[Bug]: Error on inference with LoRa request (safetensors format)

[tsvisab]

Your current environment

Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux 11 (bullseye) (x86_64)
GCC version: (Debian 10.2.1-6) 10.2.1 20210110
Clang version: Could not collect
CMake version: version 3.18.4
Libc version: glibc-2.31

Python version: 3.10.14 | packaged by conda-forge | (main, Mar 20 2024, 12:45:18) [GCC 12.3.0] (64-bit runtime)
Python platform: Linux-5.10.0-30-cloud-amd64-x86_64-with-glibc2.31
Is CUDA available: N/A
CUDA runtime version: 11.3.109
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: 
GPU 0: NVIDIA L4
GPU 1: NVIDIA L4
GPU 2: NVIDIA L4
GPU 3: NVIDIA L4

Nvidia driver version: 550.90.07
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Byte Order:                           Little Endian
Address sizes:                        46 bits physical, 48 bits virtual
CPU(s):                               48
On-line CPU(s) list:                  0-47
Thread(s) per core:                   2
Core(s) per socket:                   24
Socket(s):                            1
NUMA node(s):                         1
Vendor ID:                            GenuineIntel
CPU family:                           6
Model:                                85
Model name:                           Intel(R) Xeon(R) CPU @ 2.20GHz
Stepping:                             7
CPU MHz:                              2200.226
BogoMIPS:                             4400.45
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            768 KiB
L1i cache:                            768 KiB
L2 cache:                             24 MiB
L3 cache:                             38.5 MiB
NUMA node0 CPU(s):                    0-47
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow:   Not affected
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves arat avx512_vnni md_clear arch_capabilities

Versions of relevant libraries:
[pip3] numpy==1.24.4
[conda] numpy                     1.24.4                   pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: N/A
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      PHB     PHB     PHB     0-47    0               N/A
GPU1    PHB      X      PHB     PHB     0-47    0               N/A
GPU2    PHB     PHB      X      PHB     0-47    0               N/A
GPU3    PHB     PHB     PHB      X      0-47    0               N/A

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

🐛 Describe the bug

Llama3 8b with tuned LoraAdapter fails on inference,

Model init:

class VLLMClassifierBase(KLlmHuggingfaceClassifier):
    def __init__(self, model_config: VLLMClassifierConfig):
        logger.info(f"Model config: {model_config}")
        super().__init__()
        ray_tmp_dir = "/dev/shm/tmp/ray"
        os.makedirs(ray_tmp_dir, exist_ok=True)
        ray.init(_temp_dir=ray_tmp_dir, num_gpus=model_config.tensor_parallel_size)

        self.load_model(model_config)

    def load_model(self, model_config):

        self.lora_adapter_path = None
        logger.info(f"Lora adapter path: {model_config.gcs_lora_adapter_path}")
        if model_config.gcs_lora_adapter_path:
            local_lora_adapter_path = "/home/jupyter/huggingface-models/training/artifacts-240711_1124/"
            # copy_from_gcs_uri(model_config.gcs_lora_adapter_path, local_lora_adapter_path)
            self.lora_adapter_path = local_lora_adapter_path

        os.environ["NCCL_DEBUG"] = "INFO"

        download_dir = "/dev/shm/cache/hub"
        os.makedirs(download_dir, exist_ok=True)

        logger.info(f"Loading model: {model_config.hf_model_path}")
        engine_args = AsyncEngineArgs(
            model=model_config.hf_model_path,
            quantization=model_config.quantization if model_config.quantization else None,
            dtype="auto",
            tensor_parallel_size=model_config.tensor_parallel_size,
            enforce_eager=model_config.enforce_eager,
            disable_custom_all_reduce=model_config.disable_custom_all_reduce,
            worker_use_ray=bool(model_config.tensor_parallel_size > 1),
            engine_use_ray=bool(model_config.tensor_parallel_size > 1),
            enable_lora=bool(self.lora_adapter_path is not None),
            download_dir=download_dir,
            gpu_memory_utilization=model_config.gpu_memory_utilization,
            max_model_len=model_config.max_model_len,
            enable_prefix_caching=model_config.enable_prefix_caching,
            max_lora_rank=64,
        )

        self.model = AsyncLLMEngine.from_engine_args(engine_args, usage_context=UsageContext.API_SERVER)
        logger.info(f"Successfully loaded model: {vars(self.model)}")
        loop = asyncio.get_event_loop()

        get_tokenizer_future = asyncio.ensure_future(self.model.get_tokenizer())

        loop.run_until_complete(get_tokenizer_future)

        self.tokenizer = get_tokenizer_future.result()`

Model inference:

` async def _predict_single_instance(self, model_input: KLLMModelInput): prompt = self._to_chat_prompt(model_input.chat_history) sampling_params = SamplingParams( model_input.n, temperature=model_input.temperature, top_p=model_input.top_p, max_tokens=model_input.max_new_tokens, )

try:

    request_id = random_uuid()
    lora_request = None
    if self.lora_adapter_path:
        lora_request = LoRARequest("sql_adapter", 1, self.lora_adapter_path)

    # Run the async function within the event loop
    text_outputs = await self.get_result_from_generator(
        self.model.generate(prompt, sampling_params, request_id, lora_request=lora_request)
    )
    return KLLMModelOutput(generated_ids=text_outputs)`

Logs:

m_engine.AsyncLLMEngine object at 0x7f3117d5d600>>) handle: <Handle functools.partial(<function _log_task_completion at 0x7f2fdbc01e10>, error_callback=<bound method AsyncLLMEngine._error_callback of <vllm.engine.async_llm_engine.AsyncLLMEngine object at 0x7f3117d5d600>>)> Traceback (most recent call last): File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 43, in _log_task_completion return_value = task.result() File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 595, in run_engine_loop result = task.result() File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 538, in engine_step request_outputs = await self.engine.step.remote() # type: ignore ray.exceptions.RayTaskError(IndexError): ray::_AsyncLLMEngine.step() (pid=55187, ip=10.148.0.7, actor_id=75c03eff87f9a5f009b2815b01000000, repr=<vllm.engine.async_llm_engine._AsyncLLMEngine object at 0x7fe99db5a350>) File "/opt/conda/lib/python3.10/concurrent/futures/_base.py", line 451, in result return self.get_result() File "/opt/conda/lib/python3.10/concurrent/futures/_base.py", line 403, in get_result raise self._exception File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/engine/llm_engine.py", line 861, in step output = self.model_executor.execute_model( File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/executor/distributed_gpu_executor.py", line 76, in execute_model return self._driver_execute_model(execute_model_req) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/executor/ray_gpu_executor.py", line 224, in _driver_execute_model return self.driver_worker.execute_method("execute_model", File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/worker/worker_base.py", line 349, in execute_method raise e File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/worker/worker_base.py", line 340, in execute_method return executor(*args, *kwargs) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/worker/worker_base.py", line 271, in execute_model output = self.model_runner.execute_model( File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context return func(args, *kwargs) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/worker/model_runner.py", line 1189, in execute_model self.set_active_loras(model_input.lora_requests, File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/worker/model_runner.py", line 887, in set_active_loras self.lora_manager.set_active_loras(lora_requests, lora_mapping) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/worker_manager.py", line 138, in set_active_loras self._apply_loras(lora_requests) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/worker_manager.py", line 270, in _apply_loras self.add_lora(lora) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/worker_manager.py", line 285, in add_lora self._lora_manager.activate_lora(lora_request.lora_int_id) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/models.py", line 804, in activate_lora result = super().activate_lora(lora_id) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/models.py", line 495, in activate_lora module.set_lora(index, module_lora.lora_a, module_lora.lora_b, File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/layers.py", line 826, in set_lora lora_b = self.slice_lora_b(lora_b) File "/home/jupyter/.cache/pypoetry/virtualenvs/k-llm-huggingface-6NB9Z0db-py3.10/lib/python3.10/site-packages/vllm/lora/layers.py", line 801, in slice_lora_b lora_b_q = lora_b[0][:, self.q_proj_shard_size IndexError: too many indices for tensor of dimension 1

[tsvisab]

archive_name.tar.gz Attached is LoRa Adapter

[tsvisab]

After little investigation, seems like vllm's using safetensors to load the adapter but returns empty tensors when calling f.get_tensor() in some of the cases, still, no idea how to fix this.

More context:

i trained the Adapter using trl SFTtrainer as a workaround i'm going to use save_safetensors=False in SFTConfig

[ANYMS-A]

After little investigation, seems like vllm's using safetensors to load the adapter but returns empty tensors when calling f.get_tensor() in some of the cases, still, no idea how to fix this.

More context:

i trained the Adapter using trl SFTtrainer as a workaround i'm going to use save_safetensors=False in SFTConfig

I got a similar issue as you did I didn't met any error during loading LoRA adaptors, but it seems the LoRA weights did not make any effect on the base model's weights, the output is almost the same as the output by using the original base model.

But If I merge the LoRA weight's to base model beforehand the deployment using Peft's merge_and_unload method, the output of the model is way different from the base model.

My issuse is listed below: https://github.com/vllm-project/vllm/issues/6250

[KyonP]

I have a similar issue as well.

When LLaMA 3 8B combines LoRA and the base model using merge_and_unload, it works well, but when loaded with enable_lora in vLLM, the performance drops significantly. :cry:

[tsvisab]

I think the issue is not vllm's entirely, they should only add a check to see that adapter does not contain empty layers, the issue is why the adapter has empty layers for me, i found that using deepsped zero3 configuration, i can generate good adapters only for rank16 and lower but not for 32 and 64, it does work well with zero2, so, i am going to close the bug.