args.save_path)
# Saving the tracker file
tracker_filepath = os.path.join(args.save_path, "latest_checkpointed_iteration.txt")
with open(tracker_filepath, "w") as f:
f.write("release")
# create `release` dir in args.load_path
release_dir = os.path.join(args.save_path, "release")
os.makedirs(release_dir, exist_ok=True)
# megatron args
megatron_args = {
"orig_vocab_size": config.vocab_size,
"hidden_size": config.hidden_size,
"num_layers": config.num_hidden_layers,
"num_attention_heads": config.num_attention_heads,
"tensor_model_parallel_size": args.target_tensor_model_parallel_size,
"pipeline_model_parallel_size": args.target_pipeline_model_parallel_size,
"data_parallel_size": args.target_data_parallel_size,
"make_vocab_size_divisible_by": args.make_vocab_size_divisible_by,
"rank": 0,
"tokenizer_type": "GPT2BPETokenizer",
}
margs = types.SimpleNamespace()
for k, v in megatron_args.items():
setattr(margs, k, v)
# params dtype
if args.target_params_dtype == "fp16":
dtype = torch.float16
elif args.target_params_dtype == "bf16":
dtype = torch.bfloat16
else:
dtype = torch.float32
setattr(margs, "params_dtype", dtype)
# Convert.
print("Converting")
output_state_dict = []
for i in range(args.target_tensor_model_parallel_size):
output_state_dict.append({})
# Embedding layer
print("converting embedding layer")
word_embedding = state_dict["transformer.word_embeddings.weight"].to(dtype)
lm_head = state_dict["transformer.lm_head.weight"].to(dtype)
orig_vocab_size = config.vocab_size
#padded_vocab_size = _vocab_size_with_padding(orig_vocab_size, margs)
padded_vocab_size = orig_vocab_size
setattr(margs, "padded_vocab_size", padded_vocab_size)
# Cut out extra padding we don't need
if args.extra_num_vocabs == 0:
full_word_embed = word_embedding
full_lm_head = lm_head
else:
new_embeddings = torch.nn.Embedding(args.extra_num_vocabs, word_embedding.shape[1])
# initialize all new embeddings (in particular added tokens)
_init_embedding_weights(new_embeddings)
full_word_embed = torch.cat([word_embedding, new_embeddings.weight])
full_lm_head = torch.cat([lm_head, new_embeddings.weight])
# Split into new tensor model parallel sizes
out_word_embed = torch.chunk(full_word_embed, args.target_tensor_model_parallel_size, dim=0)
for i in range(args.target_tensor_model_parallel_size):
word_emb_dict = get_element_from_dict_by_path(
output_state_dict[i], "model.language_model.embedding.word_embeddings"
)
word_emb_dict["weight"] = out_word_embed[i]
out_lm_head = torch.chunk(full_lm_head, args.target_tensor_model_parallel_size, dim=0)
for i in range(args.target_tensor_model_parallel_size):
lm_head_dict = get_element_from_dict_by_path(
output_state_dict[i], "model.lm_head"
)
lm_head_dict["weight"] = out_lm_head[i]
# Transformer layers
print("converting transformer layers")
if config.num_hidden_layers % args.target_pipeline_model_parallel_size != 0:
raise ValueError(
f"Number of layers ({config.num_hidden_layers}) must be divisible by number of pipeline parallelism"
f" ({args.target_pipeline_model_parallel_size})"
)
num_layers = config.num_hidden_layers // args.target_pipeline_model_parallel_size
layer_re = re.compile("transformer.layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
# The number of heads.
heads = config.num_attention_heads
# The hidden_size per head.
hidden_size_per_head = config.hidden_size // config.num_attention_heads
for pp_rank in range(args.target_pipeline_model_parallel_size):
layer_offset = pp_rank * num_layers
if pp_rank > 0:
output_state_dict = []
for i in range(args.target_tensor_model_parallel_size):
output_state_dict.append({})
for layer in range(num_layers):
pp_layer_id = layer + layer_offset
layers_to_copy = [
layer_name
for layer_name in state_dict.keys()
if layer_name.startswith(f"transformer.layers.{pp_layer_id}.")
]
for layer_name in layers_to_copy:
m = layer_re.match(layer_name)
# Stop if that's not a layer
if m is None:
break
# The index of the layer.
_ = 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)
params = state_dict[layer_name].to(dtype)
# handle layernorm
if op_name.startswith("input_layernorm") or op_name.startswith("post_attention_layernorm"):
out_name = "input_layernorm" if op_name.endswith("input_layernorm") else "post_attention_layernorm"
layer_name = f"layers.{layer}.{out_name}.{weight_or_bias}"
elif op_name.startswith("self_attn.query_key_value") and weight_or_bias == "weight":
# transformers stores D X (3*D) but Megatron-LM expects (3*D) X D.
params = transformers_to_megatron_fix_query_key_value_ordering(
params,
3.0,
3,
heads,
hidden_size_per_head,
)
layer_name = f"layers.{layer}.self_attention.query_key_value.{weight_or_bias}"
elif op_name.startswith("self_attn.query_key_value") and weight_or_bias == "bias":
# transformers stores D X (3*D) but Megatron-LM expects (3*D) X D.
params = transformers_to_megatron_fix_query_key_value_ordering(
params,
3.0,
3,
heads,
hidden_size_per_head,
)
layer_name = f"layers.{layer}.self_attention.query_key_value.{weight_or_bias}"
# handle attention and mlp weights
elif weight_or_bias == "weight":
out_name = transformers_to_megatron.get(op_name, None)
if out_name is None:
continue
#params = params.transpose(0, 1)
layer_name = f"layers.{layer}.{out_name}.{weight_or_bias}"
# handle attention and mlp bias
elif weight_or_bias == 'bias':
out_name = transformers_to_megatron.get(op_name, None)
if out_name is None:
continue
layer_name = f'layers.{layer}.{out_name}.{weight_or_bias}'
# skip
else:
continue
if op_name + "." + weight_or_bias in tensor_parallel_params:
dim = 1 if op_name in ["self_attn.dense", "mlp.dense_4h_to_h"] else 0
params = torch.chunk(params, args.target_tensor_model_parallel_size, dim=dim)
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.language_model.encoder")
params_dict[layer_name] = (
params[i].clone() if (op_name + "." + weight_or_bias in tensor_parallel_params) else params.clone()
)
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i],
"model.language_model.encoder")
dense_h_to_4h_1_name = 'mlp.dense_h_to_4h_1.weight'
dense_h_to_4h_1_layer_name = f"layers.{layer}.{dense_h_to_4h_1_name}"
dense_h_to_4h_1_weight = params_dict[dense_h_to_4h_1_layer_name]
dense_h_to_4h_2_name = 'mlp.dense_h_to_4h_2.weight'
dense_h_to_4h_2_layer_name = f"layers.{layer}.{dense_h_to_4h_2_name}"
dense_h_to_4h_2_weight = params_dict[dense_h_to_4h_2_layer_name]
dense_h_to_4h_name = 'mlp.dense_h_to_4h.weight'
dense_h_to_4h_layer_name = f"layers.{layer}.{dense_h_to_4h_name}"
params_dict[dense_h_to_4h_layer_name] = torch.cat(
[dense_h_to_4h_2_weight, dense_h_to_4h_1_weight], dim=0)
del params_dict[dense_h_to_4h_1_layer_name]
del params_dict[dense_h_to_4h_2_layer_name]
if pp_rank == args.target_pipeline_model_parallel_size - 1:
# handle final layernorm
for weight_or_bias in ["weight"]:
params = state_dict[f"transformer.final_layernorm.{weight_or_bias}"].to(dtype)
layer_name = f"final_layernorm.{weight_or_bias}"
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.language_model.encoder")
params_dict[layer_name] = params.clone()
# add the LM head
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.word_embeddings_for_head")
params_dict["weight"] = out_word_embed[i].clone()
# add the LM head
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.language_model.output_layer")
params_dict["weight"] = out_lm_head[i].clone()
# saving the state dict as per the tp_rank and pp_rank
for tp_rank in range(args.target_tensor_model_parallel_size):
output_state_dict[tp_rank]["checkpoint_version"] = 3.0
output_state_dict[tp_rank]["args"] = margs
checkpoint_dir = (
f"mp_rank_{tp_rank:02d}"
if args.target_pipeline_model_parallel_size == 1
else f"mp_rank_{tp_rank:02d}_{pp_rank:03d}"
)
checkpoint_name = "model_optim_rng.pt"
checkpoint_dir = os.path.join(release_dir, checkpoint_dir)
os.makedirs(checkpoint_dir, exist_ok=True)
checkpoint_path = os.path.join(checkpoint_dir, checkpoint_name)
if args.print_checkpoint_structure:
print(
f"Checkpoint structure of model state dict shard belonging to TP rank {tp_rank} and PP rank"
f" {pp_rank}:"
)
recursive_print(None, output_state_dict[tp_rank])
torch.save(clone_state_dict(output_state_dict[tp_rank]), checkpoint_path)