[Bug] JambaAttention has no attribute `feed_forward`

[1luik]

import unsloth
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from unsloth import FastLanguageModel
model_name = "/media/1luik/46BB55AF65F351D4/5/Jamba2-Mini-int4_hf"#"/media/1luik/C2F0D801F0D7FA1D/AI21-Jamba2-Mini" # 或 qwen/Qwen-7B、internlm/internlm3-8b 等

model, tokenizer = FastLanguageModel.from_pretrained(
model_name,
load_in_4bit = True,
device_map = "balanced",
)

((venv) ) root@1luik-PC:/media/1luik/46BB55AF65F351D4/5# python
Python 3.12.11 (main, Jun 24 2025, 05:19:07) [GCC 12.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import os; os.environ['UNSLOTH_USE_MODELSCOPE'] = '1'
>>> import unsloth
^[[F🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
INFO 02-12 13:38:12 [__init__.py:216] Automatically detected platform cuda.
ERROR 02-12 13:38:12 [fa_utils.py:57] Cannot use FA version 2 is not supported due to FA2 is only supported on devices with compute capability >= 8
🦥 Unsloth Zoo will now patch everything to make training faster!
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
>>> import torch
>>> from unsloth import FastLanguageModel
>>> model_name = "/media/1luik/46BB55AF65F351D4/5/Jamba2-Mini-int4_hf"#"/media/1luik/C2F0D801F0D7FA1D/AI21-Jamba2-Mini"  # 或 qwen/Qwen-7B、internlm/internlm3-8b 等
>>> 
>>> model, tokenizer = FastLanguageModel.from_pretrained(
...     model_name,
...     load_in_4bit = True,
...     device_map = "balanced",
... )
==((====))==  Unsloth 2026.2.1: Fast Jamba patching. Transformers: 4.57.6. vLLM: 0.11.0.
   \\   /|    Tesla V100-SXM2-16GB. Num GPUs = 4. Max memory: 15.766 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.8.0+cu128. CUDA: 7.0. CUDA Toolkit: 12.8. Triton: 3.4.0
\        /    Bfloat16 = FALSE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
Loading checkpoint shards:  12%|████████████████▏                                                                                                                | 1/8 [00:01<00:12,  1.76s/it]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 543, in from_pretrained
    return FastModel.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 1292, in from_pretrained
    model, tokenizer = FastBaseModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/vision.py", line 775, in from_pretrained
    model = auto_model.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5468, in _load_pretrained_model
    _error_msgs, disk_offload_index = load_shard_file(args)
                                      ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 843, in load_shard_file
    disk_offload_index = _load_state_dict_into_meta_model(
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 774, in _load_state_dict_into_meta_model
    hf_quantizer.create_quantized_param(model, param, param_name, param_device)
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizer_bnb_4bit.py", line 190, in create_quantized_param
    module, tensor_name = get_module_from_name(model, param_name)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizers_utils.py", line 20, in get_module_from_name
    module = module.get_submodule(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 723, in get_submodule
    raise AttributeError(
AttributeError: JambaAttention has no attribute `feed_forward`
>>> 

[1luik]

Hugging Face Transformers can load the model successfully, but Unsloth cannot.

[Datta0]

Hey @1luik , when loading with transformers, did you set quantization_config = BitsAndBytesConfig(load_in_4bit=True,...)? If not, can you try removing load_in_4bit=True from unsloth load up code and try

model, tokenizer = FastLanguageModel.from_pretrained(
  model_name,
  load_in_4bit = False, # Changed this
  device_map = "balanced",
)

[1luik]

Similarly, setting both quantization_config and load_in_4bit to False results in an error

[Datta0]

No do not set quatnization_config.
Just use the code from the above comment.
If it still throws an error, please share the entire stack trace

[1luik]

>>> model, tokenizer = FastLanguageModel.from_pretrained(
...   model_name,
...   load_in_4bit = False, # Changed this
...   device_map = "balanced",
... )
==((====))==  Unsloth 2026.2.1: Fast Jamba patching. Transformers: 4.57.6. vLLM: 0.11.0.
   \\   /|    Tesla V100-SXM2-16GB. Num GPUs = 4. Max memory: 15.766 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.8.0+cu128. CUDA: 7.0. CUDA Toolkit: 12.8. Triton: 3.4.0
\        /    Bfloat16 = FALSE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
Unsloth: QLoRA and full finetuning all not selected. Switching to 16bit LoRA.
Loading checkpoint shards:  12%|█████████                                                               | 1/8 [00:01<00:12,  1.76s/it]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 543, in from_pretrained
    return FastModel.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 1292, in from_pretrained
    model, tokenizer = FastBaseModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/vision.py", line 775, in from_pretrained
    model = auto_model.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5468, in _load_pretrained_model
    _error_msgs, disk_offload_index = load_shard_file(args)
                                      ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 843, in load_shard_file
    disk_offload_index = _load_state_dict_into_meta_model(
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 774, in _load_state_dict_into_meta_model
    hf_quantizer.create_quantized_param(model, param, param_name, param_device)
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizer_bnb_4bit.py", line 190, in create_quantized_param
    module, tensor_name = get_module_from_name(model, param_name)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizers_utils.py", line 20, in get_module_from_name
    module = module.get_submodule(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 723, in get_submodule
    raise AttributeError(
AttributeError: JambaAttention has no attribute `feed_forward`
>>>

[1luik]

I am loading the model that has already been quantized and saved on the hard drive.

[Datta0]

It is very likely that your checkpoint is corrupted. I was able to load the base model with unsloth without any hassle tbh.

[Datta0]

Can you push your quantized model to huggingface hub so that I can give it a try on my setup and see if it works?

[1luik]

>>> tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
ausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="cpu",                      # 自动分配 GPU/CPU
    trust_remote_code=True,                  # 某些模型需要（如 InternLM、Qwen）
    dtype=torch.float16
)>>> model = AutoModelForCausalLM.from_pretrained(
...     model_name,
...     quantization_config=bnb_config,
...     device_map="cpu",                      # 自动分配 GPU/CPU
...     trust_remote_code=True,                  # 某些模型需要（如 InternLM、Qwen）
...     dtype=torch.float16
... )
Skipping import of cpp extensions due to incompatible torch version 2.8.0+cu128 for torchao version 0.15.0             Please see https://github.com/pytorch/ao/issues/2919 for more info
Loading checkpoint shards:   0%|                                                                                                                                       | 0/21 [00:00<?, ?it/s]
Loading checkpoint shards:   5%|█████▉                                                                                                                      | 1/21 [04:18<1:26:08, 258.41s/it]Loading checkpoint shards:   5%|█████▉                                                                                                                      | 1/21 [04:31<1:30:24, 271.23s/it]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5468, in _load_pretrained_model
    _error_msgs, disk_offload_index = load_shard_file(args)
                                      ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 843, in load_shard_file
    disk_offload_index = _load_state_dict_into_meta_model(
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 750, in _load_state_dict_into_meta_model
    param = param.to(casting_dtype)
            ^^^^^^^^^^^^^^^^^^^^^^^
KeyboardInterrupt
>>> exit()
((venv) ) 1luik@1luik-PC:/media/1luik/46BB55AF65F351D4/5$ python
Python 3.12.11 (main, Jun 24 2025, 05:19:07) [GCC 12.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import os; os.environ['UNSLOTH_USE_MODELSCOPE'] = '1'
>>> import unsloth
oModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from unsloth import FastLanguageModel
model_name = "/media/1luik/C2F0D801F0D7FA1D/AI21-Jamba2-Mini"  # 或 qwen/Qwen-7B、internlm/internlm3-8b 等🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
INFO 02-13 18:58:25 [__init__.py:216] Automatically detected platform cuda.
ERROR 02-13 18:58:26 [fa_utils.py:57] Cannot use FA version 2 is not supported due to FA2 is only supported on devices with compute capability >= 8
🦥 Unsloth Zoo will now patch everything to make training faster!
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
>>> import torch
>>> from unsloth import FastLanguageModel
>>> model_name = "/media/1luik/C2F0D801F0D7FA1D/AI21-Jamba2-Mini"  # 或 qwen/Qwen-7B、internlm/internlm3-8b 等
>>>
>>> model, tokenizer = FastLanguageModel.from_pretrained(
...   model_name,
...   #quantization_config=bnb_config,
...   load_in_4bit = True, # Changed this
...   device_map = "balanced",
... )
==((====))==  Unsloth 2026.2.1: Fast Jamba patching. Transformers: 4.57.6. vLLM: 0.11.0.
   \\   /|    Tesla V100-SXM2-16GB. Num GPUs = 4. Max memory: 15.766 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.8.0+cu128. CUDA: 7.0. CUDA Toolkit: 12.8. Triton: 3.4.0
\        /    Bfloat16 = FALSE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 543, in from_pretrained
    return FastModel.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 1292, in from_pretrained
    model, tokenizer = FastBaseModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/vision.py", line 775, in from_pretrained
    model = auto_model.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5362, in _load_pretrained_model
    model._initialize_missing_keys(missing_keys + mismatched_keys, is_quantized)
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5862, in _initialize_missing_keys
    param_or_buffer = self.get_parameter_or_buffer(key)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5946, in get_parameter_or_buffer
    module, param_name = get_module_from_name(self, target)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizers_utils.py", line 20, in get_module_from_name
    module = module.get_submodule(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 730, in get_submodule
    raise AttributeError("`" + item + "` is not an nn.Module")
AttributeError: `weight` is not an nn.Module
>>>

[1luik]

I have no issues using Transformers

[1luik]

你能把你的量化模型上传到 Hugging Face Hub 吗？这样我就可以在我的设备上试用一下，看看是否有效。
He has a 3b. I'll try to reproduce it

[1luik]

ai21labs/AI21-Jamba-Reasoning-3B
This one won't do either.

>>> model_name = "ai21labs/AI21-Jamba-Reasoning-3B"
>>>
>>> model, tokenizer = FastLanguageModel.from_pretrained(
...   model_name,
...   #quantization_config=bnb_config,
...   device_map = "balanced",
... )
Downloading Model from https://www.modelscope.cn to directory: /home/1luik/.cache/modelscope/hub/models/ai21labs/AI21-Jamba-Reasoning-3B
2026-02-13 19:03:50,555 - modelscope - INFO - Got 13 files, start to download ...
Downloading [assets/Intelligence vs Speed Jamba Reasoning 3B.png]: 100%|████████████████████████████████████████████████████████████████████████████████████| 177k/177k [00:00<00:00, 419kB/s]
Downloading [config.json]: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.00k/1.00k [00:00<00:00, 2.28kB/s]
Downloading [generation_config.json]: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 155/155 [00:00<00:00, 277B/s]
Downloading [model.safetensors.index.json]: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████| 36.5k/36.5k [00:00<00:00, 66.4kB/s]
Downloading [assets/Benchmark Performance - Jamba Reasoning 3B.png]: 100%|████████████████████████████████████████████████████████████████████████████████| 74.6k/74.6k [00:00<00:00, 124kB/s]
Downloading [configuration.json]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 73.0/73.0 [00:00<00:00, 118B/s]
Downloading [special_tokens_map.json]: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 895/895 [00:00<00:00, 2.99kB/s]
Downloading [tokenizer_config.json]: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10.9k/10.9k [00:00<00:00, 40.9kB/s]
Downloading [README.md]: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7.03k/7.03k [00:00<00:00, 11.5kB/s]
Downloading [assets/Speed vs Context Length.png]: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 143k/143k [00:00<00:00, 200kB/s]
Downloading [tokenizer.json]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7.90M/7.90M [00:02<00:00, 3.72MB/s]
Downloading [model-00002-of-00002.safetensors]: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 1.32G/1.32G [01:35<00:00, 14.9MB/s]
Downloading [model-00001-of-00002.safetensors]: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 4.63G/4.63G [05:39<00:00, 14.7MB/s]
Processing 13 items: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 13.0/13.0 [05:39<00:00, 26.1s/it]
2026-02-13 19:09:29,627 - modelscope - INFO - Download model 'ai21labs/AI21-Jamba-Reasoning-3B' successfully.
Downloading Model from https://www.modelscope.cn to directory: /home/1luik/.cache/modelscope/hub/models/ai21labs/AI21-Jamba-Reasoning-3B
==((====))==  Unsloth 2026.2.1: Fast Jamba patching. Transformers: 4.57.6. vLLM: 0.11.0.                                                                 | 13.0M/4.63G [00:02<10:59, 7.52MB/s]
   \\   /|    Tesla V100-SXM2-16GB. Num GPUs = 4. Max memory: 15.766 GB. Platform: Linux.                                                                | 1.23G/4.63G [01:35<03:22, 18.0MB/s]
O^O/ \_/ \    Torch: 2.8.0+cu128. CUDA: 7.0. CUDA Toolkit: 12.8. Triton: 3.4.0██████████████████████████████████████████████████████████████████████████▉| 4.63G/4.63G [05:39<00:00, 16.2MB/s]
\        /    Bfloat16 = FALSE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 543, in from_pretrained
    return FastModel.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 1292, in from_pretrained
    model, tokenizer = FastBaseModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/vision.py", line 775, in from_pretrained
    model = auto_model.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5362, in _load_pretrained_model
    model._initialize_missing_keys(missing_keys + mismatched_keys, is_quantized)
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5862, in _initialize_missing_keys
    param_or_buffer = self.get_parameter_or_buffer(key)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5946, in get_parameter_or_buffer
    module, param_name = get_module_from_name(self, target)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizers_utils.py", line 20, in get_module_from_name
    module = module.get_submodule(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 730, in get_submodule
    raise AttributeError("`" + item + "` is not an nn.Module")
AttributeError: `weight` is not an nn.Module
>>>

[Datta0]

Hey @1luik can you try updating transformers, unsloth and unsloth-zoo and try?

! pip uninstall unsloth unsloth-zoo transformers
! uv pip install unsloth unsloth-zoo transformers

[1luik]

The problem still occurs

((venv) ) 1luik@1luik-PC:/media/1luik/46BB55AF65F351D4/5$ uv pip install unsloth unsloth-zoo transformers
Using Python 3.12.11 environment at: /home/venv
warning: Ignoring dangling temporary directory: `/home/venv/lib/python3.12/site-packages/~nsloth-2026.2.1.dist-info`
warning: Ignoring dangling temporary directory: `/home/venv/lib/python3.12/site-packages/~nsloth_zoo-2026.2.1.dist-info`
warning: Ignoring dangling temporary directory: `/home/venv/lib/python3.12/site-packages/~ransformers-4.57.6.dist-info`
Resolved 91 packages in 2.52s
Prepared 3 packages in 1.46s
Installed 3 packages in 102ms
 + transformers==4.57.6
 + unsloth==2026.2.1
 + unsloth-zoo==2026.2.1
((venv) ) 1luik@1luik-PC:/media/1luik/46BB55AF65F351D4/5$ python
Python 3.12.11 (main, Jun 24 2025, 05:19:07) [GCC 12.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import os; os.environ['UNSLOTH_USE_MODELSCOPE'] = '1'
>>> import unsloth
ort AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from unsloth import FastLanguageModel
🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
INFO 02-13 19:57:09 [__init__.py:216] Automatically detected platform cuda.
ERROR 02-13 19:57:10 [fa_utils.py:57] Cannot use FA version 2 is not supported due to FA2 is only supported on devices with compute capability >= 8
🦥 Unsloth Zoo will now patch everything to make training faster!
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
>>> import torch
>>> from unsloth import FastLanguageModel
>>> model_name = "ai21labs/AI21-Jamba-Reasoning-3B"
>>>
>>> model, tokenizer = FastLanguageModel.from_pretrained(
...   model_name,
...   #quantization_config=bnb_config,
...   device_map = "balanced",
... )
Downloading Model from https://www.modelscope.cn to directory: /home/1luik/.cache/modelscope/hub/models/ai21labs/AI21-Jamba-Reasoning-3B
Downloading Model from https://www.modelscope.cn to directory: /home/1luik/.cache/modelscope/hub/models/ai21labs/AI21-Jamba-Reasoning-3B
==((====))==  Unsloth 2026.2.1: Fast Jamba patching. Transformers: 4.57.6. vLLM: 0.11.0.
   \\   /|    Tesla V100-SXM2-16GB. Num GPUs = 4. Max memory: 15.766 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.8.0+cu128. CUDA: 7.0. CUDA Toolkit: 12.8. Triton: 3.4.0
\        /    Bfloat16 = FALSE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 543, in from_pretrained
    return FastModel.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/loader.py", line 1292, in from_pretrained
    model, tokenizer = FastBaseModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/unsloth/models/vision.py", line 775, in from_pretrained
    model = auto_model.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5048, in from_pretrained
    ) = cls._load_pretrained_model(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5362, in _load_pretrained_model
    model._initialize_missing_keys(missing_keys + mismatched_keys, is_quantized)
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5862, in _initialize_missing_keys
    param_or_buffer = self.get_parameter_or_buffer(key)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5946, in get_parameter_or_buffer
    module, param_name = get_module_from_name(self, target)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/transformers/quantizers/quantizers_utils.py", line 20, in get_module_from_name
    module = module.get_submodule(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 730, in get_submodule
    raise AttributeError("`" + item + "` is not an nn.Module")
AttributeError: `weight` is not an nn.Module