[Usage]: vllm can't run qwen 32B inference

[kunger97]

Your current environment

PyTorch version: 2.2.0a0+git8964477
Is debug build: False
CUDA used to build PyTorch: None
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 3.30.2
Libc version: glibc-2.35

Python version: 3.10.12 (main, Jul 29 2024, 16:56:48) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-116-generic-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
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:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               10
On-line CPU(s) list:                  0-9
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8368 CPU @ 2.40GHz
CPU family:                           6
Model:                                106
Thread(s) per core:                   1
Core(s) per socket:                   10
Socket(s):                            1
Stepping:                             6
BogoMIPS:                             4799.99
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology cpuid tsc_known_freq pni pclmulqdq vmx ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves wbnoinvd arat avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid fsrm md_clear arch_capabilities
Virtualization:                       VT-x
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            320 KiB (10 instances)
L1i cache:                            320 KiB (10 instances)
L2 cache:                             40 MiB (10 instances)
L3 cache:                             16 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-9
Vulnerability Gather data sampling:   Unknown: Dependent on hypervisor status
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Reg file data sampling: 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 / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] habana-torch-dataloader==1.15.0.479
[pip3] habana-torch-plugin==1.15.0.479
[pip3] numpy==1.23.5
[pip3] pytorch-lightning==2.4.0
[pip3] torch==2.2.0a0+git8964477
[pip3] torch_tb_profiler==0.4.0
[pip3] torchaudio==2.2.0+08901ad
[pip3] torchdata==0.7.1+5e6f7b7
[pip3] torchmetrics==1.4.1
[pip3] torchtext==0.17.0+400da5c
[pip3] torchvision==0.17.0+b2383d4
[pip3] transformers==4.44.0
[pip3] triton==3.0.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.3.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
Could not collect

How would you like to use vllm

I'm try to run llm serve use gaudi2 on intel devcloud, i have install vllm-fork and i'm useing command below and it seams shows hpu oom?

PT_HPU_LAZY_MODE=1 vllm serve Qwen/Qwen1.5-32B-Chat --dtype bfloat16 --block-size 128 --device hpu

I also try qwen 13B, it works normally.
In addition, when I use optimal-habana to perform inference, it can generate text normally.

============================= HABANA PT BRIDGE CONFIGURATION =========================== 
 PT_HPU_LAZY_MODE = 1
 PT_RECIPE_CACHE_PATH = 
 PT_CACHE_FOLDER_DELETE = 0
 PT_HPU_RECIPE_CACHE_CONFIG = 
 PT_HPU_MAX_COMPOUND_OP_SIZE = 9223372036854775807
 PT_HPU_LAZY_ACC_PAR_MODE = 1
 PT_HPU_ENABLE_REFINE_DYNAMIC_SHAPES = 0
---------------------------: System Configuration :---------------------------
Num CPU Cores : 10
CPU RAM       : 100936596 KB
------------------------------------------------------------------------------
INFO 08-17 07:58:41 selector.py:85] Using HabanaAttention backend.
INFO 08-17 07:58:41 loader.py:284] Loading weights on hpu ...
Loading safetensors checkpoint shards:   0% Completed | 0/14 [00:00<?, ?it/s]
Loading safetensors checkpoint shards:   7% Completed | 1/14 [00:11<02:33, 11.83s/it]
Loading safetensors checkpoint shards:  14% Completed | 2/14 [00:23<02:21, 11.81s/it]
Loading safetensors checkpoint shards:  21% Completed | 3/14 [00:35<02:09, 11.79s/it]
Loading safetensors checkpoint shards:  29% Completed | 4/14 [00:47<01:58, 11.81s/it]
Loading safetensors checkpoint shards:  36% Completed | 5/14 [00:53<01:27,  9.71s/it]
Loading safetensors checkpoint shards:  43% Completed | 6/14 [01:04<01:21, 10.13s/it]
Loading safetensors checkpoint shards:  50% Completed | 7/14 [01:15<01:13, 10.44s/it]
Loading safetensors checkpoint shards:  57% Completed | 8/14 [01:26<01:04, 10.69s/it]
Loading safetensors checkpoint shards:  64% Completed | 9/14 [01:37<00:54, 10.93s/it]
Loading safetensors checkpoint shards:  71% Completed | 10/14 [01:49<00:44, 11.12s/it]
Loading safetensors checkpoint shards:  79% Completed | 11/14 [02:01<00:33, 11.30s/it]
Loading safetensors checkpoint shards:  86% Completed | 12/14 [02:13<00:22, 11.49s/it]
Loading safetensors checkpoint shards:  93% Completed | 13/14 [02:25<00:11, 11.72s/it]
Loading safetensors checkpoint shards: 100% Completed | 14/14 [02:38<00:00, 12.06s/it]
Loading safetensors checkpoint shards: 100% Completed | 14/14 [02:38<00:00, 11.30s/it]

INFO 08-17 08:01:20 habana_model_runner.py:433] Pre-loading model weights on hpu:0 took 60.59 GiB of device memory (60.59 GiB/94.62 GiB used) and 1.67 GiB of host memory (11.17 GiB/96.26 GiB used)
Traceback (most recent call last):
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/GaudiVllm/bin/vllm", line 33, in <module>
    sys.exit(load_entry_point('vllm', 'console_scripts', 'vllm')())
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/scripts.py", line 149, in main
    args.dispatch_function(args)
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/scripts.py", line 29, in serve
    asyncio.run(run_server(args))
  File "/usr/lib/python3.10/asyncio/runners.py", line 44, in run
    return loop.run_until_complete(main)
  File "/usr/lib/python3.10/asyncio/base_events.py", line 649, in run_until_complete
    return future.result()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/entrypoints/openai/api_server.py", line 289, in run_server
    app = await init_app(args, llm_engine)
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/entrypoints/openai/api_server.py", line 229, in init_app
    if llm_engine is not None else AsyncLLMEngine.from_engine_args(
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/engine/async_llm_engine.py", line 479, in from_engine_args
    engine = cls(
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/engine/async_llm_engine.py", line 380, in __init__
    self.engine = self._init_engine(*args, **kwargs)
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/engine/async_llm_engine.py", line 560, in _init_engine
    return engine_class(*args, **kwargs)
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/engine/llm_engine.py", line 252, in __init__
    self.model_executor = executor_class(
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/executor/executor_base.py", line 47, in __init__
    self._init_executor()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/executor/habana_executor.py", line 27, in _init_executor
    self._init_worker()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/executor/habana_executor.py", line 71, in _init_worker
    self.driver_worker.load_model()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/worker/habana_worker.py", line 121, in load_model
    self.model_runner.load_model()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/vllm-fork/vllm/worker/habana_model_runner.py", line 453, in load_model
    torch.hpu.synchronize()
  File "/home/u76dd8763cba39e0b015b7769dfb6510/GaudiWorkspace/GaudiVllm/lib/python3.10/site-packages/habana_frameworks/torch/hpu/__init__.py", line 138, in synchronize
    return _hpu_C.synchronize_device()
RuntimeError: [Rank:0] FATAL ERROR :: MODULE:PT_DEVMEM Allocation failed for size::560988160 (535)MB
inc shutdown
inc shutdown
inc shutdown

[kzawora-intel]

@afierka-intel can you check this out? I remember you've experienced a similar bug in weight loading phase of large models (llama405b or mixtral8x7b) on HPU. Should be simple to check if it's caused by the same bug.

[LeoZhao-Intel]

refer this change in #55 could fix bug, need same change in qwen at least.

like this:

diff --git a/vllm/model_executor/models/qwen2.py b/vllm/model_executor/models/qwen2.py
index f38be0e9..c42b67d4 100644
--- a/vllm/model_executor/models/qwen2.py
+++ b/vllm/model_executor/models/qwen2.py
@@ -371,3 +371,6 @@ class Qwen2ForCausalLM(nn.Module):
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
+

•        if is_hpu():
  

•            torch.hpu.synchronize()