[Bug]: RuntimeError at vllm startup, V0 engine, Llama 3.1, "The size of tensor a (50) must match the size of tensor b (56) at non-singleton dimension 0"

[akeshet]

Your current environment

The output of `python collect_env.py`

INFO 03-21 04:23:58 [__init__.py:256] Automatically detected platform cuda.
Collecting environment information...
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.12.8 (main, Dec  6 2024, 19:59:28) [Clang 18.1.8 ] (64-bit runtime)
Python platform: Linux-6.8.0-40-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA A100 80GB PCIe
GPU 1: NVIDIA A100 80GB PCIe

Nvidia driver version: 535.183.06
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        48 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               60
On-line CPU(s) list:                  0-59
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7763 64-Core Processor
CPU family:                           25
Model:                                1
Thread(s) per core:                   1
Core(s) per socket:                   30
Socket(s):                            2
Stepping:                             1
BogoMIPS:                             4890.80
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl cpuid extd_apicid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd ibrs ibpb stibp vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean pausefilter pfthreshold v_vmsave_vmload vgif umip pku ospke vaes vpclmulqdq rdpid fsrm arch_capabilities
Virtualization:                       AMD-V
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            3.8 MiB (60 instances)
L1i cache:                            3.8 MiB (60 instances)
L2 cache:                             30 MiB (60 instances)
L3 cache:                             960 MiB (60 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-29
NUMA node1 CPU(s):                    30-59
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:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Vulnerable: Safe RET, no microcode
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-ml-py==12.570.86
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] onnxruntime==1.21.0
[pip3] pyzmq==26.3.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.50.0.dev0
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      PHB     0-59    0-1             N/A
GPU1    PHB      X      0-59    0-1             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

NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

Model is meta-llama/Llama-3.1-8B-Instruct. The issue was absent on VLLM 0.7.2. VLLM is being run with VLLM_USE_V1=0 to force V0 inference engine.

VLLM crashes during cuda graph capture. Last successfully logged message:

(VllmWorkerProcess pid=1283553) INFO 03-21 04:33:32 [model_runner.py:1442] Capturing cudagraphs for decoding. This may lead to unexpected consequences if t
he model is not static. To run the model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI. If out-of-memory error occurs during
cudagraph capture, consider decreasing `gpu_memory_utilization` or switching to eager mode. You can also reduce the `max_num_seqs` as needed to decrease me
mory usage.
INFO 03-21 04:33:32 [model_runner.py:1442] Capturing cudagraphs for decoding. This may lead to unexpected consequences if the model is not static. To run t
he model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI. If out-of-memory error occurs during cudagraph capture, consider decr
easing `gpu_memory_utilization` or switching to eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
Capturing CUDA graph shapes:   0%|                                                                                                  | 0/10 [00:00<?, ?it/s]

Then exception below.

  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/utils.py", line 2216, in run_method
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/worker/worker.py", line 308, in initialize_cache
    self._warm_up_model()
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/worker/worker.py", line 338, in _warm_up_model
    self.model_runner.capture_model(self.gpu_cache)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/worker/model_runner.py", line 1509, in capture_model
    self.set_active_loras(set(), lora_mapping)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/worker/model_runner.py", line 1371, in set_active_loras
    self.lora_manager.set_active_adapters(lora_requests, lora_mapping)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/lora/worker_manager.py", line 167, in set_active_adapters
    set_active_adapters_worker(requests, mapping, self._apply_adapters,
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/adapter_commons/utils.py", line 55, in set_active_adapters_worker
    set_adapter_mapping_func(mapping)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/lora/models.py", line 688, in set_adapter_mapping
    self._last_mapping = set_adapter_mapping(mapping, self._last_mapping,
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/adapter_commons/utils.py", line 30, in set_adapter_mapping
    set_mapping_func(mapping)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/lora/models.py", line 453, in _set_adapter_mapping
    self.punica_wrapper.update_metadata(
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/lora/punica_wrapper/punica_gpu.py", line 66, in update_metadata
    self.prompt_mapping_meta.prepare_tensors(self.sampler_indices)
  File "/home/shadeform/precog/.venv/lib/python3.12/site-packages/vllm/lora/ops/triton_ops/lora_kernel_metadata.py", line 76, in prepare_tensors
    self.token_lora_mapping[:num_tokens].copy_(token_lora_mapping,
RuntimeError: The size of tensor a (50) must match the size of tensor b (56) at non-singleton dimension 0

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[jeejeelee]

Could you plz provibe your running script, I can't reproduce your issue

[akeshet]

Are you attempting the repro while forcing the V0 engine?

From my understanding from recent PRs, the codepath that my stacktrace goes through is for a kernel that was added for V1.

When I revert #14685 locally, this RuntimeError on startup goes away. So I suspect that the issue is that in v0.8.1 we are incorrectly using kernels intended for V1 in the V0 engine.

[akeshet]

@varun-sundar-rabindranath FYI

[varun-sundar-rabindranath]

Hi @akeshet - Thanks! I'll take a look.

So I suspect that the issue is that in v0.8.1 we are incorrectly using kernels intended for V1 in the V0 engine.

PR #14685 makes both the V0 and V1 engines use the same kernels. It is intentional.

[akeshet]

I also attempted merging in #15152 as it seemed related, but it did not fix the issue for me.

[varun-sundar-rabindranath]

@akeshet can you share the command you are running ? It'll help debugging greatly. Thanks !

edit: It looks like you have max_num_seqs set to 50 maybe ? Can you try setting it to 56 to see if the error goes away ? Thanks.

[varun-sundar-rabindranath]

I got a repro with the command,

VLLM_USE_V1=0 python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 200 --max-loras 4 --max-lora-rank 8 --lora-path yard1/llama-2-7b-sql-lora-test --enable-lora --max-num-seqs 50

The issue is a mismatch between, --max-num-seqs and the cudagraph capture sizes which is "cudagraph_capture_sizes":[56,48,40,32,24,16,8,4,2,1] when --max-num-seqs is set to 50. Ill put up a fix.

Thanks @akeshet 🙌

[akeshet]

Thanks for reproing! Sorry I wasn't able to share the command -- we are using the vllm engine as a library in our inference server, which has a bunch of private code too, not via the standalone entry point.

[akeshet]

In the meantime, I suppose I can fix our max_num_seqs to one of the graph capture sizes as a workaround.

[varun-sundar-rabindranath]

Thanks for reproing! Sorry I wasn't able to share the command -- we are using the vllm engine as a library in our inference server, which has a bunch of private code too, not via the standalone entry point.

Yeah. Totally understand!

In the meantime, I suppose I can fix our max_num_seqs to one of the graph capture sizes as a workaround.

Yes. Sorry for the trouble. I am also adding a test so we can catch such errors earlier.