vllm - 💡(How to fix) Fix [Bug]: Abnormally bad performance on AMD ROCM gfx1030 (W6800, V620, 6900XT 6800XT) [4 comments, 3 participants]

vllm2026-03-25 12:19:03

ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

vllm-project/vllm#38107•Fetched 2026-04-08 01:32:16

View on GitHub

Comments

Participants

Timeline

Reactions

Author

Participants

Timeline (top)

mentioned ×5subscribed ×5commented ×4labeled ×2

Fix Action

Fix / Workaround

============================== CPU Info

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 39 bits physical, 48 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) W-1370P @ 3.60GHz CPU family: 6 Model: 167 Thread(s) per core: 2 Core(s) per socket: 8 Socket(s): 1 Stepping: 1 CPU(s) scaling MHz: 46% CPU max MHz: 5200.0000 CPU min MHz: 800.0000 BogoMIPS: 7200.00 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 smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap avx512ifma clflushopt intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid fsrm md_clear flush_l1d arch_capabilities L1d cache: 384 KiB (8 instances) L1i cache: 256 KiB (8 instances) L2 cache: 4 MiB (8 instances) L3 cache: 16 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-15 Vulnerability Gather data sampling: Vulnerable Vulnerability Ghostwrite: Not affected Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Old microcode: Not affected 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 Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Not affected

Code Example

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : version 4.3.0
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.11.0+rocm7.2
Is debug build               : False
CUDA used to build PyTorch   : N/A
ROCM used to build PyTorch   : 7.2.26015

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.17.0-19-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : AMD Radeon PRO V620 (gfx1030)
Nvidia driver version        : Could not collect
cuDNN version                : Could not collect
HIP runtime version          : 7.2.26015
MIOpen runtime version       : 3.5.1
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           39 bits physical, 48 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) W-1370P @ 3.60GHz
CPU family:                              6
Model:                                   167
Thread(s) per core:                      2
Core(s) per socket:                      8
Socket(s):                               1
Stepping:                                1
CPU(s) scaling MHz:                      46%
CPU max MHz:                             5200.0000
CPU min MHz:                             800.0000
BogoMIPS:                                7200.00
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 smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap avx512ifma clflushopt intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid fsrm md_clear flush_l1d arch_capabilities
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                4 MiB (8 instances)
L3 cache:                                16 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Vulnerable
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Old microcode:             Not affected
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
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

==============================
Versions of relevant libraries
==============================
[pip3] conch-triton-kernels==1.2.1
[pip3] numpy==2.1.3
[pip3] onnx==1.19.0
[pip3] onnx-ir==0.2.0
[pip3] onnxscript==0.6.2
[pip3] onnxslim==0.1.90
[pip3] pyzmq==27.1.0
[pip3] torch==2.11.0+rocm7.2
[pip3] torchvision==0.26.0+rocm7.2
[pip3] transformers==4.57.6
[pip3] triton==3.0.0+git3889f3f3
[pip3] triton-rocm==3.6.0
[conda] No relevant packages

==============================
         vLLM Info
==============================
ROCM Version                 : 7.2.26015-fc0010cf6a
vLLM Version                 : 0.1.dev15181+g2e67fa756 (git sha: 2e67fa756)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  ============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         
GPU0   0            40           
GPU1   40           0            

================================= Hops between two GPUs ==================================
       GPU0         GPU1         
GPU0   0            2            
GPU1   2            0            

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         
GPU0   0            PCIE         
GPU1   PCIE         0            

======================================= Numa Nodes =======================================
GPU[0]          : (Topology) Numa Node: 0
GPU[0]          : (Topology) Numa Affinity: -1
GPU[1]          : (Topology) Numa Node: 0
GPU[1]          : (Topology) Numa Affinity: -1
================================== End of ROCm SMI Log ===================================

==============================
     Environment Variables
==============================
PYTORCH_ROCM_ARCH=gfx1030
LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_arli

---

vllm serve /home/arli/models/Qwen3-8B \
--gpu-memory-utilization 0.9 --max-model-len 32768 --port 8000 \
--max-num-seqs 16 -tp 2 \
--attention-backend TRITON_ATTN \
--served-model-name test

---

{
  "model": "test",
  "messages": [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "Tell me a very very long story."}
  ],
  "temperature": 1.0,
  "max_tokens": 8192,
  "stream": false,
  "top_p": 0.95,
  "top_k": 20,
  "min_p": 0.0, 
  "presence_penalty": 0.0, 
  "repetition_penalty": 1.0
}

---

(vllm-amd) (base) arli@arli-arc-server:~/vllm-amd$ ./qwen.sh 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297] 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]        █     █     █▄   ▄█
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.1.dev15181+g2e67fa756
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]   █▄█▀ █     █     █     █  model   /home/arli/models/Qwen3-8B
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297] 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:233] non-default args: {'model_tag': '/home/arli/models/Qwen3-8B', 'model': '/home/arli/models/Qwen3-8B', 'max_model_len': 32768, 'served_model_name': ['test'], 'attention_backend': 'TRITON_ATTN', 'tensor_parallel_size': 2, 'max_num_seqs': 16}
(APIServer pid=19930) INFO 03-25 05:12:13 [model.py:540] Resolved architecture: Qwen3ForCausalLM
(APIServer pid=19930) INFO 03-25 05:12:13 [model.py:1607] Using max model len 32768
(APIServer pid=19930) INFO 03-25 05:12:13 [vllm.py:750] Asynchronous scheduling is enabled.
(EngineCore pid=19992) INFO 03-25 05:12:17 [core.py:105] Initializing a V1 LLM engine (v0.1.dev15181+g2e67fa756) with config: model='/home/arli/models/Qwen3-8B', speculative_config=None, tokenizer='/home/arli/models/Qwen3-8B', skip_tokenizer_init=False, 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, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=test, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['+sparse_attn_indexer', 'none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_images_per_batch': 0, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 32, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}
(EngineCore pid=19992) WARNING 03-25 05:12:17 [multiproc_executor.py:1014] Reducing Torch parallelism from 8 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
(EngineCore pid=19992) INFO 03-25 05:12:17 [multiproc_executor.py:134] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=192.168.1.75 (local), world_size=2, local_world_size=2
(Worker pid=20046) INFO 03-25 05:12:21 [parallel_state.py:1400] world_size=2 rank=1 local_rank=1 distributed_init_method=tcp://127.0.0.1:39443 backend=nccl
(Worker pid=20045) INFO 03-25 05:12:21 [parallel_state.py:1400] world_size=2 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:39443 backend=nccl
(Worker pid=20045) INFO 03-25 05:12:21 [pynccl.py:111] vLLM is using nccl==2.27.7
(Worker pid=20045) INFO 03-25 05:12:24 [parallel_state.py:1716] rank 0 in world size 2 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(Worker_TP0 pid=20045) INFO 03-25 05:12:24 [gpu_model_runner.py:4720] Starting to load model /home/arli/models/Qwen3-8B...
(Worker_TP0 pid=20045) INFO 03-25 05:12:24 [rocm.py:457] Using AttentionBackendEnum.TRITON_ATTN backend.
(Worker_TP1 pid=20046) INFO 03-25 05:12:24 [rocm.py:457] Using AttentionBackendEnum.TRITON_ATTN backend.
(Worker_TP0 pid=20045) WARNING 03-25 05:12:25 [compilation.py:1180] Op 'sparse_attn_indexer' not present in model, enabling with '+sparse_attn_indexer' has no effect
Loading safetensors checkpoint shards:   0% Completed | 0/5 [00:00<?, ?it/s]
Loading safetensors checkpoint shards:  20% Completed | 1/5 [00:01<00:07,  1.95s/it]
Loading safetensors checkpoint shards:  40% Completed | 2/5 [00:03<00:05,  1.87s/it]
Loading safetensors checkpoint shards:  60% Completed | 3/5 [00:05<00:03,  1.83s/it]
Loading safetensors checkpoint shards:  80% Completed | 4/5 [00:06<00:01,  1.67s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:07<00:00,  1.36s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:07<00:00,  1.55s/it]
(Worker_TP0 pid=20045) 
(Worker_TP0 pid=20045) INFO 03-25 05:12:32 [default_loader.py:384] Loading weights took 7.80 seconds
(Worker_TP0 pid=20045) INFO 03-25 05:12:33 [gpu_model_runner.py:4805] Model loading took 7.7 GiB memory and 8.069006 seconds
(Worker_TP0 pid=20045) INFO 03-25 05:12:35 [backends.py:1051] Using cache directory: /home/arli/.cache/vllm/torch_compile_cache/b73e4472c5/rank_0_0/backbone for vLLM's torch.compile
(Worker_TP0 pid=20045) INFO 03-25 05:12:35 [backends.py:1111] Dynamo bytecode transform time: 2.51 s
(Worker_TP0 pid=20045) INFO 03-25 05:12:41 [backends.py:372] Cache the graph of compile range (1, 2048) for later use
(Worker_TP0 pid=20045) INFO 03-25 05:12:45 [backends.py:390] Compiling a graph for compile range (1, 2048) takes 9.91 s
(Worker_TP1 pid=20046) INFO 03-25 05:12:46 [decorators.py:303] Directly load AOT compilation from path /home/arli/.cache/vllm/torch_compile_cache/torch_aot_compile/416497db4e76784c21b9d599c8bdd6d90b61038f404cc78401c4e3d36e3c899b/rank_1_0/model
(Worker_TP0 pid=20045) INFO 03-25 05:12:46 [monitor.py:48] torch.compile took 12.77 s in total
(Worker_TP0 pid=20045) INFO 03-25 05:12:46 [decorators.py:303] Directly load AOT compilation from path /home/arli/.cache/vllm/torch_compile_cache/torch_aot_compile/416497db4e76784c21b9d599c8bdd6d90b61038f404cc78401c4e3d36e3c899b/rank_0_0/model
(Worker_TP0 pid=20045) INFO 03-25 05:12:49 [monitor.py:76] Initial profiling/warmup run took 3.05 s
(Worker_TP0 pid=20045) INFO 03-25 05:12:53 [gpu_worker.py:436] Available KV cache memory: 20.23 GiB
(EngineCore pid=19992) INFO 03-25 05:12:53 [kv_cache_utils.py:1319] GPU KV cache size: 294,624 tokens
(EngineCore pid=19992) INFO 03-25 05:12:53 [kv_cache_utils.py:1324] Maximum concurrency for 32,768 tokens per request: 8.99x
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:04<00:00,  1.42it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:05<00:00,  1.02s/it]
(Worker_TP0 pid=20045) INFO 03-25 05:13:03 [gpu_model_runner.py:6028] Graph capturing finished in 10 secs, took 0.58 GiB
(EngineCore pid=19992) INFO 03-25 05:13:03 [core.py:283] init engine (profile, create kv cache, warmup model) took 30.63 seconds
(EngineCore pid=19992) INFO 03-25 05:13:04 [vllm.py:750] Asynchronous scheduling is enabled.
(APIServer pid=19930) INFO 03-25 05:13:04 [api_server.py:590] Supported tasks: ['generate']
(APIServer pid=19930) WARNING 03-25 05:13:05 [model.py:1401] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
(APIServer pid=19930) INFO 03-25 05:13:05 [hf.py:320] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
(APIServer pid=19930) INFO 03-25 05:13:05 [api_server.py:594] Starting vLLM server on http://0.0.0.0:8000
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:37] Available routes are:
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /docs, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /tokenize, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /detokenize, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /load, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /version, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /health, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /metrics, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/models, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /ping, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /ping, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /invocations, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/completions, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/messages, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/completions/render, Methods: POST
(APIServer pid=19930) INFO:     Started server process [19930]
(APIServer pid=19930) INFO:     Waiting for application startup.
(APIServer pid=19930) INFO:     Application startup complete.
(APIServer pid=19930) INFO 03-25 05:13:35 [loggers.py:259] Engine 000: Avg prompt throughput: 2.7 tokens/s, Avg generation throughput: 1.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
(APIServer pid=19930) INFO 03-25 05:13:45 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
(APIServer pid=19930) INFO 03-25 05:13:55 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%

---

Device 0 [AMD Radeon PRO V620]                  PCIe GEN 4@ 8x RX: N/A TX: N/A
 GPU 2460MHz MEM 1000MHz  TEMP  45°C   FAN N/A   POW 197 / 250 W
 GPU[||||||||||||||||||||||||||||||99%(eff 78%)] MEM[||||||||||||||||||||||||29.374Gi/31.984Gi]

 Device 1 [AMD Radeon PRO V620]                  PCIe GEN 4@ 8x RX: N/A TX: N/A
 GPU 2413MHz MEM 1000MHz  TEMP  44°C   FAN N/A   POW 203 / 250 W
 GPU[||||||||||||||||||||||||||||||99%(eff 80%)] MEM[||||||||||||||||||||||||29.462Gi/31.984Gi]

RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary>

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : version 4.3.0
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.11.0+rocm7.2
Is debug build               : False
CUDA used to build PyTorch   : N/A
ROCM used to build PyTorch   : 7.2.26015

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.17.0-19-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : AMD Radeon PRO V620 (gfx1030)
Nvidia driver version        : Could not collect
cuDNN version                : Could not collect
HIP runtime version          : 7.2.26015
MIOpen runtime version       : 3.5.1
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           39 bits physical, 48 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) W-1370P @ 3.60GHz
CPU family:                              6
Model:                                   167
Thread(s) per core:                      2
Core(s) per socket:                      8
Socket(s):                               1
Stepping:                                1
CPU(s) scaling MHz:                      46%
CPU max MHz:                             5200.0000
CPU min MHz:                             800.0000
BogoMIPS:                                7200.00
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 smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap avx512ifma clflushopt intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid fsrm md_clear flush_l1d arch_capabilities
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                4 MiB (8 instances)
L3 cache:                                16 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Vulnerable
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Old microcode:             Not affected
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
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

==============================
Versions of relevant libraries
==============================
[pip3] conch-triton-kernels==1.2.1
[pip3] numpy==2.1.3
[pip3] onnx==1.19.0
[pip3] onnx-ir==0.2.0
[pip3] onnxscript==0.6.2
[pip3] onnxslim==0.1.90
[pip3] pyzmq==27.1.0
[pip3] torch==2.11.0+rocm7.2
[pip3] torchvision==0.26.0+rocm7.2
[pip3] transformers==4.57.6
[pip3] triton==3.0.0+git3889f3f3
[pip3] triton-rocm==3.6.0
[conda] No relevant packages

==============================
         vLLM Info
==============================
ROCM Version                 : 7.2.26015-fc0010cf6a
vLLM Version                 : 0.1.dev15181+g2e67fa756 (git sha: 2e67fa756)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  ============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         
GPU0   0            40           
GPU1   40           0            

================================= Hops between two GPUs ==================================
       GPU0         GPU1         
GPU0   0            2            
GPU1   2            0            

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         
GPU0   0            PCIE         
GPU1   PCIE         0            

======================================= Numa Nodes =======================================
GPU[0]          : (Topology) Numa Node: 0
GPU[0]          : (Topology) Numa Affinity: -1
GPU[1]          : (Topology) Numa Node: 0
GPU[1]          : (Topology) Numa Affinity: -1
================================== End of ROCm SMI Log ===================================

==============================
     Environment Variables
==============================
PYTORCH_ROCM_ARCH=gfx1030
LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_arli

</details>

🐛 Describe the bug

Performance on AMD gfx1030 is abmnormally bad. This GPU using llama.cpp or kobold.cpp will easily do in the tens of tokens/s generation speed not single digits as with vllm here. While I know these generation GPUs don't have dedicated tensor units, I don't think the performance should be this bad. Trying to understand why the performance is so bad with vllm only.

Vllm and triton rocm are installed from source following the install steps in the vllm docs. No errors during build or install using ROCm 7.2.0.

Here is the result of running Qwen3-8B on 2xAMD Radeon Pro V620 32GB:

Start command:

vllm serve /home/arli/models/Qwen3-8B \
--gpu-memory-utilization 0.9 --max-model-len 32768 --port 8000 \
--max-num-seqs 16 -tp 2 \
--attention-backend TRITON_ATTN \
--served-model-name test

API call:

{
  "model": "test",
  "messages": [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "Tell me a very very long story."}
  ],
  "temperature": 1.0,
  "max_tokens": 8192,
  "stream": false,
  "top_p": 0.95,
  "top_k": 20,
  "min_p": 0.0, 
  "presence_penalty": 0.0, 
  "repetition_penalty": 1.0
}

Output:

(vllm-amd) (base) arli@arli-arc-server:~/vllm-amd$ ./qwen.sh 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297] 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]        █     █     █▄   ▄█
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.1.dev15181+g2e67fa756
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]   █▄█▀ █     █     █     █  model   /home/arli/models/Qwen3-8B
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:297] 
(APIServer pid=19930) INFO 03-25 05:12:13 [utils.py:233] non-default args: {'model_tag': '/home/arli/models/Qwen3-8B', 'model': '/home/arli/models/Qwen3-8B', 'max_model_len': 32768, 'served_model_name': ['test'], 'attention_backend': 'TRITON_ATTN', 'tensor_parallel_size': 2, 'max_num_seqs': 16}
(APIServer pid=19930) INFO 03-25 05:12:13 [model.py:540] Resolved architecture: Qwen3ForCausalLM
(APIServer pid=19930) INFO 03-25 05:12:13 [model.py:1607] Using max model len 32768
(APIServer pid=19930) INFO 03-25 05:12:13 [vllm.py:750] Asynchronous scheduling is enabled.
(EngineCore pid=19992) INFO 03-25 05:12:17 [core.py:105] Initializing a V1 LLM engine (v0.1.dev15181+g2e67fa756) with config: model='/home/arli/models/Qwen3-8B', speculative_config=None, tokenizer='/home/arli/models/Qwen3-8B', skip_tokenizer_init=False, 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, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=test, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['+sparse_attn_indexer', 'none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_images_per_batch': 0, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 32, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}
(EngineCore pid=19992) WARNING 03-25 05:12:17 [multiproc_executor.py:1014] Reducing Torch parallelism from 8 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
(EngineCore pid=19992) INFO 03-25 05:12:17 [multiproc_executor.py:134] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=192.168.1.75 (local), world_size=2, local_world_size=2
(Worker pid=20046) INFO 03-25 05:12:21 [parallel_state.py:1400] world_size=2 rank=1 local_rank=1 distributed_init_method=tcp://127.0.0.1:39443 backend=nccl
(Worker pid=20045) INFO 03-25 05:12:21 [parallel_state.py:1400] world_size=2 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:39443 backend=nccl
(Worker pid=20045) INFO 03-25 05:12:21 [pynccl.py:111] vLLM is using nccl==2.27.7
(Worker pid=20045) INFO 03-25 05:12:24 [parallel_state.py:1716] rank 0 in world size 2 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(Worker_TP0 pid=20045) INFO 03-25 05:12:24 [gpu_model_runner.py:4720] Starting to load model /home/arli/models/Qwen3-8B...
(Worker_TP0 pid=20045) INFO 03-25 05:12:24 [rocm.py:457] Using AttentionBackendEnum.TRITON_ATTN backend.
(Worker_TP1 pid=20046) INFO 03-25 05:12:24 [rocm.py:457] Using AttentionBackendEnum.TRITON_ATTN backend.
(Worker_TP0 pid=20045) WARNING 03-25 05:12:25 [compilation.py:1180] Op 'sparse_attn_indexer' not present in model, enabling with '+sparse_attn_indexer' has no effect
Loading safetensors checkpoint shards:   0% Completed | 0/5 [00:00<?, ?it/s]
Loading safetensors checkpoint shards:  20% Completed | 1/5 [00:01<00:07,  1.95s/it]
Loading safetensors checkpoint shards:  40% Completed | 2/5 [00:03<00:05,  1.87s/it]
Loading safetensors checkpoint shards:  60% Completed | 3/5 [00:05<00:03,  1.83s/it]
Loading safetensors checkpoint shards:  80% Completed | 4/5 [00:06<00:01,  1.67s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:07<00:00,  1.36s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:07<00:00,  1.55s/it]
(Worker_TP0 pid=20045) 
(Worker_TP0 pid=20045) INFO 03-25 05:12:32 [default_loader.py:384] Loading weights took 7.80 seconds
(Worker_TP0 pid=20045) INFO 03-25 05:12:33 [gpu_model_runner.py:4805] Model loading took 7.7 GiB memory and 8.069006 seconds
(Worker_TP0 pid=20045) INFO 03-25 05:12:35 [backends.py:1051] Using cache directory: /home/arli/.cache/vllm/torch_compile_cache/b73e4472c5/rank_0_0/backbone for vLLM's torch.compile
(Worker_TP0 pid=20045) INFO 03-25 05:12:35 [backends.py:1111] Dynamo bytecode transform time: 2.51 s
(Worker_TP0 pid=20045) INFO 03-25 05:12:41 [backends.py:372] Cache the graph of compile range (1, 2048) for later use
(Worker_TP0 pid=20045) INFO 03-25 05:12:45 [backends.py:390] Compiling a graph for compile range (1, 2048) takes 9.91 s
(Worker_TP1 pid=20046) INFO 03-25 05:12:46 [decorators.py:303] Directly load AOT compilation from path /home/arli/.cache/vllm/torch_compile_cache/torch_aot_compile/416497db4e76784c21b9d599c8bdd6d90b61038f404cc78401c4e3d36e3c899b/rank_1_0/model
(Worker_TP0 pid=20045) INFO 03-25 05:12:46 [monitor.py:48] torch.compile took 12.77 s in total
(Worker_TP0 pid=20045) INFO 03-25 05:12:46 [decorators.py:303] Directly load AOT compilation from path /home/arli/.cache/vllm/torch_compile_cache/torch_aot_compile/416497db4e76784c21b9d599c8bdd6d90b61038f404cc78401c4e3d36e3c899b/rank_0_0/model
(Worker_TP0 pid=20045) INFO 03-25 05:12:49 [monitor.py:76] Initial profiling/warmup run took 3.05 s
(Worker_TP0 pid=20045) INFO 03-25 05:12:53 [gpu_worker.py:436] Available KV cache memory: 20.23 GiB
(EngineCore pid=19992) INFO 03-25 05:12:53 [kv_cache_utils.py:1319] GPU KV cache size: 294,624 tokens
(EngineCore pid=19992) INFO 03-25 05:12:53 [kv_cache_utils.py:1324] Maximum concurrency for 32,768 tokens per request: 8.99x
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:04<00:00,  1.42it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:05<00:00,  1.02s/it]
(Worker_TP0 pid=20045) INFO 03-25 05:13:03 [gpu_model_runner.py:6028] Graph capturing finished in 10 secs, took 0.58 GiB
(EngineCore pid=19992) INFO 03-25 05:13:03 [core.py:283] init engine (profile, create kv cache, warmup model) took 30.63 seconds
(EngineCore pid=19992) INFO 03-25 05:13:04 [vllm.py:750] Asynchronous scheduling is enabled.
(APIServer pid=19930) INFO 03-25 05:13:04 [api_server.py:590] Supported tasks: ['generate']
(APIServer pid=19930) WARNING 03-25 05:13:05 [model.py:1401] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
(APIServer pid=19930) INFO 03-25 05:13:05 [hf.py:320] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
(APIServer pid=19930) INFO 03-25 05:13:05 [api_server.py:594] Starting vLLM server on http://0.0.0.0:8000
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:37] Available routes are:
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /docs, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /tokenize, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /detokenize, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /load, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /version, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /health, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /metrics, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/models, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /ping, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /ping, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /invocations, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/completions, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/messages, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
(APIServer pid=19930) INFO 03-25 05:13:05 [launcher.py:46] Route: /v1/completions/render, Methods: POST
(APIServer pid=19930) INFO:     Started server process [19930]
(APIServer pid=19930) INFO:     Waiting for application startup.
(APIServer pid=19930) INFO:     Application startup complete.
(APIServer pid=19930) INFO 03-25 05:13:35 [loggers.py:259] Engine 000: Avg prompt throughput: 2.7 tokens/s, Avg generation throughput: 1.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
(APIServer pid=19930) INFO 03-25 05:13:45 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
(APIServer pid=19930) INFO 03-25 05:13:55 [loggers.py:259] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%

GPU stats during inference looks normal?

 Device 0 [AMD Radeon PRO V620]                  PCIe GEN 4@ 8x RX: N/A TX: N/A
 GPU 2460MHz MEM 1000MHz  TEMP  45°C   FAN N/A   POW 197 / 250 W
 GPU[||||||||||||||||||||||||||||||99%(eff 78%)] MEM[||||||||||||||||||||||||29.374Gi/31.984Gi]

 Device 1 [AMD Radeon PRO V620]                  PCIe GEN 4@ 8x RX: N/A TX: N/A
 GPU 2413MHz MEM 1000MHz  TEMP  44°C   FAN N/A   POW 203 / 250 W
 GPU[||||||||||||||||||||||||||||||99%(eff 80%)] MEM[||||||||||||||||||||||||29.462Gi/31.984Gi]

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.

extent analysis

Fix Plan

To improve the performance of vLLM on AMD Radeon PRO V620, consider the following steps:

Update ROCm and drivers: Ensure that the latest versions of ROCm and AMD drivers are installed, as updates often include performance optimizations.
Optimize memory allocation: The current memory utilization is around 29 GB out of 32 GB per GPU. Consider reducing the --gpu-memory-utilization parameter to prevent potential memory bottlenecks.
Adjust the --max-num-seqs parameter: The current setting of 16 might be too high, causing contention and reducing performance. Try reducing this value to 8 or 4 to see if it improves throughput.
Experiment with different attention backends: The current setting is TRITON_ATTN. Try switching to other attention backends, such as FLASH_ATTN or CUDA_ATTN, to see if they offer better performance.
Profile and optimize the model: Use tools like torch.profiler to identify performance bottlenecks in the model and optimize them accordingly.

Example code to update the vllm command with adjusted parameters:

vllm serve /home/arli/models/Qwen3-8B \
--gpu-memory-utilization 0.8 --max-model-len 32768 --port 8000 \
--max-num-seqs 8 -tp 2 \
--attention-backend FLASH_ATTN \
--served-model-name test

Verification

To verify the effectiveness of these changes, monitor the GPU stats and throughput metrics during inference. Compare the results with the original configuration to determine if the adjustments have improved performance.

Extra Tips

Regularly update ROCm and drivers to ensure you have the latest performance optimizations.
Experiment with different model configurations and attention backends to find the optimal setup for your specific use case.
Consider using tools like torch.profiler to identify and optimize performance bottlenecks in the model.

Vote matrix · Quick signals

Works

Did the solution work? Tap to confirm.

Easy Fix

Was it a quick fix?

Time Saver

Did it save you time?

Blocking

Was it severely blocking?

Common Issue

Are others likely hitting this too?

Flaky / Intermittent

Is it intermittent?

Verified / Reproducible

Can you reproduce it reliably?

#api #memory optimization #batch processing #model loading #environment variable

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

vllm - 💡(How to fix) Fix [Bug]: Abnormally bad performance on AMD ROCM gfx1030 (W6800, V620, 6900XT 6800XT) [4 comments, 3 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fix / Workaround

============================== CPU Info

Code Example

Your current environment

🐛 Describe the bug

Before submitting a new issue...

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

TRENDING

vllm - 💡(How to fix) Fix [Bug]: Abnormally bad performance on AMD ROCM gfx1030 (W6800, V620, 6900XT 6800XT) [4 comments, 3 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fix / Workaround

============================== CPU Info

Code Example

Your current environment

🐛 Describe the bug

Before submitting a new issue...

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

RELATED_DISCOVERY

TRENDING