search

vllm-project / vllm

A high-throughput and memory-efficient inference and serving engine for LLMs
https://docs.vllm.ai
Apache License 2.0
25.56k stars 3.7k forks source link

[Bug]: OutOfMemoryError when loading a small model with a huge context length #5847

Open alugowski opened 2 months ago

alugowski commented 2 months ago

Your current environment

Collecting environment information...
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.29.5
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-107-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
GPU 2: NVIDIA H100 80GB HBM3
GPU 3: NVIDIA H100 80GB HBM3
GPU 4: NVIDIA H100 80GB HBM3
GPU 5: NVIDIA H100 80GB HBM3
GPU 6: NVIDIA H100 80GB HBM3
GPU 7: NVIDIA H100 80GB HBM3

Nvidia driver version: 555.42.02
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Address sizes:                      46 bits physical, 57 bits virtual
Byte Order:                         Little Endian
CPU(s):                             224
On-line CPU(s) list:                0-111
Off-line CPU(s) list:               112-223
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) Platinum 8480+
CPU family:                         6
Model:                              143
Thread(s) per core:                 1
Core(s) per socket:                 56
Socket(s):                          2
Stepping:                           8
Frequency boost:                    enabled
CPU max MHz:                        2001.0000
CPU min MHz:                        0.0000
BogoMIPS:                           4000.00
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 intel_ppin 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 split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          5.3 MiB (112 instances)
L1i cache:                          3.5 MiB (112 instances)
L2 cache:                           224 MiB (112 instances)
L3 cache:                           210 MiB (2 instances)
NUMA node(s):                       2
NUMA node0 CPU(s):                  0-55
NUMA node1 CPU(s):                  56-111
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: 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; BHI BHI_DIS_S
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] mypy==1.9.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] sentence-transformers==3.0.1
[pip3] torch==2.3.0
[pip3] torchvision==0.18.1
[pip3] transformers==4.41.2
[pip3] triton==2.3.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.0.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    NIC0    NIC1    NIC2    NIC3    NIC4    NIC5    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X  NV18    NV18    NV18    NV18    NV18    NV18    NV18    PIX NODE    NODE    NODE    SYS SYS 0-55    0       N/A
GPU1    NV18     X  NV18    NV18    NV18    NV18    NV18    NV18    NODE    NODE    NODE    NODE    SYS SYS 0-55    0       N/A
GPU2    NV18    NV18     X  NV18    NV18    NV18    NV18    NV18    NODE    PIX NODE    NODE    SYS SYS 0-55    0       N/A
GPU3    NV18    NV18    NV18     X  NV18    NV18    NV18    NV18    NODE    NODE    NODE    NODE    SYS SYS 0-55    0       N/A
GPU4    NV18    NV18    NV18    NV18     X  NV18    NV18    NV18    SYS SYS SYS SYS PIX NODE    56-111  1       N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X  NV18    NV18    SYS SYS SYS SYS NODE    NODE    56-111  1       N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X  NV18    SYS SYS SYS SYS NODE    PIX 56-111  1       N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X  SYS SYS SYS SYS NODE    NODE    56-111  1       N/A
NIC0    PIX NODE    NODE    NODE    SYS SYS SYS SYS  X  NODE    NODE    NODE    SYS SYS
NIC1    NODE    NODE    PIX NODE    SYS SYS SYS SYS NODE     X  NODE    NODE    SYS SYS
NIC2    NODE    NODE    NODE    NODE    SYS SYS SYS SYS NODE    NODE     X  PIX SYS SYS
NIC3    NODE    NODE    NODE    NODE    SYS SYS SYS SYS NODE    NODE    PIX  X  SYS SYS
NIC4    SYS SYS SYS SYS PIX NODE    NODE    NODE    SYS SYS SYS SYS  X  NODE
NIC5    SYS SYS SYS SYS NODE    NODE    PIX NODE    SYS SYS SYS SYS NODE     X

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

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5

🐛 Describe the bug

Loading a model with a huge context length (100k+) causes an out of memory error when allocating space for the KV cache. While it is correct that technically the model does not fit, such models should still start with reduced context size.

An easily reproducible example is gradientai/Llama-3-8B-Instruct-Gradient-1048k which is 8B with a 1M context. It will fail to start on any single GPU:

python -m vllm.entrypoints.openai.api_server --model gradientai/Llama-3-8B-Instruct-Gradient-1048k

Results in:

INFO 06-26 07:48:15 model_runner.py:160] Loading model weights took 15.2075 GB
[rank0]: Traceback (most recent call last):
[rank0]:   File "/usr/lib/python3.10/runpy.py", line 196, in _run_module_as_main
[rank0]:     return _run_code(code, main_globals, None,
[rank0]:   File "/usr/lib/python3.10/runpy.py", line 86, in _run_code
[rank0]:     exec(code, run_globals)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/entrypoints/openai/api_server.py", line 196, in <module>
[rank0]:     engine = AsyncLLMEngine.from_engine_args(
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/engine/async_llm_engine.py", line 415, in from_engine_args
[rank0]:     engine = cls(
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/engine/async_llm_engine.py", line 355, in __init__
[rank0]:     self.engine = self._init_engine(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/engine/async_llm_engine.py", line 490, in _init_engine
[rank0]:     return engine_class(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/engine/llm_engine.py", line 243, in __init__
[rank0]:     self._initialize_kv_caches()
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/engine/llm_engine.py", line 326, in _initialize_kv_caches
[rank0]:     self.model_executor.determine_num_available_blocks())
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/executor/gpu_executor.py", line 75, in determine_num_available_blocks
[rank0]:     return self.driver_worker.determine_num_available_blocks()
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
[rank0]:     return func(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/worker/worker.py", line 163, in determine_num_available_blocks
[rank0]:     self.model_runner.profile_run()
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
[rank0]:     return func(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/worker/model_runner.py", line 844, in profile_run
[rank0]:     self.execute_model(seqs, kv_caches)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
[rank0]:     return func(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/worker/model_runner.py", line 749, in execute_model
[rank0]:     hidden_states = model_executable(
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/models/llama.py", line 371, in forward
[rank0]:     hidden_states = self.model(input_ids, positions, kv_caches,
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/models/llama.py", line 288, in forward
[rank0]:     hidden_states, residual = layer(
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/models/llama.py", line 237, in forward
[rank0]:     hidden_states = self.mlp(hidden_states)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/models/llama.py", line 79, in forward
[rank0]:     gate_up, _ = self.gate_up_proj(x)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/home/adam/venvs/vllm/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/layers/linear.py", line 298, in forward
[rank0]:     output_parallel = self.quant_method.apply(self, input_, bias)
[rank0]:   File "/home/adam/work/vllm-parasail/vllm/model_executor/layers/linear.py", line 111, in apply
[rank0]:     return F.linear(x, weight, bias)
[rank0]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 56.00 GiB. GPU
alugowski commented 2 months ago

This error is in the profile_run used to determine memory usage. If I patch the code to ignore the error, and a few lines below patch the model length to make raise_if_cache_size_invalid happy, the model starts. Sure it won't reach the full 1M context, but it will work with 200k on an 80GB GPU.

This will be more pressing for users of small GPUs as the popular models increase their context lengths beyond 8k.

I'm happy to submit my monkey patches if there isn't already a plan to support large context models.

DarkLight1337 commented 2 months ago

You can manually set --max-model-len to reduce the context length.

Not sure whether it's a good idea to automatically limit the context length based on available memory. @simon-mo any thoughts?

alugowski commented 2 months ago

You can manually set --max-model-len to reduce the context length.

Not sure whether it's a good idea to automatically limit the context length based on available memory. @simon-mo any thoughts?

Agreed that a purely automatic setting may give folks the wrong impression that they can use the full context of the model even if their hardware won't allow it. One alternative is --max-model-len max that would start the model no matter what and report the actual max context in the logs.

Right now someone must start vllm, see the crash, parse out the max context size from the log, and set that with --max-model-len. But that's only if the profile_run() doesn't OOM with the exception in the OP, in that case the user must guess at the max model len (the log message with the actual max is printed later, and depends on profile_run() succeeding).