[Bug]: TTFT increased especially in some Distill Models with small BatchSize in v0.10.0 compared to v0.9.2 #21983
Description
Your current environment
After Update (v0.10.0):
The output of python collect_env.py
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.2 LTS (x86_64)
GCC version : (Ubuntu 11.3.0-1ubuntu1~22.04.1) 11.3.0
Clang version : Could not collect
CMake version : version 3.26.4
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.7.1+cu126
Is debug build : False
CUDA used to build PyTorch : 12.6
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.10.6 (main, Oct 24 2022, 16:07:47) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.4.0-144-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.1.105
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration :
GPU 0: NVIDIA GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
GPU 2: NVIDIA GeForce RTX 4090
GPU 3: NVIDIA GeForce RTX 4090
Nvidia driver version : 550.54.15
cuDNN version : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.3
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
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): 64
On-line CPU(s) list: 0-63
Vendor ID: GenuineIntel
Model name: INTEL(R) XEON(R) GOLD 6530
CPU family: 6
Model: 207
Thread(s) per core: 1
Core(s) per socket: 32
Socket(s): 2
Stepping: 2
CPU max MHz: 4000.0000
CPU min MHz: 800.0000
BogoMIPS: 4200.00
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 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 avx2 smep bmi2 erms invpcid 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 avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid cldemote movdiri movdir64b md_clear pconfig flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 3 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 128 MiB (64 instances)
L3 cache: 320 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-31
NUMA node1 CPU(s): 32-63
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 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 IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pyzmq==27.0.0
[pip3] torch==2.7.1
[pip3] torchaudio==2.7.1
[pip3] torchvision==0.22.1
[pip3] transformers==4.54.1
[pip3] triton==3.3.1
[conda] numpy 2.2.6 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.6.4.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.6.80 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.6.77 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.6.77 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.5.1.17 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.0.4 pypi_0 pypi
[conda] nvidia-cufile-cu12 1.11.1.6 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.7.77 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.1.2 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.4.2 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.6.3 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.26.2 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.6.85 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.6.77 pypi_0 pypi
[conda] pyzmq 27.0.0 pypi_0 pypi
[conda] torch 2.7.1 pypi_0 pypi
[conda] torchaudio 2.7.1 pypi_0 pypi
[conda] torchvision 0.22.1 pypi_0 pypi
[conda] transformers 4.54.1 pypi_0 pypi
[conda] triton 3.3.1 pypi_0 pypi
==============================
vLLM Info
==============================
ROCM Version : Could not collect
Neuron SDK Version : N/A
vLLM Version : 0.10.0
vLLM Build Flags:
CUDA Archs: 5.2 6.0 6.1 7.0 7.5 8.0 8.6 9.0+PTX; ROCm: Disabled; Neuron: Disabled
GPU Topology:
�[4mGPU0 GPU1 GPU2 GPU3 CPU Affinity NUMA Affinity GPU NUMA ID�[0m
GPU0 X SYS SYS SYS 0-31 0 N/A
GPU1 SYS X SYS SYS 0-31 0 N/A
GPU2 SYS SYS X SYS 32-63 1 N/A
GPU3 SYS SYS SYS X 32-63 1 N/A
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
==============================
Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=GPU-066fb607-7eac-9deb-189e-45f7a47a8d6f,GPU-7ba2cd7b-26af-f6a3-98c0-784e5291c5aa,GPU-c4a62fdc-f985-9701-fe47-cc341a8dd0d3,GPU-cc612b54-3513-d276-331e-ff46e6315b96
CUBLAS_VERSION=12.1.3.1
NVIDIA_REQUIRE_CUDA=cuda>=9.0
CUDA_CACHE_DISABLE=1
TORCH_CUDA_ARCH_LIST=5.2 6.0 6.1 7.0 7.5 8.0 8.6 9.0+PTX
NCCL_VERSION=2.18.3
NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
VLLM_USE_MODELSCOPE=True
NVIDIA_PRODUCT_NAME=PyTorch
CUDA_VERSION=12.1.1.009
PYTORCH_VERSION=2.1.0a0+b5021ba
PYTORCH_BUILD_NUMBER=0
CUDNN_VERSION=8.9.3.28
PYTORCH_HOME=/opt/pytorch/pytorch
LD_LIBRARY_PATH=/usr/local/nvidia/lib64
NVIDIA_BUILD_ID=63867923
CUDA_DRIVER_VERSION=530.30.02
PYTORCH_BUILD_VERSION=2.1.0a0+b5021ba
CUDA_HOME=/usr/local/cuda
CUDA_HOME=/usr/local/cuda
CUDA_MODULE_LOADING=LAZY
NVIDIA_REQUIRE_JETPACK_HOST_MOUNTS=
NVIDIA_PYTORCH_VERSION=23.07
TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
Before Update (v0.9.2):
The output of python collect_env.py
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.2 LTS (x86_64)
GCC version : (Ubuntu 11.3.0-1ubuntu1~22.04.1) 11.3.0
Clang version : Could not collect
CMake version : version 3.26.4
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.7.0+cu126
Is debug build : False
CUDA used to build PyTorch : 12.6
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.10.6 (main, Oct 24 2022, 16:07:47) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.4.0-144-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.1.105
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration :
GPU 0: NVIDIA GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
GPU 2: NVIDIA GeForce RTX 4090
GPU 3: NVIDIA GeForce RTX 4090
Nvidia driver version : 550.54.15
cuDNN version : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.3
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.3
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
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): 64
On-line CPU(s) list: 0-63
Vendor ID: GenuineIntel
Model name: INTEL(R) XEON(R) GOLD 6530
CPU family: 6
Model: 207
Thread(s) per core: 1
Core(s) per socket: 32
Socket(s): 2
Stepping: 2
CPU max MHz: 4000.0000
CPU min MHz: 800.0000
BogoMIPS: 4200.00
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 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 avx2 smep bmi2 erms invpcid 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 avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid cldemote movdiri movdir64b md_clear pconfig flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 3 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 128 MiB (64 instances)
L3 cache: 320 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-31
NUMA node1 CPU(s): 32-63
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 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 IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pyzmq==27.0.0
[pip3] torch==2.7.0
[pip3] torchaudio==2.7.0
[pip3] torchvision==0.22.0
[pip3] transformers==4.53.3
[pip3] triton==3.3.0
[conda] numpy 2.2.6 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.6.4.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.6.80 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.6.77 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.6.77 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.5.1.17 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.0.4 pypi_0 pypi
[conda] nvidia-cufile-cu12 1.11.1.6 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.7.77 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.1.2 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.4.2 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.6.3 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.26.2 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.6.85 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.6.77 pypi_0 pypi
[conda] pyzmq 27.0.0 pypi_0 pypi
[conda] torch 2.7.0 pypi_0 pypi
[conda] torchaudio 2.7.0 pypi_0 pypi
[conda] torchvision 0.22.0 pypi_0 pypi
[conda] transformers 4.53.3 pypi_0 pypi
[conda] triton 3.3.0 pypi_0 pypi
==============================
vLLM Info
==============================
ROCM Version : Could not collect
Neuron SDK Version : N/A
vLLM Version : 0.9.2
vLLM Build Flags:
CUDA Archs: 5.2 6.0 6.1 7.0 7.5 8.0 8.6 9.0+PTX; ROCm: Disabled; Neuron: Disabled
GPU Topology:
�[4mGPU0 GPU1 GPU2 GPU3 CPU Affinity NUMA Affinity GPU NUMA ID�[0m
GPU0 X SYS SYS SYS 0-31 0 N/A
GPU1 SYS X SYS SYS 0-31 0 N/A
GPU2 SYS SYS X SYS 32-63 1 N/A
GPU3 SYS SYS SYS X 32-63 1 N/A
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
==============================
Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=GPU-066fb607-7eac-9deb-189e-45f7a47a8d6f,GPU-7ba2cd7b-26af-f6a3-98c0-784e5291c5aa,GPU-c4a62fdc-f985-9701-fe47-cc341a8dd0d3,GPU-cc612b54-3513-d276-331e-ff46e6315b96
CUBLAS_VERSION=12.1.3.1
NVIDIA_REQUIRE_CUDA=cuda>=9.0
CUDA_CACHE_DISABLE=1
TORCH_CUDA_ARCH_LIST=5.2 6.0 6.1 7.0 7.5 8.0 8.6 9.0+PTX
NCCL_VERSION=2.18.3
NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
VLLM_USE_MODELSCOPE=True
NVIDIA_PRODUCT_NAME=PyTorch
CUDA_VERSION=12.1.1.009
PYTORCH_VERSION=2.1.0a0+b5021ba
PYTORCH_BUILD_NUMBER=0
CUDNN_VERSION=8.9.3.28
PYTORCH_HOME=/opt/pytorch/pytorch
LD_LIBRARY_PATH=/usr/local/nvidia/lib64
NVIDIA_BUILD_ID=63867923
CUDA_DRIVER_VERSION=530.30.02
PYTORCH_BUILD_VERSION=2.1.0a0+b5021ba
CUDA_HOME=/usr/local/cuda
CUDA_HOME=/usr/local/cuda
CUDA_MODULE_LOADING=LAZY
NVIDIA_REQUIRE_JETPACK_HOST_MOUNTS=
NVIDIA_PYTORCH_VERSION=23.07
TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
🐛 Describe the bug
When running some distilled models (e.g. DeepSeek-R1-Distill-Llama-70B(W8A8) and DeepSeek-R1-Distill-Qwen-32B(BF16) ), we observe a significant increase in Time-To-First-Token (TTFT) after upgrading to vLLM v0.10.0 compared to vLLM v0.9.2.
The degradation is most pronounced with small batch sizes (bs=1), where TTFT increases by 33.28% to 96.73% (measured across multiple test cases). However, as the batch size gradually increases to bs=10, the TTFT difference diminishes and eventually aligns with the performance of the previous version (v0.9.2).
If benchmark code or additional testing is needed to investigate this issue, feel free to contact me.
Example Case: DeepSeek-R1-Distill-Llama-70B-quantized.w8a8
Before Update:
Server Log:
(vllm092) root@test:/workspace$ vllm serve /root/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 --tensor-parallel-size 4 --max-num-seqs 1 --max-model-len 10240 --trust-remote-code --gpu-memory-utilization 0.9
INFO 07-25 09:36:40 [__init__.py:244] Automatically detected platform cuda.
INFO 07-25 09:36:43 [api_server.py:1395] vLLM API server version 0.9.2
INFO 07-25 09:36:43 [cli_args.py:325] non-default args: {'model': '/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', 'trust_remote_code': True, 'max_model_len': 10240, 'tensor_parallel_size': 4, 'gpu_memory_utilization': 0.85, 'max_num_seqs': 1}
INFO 07-25 09:36:49 [config.py:841] This model supports multiple tasks: {'generate', 'classify', 'reward', 'embed'}. Defaulting to 'generate'.
INFO 07-25 09:36:49 [config.py:1472] Using max model len 10240
INFO 07-25 09:36:50 [config.py:2285] Chunked prefill is enabled with max_num_batched_tokens=2048.
INFO 07-25 09:36:55 [__init__.py:244] Automatically detected platform cuda.
INFO 07-25 09:36:57 [core.py:526] Waiting for init message from front-end.
INFO 07-25 09:36:58 [core.py:69] Initializing a V1 LLM engine (v0.9.2) with config: model='/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', speculative_config=None, tokenizer='/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=10240, download_dir=None, load_format=auto, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=compressed-tensors, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
WARNING 07-25 09:36:58 [multiproc_worker_utils.py:307] Reducing Torch parallelism from 28 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
INFO 07-25 09:36:58 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 16777216, 10, 'psm_5f9a28ff'), local_subscribe_addr='ipc:///tmp/99467b4f-ebc3-406a-9294-239ba64464fc', remote_subscribe_addr=None, remote_addr_ipv6=False)
INFO 07-25 09:37:02 [__init__.py:244] Automatically detected platform cuda.
INFO 07-25 09:37:02 [__init__.py:244] Automatically detected platform cuda.
INFO 07-25 09:37:02 [__init__.py:244] Automatically detected platform cuda.
INFO 07-25 09:37:02 [__init__.py:244] Automatically detected platform cuda.
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:05 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_280f26ff'), local_subscribe_addr='ipc:///tmp/be1de22e-0330-4710-8b6c-71cf26485c90', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:05 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e8877851'), local_subscribe_addr='ipc:///tmp/4cedf191-9498-42f5-af06-701a14621951', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:05 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_7bef00b9'), local_subscribe_addr='ipc:///tmp/c290b828-dfb0-46c1-bf6e-d69f9e2ee1db', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:05 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_1d967ed8'), local_subscribe_addr='ipc:///tmp/c175603e-dd9a-4857-8aa0-596f69a9f12e', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:06 [__init__.py:1152] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:06 [__init__.py:1152] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:06 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:06 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:06 [__init__.py:1152] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:06 [__init__.py:1152] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:06 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:06 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=3 pid=6465)[0;0m WARNING 07-25 09:37:07 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=2 pid=6464)[0;0m WARNING 07-25 09:37:07 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=1 pid=6463)[0;0m WARNING 07-25 09:37:07 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=0 pid=6462)[0;0m WARNING 07-25 09:37:07 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:07 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_389b7df6'), local_subscribe_addr='ipc:///tmp/e50dfec3-1ac2-4496-a39a-5557cd6cccee', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:07 [parallel_state.py:1076] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3, EP rank 3
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:07 [parallel_state.py:1076] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1, EP rank 1
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:07 [parallel_state.py:1076] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2, EP rank 2
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:07 [parallel_state.py:1076] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
[1;36m(VllmWorker rank=3 pid=6465)[0;0m WARNING 07-25 09:37:07 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=2 pid=6464)[0;0m WARNING 07-25 09:37:07 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=1 pid=6463)[0;0m WARNING 07-25 09:37:07 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=0 pid=6462)[0;0m WARNING 07-25 09:37:07 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1770] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1770] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1770] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1770] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1775] Loading model from scratch...
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:07 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1775] Loading model from scratch...
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:07 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1775] Loading model from scratch...
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:07 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:07 [cuda.py:284] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:07 [cuda.py:284] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:07 [gpu_model_runner.py:1775] Loading model from scratch...
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:07 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:08 [cuda.py:284] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:08 [cuda.py:284] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:17 [default_loader.py:272] Loading weights took 9.09 seconds
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:17 [default_loader.py:272] Loading weights took 9.09 seconds
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:17 [default_loader.py:272] Loading weights took 9.45 seconds
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:17 [gpu_model_runner.py:1801] Model loading took 16.9619 GiB and 9.506413 seconds
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:17 [default_loader.py:272] Loading weights took 9.54 seconds
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:18 [gpu_model_runner.py:1801] Model loading took 16.9619 GiB and 9.523897 seconds
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:18 [gpu_model_runner.py:1801] Model loading took 16.9619 GiB and 9.862348 seconds
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:18 [gpu_model_runner.py:1801] Model loading took 16.9619 GiB and 9.959446 seconds
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:39 [backends.py:508] Using cache directory: /root/.cache/vllm/torch_compile_cache/84c3f599df/rank_2_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:39 [backends.py:519] Dynamo bytecode transform time: 20.15 s
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:39 [backends.py:508] Using cache directory: /root/.cache/vllm/torch_compile_cache/84c3f599df/rank_3_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:39 [backends.py:519] Dynamo bytecode transform time: 20.38 s
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:39 [backends.py:508] Using cache directory: /root/.cache/vllm/torch_compile_cache/84c3f599df/rank_1_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:39 [backends.py:519] Dynamo bytecode transform time: 20.57 s
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:39 [backends.py:508] Using cache directory: /root/.cache/vllm/torch_compile_cache/84c3f599df/rank_0_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:39 [backends.py:519] Dynamo bytecode transform time: 20.82 s
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:52 [backends.py:155] Directly load the compiled graph(s) for shape None from the cache, took 12.590 s
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:52 [backends.py:155] Directly load the compiled graph(s) for shape None from the cache, took 12.477 s
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:53 [backends.py:155] Directly load the compiled graph(s) for shape None from the cache, took 12.533 s
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:53 [backends.py:155] Directly load the compiled graph(s) for shape None from the cache, took 12.583 s
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:57 [monitor.py:34] torch.compile takes 20.38 s in total
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:57 [monitor.py:34] torch.compile takes 20.82 s in total
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:37:57 [monitor.py:34] torch.compile takes 20.15 s in total
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:57 [monitor.py:34] torch.compile takes 20.57 s in total
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:37:59 [gpu_worker.py:232] Available KV cache memory: 2.59 GiB
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:37:59 [gpu_worker.py:232] Available KV cache memory: 2.59 GiB
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:37:59 [gpu_worker.py:232] Available KV cache memory: 2.59 GiB
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:38:00 [gpu_worker.py:232] Available KV cache memory: 2.59 GiB
INFO 07-25 09:38:00 [kv_cache_utils.py:716] GPU KV cache size: 33,984 tokens
INFO 07-25 09:38:00 [kv_cache_utils.py:720] Maximum concurrency for 10,240 tokens per request: 3.32x
INFO 07-25 09:38:00 [kv_cache_utils.py:716] GPU KV cache size: 33,984 tokens
INFO 07-25 09:38:00 [kv_cache_utils.py:720] Maximum concurrency for 10,240 tokens per request: 3.32x
INFO 07-25 09:38:00 [kv_cache_utils.py:716] GPU KV cache size: 33,984 tokens
INFO 07-25 09:38:00 [kv_cache_utils.py:720] Maximum concurrency for 10,240 tokens per request: 3.32x
INFO 07-25 09:38:00 [kv_cache_utils.py:716] GPU KV cache size: 33,984 tokens
INFO 07-25 09:38:00 [kv_cache_utils.py:720] Maximum concurrency for 10,240 tokens per request: 3.32x
[1;36m(VllmWorker rank=1 pid=6463)[0;0m INFO 07-25 09:38:49 [gpu_model_runner.py:2326] Graph capturing finished in 49 secs, took 1.82 GiB
[1;36m(VllmWorker rank=3 pid=6465)[0;0m INFO 07-25 09:38:49 [gpu_model_runner.py:2326] Graph capturing finished in 49 secs, took 1.82 GiB
[1;36m(VllmWorker rank=0 pid=6462)[0;0m INFO 07-25 09:38:49 [gpu_model_runner.py:2326] Graph capturing finished in 49 secs, took 1.82 GiB
[1;36m(VllmWorker rank=2 pid=6464)[0;0m INFO 07-25 09:38:49 [gpu_model_runner.py:2326] Graph capturing finished in 49 secs, took 1.82 GiB
INFO 07-25 09:38:49 [core.py:172] init engine (profile, create kv cache, warmup model) took 90.82 seconds
INFO 07-25 09:38:50 [loggers.py:137] Engine 000: vllm cache_config_info with initialization after num_gpu_blocks is: 2124
WARNING 07-25 09:38:50 [config.py:1392] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
INFO 07-25 09:38:50 [serving_chat.py:125] Using default chat sampling params from model: {'temperature': 0.6, 'top_p': 0.95}
INFO 07-25 09:38:50 [serving_completion.py:72] Using default completion sampling params from model: {'temperature': 0.6, 'top_p': 0.95}
INFO 07-25 09:38:50 [api_server.py:1457] Starting vLLM API server 0 on http://0.0.0.0:8000
INFO 07-25 09:38:50 [launcher.py:29] Available routes are:
INFO 07-25 09:38:50 [launcher.py:37] Route: /openapi.json, Methods: GET, HEAD
INFO 07-25 09:38:50 [launcher.py:37] Route: /docs, Methods: GET, HEAD
INFO 07-25 09:38:50 [launcher.py:37] Route: /docs/oauth2-redirect, Methods: GET, HEAD
INFO 07-25 09:38:50 [launcher.py:37] Route: /redoc, Methods: GET, HEAD
INFO 07-25 09:38:50 [launcher.py:37] Route: /health, Methods: GET
INFO 07-25 09:38:50 [launcher.py:37] Route: /load, Methods: GET
INFO 07-25 09:38:50 [launcher.py:37] Route: /ping, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /ping, Methods: GET
INFO 07-25 09:38:50 [launcher.py:37] Route: /tokenize, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /detokenize, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/models, Methods: GET
INFO 07-25 09:38:50 [launcher.py:37] Route: /version, Methods: GET
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/chat/completions, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/completions, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/embeddings, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /pooling, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /classify, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /score, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/score, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/audio/transcriptions, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/audio/translations, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /rerank, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v1/rerank, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /v2/rerank, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /invocations, Methods: POST
INFO 07-25 09:38:50 [launcher.py:37] Route: /metrics, Methods: GET
......(Prompts)
INFO: 127.0.0.1:45180 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 07-25 09:40:08 [async_llm.py:270] Added request cmpl-54cba8629a6f459b84487918dfe3e635-0.
INFO 07-25 09:40:10 [loggers.py:118] Engine 000: Avg prompt throughput: 204.8 tokens/s, Avg generation throughput: 31.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 49.6%
INFO 07-25 09:40:20 [loggers.py:118] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 49.6%
INFO 07-25 09:40:30 [loggers.py:118] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 49.6%
INFO 07-25 09:40:40 [loggers.py:118] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 49.6%
INFO 07-25 09:40:50 [loggers.py:118] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 49.6%
INFO 07-25 09:41:00 [loggers.py:118] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 49.6%
Results:
(vllm092) root@test:/workspace$ python3 benchmark_serving.py --num-prompts 1 --model /root/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 --input-tokens 2048 --output-tokens 1024 --host 0.0.0.0 --port 8000
INFO 07-28 06:40:30 [__init__.py:235] Automatically detected platform cuda.
============ Serving Benchmark Result ============
Successful requests: 1
Benchmark duration (s): 32.87
Total input tokens: 2048
Total generated tokens: 1024
Request throughput (req/s): 0.03
Output token throughput (tok/s): 31.15
Total Token throughput (tok/s): 93.46
---------------Time to First Token----------------
Min TTFT (ms): 57.83
Mean TTFT (ms): 57.83
Median TTFT (ms): 57.83
P99 TTFT (ms): 57.83
After Update
Server Log:
(vllm0100) root@test:/workspace$ vllm serve /root/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 --tensor-parallel-size 4 --max-num-seqs 1 --max-model-len 10240 --trust-remote-code --gpu-memory-utilization 0.9
INFO 07-28 06:35:51 [__init__.py:235] Automatically detected platform cuda.
INFO 07-28 06:35:54 [api_server.py:1755] vLLM API server version 0.10.0
INFO 07-28 06:35:54 [cli_args.py:261] non-default args: {'model_tag': '/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', 'model': '/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', 'trust_remote_code': True, 'max_model_len': 10240, 'tensor_parallel_size': 4, 'gpu_memory_utilization': 0.95, 'max_num_seqs': 1}
INFO 07-28 06:36:00 [config.py:1604] Using max model len 10240
INFO 07-28 06:36:01 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=2048.
INFO 07-28 06:36:05 [__init__.py:235] Automatically detected platform cuda.
INFO 07-28 06:36:08 [core.py:572] Waiting for init message from front-end.
INFO 07-28 06:36:08 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', speculative_config=None, tokenizer='/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=10240, download_dir=None, load_format=auto, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=compressed-tensors, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=/root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":4,"local_cache_dir":null}
WARNING 07-28 06:36:08 [multiproc_worker_utils.py:307] Reducing Torch parallelism from 28 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
INFO 07-28 06:36:08 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 16777216, 10, 'psm_71d626d5'), local_subscribe_addr='ipc:///tmp/e0ce8fe2-8b3e-4d47-9474-874cd49dde9b', remote_subscribe_addr=None, remote_addr_ipv6=False)
INFO 07-28 06:36:12 [__init__.py:235] Automatically detected platform cuda.
INFO 07-28 06:36:12 [__init__.py:235] Automatically detected platform cuda.
INFO 07-28 06:36:12 [__init__.py:235] Automatically detected platform cuda.
INFO 07-28 06:36:12 [__init__.py:235] Automatically detected platform cuda.
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:16 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b6df00f5'), local_subscribe_addr='ipc:///tmp/014ede7a-1b15-4175-91f4-dd86701cf36f', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:16 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3bee758b'), local_subscribe_addr='ipc:///tmp/8f4e0eb8-bb11-4479-b43d-1a41ff125519', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:16 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_4ae32e16'), local_subscribe_addr='ipc:///tmp/a18a3aee-da26-4e38-b7b1-6b0b36ffa96d', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:16 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a92bf551'), local_subscribe_addr='ipc:///tmp/e84bd9fb-a253-4abc-ac0f-69429861d4c3', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:17 [__init__.py:1375] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:17 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:17 [__init__.py:1375] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:17 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:17 [__init__.py:1375] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:17 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:17 [__init__.py:1375] Found nccl from library libnccl.so.2
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:17 [pynccl.py:70] vLLM is using nccl==2.26.2
[1;36m(VllmWorker rank=0 pid=14480)[0;0m WARNING 07-28 06:36:18 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=2 pid=14482)[0;0m WARNING 07-28 06:36:18 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=3 pid=14483)[0;0m WARNING 07-28 06:36:18 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=1 pid=14481)[0;0m WARNING 07-28 06:36:18 [custom_all_reduce.py:137] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:18 [shm_broadcast.py:289] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_2fb20f5d'), local_subscribe_addr='ipc:///tmp/2e2440eb-19c0-4551-b19c-16f0f31326b3', remote_subscribe_addr=None, remote_addr_ipv6=False)
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:18 [parallel_state.py:1102] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1, EP rank 1
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:18 [parallel_state.py:1102] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:18 [parallel_state.py:1102] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2, EP rank 2
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:18 [parallel_state.py:1102] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3, EP rank 3
[1;36m(VllmWorker rank=1 pid=14481)[0;0m WARNING 07-28 06:36:18 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=0 pid=14480)[0;0m WARNING 07-28 06:36:18 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=3 pid=14483)[0;0m WARNING 07-28 06:36:18 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:18 [gpu_model_runner.py:1843] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=2 pid=14482)[0;0m WARNING 07-28 06:36:18 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:18 [gpu_model_runner.py:1843] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:18 [gpu_model_runner.py:1843] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:18 [gpu_model_runner.py:1843] Starting to load model /root/fshare/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8...
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:18 [gpu_model_runner.py:1875] Loading model from scratch...
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:18 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:19 [gpu_model_runner.py:1875] Loading model from scratch...
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:19 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:19 [gpu_model_runner.py:1875] Loading model from scratch...
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:19 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:19 [gpu_model_runner.py:1875] Loading model from scratch...
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:19 [compressed_tensors_w8a8_int8.py:52] Using CutlassScaledMMLinearKernel for CompressedTensorsW8A8Int8
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:19 [cuda.py:290] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:19 [cuda.py:290] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:19 [cuda.py:290] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:19 [cuda.py:290] Using Flash Attention backend on V1 engine.
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:57 [default_loader.py:262] Loading weights took 37.80 seconds
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:57 [default_loader.py:262] Loading weights took 37.94 seconds
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:57 [default_loader.py:262] Loading weights took 37.99 seconds
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:57 [default_loader.py:262] Loading weights took 38.13 seconds
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:36:57 [gpu_model_runner.py:1892] Model loading took 16.9619 GiB and 38.408816 seconds
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:36:58 [gpu_model_runner.py:1892] Model loading took 16.9619 GiB and 38.360333 seconds
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:36:58 [gpu_model_runner.py:1892] Model loading took 16.9619 GiB and 38.230198 seconds
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:36:58 [gpu_model_runner.py:1892] Model loading took 16.9619 GiB and 38.559743 seconds
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:37:19 [backends.py:530] Using cache directory: /root/.cache/vllm/torch_compile_cache/310a2dced5/rank_1_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:37:19 [backends.py:530] Using cache directory: /root/.cache/vllm/torch_compile_cache/310a2dced5/rank_3_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:37:19 [backends.py:541] Dynamo bytecode transform time: 20.41 s
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:37:19 [backends.py:541] Dynamo bytecode transform time: 20.56 s
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:37:19 [backends.py:530] Using cache directory: /root/.cache/vllm/torch_compile_cache/310a2dced5/rank_0_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:37:19 [backends.py:541] Dynamo bytecode transform time: 20.74 s
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:37:19 [backends.py:530] Using cache directory: /root/.cache/vllm/torch_compile_cache/310a2dced5/rank_2_0/backbone for vLLM's torch.compile
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:37:19 [backends.py:541] Dynamo bytecode transform time: 20.78 s
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:37:24 [backends.py:194] Cache the graph for dynamic shape for later use
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:37:24 [backends.py:194] Cache the graph for dynamic shape for later use
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:37:25 [backends.py:194] Cache the graph for dynamic shape for later use
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:37:25 [backends.py:194] Cache the graph for dynamic shape for later use
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:38:34 [backends.py:215] Compiling a graph for dynamic shape takes 74.77 s
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:38:35 [backends.py:215] Compiling a graph for dynamic shape takes 75.42 s
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:38:36 [backends.py:215] Compiling a graph for dynamic shape takes 76.43 s
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:38:37 [backends.py:215] Compiling a graph for dynamic shape takes 77.00 s
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:38:54 [monitor.py:34] torch.compile takes 95.18 s in total
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:38:54 [monitor.py:34] torch.compile takes 97.74 s in total
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:38:54 [monitor.py:34] torch.compile takes 97.21 s in total
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:38:54 [monitor.py:34] torch.compile takes 95.98 s in total
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:38:56 [gpu_worker.py:255] Available KV cache memory: 4.95 GiB
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:38:56 [gpu_worker.py:255] Available KV cache memory: 4.95 GiB
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:38:56 [gpu_worker.py:255] Available KV cache memory: 4.95 GiB
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:38:57 [gpu_worker.py:255] Available KV cache memory: 4.95 GiB
INFO 07-28 06:38:57 [kv_cache_utils.py:833] GPU KV cache size: 64,848 tokens
INFO 07-28 06:38:57 [kv_cache_utils.py:837] Maximum concurrency for 10,240 tokens per request: 6.33x
INFO 07-28 06:38:57 [kv_cache_utils.py:833] GPU KV cache size: 64,848 tokens
INFO 07-28 06:38:57 [kv_cache_utils.py:837] Maximum concurrency for 10,240 tokens per request: 6.33x
INFO 07-28 06:38:57 [kv_cache_utils.py:833] GPU KV cache size: 64,848 tokens
INFO 07-28 06:38:57 [kv_cache_utils.py:837] Maximum concurrency for 10,240 tokens per request: 6.33x
INFO 07-28 06:38:57 [kv_cache_utils.py:833] GPU KV cache size: 64,848 tokens
INFO 07-28 06:38:57 [kv_cache_utils.py:837] Maximum concurrency for 10,240 tokens per request: 6.33x
[1;36m(VllmWorker rank=1 pid=14481)[0;0m INFO 07-28 06:38:58 [gpu_model_runner.py:2485] Graph capturing finished in 1 secs, took 0.08 GiB
[1;36m(VllmWorker rank=2 pid=14482)[0;0m INFO 07-28 06:38:58 [gpu_model_runner.py:2485] Graph capturing finished in 1 secs, took 0.08 GiB
[1;36m(VllmWorker rank=0 pid=14480)[0;0m INFO 07-28 06:38:58 [gpu_model_runner.py:2485] Graph capturing finished in 1 secs, took 0.08 GiB
[1;36m(VllmWorker rank=3 pid=14483)[0;0m INFO 07-28 06:38:58 [gpu_model_runner.py:2485] Graph capturing finished in 1 secs, took 0.08 GiB
INFO 07-28 06:38:58 [core.py:193] init engine (profile, create kv cache, warmup model) took 120.44 seconds
INFO 07-28 06:38:59 [loggers.py:141] Engine 000: vllm cache_config_info with initialization after num_gpu_blocks is: 4053
WARNING 07-28 06:38:59 [config.py:1528] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
INFO 07-28 06:38:59 [serving_responses.py:89] Using default chat sampling params from model: {'temperature': 0.6, 'top_p': 0.95}
INFO 07-28 06:38:59 [serving_chat.py:122] Using default chat sampling params from model: {'temperature': 0.6, 'top_p': 0.95}
INFO 07-28 06:38:59 [serving_completion.py:77] Using default completion sampling params from model: {'temperature': 0.6, 'top_p': 0.95}
INFO 07-28 06:38:59 [api_server.py:1818] Starting vLLM API server 0 on http://0.0.0.0:8000
INFO 07-28 06:38:59 [launcher.py:29] Available routes are:
INFO 07-28 06:38:59 [launcher.py:37] Route: /openapi.json, Methods: GET, HEAD
INFO 07-28 06:38:59 [launcher.py:37] Route: /docs, Methods: GET, HEAD
INFO 07-28 06:38:59 [launcher.py:37] Route: /docs/oauth2-redirect, Methods: GET, HEAD
INFO 07-28 06:38:59 [launcher.py:37] Route: /redoc, Methods: GET, HEAD
INFO 07-28 06:38:59 [launcher.py:37] Route: /health, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /load, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /ping, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /ping, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /tokenize, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /detokenize, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/models, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /version, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/responses, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/responses/{response_id}, Methods: GET
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/responses/{response_id}/cancel, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/chat/completions, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/completions, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/embeddings, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /pooling, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /classify, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /score, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/score, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/audio/transcriptions, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/audio/translations, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /rerank, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v1/rerank, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /v2/rerank, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /scale_elastic_ep, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /is_scaling_elastic_ep, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /invocations, Methods: POST
INFO 07-28 06:38:59 [launcher.py:37] Route: /metrics, Methods: GET
……(Prompts)
INFO: 127.0.0.1:38564 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 07-28 06:41:04 [async_llm.py:269] Added request cmpl-3171a08b6ba24918876820150e297272-0.
INFO 07-28 06:41:10 [loggers.py:122] Engine 000: Avg prompt throughput: 204.8 tokens/s, Avg generation throughput: 30.9 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 49.6%
INFO 07-28 06:41:20 [loggers.py:122] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 49.6%
INFO 07-28 06:41:30 [loggers.py:122] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 31.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 49.6%
INFO 07-28 06:41:40 [loggers.py:122] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 23.7 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 49.6%
INFO 07-28 06:41:50 [loggers.py:122] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 49.6%
Results:
(vllm0100) root@test:/workspace$ python3 benchmark_serving.py --num-prompts 1 --model /root/models/neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 --input-tokens 2048 --output-tokens 1024 --host 0.0.0.0 --port 8000
INFO 07-28 06:40:30 [__init__.py:235] Automatically detected platform cuda.
============ Serving Benchmark Result ============
Successful requests: 1
Benchmark duration (s): 32.78
Total input tokens: 2048
Total generated tokens: 1024
Request throughput (req/s): 0.03
Output token throughput (tok/s): 31.24
Total Token throughput (tok/s): 93.73
---------------Time to First Token----------------
Min TTFT (ms): 113.77
Mean TTFT (ms): 113.77
Median TTFT (ms): 113.77
P99 TTFT (ms): 113.77
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.