unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-dynamic-bnb-4bit loading error

[THEOLIN]

I use Colab (A100, 80GB VRAM) to fine-tune Llama-4 model.

When I load "unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-dynamic-bnb-4bit" model, it shows "AttributeError: SequentialLlama4TextExperts has no attribute down_proj".

Same promblem in other unsloth/Llama-4 series models I tried:
unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-bnb-4bit
unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-bnb-8bit

error detal (see below):

---

from unsloth import FastLanguageModel
import torch
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-dynamic-bnb-4bit",
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
🦥 Unsloth Zoo will now patch everything to make training faster!
==((====))== Unsloth 2025.11.3: Fast Llama4 patching. Transformers: 4.57.1.
\ /| NVIDIA A100-SXM4-80GB. Num GPUs = 1. Max memory: 79.318 GB. Platform: Linux.
O^O/ _/ \ Torch: 2.9.0+cu128. CUDA: 8.0. CUDA Toolkit: 12.8. Triton: 3.5.0
\ / Bfloat16 = TRUE. FA [Xformers = 0.0.33.post1. FA2 = False]
"-____-" Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
model.safetensors.index.json:  382k/? [00:00<00:00, 39.5MB/s]model-00001-of-00013.safetensors: 100% 5.00G/5.00G [00:26<00:00, 105MB/s]model-00002-of-00013.safetensors: 100% 4.81G/4.81G [00:44<00:00, 80.6MB/s]model-00003-of-00013.safetensors: 100% 4.82G/4.82G [00:26<00:00, 426MB/s]model-00004-of-00013.safetensors: 100% 4.98G/4.98G [00:44<00:00, 98.3MB/s]model-00005-of-00013.safetensors: 100% 4.73G/4.73G [00:09<00:00, 171MB/s]model-00006-of-00013.safetensors: 100% 4.73G/4.73G [00:11<00:00, 321MB/s]model-00007-of-00013.safetensors: 100% 4.89G/4.89G [00:12<00:00, 600MB/s]model-00008-of-00013.safetensors: 100% 4.98G/4.98G [00:13<00:00, 388MB/s]model-00009-of-00013.safetensors: 100% 4.74G/4.74G [00:12<00:00, 673MB/s]model-00010-of-00013.safetensors: 100% 4.98G/4.98G [00:23<00:00, 60.1MB/s]model-00011-of-00013.safetensors: 100% 4.89G/4.89G [00:42<00:00, 93.2MB/s]model-00012-of-00013.safetensors: 100% 4.99G/4.99G [00:15<00:00, 180MB/s]model-00013-of-00013.safetensors: 100% 3.17G/3.17G [00:13<00:00, 148MB/s]Loading checkpoint shards:   0% 0/13 [00:00<?, ?it/s]---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
/tmp/ipython-input-1968070144.py in <cell line: 0>()
5 load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
6
----> 7 model, tokenizer = FastLanguageModel.from_pretrained(
8 model_name = "unsloth/Llama-4-Scout-17B-16E-Instruct-unsloth-dynamic-bnb-4bit",
9 max_seq_length = max_seq_length,

12 frames/usr/local/lib/python3.12/dist-packages/unsloth/models/loader.py in from_pretrained(model_name, max_seq_length, dtype, load_in_4bit, load_in_8bit, load_in_16bit, full_finetuning, token, device_map, rope_scaling, fix_tokenizer, trust_remote_code, use_gradient_checkpointing, resize_model_vocab, revision, use_exact_model_name, offload_embedding, float32_mixed_precision, fast_inference, gpu_memory_utilization, float8_kv_cache, random_state, max_lora_rank, disable_log_stats, qat_scheme, *args, **kwargs)
449 # dispatch_model = FastGraniteModel
450 else:
--> 451 return FastModel.from_pretrained(
452 model_name = old_model_name,
453 max_seq_length = max_seq_length,

/usr/local/lib/python3.12/dist-packages/unsloth/models/loader.py in from_pretrained(model_name, max_seq_length, dtype, load_in_4bit, load_in_8bit, load_in_16bit, full_finetuning, token, device_map, rope_scaling, fix_tokenizer, trust_remote_code, use_gradient_checkpointing, resize_model_vocab, revision, return_logits, fullgraph, use_exact_model_name, auto_model, whisper_language, whisper_task, unsloth_force_compile, offload_embedding, float32_mixed_precision, fast_inference, gpu_memory_utilization, float8_kv_cache, random_state, max_lora_rank, disable_log_stats, qat_scheme, *args, **kwargs)
1063 auto_model = AutoModelForVision2Seq if is_vlm else AutoModelForCausalLM
1064
-> 1065 model, tokenizer = FastBaseModel.from_pretrained(
1066 model_name = model_name,
1067 max_seq_length = max_seq_length,

/usr/local/lib/python3.12/dist-packages/unsloth/models/vision.py in from_pretrained(model_name, max_seq_length, dtype, load_in_4bit, load_in_8bit, load_in_16bit, full_finetuning, token, device_map, trust_remote_code, model_types, tokenizer_name, auto_model, use_gradient_checkpointing, supports_sdpa, whisper_language, whisper_task, auto_config, offload_embedding, float32_mixed_precision, fast_inference, gpu_memory_utilization, float8_kv_cache, random_state, max_lora_rank, disable_log_stats, unsloth_vllm_standby, **kwargs)
647 raise_handler = RaiseUninitialized()
648 if not fast_inference:
--> 649 model = auto_model.from_pretrained(
650 model_name,
651 device_map = device_map,

/usr/local/lib/python3.12/dist-packages/transformers/models/auto/auto_factory.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
602 if model_class.config_class == config.sub_configs.get("text_config", None):
603 config = config.get_text_config()
--> 604 return model_class.from_pretrained(
605 pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
606 )

/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py in _wrapper(*args, **kwargs)
275 old_dtype = torch.get_default_dtype()
276 try:
--> 277 return func(*args, **kwargs)
278 finally:
279 torch.set_default_dtype(old_dtype)

/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py in from_pretrained(cls, pretrained_model_name_or_path, config, cache_dir, ignore_mismatched_sizes, force_download, local_files_only, token, revision, use_safetensors, weights_only, *model_args, **kwargs)
5046 offload_index,
5047 error_msgs,
-> 5048 ) = cls._load_pretrained_model(
5049 model,
5050 state_dict,

/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py in _load_pretrained_model(cls, model, state_dict, checkpoint_files, pretrained_model_name_or_path, ignore_mismatched_sizes, sharded_metadata, device_map, disk_offload_folder, dtype, hf_quantizer, keep_in_fp32_regex, device_mesh, key_mapping, weights_only)
5466
5467 for args in args_list:
-> 5468 _error_msgs, disk_offload_index = load_shard_file(args)
5469 error_msgs += _error_msgs
5470

/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py in load_shard_file(args)
841 # Skip it with fsdp on ranks other than 0
842 elif not (is_fsdp_enabled() and not is_local_dist_rank_0() and not is_quantized):
--> 843 disk_offload_index = _load_state_dict_into_meta_model(
844 model,
845 state_dict,

/usr/local/lib/python3.12/dist-packages/torch/utils/_contextlib.py in decorate_context(*args, **kwargs)
118 def decorate_context(*args, **kwargs):
119 with ctx_factory():
--> 120 return func(*args, **kwargs)
121
122 return decorate_context

/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py in _load_state_dict_into_meta_model(model, state_dict, shard_file, reverse_renaming_mapping, device_map, disk_offload_folder, disk_offload_index, hf_quantizer, keep_in_fp32_regex, device_mesh)
772 else:
773 # TODO naming is stupid it loads it as well
--> 774 hf_quantizer.create_quantized_param(model, param, param_name, param_device)
775
776 # For quantized modules with FSDP/DeepSpeed Stage 3, we need to quantize the parameter on the GPU

/usr/local/lib/python3.12/dist-packages/transformers/quantizers/quantizer_bnb_4bit.py in create_quantized_param(self, model, param_value, param_name, target_device, **kwargs)
188 # update param name to get the weights instead of the quantized stats
189 param_name = self.get_param_name(param_name)
--> 190 module, tensor_name = get_module_from_name(model, param_name)
191
192 # torch.Tensor.to(<int num>) is not supported by torch_npu (see this issue).

/usr/local/lib/python3.12/dist-packages/transformers/quantizers/quantizers_utils.py in get_module_from_name(module, tensor_name)
18 if "." in tensor_name:
19 module_name, tensor_name = tensor_name.rsplit(".", 1)
---> 20 module = module.get_submodule(module_name)
21 return module, tensor_name

/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py in get_submodule(self, target)
723 for item in atoms:
724 if not hasattr(mod, item):
--> 725 raise AttributeError(
726 mod._get_name() + " has no attribute " + item + ""
727 )

AttributeError: SequentialLlama4TextExperts has no attribute down_proj