[Issue]: failed deploying QWen-VL on AMD MI308X GPU using vllm backend, #2105
Description
Problem Description
Script
CUDA_VISIBLE_DEVICES=0 vllm serve /home/admin/HF/Qwen2.5-VL-7B \
--host localhost \
--port 1234 \
--limit-mm-per-prompt '{"image":8,"video":1}'Bug detail
Error log:
(EngineCore_DP0 pid=50949) DEBUG 10-15 17:21:00 [config/compilation.py:824] disabled custom ops: Counter({'rms_norm': 122, 'column_parallel_linear': 121, 'row_parallel_linear': 121, 'silu_and_mul': 29, 'vocab_parallel_embedding': 1, 'rotary_embedding': 1, 'parallel_lm_head': 1, 'logits_processor': 1})
(EngineCore_DP0 pid=50949) DEBUG 10-15 17:21:00 [model_executor/model_loader/base_loader.py:53] Loading weights on cuda ...
Loading safetensors checkpoint shards: 0% Completed | 0/5 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 20% Completed | 1/5 [00:01<00:05, 1.29s/it]
Loading safetensors checkpoint shards: 40% Completed | 2/5 [00:02<00:04, 1.38s/it]
Loading safetensors checkpoint shards: 60% Completed | 3/5 [00:04<00:02, 1.41s/it]
(EngineCore_DP0 pid=50949) DEBUG 10-15 17:21:05 [model_executor/models/utils.py:204] Loaded weight lm_head.weight with shape torch.Size([152064, 3584])
Loading safetensors checkpoint shards: 80% Completed | 4/5 [00:04<00:01, 1.07s/it]
(APIServer pid=50822) DEBUG 10-15 17:21:06 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:06<00:00, 1.16s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:06<00:00, 1.21s/it]
(EngineCore_DP0 pid=50949)
(EngineCore_DP0 pid=50949) INFO 10-15 17:21:06 [model_executor/model_loader/default_loader.py:314] Loading weights took 6.19 seconds
(EngineCore_DP0 pid=50949) INFO 10-15 17:21:07 [v1/worker/gpu_model_runner.py:2906] Model loading took 15.8145 GiB and 6.607936 seconds
(EngineCore_DP0 pid=50949) INFO 10-15 17:21:07 [v1/worker/gpu_model_runner.py:3672] Encoder cache will be initialized with a budget of 114688 tokens, and profiled with 1 video items of the maximum feature size.
MIOpen(HIP): miopenStatus_t miopenCreateTensorDescriptor(miopenTensorDescriptor_t *){
MIOpen(HIP): tensorDesc = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenSetTensorDescriptor(miopenTensorDescriptor_t, miopenDataType_t, int, const int *, const int *){
MIOpen(HIP): tensorDesc = {}, {}, packed,
MIOpen(HIP): dataType = 5
MIOpen(HIP): nbDims = 5
MIOpen(HIP): dim.values = { 458752 3 2 14 14 }
MIOpen(HIP): stride.values = { 1176 392 196 14 1 }
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenCreateTensorDescriptor(miopenTensorDescriptor_t *){
MIOpen(HIP): tensorDesc = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenSetTensorDescriptor(miopenTensorDescriptor_t, miopenDataType_t, int, const int *, const int *){
MIOpen(HIP): tensorDesc = {}, {}, packed,
MIOpen(HIP): dataType = 5
MIOpen(HIP): nbDims = 5
MIOpen(HIP): dim.values = { 1280 3 2 14 14 }
MIOpen(HIP): stride.values = { 1176 392 196 14 1 }
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenCreateTensorDescriptor(miopenTensorDescriptor_t *){
MIOpen(HIP): tensorDesc = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenSetTensorDescriptor(miopenTensorDescriptor_t, miopenDataType_t, int, const int *, const int *){
MIOpen(HIP): tensorDesc = {}, {}, packed,
MIOpen(HIP): dataType = 5
MIOpen(HIP): nbDims = 5
MIOpen(HIP): dim.values = { 458752 1280 1 1 1 }
MIOpen(HIP): stride.values = { 1280 1 1 1 1 }
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenCreateConvolutionDescriptor(miopenConvolutionDescriptor_t *){
MIOpen(HIP): convDesc = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenInitConvolutionNdDescriptor(miopenConvolutionDescriptor_t, int, const int *, const int *, const int *, miopenConvolutionMode_t){
MIOpen(HIP): convDesc = conv2d, miopenConvolution, miopenPaddingDefault, {0, 0}, {1, 1}, {1, 1},
MIOpen(HIP): spatialDim = 3
MIOpen(HIP): pads = { 0 0 0 }
MIOpen(HIP): strides = { 2 14 14 }
MIOpen(HIP): dilations = { 1 1 1 }
MIOpen(HIP): c_mode = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenSetConvolutionGroupCount(miopenConvolutionDescriptor_t, int){
MIOpen(HIP): convDesc = conv3d, miopenConvolution, miopenPaddingDefault, {0, 0, 0}, {2, 14, 14}, {1, 1, 1},
MIOpen(HIP): groupCount = 1
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenSetConvolutionAttribute(miopenConvolutionDescriptor_t, const miopenConvolutionAttrib_t, const int){
MIOpen(HIP): convDesc = conv3d, miopenConvolution, miopenPaddingDefault, {0, 0, 0}, {2, 14, 14}, {1, 1, 1},
MIOpen(HIP): attr = 1
MIOpen(HIP): value = 0
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenConvolutionForwardGetWorkSpaceSize(miopenHandle_t, const miopenTensorDescriptor_t, const miopenTensorDescriptor_t, const miopenConvolutionDescriptor_t, const miopenTensorDescriptor_t, size_t *){
MIOpen(HIP): handle = stream: 0, device_id: 0
MIOpen(HIP): wDesc = {1280, 3, 2, 14, 14}, {1176, 392, 196, 14, 1}, packed,
MIOpen(HIP): xDesc = {458752, 3, 2, 14, 14}, {1176, 392, 196, 14, 1}, packed,
MIOpen(HIP): convDesc = conv3d, miopenConvolution, miopenPaddingDefault, {0, 0, 0}, {2, 14, 14}, {1, 1, 1},
MIOpen(HIP): yDesc = {458752, 1280, 1, 1, 1}, {1280, 1, 1, 1, 1}, packed,
MIOpen(HIP): }
MIOpen(HIP): miopenStatus_t miopenFindConvolutionForwardAlgorithm(miopenHandle_t, const miopenTensorDescriptor_t, const void *, const miopenTensorDescriptor_t, const void *, const miopenConvolutionDescriptor_t, const miopenTensorDescriptor_t, void *, const int, int *, miopenConvAlgoPerf_t *, void *, size_t, bool){
MIOpen(HIP): handle = stream: 0, device_id: 0
MIOpen(HIP): xDesc = {458752, 3, 2, 14, 14}, {1176, 392, 196, 14, 1}, packed,
MIOpen(HIP): x = 0x7f1ddc800000
MIOpen(HIP): wDesc = {1280, 3, 2, 14, 14}, {1176, 392, 196, 14, 1}, packed,
MIOpen(HIP): w = 0x7f230aa00000
MIOpen(HIP): convDesc = conv3d, miopenConvolution, miopenPaddingDefault, {0, 0, 0}, {2, 14, 14}, {1, 1, 1},
MIOpen(HIP): yDesc = {458752, 1280, 1, 1, 1}, {1280, 1, 1, 1, 1}, packed,
MIOpen(HIP): y = 0x7f1d96600000
MIOpen(HIP): requestAlgoCount = 1
MIOpen(HIP): returnedAlgoCount = 0
MIOpen(HIP): perfResults =
MIOpen(HIP): workSpace = 0x7f1b95200000
MIOpen(HIP): workSpaceSize = 2256400384
MIOpen(HIP): exhaustiveSearch = 0
MIOpen(HIP): }
MIOpen(HIP): Command [LogCmdFindConvolution] ./bin/MIOpenDriver convbfp16 -n 458752 -c 3 --in_d 2 -H 14 -W 14 -k 1280 --fil_d 2 -y 14 -x 14 --pad_d 0 -p 0 -q 0 --conv_stride_d 2 -u 14 -v 14 --dilation_d 1 -l 1 -j 1 --spatial_dim 3 -m conv -g 1 -F 1 -t 1
MIOpen(HIP): Error [EvaluateInvokers] /longer_pathname_so_that_rpms_can_support_packaging_the_debug_info_for_all_os_profiles/src/rocm-libraries/projects/miopen/src/hipoc/hipoc_kernel.cpp:128: Failed to launch kernel: invalid configuration argument
MIOpen(HIP): auto miopen::solver::conv::GemmFwdRest::GetSolution(const ExecutionContext &, const ProblemDescription &)::(anonymous class)::operator()(const std::vector<Kernel> &)::(anonymous class)::operator()(const Handle &, const AnyInvokeParams &) const{
MIOpen(HIP): name + ", non 1x1" = convolution, non 1x1
MIOpen(HIP): }
(APIServer pid=50822) DEBUG 10-15 17:21:16 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:21:26 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:21:36 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:21:46 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:21:56 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:06 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:16 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:26 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:36 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:46 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:22:56 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:23:06 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=50822) DEBUG 10-15 17:23:16 [v1/engine/utils.py:863] Waiting for 1 local, 0 remote core engine proc(s) to start.
MIOpen(HIP): auto miopen::solver::conv::GemmFwdRest::GetSolution(const ExecutionContext &, const ProblemDescription &)::(anonymous class)::operator()(const std::vector<Kernel> &)::(anonymous class)::operator()(const Handle &, const AnyInvokeParams &) const{
MIOpen(HIP): name + ", non 1x1" = convolution, non 1x1
MIOpen(HIP): }
Current environment
details
Collecting environment information...
==============================
System Info
==============================
OS : Red Hat Enterprise Linux 8.8(Ootpa) (x86_64)
GCC version : (GCC) 10.2.1 20200825
Clang version : Could not collect
CMake version : version 4.1.0
Libc version : glibc-2.32
==============================
PyTorch Info
==============================
PyTorch version : 2.8.0+rocm6.4
Is debug build : False
CUDA used to build PyTorch : N/A
ROCM used to build PyTorch : 6.4.43482-0f2d60242
==============================
Python Environment
==============================
Python version : 3.10.13 (main, Sep 11 2023, 13:44:35) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.10.134-16.3.al8.x86_64-x86_64-with-glibc2.32
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : Could not collect
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration : AMD Instinct MI308X (gfx942:sramecc+:xnack-)
Nvidia driver version : Could not collect
cuDNN version : Could not collect
HIP runtime version : 6.4.43482
MIOpen runtime version : 3.4.0
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 192
On-line CPU(s) list: 8-47,50-55,104-143,146-151
Off-line CPU(s) list: 0-7,48,49,56-103,144,145,152-191
Thread(s) per core: 0
Core(s) per socket: 48
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 207
Model name: INTEL(R) XEON(R) PLATINUM 8575C
Stepping: 2
CPU MHz: 3200.008
CPU max MHz: 4000.0000
CPU min MHz: 800.0000
BogoMIPS: 5600.00
Virtualization: VT-x
L1d cache: 48K
L1i cache: 32K
L2 cache: 2048K
L3 cache: 327680K
NUMA node0 CPU(s): 0-47,96-143
NUMA node1 CPU(s): 48-95,144-191
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm uintr md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
==============================
Versions of relevant libraries
==============================
[pip3] conch-triton-kernels==1.2.1
[pip3] numpy==1.26.4
[pip3] nvidia-ml-py3==7.352.0
[pip3] pytorch-triton-rocm==3.4.0
[pip3] pyzmq==26.4.0
[pip3] torch==2.8.0+rocm6.4
[pip3] torchaudio==2.6.0+rocm6.4.1.gitd8831425
[pip3] torchdata==0.11.0
[pip3] torchmetrics==1.7.2
[pip3] torchvision==0.23.0+rocm6.4
[pip3] transformers==4.57.0
[pip3] triton==3.2.0
[conda] mkl 2025.1.0 pypi_0 pypi
[conda] mkl-include 2025.1.0 pypi_0 pypi
[conda] numpy 1.26.4 pypi_0 pypi
[conda] nvidia-ml-py3 7.352.0 pypi_0 pypi
[conda] pytorch-triton-rocm 3.2.0+rocm6.4.1.git6da9e660 pypi_0 pypi
[conda] pyzmq 26.4.0 pypi_0 pypi
[conda] torch 2.6.0+rocm6.4.1.git5c42a370 pypi_0 pypi
[conda] torchaudio 2.6.0+rocm6.4.1.gitd8831425 pypi_0 pypi
[conda] torchdata 0.11.0 pypi_0 pypi
[conda] torchmetrics 1.7.2 pypi_0 pypi
[conda] torchvision 0.21.0+rocm6.4.1.git4040d51f pypi_0 pypi
[conda] transformers 4.38.2 pypi_0 pypi
[conda] triton 3.2.0 pypi_0 pypi
==============================
vLLM Info
==============================
ROCM Version : 6.4.43483-a187df25c
vLLM Version : 0.11.0rc2.dev436+g98f30b8cb (git sha: 98f30b8cb)
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Enabled
GPU Topology:
============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
GPU0 GPU1 GPU2 GPU3
GPU0 0 15 15 15
GPU1 15 0 15 15
GPU2 15 15 0 15
GPU3 15 15 15 0
================================= Hops between two GPUs ==================================
GPU0 GPU1 GPU2 GPU3
GPU0 0 1 1 1
GPU1 1 0 1 1
GPU2 1 1 0 1
GPU3 1 1 1 0
=============================== Link Type between two GPUs ===============================
GPU0 GPU1 GPU2 GPU3
GPU0 0 XGMI XGMI XGMI
GPU1 XGMI 0 XGMI XGMI
GPU2 XGMI XGMI 0 XGMI
GPU3 XGMI XGMI XGMI 0
======================================= Numa Nodes =======================================
GPU[0] : (Topology) Numa Node: 0
GPU[0] : (Topology) Numa Affinity: 0
GPU[1] : (Topology) Numa Node: 0
GPU[1] : (Topology) Numa Affinity: 0
GPU[2] : (Topology) Numa Node: 0
GPU[2] : (Topology) Numa Affinity: 0
GPU[3] : (Topology) Numa Node: 0
GPU[3] : (Topology) Numa Affinity: 0
================================== End of ROCm SMI Log ===================================
==============================
Environment Variables
==============================
LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/gcc-5.3.0/lib64:/usr/local/lib:/usr/local/lib64:/usr/local/lib64/boost:/opt/taobao/java/jre/lib/amd64/server:/hadoop_java/java/jdk/jre/lib/amd64/server:/worker/:/worker/lib:/apsara_lib64/:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/opt/conda/envs/python3.10.13/lib/python3.10/site-packages/:/lib:/opt/rocm/lib:/opt/amdgpu/lib64:/usr/local/lib64:/lib64::/usr/lib64:/usr/local/gcc75/lib:/usr/local/gcc75/lib64:/pu:/opt/taobao/java/jre/lib/amd64/server:/apsara/alicpp/built/gcc-4.9.2/glog-0.3.4/lib:/apsara/alicpp/built/gcc-4.9.2/gflags-2.1.2/lib:/apsara/alicpp/built/gcc-4.9.2/protobuf-2.4.1.ali/lib:/apsara/alicpp/built/gcc-4.9.2/odps-cryptography-1.0.0/lib:/apsara/alicpp/built/gcc-4.9.2/boost-1.58.0.fix.thread/lib:/apsara/alicpp/built/gcc-4.9.2/openssl-1.0.2a/lib:/apsara/alicpp/built/gcc-4.9.2/mysql-connector-c-6.1.6/lib:/apsara/alicpp/built/gcc-4.9.2/arrow-0.16.0/lib64:/apsara/alicpp/built/gcc-4.9.2/bzip2-1.0.6/lib64:/apsara/alicpp/built/gcc-4.9.2/zstd-1.4.4/lib:/apsara/alicpp/built/gcc-4.9.2/libevent-2.0.22.stable/lib64:/worker:/worker/lib:/opt/conda/envs/python3.10.13/lib
NCCL_IB_GID_INDEX=3
TORCH_EXTENSIONS_DIR=:/usr/local/ninja
NVIDIA_VISIBLE_DEVICES=
NVIDIA_DRIVER_CAPABILITIES=all
NCCL_NET_PLUGIN=none
NCCL_DEBUG=INFO
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY
Operating System
Red Hat Enterprise Linux 8.8(Ootpa) (x86_64)
CPU
INTEL(R) XEON(R) PLATINUM 8575C
GPU
AMD Instinct MI308X (gfx942:sramecc+:xnack-)
ROCm Version
ROCm 6.4.1
ROCm Component
MIOpen
Steps to Reproduce
No response
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response