vllm-project
[Bug]: Invalid Device Ordinal on ROCm
Your current environment
PyTorch version: 2.4.0.dev20240415+rocm6.0
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.0.32830-d62f6a171
OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.29.2
Libc version: glibc-2.35
Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-102-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: AMD Instinct MI300X (gfx942:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.0.32830
MIOpen runtime version: 3.0.0
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 384
On-line CPU(s) list: 0-383
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9654 96-Core Processor
CPU family: 25
Model: 17
Thread(s) per core: 2
Core(s) per socket: 96
Socket(s): 2
Stepping: 1
Frequency boost: enabled
CPU max MHz: 3707.8120
CPU min MHz: 1500.0000
BogoMIPS: 4792.43
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 constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization: AMD-V
L1d cache: 6 MiB (192 instances)
L1i cache: 6 MiB (192 instances)
L2 cache: 192 MiB (192 instances)
L3 cache: 768 MiB (24 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-95,192-287
NUMA node1 CPU(s): 96-191,288-383
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 Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; safe RET
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; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-triton-rocm==3.0.0+0a22a91d04
[pip3] torch==2.4.0.dev20240415+rocm6.0
[pip3] torchvision==0.19.0.dev20240415+rocm6.0
[conda] No relevant packagesROCM Version: 6.1.33591-3a954afdc
Neuron SDK Version: N/A
vLLM Version: 0.4.0.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
Could not collect
🐛 Describe the bug
Issue with invalid ordinal when running with tp=2 on ROCm:
python benchmarks/benchmark_throughput.py --input-len=50 --output-len=100 --model=mistralai/Mistral-7B-v0.1 --tensor-parallel-size=2 --enforce-eager
Namespace(backend='vllm', dataset=None, input_len=50, output_len=100, model='mistralai/Mistral-7B-v0.1', tokenizer='mistralai/Mistral-7B-v0.1', quantization=None, tensor_parallel_size=2, n=1, use_beam_search=False, num_prompts=1000, seed=0, hf_max_batch_size=None, trust_remote_code=False, max_model_len=None, dtype='auto', gpu_memory_utilization=0.9, enforce_eager=True, kv_cache_dtype='auto', quantization_param_path=None, device='cuda', enable_prefix_caching=False, enable_chunked_prefill=False, max_num_batched_tokens=None, download_dir=None)
INFO 04-16 23:49:56 pynccl.py:58] Loading nccl from library librccl.so.1
INFO 04-16 23:49:56 config.py:523] Disabled the custom all-reduce kernel because it is not supported on AMD GPUs.
2024-04-16 23:49:58,704 INFO worker.py:1724 -- Started a local Ray instance.
INFO 04-16 23:50:00 llm_engine.py:87] Initializing an LLM engine (v0.4.0.post1) with config: model='mistralai/Mistral-7B-v0.1', speculative_config=None, tokenizer='mistralai/Mistral-7B-v0.1', tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=2, disable_custom_all_reduce=True, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), seed=0)
(pid=1376632) INFO 04-16 23:50:01 pynccl.py:58] Loading nccl from library librccl.so.1
INFO 04-16 23:50:04 selector.py:38] Using ROCmFlashAttention backend.
(RayWorkerVllm pid=1376787) INFO 04-16 23:50:04 selector.py:38] Using ROCmFlashAttention backend.
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] Error executing method init_device. This might cause deadlock in distributed execution.
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] Traceback (most recent call last):
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] File "/home/vllm_install/vllm/vllm/engine/ray_utils.py", line 43, in execute_method
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] return executor(*args, **kwargs)
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] File "/home/vllm_install/vllm/vllm/worker/worker.py", line 97, in init_device
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] torch.cuda.set_device(self.device)
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] File "/home/vllm_install/venv/lib/python3.10/site-packages/torch/cuda/__init__.py", line 399, in set_device
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] torch._C._cuda_setDevice(device)
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] RuntimeError: HIP error: invalid device ordinal
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] For debugging consider passing AMD_SERIALIZE_KERNEL=3.
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50] Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.
(RayWorkerVllm pid=1376787) ERROR 04-16 23:50:04 ray_utils.py:50]
This with building on latest main. Was hoping this was fixed with https://github.com/vllm-project/vllm/pull/3770, but no amount of environmental configuration has helped either (CUDA_VISIBLE_DEVICES, etc).
I get same error
(vllm) ehartford@tw003:~/models/dolphin-2.9.2-qwen2-72b$ python -m vllm.entrypoints.openai.api_server --trust-remote-code --tensor-parallel-size 8 --model /home/ehartford/models/dolphin-2.9.2-qwen2-72b
INFO 05-27 20:36:35 config.py:569] Disabled the custom all-reduce kernel because it is not supported on AMD GPUs.
2024-05-27 20:36:37,825 INFO worker.py:1749 -- Started a local Ray instance.
INFO 05-27 20:36:40 llm_engine.py:103] Initializing an LLM engine (v0.4.2) with config: model='/home/ehartford/models/dolphin-2.9.2-qwen2-72b', speculative_config=None, tokenizer='/home/ehartford/models/dolphin-2.9.2-qwen2-72b', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, rope_scaling=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=131072, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=8, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), seed=0, served_model_name=/home/ehartford/models/dolphin-2.9.2-qwen2-72b)
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
INFO 05-27 20:36:57 selector.py:56] Using ROCmFlashAttention backend.
(RayWorkerWrapper pid=2194398) INFO 05-27 20:36:57 selector.py:56] Using ROCmFlashAttention backend.
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] Error executing method init_device. This might cause deadlock in distributed execution.
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] Traceback (most recent call last):
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] File "/home/ehartford/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm-0.4.2+rocm613-py3.11-linux-x86_64.egg/vllm/worker/worker_base.py", line 138, in execute_method
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] return executor(*args, **kwargs)
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] ^^^^^^^^^^^^^^^^^^^^^^^^^
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] File "/home/ehartford/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm-0.4.2+rocm613-py3.11-linux-x86_64.egg/vllm/worker/worker.py", line 105, in init_device
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] torch.cuda.set_device(self.device)
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] File "/home/ehartford/miniconda3/envs/vllm/lib/python3.11/site-packages/torch/cuda/__init__.py", line 404, in set_device
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] torch._C._cuda_setDevice(device)
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] RuntimeError: HIP error: invalid device ordinal
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] For debugging consider passing HIP_LAUNCH_BLOCKING=1.
(RayWorkerWrapper pid=2194149) ERROR 05-27 20:36:57 worker_base.py:146] Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.
ehartford
Hello Eric, I just encountered the same issue as you. Have you resolved it?
did you solve this?
Did you solve this? Several of my machine encountered this issue at the same time.
did you solve this?
Did you solve this? Several of my machine encountered this issue at the same time.
Reinstalling the conda env can solve this.
did you solve this?
Did you solve this? Several of my machine encountered this issue at the same time.
Reinstalling the conda env can solve this.
how to do this?
did you solve this?
Did you solve this? Several of my machine encountered this issue at the same time.
Reinstalling the conda env can solve this.
I will try this if I encounter it again
Closing this issue as the issue should have been resolved. Please open a new one if you run into the similar issue again
I believe this issue still exists.
First, PyTorch ROCm build does not obey to CUDA_DEVICE_ORDER="PCI_BUS_ID".
This can easily be verified with CUDA_DEVICE_ORDER="PCI_BUS_ID" python -c "import torch; import time; a = torch.rand(4000, 300000, device='cuda:0'); time.sleep(20);", giving:
It appears that vLLM with ray backend assigns some devices to each worker using CUDA_VISIBLE_DEVICES, as can be checked with ray.get_gpu_ids().
In my case, I get:
INFO 01-14 13:34:36 ray_gpu_executor.py:248] all_args_to_update_environment_variables [({'CUDA_VISIBLE_DEVICES': '0,1', 'VLLM_TRACE_FUNCTION': '0'},), ({'CUDA_VISIBLE_DEVICES': '0,1', 'VLLM_TRACE_FUNCTION': '0'},)]
...
INFO 01-14 13:17:49 worker.py:167] [rank 0] self.device cuda:0
INFO 01-14 13:17:49 worker.py:168] [rank 0] torch.cuda.device_count() 2
INFO 01-14 13:17:49 worker.py:170] [rank 0] ray.get_gpu_ids []
INFO 01-14 13:17:49 worker.py:174] [rank 0] can use cuda:0
INFO 01-14 13:17:49 worker.py:174] [rank 0] can use cuda:1
(RayWorkerWrapper pid=26711) INFO 01-14 13:17:49 worker.py:167] [rank 1] self.device cuda:1
(RayWorkerWrapper pid=26711) INFO 01-14 13:17:49 worker.py:168] [rank 1] torch.cuda.device_count() 2
(RayWorkerWrapper pid=26711) INFO 01-14 13:17:49 worker.py:170] [rank 1] ray.get_gpu_ids [1]
(RayWorkerWrapper pid=26711) INFO 01-14 13:17:49 worker.py:174] [rank 1] can use cuda:0
(RayWorkerWrapper pid=26711) INFO 01-14 13:17:49 worker.py:176] [rank 1] exception trying to use cuda:1, HIP error: invalid device ordinal
Sounds like a bug in ray to be fair https://github.com/ray-project/ray/issues/49260
Actually, pytorch rocm not respecting the PCI_BUS_ID order might not be the issue here. The issue seems more to be a conflict between CUDA_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES env variables.
Adding
if 'ROCR_VISIBLE_DEVICES' in os.environ:
del os.environ['ROCR_VISIBLE_DEVICES']
to https://github.com/ROCm/vllm/blob/113274a0e740ac779ac713de643e12856b10ce05/vllm/utils.py#L654-L660 fixed the issue for me.
Note that I was working on https://github.com/ROCm/vllm/commit/c040f0eb9ccab817363ad351f1f360e6b4cf81c2 with ray 2.40.0, the issue might be different / not exist in upstream vllm.
An other colleague had the same issue. I'll see if I can repro in vLLM upstream and tentatively fix this.
This should of course go without saying, and is unrelated to the actual issue on this thread, but for other dummies like me: Make sure you haven't accidentally set --tensor-parallel-size to a higher number than the actual number of GPUs that are connected to your machine.
