[Bug]: V1 pre-compiled graph loading much slower than V0 #20342
Description
Your current environment
The output of python collect_env.py
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.5 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version : 19.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.4.1 25184 c87081df219c42dc27c5b6d86c0525bc7d01f727)
CMake version : version 3.31.6
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.7.0+gitf717b2a
Is debug build : False
CUDA used to build PyTorch : N/A
ROCM used to build PyTorch : 6.4.43483-a187df25c
==============================
Python Environment
==============================
Python version : 3.12.11 (main, Jun 4 2025, 08:56:18) [GCC 11.4.0] (64-bit runtime)
Python platform : Linux-6.8.0-52-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : Could not collect
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration : AMD Instinct MI250X/MI250 (gfx90a:sramecc+:xnack-)
Nvidia driver version : Could not collect
cuDNN version : Could not collect
HIP runtime version : 6.4.43483
MIOpen runtime version : 3.4.0
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7713 64-Core Processor
CPU family: 25
Model: 1
Thread(s) per core: 1
Core(s) per socket: 64
Socket(s): 2
Stepping: 1
Frequency boost: enabled
CPU max MHz: 3720.7029
CPU min MHz: 1500.0000
BogoMIPS: 3992.52
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin brs arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm debug_swap
Virtualization: AMD-V
L1d cache: 4 MiB (128 instances)
L1i cache: 4 MiB (128 instances)
L2 cache: 64 MiB (128 instances)
L3 cache: 512 MiB (16 instances)
NUMA node(s): 8
NUMA node0 CPU(s): 0-15
NUMA node1 CPU(s): 16-31
NUMA node2 CPU(s): 32-47
NUMA node3 CPU(s): 48-63
NUMA node4 CPU(s): 64-79
NUMA node5 CPU(s): 80-95
NUMA node6 CPU(s): 96-111
NUMA node7 CPU(s): 112-127
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; Safe RET
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; Retpolines; IBPB conditional; IBRS_FW; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] pyzmq==27.0.0
[pip3] torch==2.7.0+gitf717b2a
[pip3] torchvision==0.21.0+7af6987
[pip3] transformers==4.53.0
[pip3] triton==3.2.0+gite5be006a
[conda] Could not collect
==============================
vLLM Info
==============================
ROCM Version : 6.4.43483-a187df25c
Neuron SDK Version : N/A
vLLM Version : 0.1.dev7354+g16f89f4.d20250628 (git sha: 16f89f4, date: 20250628)
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7
GPU0 0 15 15 30 30 30 15 30
GPU1 15 0 30 15 30 15 30 45
GPU2 15 30 0 15 15 30 30 30
GPU3 30 15 15 0 30 45 30 15
GPU4 30 30 15 30 0 15 15 30
GPU5 30 15 30 45 15 0 30 15
GPU6 15 30 30 30 15 30 0 15
GPU7 30 45 30 15 30 15 15 0
================================= Hops between two GPUs ==================================
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7
GPU0 0 1 1 1 1 1 1 1
GPU1 1 0 1 1 1 1 1 1
GPU2 1 1 0 1 1 1 1 1
GPU3 1 1 1 0 1 1 1 1
GPU4 1 1 1 1 0 1 1 1
GPU5 1 1 1 1 1 0 1 1
GPU6 1 1 1 1 1 1 0 1
GPU7 1 1 1 1 1 1 1 0
=============================== Link Type between two GPUs ===============================
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7
GPU0 0 XGMI XGMI XGMI XGMI XGMI XGMI XGMI
GPU1 XGMI 0 XGMI XGMI XGMI XGMI XGMI XGMI
GPU2 XGMI XGMI 0 XGMI XGMI XGMI XGMI XGMI
GPU3 XGMI XGMI XGMI 0 XGMI XGMI XGMI XGMI
GPU4 XGMI XGMI XGMI XGMI 0 XGMI XGMI XGMI
GPU5 XGMI XGMI XGMI XGMI XGMI 0 XGMI XGMI
GPU6 XGMI XGMI XGMI XGMI XGMI XGMI 0 XGMI
GPU7 XGMI XGMI XGMI XGMI XGMI XGMI XGMI 0
======================================= Numa Nodes =======================================
GPU[0] : (Topology) Numa Node: 3
GPU[0] : (Topology) Numa Affinity: 3
GPU[1] : (Topology) Numa Node: 3
GPU[1] : (Topology) Numa Affinity: 3
GPU[2] : (Topology) Numa Node: 2
GPU[2] : (Topology) Numa Affinity: 2
GPU[3] : (Topology) Numa Node: 2
GPU[3] : (Topology) Numa Affinity: 2
GPU[4] : (Topology) Numa Node: 7
GPU[4] : (Topology) Numa Affinity: 7
GPU[5] : (Topology) Numa Node: 7
GPU[5] : (Topology) Numa Affinity: 7
GPU[6] : (Topology) Numa Node: 6
GPU[6] : (Topology) Numa Affinity: 6
GPU[7] : (Topology) Numa Node: 6
GPU[7] : (Topology) Numa Affinity: 6
================================== End of ROCm SMI Log ===================================
==============================
Environment Variables
==============================
PYTORCH_ROCM_ARCH=gfx90a;gfx942
LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
PYTORCH_TUNABLEOP_FILENAME=/app/afo_tune_device_%d_full.csv
PYTORCH_TUNABLEOP_TUNING=0
PYTORCH_TUNABLEOP_ENABLED=1
VLLM_TARGET_DEVICE=rocm
VERBOSE=1
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY
🐛 Describe the bug
When using V0, loading compilation artifacts typically takes single-digit seconds. However, with V1 it takes around 30x longer, e.g. 70s for the graph for a single shape for a larger model. This can be tested by observing the compilation graph load times when running the same command a 2nd time (allowing vLLM to load the pre-compiled graphs from the cache rather than compiling from scratch):
export VLLM_USE_V1=1
MODEL_NAME="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
echo "Running with compilation config "
python3 -m vllm.entrypoints.openai.api_server \
--port 8080 \
--model $MODEL_NAME \
--served-model-name $MODEL_NAME \
--gpu-memory-utilization 0.95 \
--disable-custom-all-reduce \
--tensor-parallel-size 1 \
--enable-chunked-prefill \
--disable-log-requests \
--enable-reasoning \
--compilation-config '{"compile_sizes": [1], "level": 3, "cudagraph_capture_sizes": [256, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 144, 136, 16, 152, 24, 128, 160, 32, 168, 40, 176, 48, 184, 56, 192, 64, 200, 72, 208, 80, 216, 88, 120, 224, 96, 232, 104, 240, 112, 248, 248]}' \
--reasoning-parser deepseek_r1If one sets export VLLM_USE_V1=0 and re-runs the above twice, it will load the compilation artifacts much faster. From turning debugging logs on, one can see that V1 loads 28x times more graphs compared to V0, which explains why it's so slow to initialize.
V0:
DEBUG 07-01 21:11:24 [backends.py:123] Directly load the 0-th graph for shape 1 from inductor via handle ('fn7nwql5utlkjsob3xnghodbg4wpmfszfp4t5rbntqa4tjau3wp5', '/root/.cache/vllm/torch_compile_cache/469366c8cf/rank_0_0/inductor_cache/x3/cx37k6oulv3pi5qbdgh4xvj5h2emyzprlhb36dgljhlhfckglrsp.py')
INFO 07-01 21:11:24 [backends.py:167] Directly load the compiled graph(s) for shape 1 from the cache, took 4.428 sV1:
DEBUG 07-01 21:20:33 [backends.py:151] Runtime shape: 1
DEBUG 07-01 21:20:33 [backends.py:123] Directly load the 28-th graph for shape 1 from inductor via handle ('fhu67rrnvc35buqntqu3cckyijpmigkuhnw2f4czeby4womoyu74', '/root/.cache/vllm/torch_compile_cache/250015fb60/rank_0_0/inductor_cache/cd/ccdqxkqebrqnhlnyzlmblhaolao245amjndrrhp2akftnnmo6nki.py')
INFO 07-01 21:20:33 [backends.py:155] Successfully loaded graph 28Compared to V0, compilation gives similar speedups compared to vanilla V1 for a given batch size, which suggests to me that the extra loaded graphs are not necessary for performance, although this has not been tested.
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.