[BUG] DeepSpeed Inference Error for Llama 3 Models AssertionError: Merging tensors is not allowed here!

Describe the bug
Running inference on Llama 3 models gives an error

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/replace_module.py in _replace_module(model, policies, prefix, layer_id, level_id, state_dict)
    687     for name, child in model.named_children():
    688         if child.__class__ in policies:
--> 689             replaced_module = policies[child.__class__][0](child,
    690                                                            policies[child.__class__][-1],
    691                                                            layer_id,

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/replace_module.py in replace_fn(child, _policy, layer_id, prefix, state_dict)
    325             # copy relevant state from child -> new module
    326             if not is_autotp_training_mode() and config.replace_with_kernel_inject:
--> 327                 new_module = replace_with_policy(child,
    328                                                  _policy,
    329                                                  config.triangular_masking,

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/replace_module.py in replace_with_policy(child, policy_cls, triangular_masking, inference, layer_id)
    252
    253         # 9. deal with tensor parallelism.
--> 254         _container.apply_tensor_parallelism(mp_replace)
    255
    256         # 10. copy the tensors from the model-specific container to the new module

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/containers/features/meta_tensor.py in apply_tensor_parallelism(self, mp_replace, **kwargs)
     34                 self.module.attention.attn_ob = None
     35         else:
---> 36             super().apply_tensor_parallelism(mp_replace, **kwargs)
     37
     38     def copy_data_to_new_module(self):

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/containers/features/hybrid_engine.py in apply_tensor_parallelism(self, mp_replace, reversed_dim)
     87         """
     88         # Setup the new Attention module
---> 89         self.attention_qkv_mp(mp_replace, reversed_dim=reversed_dim)
     90         self.attention_o_mp(mp_replace, reversed_dim=reversed_dim)
     91

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/containers/features/split_qkv.py in attention_qkv_mp(self, mp_replace, reversed_dim)
     47                     allocate_tensor=reversed_dim) if src is not None else None
     48         else:
---> 49             super().attention_qkv_mp(mp_replace)
     50
     51     def release_qkv(self):

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/containers/base.py in attention_qkv_mp(self, mp_replace, reversed_dim)
    238
    239     def attention_qkv_mp(self, mp_replace, reversed_dim=False):
--> 240         self.module.attention.attn_qkvw = mp_replace.strided_copy(self.module.attention.attn_qkvw,
    241                                                                   self.qkvw,
    242                                                                   num_splits=3,

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/auto_tp.py in strided_copy(self, dst, src, num_splits, int8, allocate_tensor)
     75                     dst.scale = src.scale
     76                 return dst
---> 77             self.merge_assert(src_shape[outer_dim], dst_shape[self.out_dim])
     78             qkv_size = dst_shape[self.out_dim] // num_splits
     79             qkv_split = [torch.split(src_s, qkv_size, dim=outer_dim) for src_s in src_split]

/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed/module_inject/auto_tp.py in merge_assert(self, dim1, dim2)
     42
     43     def merge_assert(self, dim1, dim2):
---> 44         assert dim1 > dim2, \
     45             'Merging tensors is not allowed here! Please use deepspeed load_checkpoint\
     46             for merging your checkpoints before replacing the transformer layer with\

AssertionError: Merging tensors is not allowed here! Please use deepspeed load_checkpoint            for merging your checkpoints before replacing the transformer layer with            inference-kernels

To Reproduce

# newer transformers have this issue for some models: https://github.com/unslothai/unsloth/issues/1476
pip install transformers==4.47.1
pip install deepspeed==0.16.7

and run

import transformers
from transformers import pipeline
from transformers import AutoModelForCausalLM, AutoTokenizer


import torch
import deepspeed


not_working_models = [
    "HuggingFaceTB/SmolLM2-135M",
    "meta-llama/Llama-3.2-1B",
    "meta-llama/Llama-3.1-8B",
]

model_index = 0
model_name = not_working_models[model_index]

model = AutoModelForCausalLM.from_pretrained(model_name, low_cpu_mem_usage=True, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
# Initialize the DeepSpeed-Inference engine
pipe.model = deepspeed.init_inference(
    pipe.model,
    dtype=torch.float16,
    replace_method='auto',
    replace_with_kernel_inject=True,
)
output = pipe('Hello')

Expected behavior
Inference works, generates texts but faster than transformers.

ds_report output

--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
dc ..................... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
 [WARNING]  NVIDIA Inference is only supported on Ampere and newer architectures
 [WARNING]  FP Quantizer is using an untested triton version (3.2.0), only 2.3.(0, 1) and 3.0.0 are known to be compatible with these kernels
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
 [WARNING]  gds requires the dev libaio .so object and headers but these were not found.
 [WARNING]  gds: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
gds .................... [NO] ....... [NO]
transformer_inference .. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.6
 [WARNING]  using untested triton version (3.2.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/src/chat-ai/venv3.10/lib/python3.10/site-packages/torch']
torch version .................... 2.6.0+cu124
deepspeed install path ........... ['/src/chat-ai/venv3.10/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.16.7, unknown, unknown
torch cuda version ............... 12.4
torch hip version ................ None
nvcc version ..................... 11.7
deepspeed wheel compiled w. ...... torch 2.6, cuda 12.4
shared memory (/dev/shm) size .... 15.44 GB

System info (please complete the following information):

• OS: Ubuntu 22.04
• GPU count and types 1 GPU T4
• transformers 4.47.1
• deepspeed 0.16.7
• Python version 3.10.14

Similar Issues

https://github.com/deepspeedai/DeepSpeed/issues/3960
https://github.com/deepspeedai/DeepSpeed/issues/2691

Start at deepspeed/module_inject/auto_tp.py and trace the strided_copy and merge_assert entry points shown in the traceback, then follow their callers through replace_module.py and the tensor-parallel container files. Reproduce with the listed Transformers and DeepSpeed versions using deepspeed.init_inference. Done means the listed Llama and SmolLM2 models initialize inference and generate text without the merging assertion.