vllm推理报错:无法在rope_scaling中获取factor字段 #96
Comments
|
@Potato-wll 您好，这是由于您使用的vllm版本不匹配导致的，具体可参考 #35 。 |
|
我用vllm启动后报错，FlashAttention only supports Ampere GPUs or newer.我的显卡是T4，用不了flashatt，怎么在哪关 |
@Potato-wll 您好，我们更新了vllm代码 以及相应的镜像，在不支持flash-attn的情况下使用xformers进行推理，请更新到最新的代码/镜像然后重试。 |
用你们最新的vllm代码安装后还是有个这个错： |
你好，根据错误信息，这应该不是和https://github.com/fyabc/vllm/tree/add_qwen2_vl_new一致的最新版本，麻烦检查下git commit id是否正确。 |
好的谢谢，切了分支，目前正在重新安装，请问下这个vllm版本的话，qwen2vl支持单请求多图调用吗？ |
你好，切了这个分支还是报这个错，只是行数不一样了，请帮忙看看： |
Qwen2-VL支持单条请求多个图片，具体调用方式请参考这里 |
可以提供一下您下载的模型文件中的 |
{
"architectures": [
"Qwen2VLForConditionalGeneration"
],
"attention_dropout": 0.0,
"bos_token_id": 151643,
"eos_token_id": 151645,
"vision_start_token_id": 151652,
"vision_end_token_id": 151653,
"vision_token_id": 151654,
"image_token_id": 151655,
"video_token_id": 151656,
"hidden_act": "silu",
"hidden_size": 3584,
"initializer_range": 0.02,
"intermediate_size": 18944,
"max_position_embeddings": 32768,
"max_window_layers": 28,
"model_type": "qwen2_vl",
"num_attention_heads": 28,
"num_hidden_layers": 28,
"num_key_value_heads": 4,
"rms_norm_eps": 1e-06,
"rope_theta": 1000000.0,
"sliding_window": 32768,
"tie_word_embeddings": false,
"torch_dtype": "bfloat16",
"transformers_version": "4.41.2",
"use_cache": true,
"use_sliding_window": false,
"vision_config": {
"depth": 32,
"embed_dim": 1280,
"mlp_ratio": 4,
"num_heads": 16,
"in_chans": 3,
"hidden_size": 3584,
"patch_size": 14,
"spatial_merge_size": 2,
"spatial_patch_size": 14,
"temporal_patch_size": 2
},
"rope_scaling": {
"type": "mrope",
"mrope_section": [
16,
24,
24
]
},
"vocab_size": 152064
} |
我的安装方法如下，不知道有没有问题？： |
我也是 按照官方fork的vllm版本，但是还是会报这个错误 |
请问是这里描述的这个问题导致的吗？：vllm-project/vllm#7905 (comment) |
|
问题+1，exact same problem |
|
@xyfZzz @docShen @Potato-wll 您好，这应当是transformers最新版本的一个bug，我已经提交了相关issue，目前请先使用如下方式安装没有bug的版本： pip install git+https://github.com/huggingface/transformers@21fac7abba2a37fae86106f87fcf9974fd1e3830 |
|
在config文件rope_scaling中加一个factor字段；另在https://github.com/fyabc/vllm/blob/6f3116c9dad0537b2858af703938aa9bf6c25bcf/vllm/model_executor/models/qwen2.py#L175 前面加一行if rope_scaling and rope_scaling.get('type','default') == 'default': rope_scaling['type'] = 'mrope'可暂时解决问题 |
@lilin-git 感谢说明，这里提到的前一个方法是可行的；后一个方法不建议（由于 @xyfZzz @Potato-wll @docShen 如果重新安装transformers较为麻烦，也可使用上面提到的方法。 |
正常运行了，感谢！ |
|
加了factor，改了qwen2.py后运行python -m vllm.entrypoints.openai.api_server --model /data/Qwen2-VL-7B-Instruct --served-model-name Qwen2-VL-7B --port 10000 |
您好，请检查一下您使用的vllm版本，似乎不是正确的版本？ |
0.6.0和0.5.5都报这个错误 |
额，看链接。不是官方版本，官方版本还不支持。这个项目：https://github.com/fyabc/vllm/tree/add_qwen2_vl_new |
@wuzhizhige 目前Qwen2-VL vllm支持尚未合并到官方，请使用这个版本：https://github.com/fyabc/vllm/tree/add_qwen2_vl_new |
你好，我安装了你指定的这个版本的transformers还是有这个问题，请问在config.json中增加factor具体是怎么加呢？ |
@azuercici 可以修改如下 {
...
"rope_scaling": {
"type": "mrope",
"factor": 1,
"mrope_section": [
16,
24,
24
]
},
} |
|
配置文件和vllm我都修改了但是还是报错：: NameError: name 'rod_scaling' is not defined |
注意名称应当是'rope_scaling'而非'rod_scaling' |
目前使用这个版本的transformer还是会报原来的错误 |
|
|
|
vLLM 初始化时传入 rope_scaling 参数覆盖原有 config，可以临时解决。 llm = LLM(
model=model_dir,
rope_scaling={
"type": "mrope",
"mrope_section": [
16,
24,
24
],
},
) |
这是我的运行代码：
python -m vllm.entrypoints.openai.api_server --served-model-name Qwen2-VL-7B-Instruct --model /home/wangll/llm/model_download_demo/models/Qwen/Qwen2-VL-7B-Instruct
以下是报错信息：
INFO 09-03 18:48:04 api_server.py:440] vLLM API server version 0.5.5
INFO 09-03 18:48:04 api_server.py:441] args: Namespace(host=None, port=8000, uvicorn_log_level='info', allow_credentials=False, allowed_origins=[''], allowed_methods=[''], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template=None, response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=False, model='/home/wangll/llm/model_download_demo/models/Qwen/Qwen2-VL-7B-Instruct', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=False, download_dir=None, load_format='auto', dtype='float16', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=None, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=1, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=False, disable_sliding_window=False, use_v2_block_manager=False, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=None, max_num_seqs=256, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=False, max_context_len_to_capture=None, max_seq_len_to_capture=8192, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=1, scheduler_delay_factor=0.0, enable_chunked_prefill=None, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['Qwen2-VL-7B-Instruct'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, engine_use_ray=False, disable_log_requests=False, max_log_len=None)
Traceback (most recent call last):
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/runpy.py", line 196, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/runpy.py", line 86, in _run_code
exec(code, run_globals)
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 476, in
asyncio.run(run_server(args))
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/asyncio/runners.py", line 44, in run
return loop.run_until_complete(main)
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/asyncio/base_events.py", line 649, in run_until_complete
return future.result()
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 443, in run_server
async with build_async_engine_client(args) as async_engine_client:
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/contextlib.py", line 199, in aenter
return await anext(self.gen)
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 117, in build_async_engine_client
if (model_is_embedding(args.model, args.trust_remote_code,
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 71, in model_is_embedding
return ModelConfig(model=model_name,
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/config.py", line 214, in init
self.max_model_len = _get_and_verify_max_len(
File "/home/wangll/.conda/envs/Qwen2vl/lib/python3.10/site-packages/vllm/config.py", line 1650, in _get_and_verify_max_len
assert "factor" in rope_scaling
AssertionError
我去看了模型的配置文件config.json,里面的rope_scaling确实没有factor字段，
"rope_scaling": {
"type": "mrope",
"mrope_section": [
16,
24,
24
]
},
"vocab_size": 152064
}
The text was updated successfully, but these errors were encountered: