AssertionError: g_tokens must be None if l_tokens is None

I have an issue with trying to train sd3.5L when sampling during training, please help :(

```
Traceback (most recent call last):
File "C:\sd-scripts\sd3_train.py", line 1074, in
train(args)
File "C:\sd-scripts\sd3_train.py", line 949, in train
sd3_train_utils.sample_images(
File "C:\sd-scripts\library\sd3_train_utils.py", line 429, in sample_images
sample_image_inference(
File "C:\sd-scripts\library\sd3_train_utils.py", line 550, in sample_image_inference
lg_out, t5_out, pooled, l_attn_mask, g_attn_mask, t5_attn_mask = encode_prompt(prompt)
^^^^^^^^^^^^^^^^^^^^^
File "C:\sd-scripts\library\sd3_train_utils.py", line 538, in encode_prompt
encoded_text_encoder_conds = encoding_strategy.encode_tokens(tokenize_strategy, text_encoders, tokens_and_masks)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\sd-scripts\library\strategy_sd3.py", line 97, in encode_tokens
assert g_tokens is None, "g_tokens must be None if l_tokens is None"
^^^^^^^^^^^^^^^^
AssertionError: g_tokens must be None if l_tokens is None

```
Accelerate launch --mixed_precision bf16 --num_processes 1 --num_machines 1 --num_cpu_threads_per_process 2 C:/sd-scripts/sd3_train.py --config_file C:/train/test.toml

cfg:
```
clip_l = "C:/train/sd3.5/text_encoder/model.safetensors"
clip_g = "C:/train/sd3.5/text_encoder_2/model.safetensors"
t5xxl = "C:/train/sd3.5/text_encoders/t5xxl_fp16.safetensors"
pretrained_model_name_or_path = "C:/train/sd3.5/sd3.5_large.safetensors"
skip_cache_check = false
blockwise_fused_optimizers = false
fused_backward_pass = true
disable_mmap_load_safetensors = true
highvram = true
cache_latents = true
cache_latents_to_disk = true
cache_text_encoder_outputs = true
cache_text_encoder_outputs_to_disk = true
caption_extension = ".txt"
dataset_config = "C:/train/test.toml"
gradient_accumulation_steps = 1
gradient_checkpointing = true
#enable_scaled_pos_embed = true
max_grad_norm = 0.0
huber_c = 0.1
huber_schedule = "snr"
logging_dir = "C:/train/tensorboard"
loss_type = "l2"
lr_scheduler = "constant_with_warmup"
lr_scheduler_args = []
max_timestep = 1000
min_snr_gamma = 5
ip_noise_gamma = 0.1
ip_noise_gamma_random_strength = true
noise_offset_type = "Original"
output_dir = "C:/train"
output_name = "last"
persistent_data_loader_workers = true
max_data_loader_n_workers = 2
sample_every_n_epochs = 1
sample_prompts = "C:/train/sample/test_prompt.txt"
sample_sampler = "euler_a"
save_every_n_epochs = 100
save_model_as = "diffusers"
save_precision = "bf16"
save_state = true
mixed_precision = "bf16"
sdpa = true
seed = 1234
max_train_epochs = 1000
optimizer_args = [ "relative_step=False", "scale_parameter=True", "warmup_init=False", "weight_decay=0.05"]
optimizer_type = "Adafactor"
lr_warmup_steps = 50
learning_rate = 1e-5
blocks_to_swap = 20
train_batch_size = 12
train_blocks = "all"
wandb_run_name = "last"
```

ds:
```
[general]
# define common settings here
flip_aug = false
color_aug = false
shuffle_caption = false
caption_tag_dropout_rate = 0
caption_extension = ".txt"

[[datasets]]
# define the first resolution here
enable_bucket = true
min_bucket_reso = 64
max_bucket_reso = 512
bucket_reso_steps = 32
bucket_no_upscale = true
resolution = [512, 512]

[[datasets.subsets]]
image_dir = "F:/1_testp1"
num_repeats = 1

[[datasets]]
# define the first resolution here
enable_bucket = true
min_bucket_reso = 64
max_bucket_reso = 512
bucket_reso_steps = 32
bucket_no_upscale = true
resolution = [512, 512]

[[datasets.subsets]]
image_dir = "F:/1_testp2"
num_repeats = 1
```

Start with library/strategy_sd3.py at encode_tokens and trace its callers in library/sd3_train_utils.py, especially sample_image_inference and sample_images. Reproduce the failure with the supplied SD3.5 training configuration and sampling prompt, then verify that sampling during training no longer raises the reported assertion.

Written by the indexing model from the issue text.