DeepSpeed
DeepSpeed copied to clipboard
microsoft
[BUG] AutoTP fails for Qwen2.5 models when tp size > 2
Describe the bug
Tensor size mismatch for k_proj using Qwen/Qwen2.5-3B-Instruct when running with 8 GPUs, TP=8, DP=1, error logs:
File "openrlhf/trainer/ray/launcher.py", line 98, in execute_batch
result = func(**sample_kwargs)
^^^^^^^^^^^^^^^^^^^^^
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "openrlhf/trainer/ray/ppo_actor.py", line 486, in forward
action_log_probs = self.actor(
^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "openrlhf/models/actor.py", line 195, in forward
output = self.model(sequences, attention_mask=foward_attention_mask, position_ids=position_ids)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "deepspeed/utils/nvtx.py", line 20, in wrapped_fn
ret_val = func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "deepspeed/runtime/engine.py", line 2054, in forward
loss = self.module(*inputs, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1845, in _call_impl
return inner()
^^^^^^^
File "torch/nn/modules/module.py", line 1793, in inner
result = forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/utils/generic.py", line 965, in wrapper
output = func(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/utils/deprecation.py", line 172, in wrapped_func
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "transformers/models/qwen2/modeling_qwen2.py", line 824, in forward
outputs: BaseModelOutputWithPast = self.model(
^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/utils/generic.py", line 965, in wrapper
output = func(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/models/qwen2/modeling_qwen2.py", line 550, in forward
layer_outputs = decoder_layer(
^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/models/qwen2/modeling_qwen2.py", line 263, in forward
hidden_states, self_attn_weights = self.self_attn(
^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/models/qwen2/modeling_qwen2.py", line 166, in forward
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: shape '[1, 1862, -1, 128]' is invalid for input of size 59584
To Reproduce
I'm using OpenRLHF with AutoTP, but I think this should not be directly related to code in OpenRLHF and should be general enough to reproduce.
Expected behavior No error thrown
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-devel package with yum
[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] 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-devel package with yum
[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 version .................... 2.6.0+cu124
deepspeed info ................... 0.16.6, unknown, unknown
torch cuda version ............... 12.4
torch hip version ................ None
nvcc version ..................... 12.4
deepspeed wheel compiled w. ...... torch 2.6, cuda 12.4
shared memory (/dev/shm) size .... 251.75 GB
System info (please complete the following information):
- OS: RHEL 9.5
- 1 machines with x4 A100s each
- transformers==4.51.1
- flash_attn==2.7.4.post1
- Python version: 3.12.9
- OpenRLHF version: 0.7.4.post2
In Qwen2.5-3B, the kv_head number is 2. If you want to set kv_head > 2, you would need to replicate the KV heads, and if you only do inference, the KV cache would also be duplicated.
Although a similar replication approach has been implemented for ChatGLM, it seems that Qwen2.5 has not yet supported this. At this stage, I suggest we keep kv_head divisible by tp_size.
