[Bug]: CohereForAI/c4ai-command-r-v01 : ValueError: User-specified max_model_len (131072) is greater than the derived max_model_len (None=8192 in model's config.json). This may lead to incorrect model outputs or CUDA errors. Make sure the value is correct and within the model >

[pseudotensor]

Your current environment

Head of main after various cohere updates/fixes.

Issues:

• Default is 8192 if don't do anything, but model card says 128k context: https://huggingface.co/CohereForAI/c4ai-command-r-v01
• Trying to change --max-model-len to 131072 fails with this error.

sudo apt update
sudo apt install libnccl2 libnccl-dev

wget https://developer.download.nvidia.com/compute/cuda/12.1.0/local_installers/cuda_12.1.0_530.30.02_linux.run
sudo sh cuda_12.1.0_530.30.02_linux.run
sudo chmod -R a+rwx /usr/local/

export CUDA_HOME=/usr/local/cuda-12.1
export PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cu121"

conda create -n vllm_cuda12.1 -y
conda activate vllm_cuda12.1
conda install python=3.10 -y

pip install git+https://github.com/vllm-project/vllm.git
pip install hf_transfer
pip install tiktoken accelerate flash_attn

export HF_HUB_ENABLE_HF_TRANSFER=1
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/lib64:$HOME/extras/CUPTI/lib64
export PATH=$PATH:$CUDA_HOME/bin

export CUDA_VISIBLE_DEVICES="0,1"
python -m vllm.entrypoints.openai.api_server --port=5005 --host=0.0.0.0 --model CohereForAI/c4ai-command-r-v01 --seed 1234 --tensor-parallel-size=2 --max-num-batched-tokens=131072 --max-log-len=100  --max-model-len 131072
# --trust-remote-code

have to comment out trust-remote-code due to a bug in their model that has a PR for registration of the model name that isn't merged yet.

🐛 Describe the bug

INFO 03-28 08:14:18 api_server.py:147] vLLM API server version 0.3.3
INFO 03-28 08:14:18 api_server.py:148] args: Namespace(host='0.0.0.0', port=5005, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, served_model_name=Non>
Traceback (most recent call last):
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 156, in <module>
    engine = AsyncLLMEngine.from_engine_args(engine_args)
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 327, in from_engine_args
    engine_configs = engine_args.create_engine_configs()
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/site-packages/vllm/engine/arg_utils.py", line 362, in create_engine_configs
    model_config = ModelConfig(
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/site-packages/vllm/config.py", line 124, in __init__
    self.max_model_len = _get_and_verify_max_len(self.hf_text_config,
  File "/home/fsuser/miniconda3/envs/vllm_cuda12.1/lib/python3.10/site-packages/vllm/config.py", line 791, in _get_and_verify_max_len
    raise ValueError(
ValueError: User-specified max_model_len (131072) is greater than the derived max_model_len (None=8192 in model's config.json). This may lead to incorrect model outputs or CUDA errors. Make sure the value is correct and within the model >

[youkaichao]

cc @zeppombal

[mwbyeon]

https://huggingface.co/CohereForAI/c4ai-command-r-v01/blob/9c33b0976099d0f406f0d007613676fe42b78e3b/config.json#L16

derived_max_model_len is limited to 8192 because max_position_embeddings is 8192 in config.json

vllm/vllm/config.py

Lines 749 to 772 in 14ccd94

	def _get_and_verify_max_len(
	hf_config: PretrainedConfig,
	max_model_len: Optional[int],
	) -> int:
	"""Get and verify the model's maximum length."""
	derived_max_model_len = float("inf")
	possible_keys = [
	# OPT
	"max_position_embeddings",
	# GPT-2
	"n_positions",
	# MPT
	"max_seq_len",
	# ChatGLM2
	"seq_length",
	# Others
	"max_sequence_length",
	"max_seq_length",
	"seq_len",
	]
	for key in possible_keys:
	max_len_key = getattr(hf_config, key, None)
	if max_len_key is not None:
	derived_max_model_len = min(derived_max_model_len, max_len_key)

If you want to change max_model_len to 131072, you must also change max_position_embeddings to the same value.

But I can't think of a good way to solve this.

[pseudotensor]

But I thought they use rope scaling, so should be accounted for in vllm. Shouldn't change embedding size I'd think.

[mwbyeon]

@pseudotensor that makes sense.

[ywang96]

@pseudotensor If you look at the code right below - it'll scale the derived max model length by the factor if it exists in the config.json, but people may not want to used a rope-scaled model all the time.

vllm/vllm/config.py

Lines 786 to 796 in 14ccd94

	rope_scaling = getattr(hf_config, "rope_scaling", None)
	if rope_scaling is not None:
	assert "factor" in rope_scaling
	scaling_factor = rope_scaling["factor"]
	if rope_scaling["type"] == "yarn":
	derived_max_model_len = rope_scaling[
	"original_max_position_embeddings"]
	derived_max_model_len *= scaling_factor
	if max_model_len is None:
	max_model_len = derived_max_model_len

@saurabhdash What's your opinion on this? I guess we should add args on EngineArgs to allow custom rope settings, correct? Please ignore - I misunderstood the issue here.

[pseudotensor]

But less than the maximum should be controlled by passing model_max_length to vLLM. I'm not aware of any other model that fails int his way with rope scaling.

[ywang96]

But less than the maximum should be controlled by passing model_max_length to vLLM. I'm not aware of any other model that fails int his way with rope scaling.

You're right - I did a bit of more research and found https://huggingface.co/CohereForAI/c4ai-command-r-v01/discussions/12, and it seems that the "model_max_length" is added after discussion in this thread

This is indeed a bug we should fix then - although going forward I'm sure if we should just take model_max_length is the truth context window whenever we see it in the config. I will make a PR for this.

[pseudotensor]

At least from that discussion, a default of low is ok, but shouldn't require editing the model config.json in order to go up to its maximum.

Although normally I think vLLM always does maximum by default unless make smaller, different than what that person said HF does.

[ywang96]

@pseudotensor Yep - I had the same thoughts in #3727. Please take a look, thanks!

[saurabhdash]

The max_model_length was added to support llama.cpp (ggml-org/llama.cpp#6033 (comment))