[Bug]: Can't load any LLM with v0.8.*

[Salihcan29]

Your current environment

The output of `python collect_env.py`

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: 3.11.0rc1 (main, Aug 12 2022, 10:02:14) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-117-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 H100 NVL
Nvidia driver version: 535.183.01
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:                        46 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               16
On-line CPU(s) list:                  0-15
Vendor ID:                            GenuineIntel
Model name:                           INTEL(R) XEON(R) GOLD 6542Y
CPU family:                           6
Model:                                207
Thread(s) per core:                   1
Core(s) per socket:                   8
Socket(s):                            2
Stepping:                             2
BogoMIPS:                             5800.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology cpuid tsc_known_freq pni pclmulqdq dtes64 vmx ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                       VT-x
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            512 KiB (16 instances)
L1i cache:                            512 KiB (16 instances)
L2 cache:                             64 MiB (16 instances)
L3 cache:                             32 MiB (2 instances)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-15
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:   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; BHI SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Mitigation; TSX disabled

Versions of relevant libraries:
[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-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.3.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.50.2
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.2
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	0-15	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

NVIDIA_REQUIRE_CUDA=cuda>=12.6 brand=unknown,driver>=470,driver<471 brand=grid,driver>=470,driver<471 brand=tesla,driver>=470,driver<471 brand=nvidia,driver>=470,driver<471 brand=quadro,driver>=470,driver<471 brand=quadrortx,driver>=470,driver<471 brand=nvidiartx,driver>=470,driver<471 brand=vapps,driver>=470,driver<471 brand=vpc,driver>=470,driver<471 brand=vcs,driver>=470,driver<471 brand=vws,driver>=470,driver<471 brand=cloudgaming,driver>=470,driver<471 brand=unknown,driver>=535,driver<536 brand=grid,driver>=535,driver<536 brand=tesla,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=vapps,driver>=535,driver<536 brand=vpc,driver>=535,driver<536 brand=vcs,driver>=535,driver<536 brand=vws,driver>=535,driver<536 brand=cloudgaming,driver>=535,driver<536 brand=unknown,driver>=550,driver<551 brand=grid,driver>=550,driver<551 brand=tesla,driver>=550,driver<551 brand=nvidia,driver>=550,driver<551 brand=quadro,driver>=550,driver<551 brand=quadrortx,driver>=550,driver<551 brand=nvidiartx,driver>=550,driver<551 brand=vapps,driver>=550,driver<551 brand=vpc,driver>=550,driver<551 brand=vcs,driver>=550,driver<551 brand=vws,driver>=550,driver<551 brand=cloudgaming,driver>=550,driver<551
CUDA_VERSION=12.6.2
LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NVIDIA_VISIBLE_DEVICES=all
NVIDIA_DRIVER_CAPABILITIES=compute,utility
VLLM_USE_V1=0
NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

The simple code i run:

# built on nvidia/cuda:12.6.2-base-ubuntu22.04
from vllm import LLM, SamplingParams

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

llm = LLM(model="google/gemma-3-1b-it") # or any gemma model

Here is the output. Right after "Model loading took" pod is dying with status code 139. When i run these codes in notebook kernel dies immidiately. I don't know how can i track any other error. I couldn't catch but the setup was very simple.

Fetching 10 files:   0%|          | 0/10 [00:00<?, ?it/s]
Fetching 10 files: 100%|██████████| 10/10 [00:00<00:00, 3725.29it/s]
INFO 03-27 15:53:05 [config.py:585] This model supports multiple tasks: {'classify', 'generate', 'embed', 'reward', 'score'}. Defaulting to 'generate'.
INFO 03-27 15:53:05 [llm_engine.py:241] Initializing a V0 LLM engine (v0.8.2) with config: model='google/gemma-3-1b-it', speculative_config=None, tokenizer='google/gemma-3-1b-it', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='xgrammar', reasoning_backend=None), observability_config=ObservabilityConfig(show_hidden_metrics=False, otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=None, served_model_name=google/gemma-3-1b-it, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=False, use_async_output_proc=True, disable_mm_preprocessor_cache=False, mm_processor_kwargs=None, pooler_config=None, compilation_config={"splitting_ops":[],"compile_sizes":[],"cudagraph_capture_sizes":[256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":256}, use_cached_outputs=False, 
INFO 03-27 15:53:07 [cuda.py:291] Using Flash Attention backend.
INFO 03-27 15:53:08 [parallel_state.py:954] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0
INFO 03-27 15:53:08 [model_runner.py:1110] Starting to load model google/gemma-3-1b-it...
INFO 03-27 15:53:08 [weight_utils.py:265] Using model weights format ['*.safetensors']
INFO 03-27 15:53:09 [weight_utils.py:315] No model.safetensors.index.json found in remote.

Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]

Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00,  1.06s/it]
2025-03-27T15:53:10.214491613Z 
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00,  1.06s/it]
2025-03-27T15:53:10.214493681Z 
INFO 03-27 15:53:10 [loader.py:447] Loading weights took 1.11 seconds
INFO 03-27 15:53:10 [model_runner.py:1146] Model loading took 1.9148 GB and 1.945495 seconds

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.

[Salihcan29]

This is probably not a bug, But I wasn't sure if I should open it as usage. Even though everything seemed correct, I got an error so I opened it as a bug.

[hmellor]

Can you try with VLLM_LOGGING_LEVEL=DEBUG to see what is happening immediately before it crashes?

[Salihcan29]

Can you try with VLLM_LOGGING_LEVEL=DEBUG to see what is happening immediately before it crashes?

Here is the output with vllm debug mode

DEBUG 03-27 18:45:47 [__init__.py:28] No plugins for group vllm.platform_plugins found.
2025-03-27T18:45:47.076308378Z DEBUG 03-27 18:45:47 [__init__.py:34] Checking if TPU platform is available.
2025-03-27T18:45:47.076309962Z DEBUG 03-27 18:45:47 [__init__.py:44] TPU platform is not available because: No module named 'libtpu'
DEBUG 03-27 18:45:47 [__init__.py:52] Checking if CUDA platform is available.
2025-03-27T18:45:47.087641635Z DEBUG 03-27 18:45:47 [__init__.py:72] Confirmed CUDA platform is available.
2025-03-27T18:45:47.089829729Z DEBUG 03-27 18:45:47 [__init__.py:100] Checking if ROCm platform is available.
2025-03-27T18:45:47.089838016Z DEBUG 03-27 18:45:47 [__init__.py:114] ROCm platform is not available because: No module named 'amdsmi'
2025-03-27T18:45:47.089839217Z DEBUG 03-27 18:45:47 [__init__.py:122] Checking if HPU platform is available.
2025-03-27T18:45:47.090087304Z DEBUG 03-27 18:45:47 [__init__.py:129] HPU platform is not available because habana_frameworks is not found.
2025-03-27T18:45:47.090095486Z DEBUG 03-27 18:45:47 [__init__.py:140] Checking if XPU platform is available.
2025-03-27T18:45:47.090096882Z DEBUG 03-27 18:45:47 [__init__.py:150] XPU platform is not available because: No module named 'intel_extension_for_pytorch'
2025-03-27T18:45:47.090097958Z DEBUG 03-27 18:45:47 [__init__.py:158] Checking if CPU platform is available.
2025-03-27T18:45:47.090824423Z DEBUG 03-27 18:45:47 [__init__.py:180] Checking if Neuron platform is available.
2025-03-27T18:45:47.090832527Z DEBUG 03-27 18:45:47 [__init__.py:187] Neuron platform is not available because: No module named 'transformers_neuronx'
2025-03-27T18:45:47.090833713Z DEBUG 03-27 18:45:47 [__init__.py:52] Checking if CUDA platform is available.
2025-03-27T18:45:47.093444928Z DEBUG 03-27 18:45:47 [__init__.py:72] Confirmed CUDA platform is available.

# some custom outputs

Fetching 10 files:   0%|          | 0/10 [00:00<?, ?it/s]
Fetching 10 files: 100%|██████████| 10/10 [00:00<00:00, 11557.74it/s]
DEBUG 03-27 18:46:17 [__init__.py:28] No plugins for group vllm.general_plugins found.
INFO 03-27 18:46:22 [config.py:585] This model supports multiple tasks: {'embed', 'classify', 'generate', 'reward', 'score'}. Defaulting to 'generate'.
INFO 03-27 18:46:22 [llm_engine.py:241] Initializing a V0 LLM engine (v0.8.2) with config: model='google/gemma-3-1b-it', speculative_config=None, tokenizer='google/gemma-3-1b-it', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='xgrammar', reasoning_backend=None), observability_config=ObservabilityConfig(show_hidden_metrics=False, otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=None, served_model_name=google/gemma-3-1b-it, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=False, use_async_output_proc=True, disable_mm_preprocessor_cache=False, mm_processor_kwargs=None, pooler_config=None, compilation_config={"splitting_ops":[],"compile_sizes":[],"cudagraph_capture_sizes":[256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":256}, use_cached_outputs=False, 
INFO 03-27 18:46:24 [cuda.py:291] Using Flash Attention backend.
2025-03-27T18:46:24.882902171Z DEBUG 03-27 18:46:24 [config.py:3680] enabled custom ops: Counter()
DEBUG 03-27 18:46:24 [config.py:3682] disabled custom ops: Counter()
DEBUG 03-27 18:46:25 [parallel_state.py:817] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.244.231:48867 backend=nccl
INFO 03-27 18:46:25 [parallel_state.py:954] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0
2025-03-27T18:46:25.562273729Z DEBUG 03-27 18:46:25 [config.py:3680] enabled custom ops: Counter()
2025-03-27T18:46:25.562275827Z DEBUG 03-27 18:46:25 [config.py:3682] disabled custom ops: Counter()
INFO 03-27 18:46:25 [model_runner.py:1110] Starting to load model google/gemma-3-1b-it...
DEBUG 03-27 18:46:25 [decorators.py:109] Inferred dynamic dimensions for forward method of <class 'vllm.model_executor.models.gemma3.Gemma3Model'>: ['input_ids', 'positions', 'intermediate_tensors', 'inputs_embeds']
DEBUG 03-27 18:46:25 [config.py:3680] enabled custom ops: Counter({'gemma_rms_norm': 157, 'gelu_and_mul': 26, 'rotary_embedding': 2})
DEBUG 03-27 18:46:25 [config.py:3682] disabled custom ops: Counter()
INFO 03-27 18:46:25 [weight_utils.py:265] Using model weights format ['*.safetensors']
INFO 03-27 18:46:26 [weight_utils.py:315] No model.safetensors.index.json found in remote.

Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]

Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00,  1.11s/it]

Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00,  1.11s/it]
2025-03-27T18:46:27.358123409Z 
INFO 03-27 18:46:27 [loader.py:447] Loading weights took 1.15 seconds
INFO 03-27 18:46:27 [model_runner.py:1146] Model loading took 1.9148 GB and 1.841506 seconds

[hmellor]

Hmm, so the last thing you see is the model loading info log. What does your system memory look like when this happens? Is your host going OOM?

[Salihcan29]

Hmm, so the last thing you see is the model loading info log. What does your system memory look like when this happens? Is your host going OOM?

RAM and VRAM usage looks fine. As i see i can load 1b model but not 4 12 and 27. They are vision models. So the main problem can be load vision lms.

[nskpro-cmd]

+1

[hmellor]

Do you have VLLM_USE_V1=0 in your environment? When I run your repro it uses V1 engine by default but your logs indicate you're using the V0 engine.

[Salihcan29]

Yes i use v1=0 because when i use v1 it stucks without giving error. It doesn't even load. And i can not see even any logs that time.

[hmellor]

If you weren't using debug logs the last time you used V1, it could be quietly compiling graphs rather than hanging.

In general we are more interested in making things work with V1 because we hope to fully deprecate V0 as soon as possible.

[Salihcan29]

I resolved the issue by making the following changes:

• Upgraded Python from 3.11 to 3.12.
• Switched the Docker base image from nvidia/cuda:12.6.2-base-ubuntu22.04 to nvidia/cuda:12.6.2-base-ubuntu24.04.