Spaces:
Build error
Build error
File size: 5,238 Bytes
28451f7 |
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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 |
# 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.
from typing import List, Tuple, Union
import torch
import transformers
from transformers import T5EncoderModel, T5TokenizerFast
from cosmos_predict1.utils import log
transformers.logging.set_verbosity_error()
class CosmosT5TextEncoder(torch.nn.Module):
"""Handles T5 text encoding operations."""
def __init__(self, model_name: str = "google-t5/t5-11b", device: str = "cuda", cache_dir: str = "~/.cache"):
"""Initializes the T5 tokenizer and encoder.
Args:
model_name: The name of the T5 model to use.
device: The device to use for computations.
"""
super().__init__()
try:
self.tokenizer = T5TokenizerFast.from_pretrained(cache_dir, cache_dir=cache_dir)
self.text_encoder = T5EncoderModel.from_pretrained(cache_dir, cache_dir=cache_dir).to(device)
except Exception as e:
log.warning(f"Failed to load T5 model using cache_dir '{cache_dir}', falling back to default location: {e}")
self.tokenizer = T5TokenizerFast.from_pretrained(model_name)
self.text_encoder = T5EncoderModel.from_pretrained(model_name).to(device)
self.text_encoder.eval()
self.device = device
@torch.inference_mode()
def encode_prompts(
self, prompts: Union[str, List[str]], max_length: int = 512
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Encodes text prompts into hidden state representations using a T5 encoder.
This function tokenizes the input prompts, processes them through a T5 text encoder,
and returns the last hidden states. The encoded outputs beyond the actual sequence
length are zero-padded. All prompts in a batch are padded to max_length.
Args:
prompts: Input text to encode. Can be a single string or a list of strings.
max_length: Maximum sequence length for tokenization and padding. Longer
sequences will be truncated. Defaults to 512.
return_mask: If True, returns the attention mask along with encoded text.
Defaults to False.
Returns:
If return_mask is False:
torch.Tensor: Encoded text embeddings of shape (batch_size, max_length, hidden_size).
If return_mask is True:
tuple[torch.Tensor, torch.Tensor]: A tuple containing:
- Encoded text embeddings of shape (batch_size, max_length, hidden_size)
- Attention mask of shape (batch_size, max_length) as boolean tensor
Raises:
ValueError: If the input prompts list is empty.
Example:
>>> encoder = CosmosT5TextEncoder()
>>> prompts = ["Hello world", "Another example"]
>>> embeddings = encoder.encode_prompts(prompts, max_length=128)
"""
if isinstance(prompts, str):
prompts = [prompts]
if not prompts:
raise ValueError("The input prompt list is empty.")
batch_encoding = self.tokenizer.batch_encode_plus(
prompts,
return_tensors="pt",
truncation=True,
padding="max_length",
max_length=max_length,
return_length=True,
return_offsets_mapping=False,
)
input_ids = batch_encoding.input_ids.to(self.device)
attn_mask = batch_encoding.attention_mask.to(self.device)
outputs = self.text_encoder(input_ids=input_ids, attention_mask=attn_mask)
encoded_text = outputs.last_hidden_state
lengths = attn_mask.sum(dim=1).cpu()
for batch_id in range(encoded_text.shape[0]):
encoded_text[batch_id][lengths[batch_id] :] = 0
return encoded_text, attn_mask
class DummyT5TextEncoder(torch.nn.Module):
def __init__(self, device: str = "cuda"):
super().__init__()
self.device = device
@torch.inference_mode()
def encode_prompts(
self, prompts: Union[str, List[str]], max_length: int = 512
) -> Tuple[torch.Tensor, torch.Tensor]:
if isinstance(prompts, str):
prompts = [prompts]
if not prompts:
raise ValueError("The input prompt list is empty.")
batch_size = len(prompts)
dummy_text_embedding = torch.zeros(batch_size, max_length, 1024, device=self.device)
dummy_text_mask = torch.zeros(batch_size, max_length, device=self.device, dtype=torch.bool)
dummy_text_mask[0] = True
return dummy_text_embedding, dummy_text_mask
|