File size: 2,286 Bytes
ecf8cbe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 |
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.
"""Utility to convert weights to safetensors."""
import argparse
import torch
from .configuration_embed1 import CosmosEmbed1Config
from .modeling_embed1 import CosmosEmbed1
def parse_args():
parser = argparse.ArgumentParser(description="Save model weights with optional format conversion and sharding.")
parser.add_argument("--input_weights", type=str, required=True, help="Path to the input .pt weights file")
parser.add_argument(
"--output_weights",
type=str,
required=True,
help="Path to the output directory where safetensors weights will be saved",
)
return parser.parse_args()
def main():
args = parse_args()
model = CosmosEmbed1(CosmosEmbed1Config()).to("cuda", dtype=torch.bfloat16)
# remove tensor sharing
model.qformer.cls.predictions.decoder.weight = torch.nn.Parameter(
model.qformer.cls.predictions.decoder.weight.clone()
)
model.qformer.bert.embeddings.word_embeddings.weight = torch.nn.Parameter(
model.qformer.bert.embeddings.word_embeddings.weight.clone()
)
model.qformer.cls.predictions.decoder.bias = torch.nn.Parameter(model.qformer.cls.predictions.decoder.bias.clone())
model.qformer.cls.predictions.bias = torch.nn.Parameter(model.qformer.cls.predictions.bias.clone())
with open(args.input_weights, "rb") as fp:
state_dict = torch.load(fp)
model.load_state_dict(state_dict, strict=True)
model.save_pretrained(
args.output_weights,
safe_serialization=True,
max_shard_size="500MB",
)
if __name__ == "__main__":
main()
|