CUDA OOM in the middle of prompt processing

[jurov]

I am trying to process large prompt. I tuned kv cache quantization, offloading as many layers as possible to GPU, it starts processing and all looks fine...and after few hours it fails with OOM.

nvidia-smi shows GPU memory usage by llama.cpp steadily creeping up. Is this expected behavior and if so how much reserve should I keep? Seems like 10% is needed.

Searched for similar discussions but the topic was allocation failure before submitting any prompt. This happens in the middle of processing.

[slaren]

There should be no allocations after the first few evaluations. Please include a log that shows the OOM error.

[jurov]

./llama3server supernova/SuperNova-Medius-IQ4_XS.gguf                                                                                                                                               
Running llama-server  --no-mmap --metrics --props --parallel 1 --slots --slot-save-path slots --sampling-seq mx --min-p 0.02 --flash-attn --ctx_size 131072 -ctk iq4_nl -ctv iq4_nl --batch-size 512 --n-gpu-la
yers 41 -m supernova/SuperNova-Medius-IQ4_XS.gguf                                                                                                                                                              
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no                                                                                                                                                                     
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no                                                                                                                                                                     
ggml_cuda_init: found 1 CUDA devices:                                                                                                                                                                          
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes                                                                                                                                          
build: 4059 (d102caa1) with cc (Gentoo 12.4.0 p3) 12.4.0 for x86_64-pc-linux-gnu                                                                                                                               
system info: n_threads = 16, n_threads_batch = 16, total_threads = 16                                                                                                                                          
                                                                                                                                                                                                               
system_info: n_threads = 16 (n_threads_batch = 16) / 16 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | AMX_INT8 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | 
ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |                                                          
                                                                                                                                                                                                               
main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 15                                                                                                                              
main: loading model                                                                                                                                                                                            
llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 3060) - 11926 MiB free
llama_model_loader: loaded meta data with 36 key-value pairs and 579 tensors from supernova/SuperNova-Medius-IQ4_XS.gguf (version GGUF V3 (latest))                                                            
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.                                                                                                              
llama_model_loader: - kv   0:                       general.architecture str              = qwen2                                                                                                              
llama_model_loader: - kv   1:                               general.type str              = model                                                                                                              
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 14B                                                                                                        
llama_model_loader: - kv   3:                       general.organization str              = Qwen                                                                                                               
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5                                                                                                            
llama_model_loader: - kv   5:                         general.size_label str              = 14B                                                                                                                
llama_model_loader: - kv   6:                            general.license str              = apache-2.0                                                                                                         
llama_model_loader: - kv   7:                   general.base_model.count u32              = 1                                                                                                                  
llama_model_loader: - kv   8:                  general.base_model.0.name str              = Qwen2.5 14B                                                                                                        
llama_model_loader: - kv   9:          general.base_model.0.organization str              = Qwen                                                                                                               
llama_model_loader: - kv  10:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-14B                                                                            
llama_model_loader: - kv  11:                               general.tags arr[str,2]       = ["mergekit", "merge"]                                                                                              
llama_model_loader: - kv  12:                          qwen2.block_count u32              = 48                                                                                                                 
llama_model_loader: - kv  13:                       qwen2.context_length u32              = 131072                                                                                                             
llama_model_loader: - kv  14:                     qwen2.embedding_length u32              = 5120                                                                                                               
llama_model_loader: - kv  15:                  qwen2.feed_forward_length u32              = 13824                                                                                                              
llama_model_loader: - kv  16:                 qwen2.attention.head_count u32              = 40                                                                                                                 
llama_model_loader: - kv  17:              qwen2.attention.head_count_kv u32              = 8                                                                                                                  
llama_model_loader: - kv  18:                       qwen2.rope.freq_base f32              = 1000000.000000                                                                                                     
llama_model_loader: - kv  19:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010                                                                                                           
llama_model_loader: - kv  20:                          general.file_type u32              = 30                                                                                                                 
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = gpt2                                                                                                               
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = qwen2                                                                                                              
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...                                                                           
llama_model_loader: - kv  24:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...                                                                           
llama_model_loader: - kv  25:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...                                                                                  
llama_model_loader: - kv  26:                tokenizer.ggml.eos_token_id u32              = 151645                                                                                                             
llama_model_loader: - kv  27:            tokenizer.ggml.padding_token_id u32              = 151643                                                                                                             
llama_model_loader: - kv  28:                tokenizer.ggml.bos_token_id u32              = 151643                                                                                                             
llama_model_loader: - kv  29:               tokenizer.ggml.add_bos_token bool             = false                                                                                                              
llama_model_loader: - kv  30:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...                                                                          
llama_model_loader: - kv  31:               general.quantization_version u32              = 2                                                                                                                  
llama_model_loader: - kv  32:                      quantize.imatrix.file str              = /models_out/SuperNova-Medius-GGUF/Sup...                                                                           
llama_model_loader: - kv  33:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt                                                                               
llama_model_loader: - kv  34:             quantize.imatrix.entries_count i32              = 336                                                                                                                
llama_model_loader: - kv  35:              quantize.imatrix.chunks_count i32              = 128                                                                                                                
llama_model_loader: - type  f32:  241 tensors                                                                                                                                                                  
llama_model_loader: - type q5_K:   48 tensors                                                                                                                                                                  
llama_model_loader: - type q6_K:    1 tensors                                                                                                                                                                  
llama_model_loader: - type iq4_xs:  289 tensors                                                                                                                                                                
llm_load_vocab: special tokens cache size = 22                                                                                                                                                                 
llm_load_vocab: token to piece cache size = 0.9310 MB                                                                                                                                                          
llm_load_print_meta: format           = GGUF V3 (latest)                                                                                                                                                       
llm_load_print_meta: arch             = qwen2                                                                                                                                                                  
llm_load_print_meta: vocab type       = BPE                                                                                                                                                                    
llm_load_print_meta: n_vocab          = 152064                                                                                                                                                                 
llm_load_print_meta: n_merges         = 151387                                                                                                                                                                 
llm_load_print_meta: vocab_only       = 0                                                                                                                                                                      
llm_load_print_meta: n_ctx_train      = 131072                                                                                                                                                                 
llm_load_print_meta: n_embd           = 5120                                                                                                                                                                   
llm_load_print_meta: n_layer          = 48                                                                                                                                                                     
llm_load_print_meta: n_head           = 40     
lm_load_print_meta: n_head_kv        = 8                                                                                                                                                                      
llm_load_print_meta: n_rot            = 128                                                                                                                                                                    
llm_load_print_meta: n_swa            = 0                                                                                                                                                                      
llm_load_print_meta: n_embd_head_k    = 128                                                                                                                                                                    
llm_load_print_meta: n_embd_head_v    = 128                                                                                                                                                                    
llm_load_print_meta: n_gqa            = 5                                                                                                                                                                      
llm_load_print_meta: n_embd_k_gqa     = 1024                                                                                                                                                                   
llm_load_print_meta: n_embd_v_gqa     = 1024                                                                                                                                                                   
llm_load_print_meta: f_norm_eps       = 0.0e+00                                                                                                                                                                
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05                                                                                                                                                                
llm_load_print_meta: f_clamp_kqv      = 0.0e+00                                                                                                                                                                
llm_load_print_meta: f_max_alibi_bias = 0.0e+00                                                                                                                                                                
llm_load_print_meta: f_logit_scale    = 0.0e+00                                                                                                                                                                
llm_load_print_meta: n_ff             = 13824                                                                                                                                                                  
llm_load_print_meta: n_expert         = 0                                                                                                                                                                      
llm_load_print_meta: n_expert_used    = 0                                                                                                                                                                      
llm_load_print_meta: causal attn      = 1                                                                                                                                                                      
llm_load_print_meta: pooling type     = 0                                                                                                                                                                      
llm_load_print_meta: rope type        = 2                                                                                                                                                                      
llm_load_print_meta: rope scaling     = linear                                                                                                                                                                 
llm_load_print_meta: freq_base_train  = 1000000.0                                                                                                                                                              
llm_load_print_meta: freq_scale_train = 1                                                                                                                                                                      
llm_load_print_meta: n_ctx_orig_yarn  = 131072                                                                                                                                                                 
llm_load_print_meta: rope_finetuned   = unknown                                                                                                                                                                
llm_load_print_meta: ssm_d_conv       = 0                                                                                                                                                                      
llm_load_print_meta: ssm_d_inner      = 0                                                                                                                                                                      
llm_load_print_meta: ssm_d_state      = 0                                                                                                                                                                      
llm_load_print_meta: ssm_dt_rank      = 0                                                                                                                                                                      
llm_load_print_meta: ssm_dt_b_c_rms   = 0                                                                                                                                                                      
llm_load_print_meta: model type       = ?B                                                                                                                                                                     
llm_load_print_meta: model ftype      = IQ4_XS - 4.25 bpw                                                                                                                                                      
llm_load_print_meta: model params     = 14.77 B                                                                                                                                                                
llm_load_print_meta: model size       = 7.56 GiB (4.39 BPW)                                                                                                                                                    
llm_load_print_meta: general.name     = Qwen2.5 14B                                                                                                                                                            
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'                                                                                                                                                 
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'                                                                                                                                                    
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'                                                                                                                                                    
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'                                                                                                                                                 
llm_load_print_meta: LF token         = 148848 'ÄĬ'                                                                                                                                                            
llm_load_print_meta: FIM PRE token    = 151659 '<|fim_prefix|>'                                                                                                                                                
llm_load_print_meta: FIM SUF token    = 151661 '<|fim_suffix|>'                                                                                                                                                
llm_load_print_meta: FIM MID token    = 151660 '<|fim_middle|>'                                                                                                                                                
llm_load_print_meta: FIM PAD token    = 151662 '<|fim_pad|>'                                                                                                                                                   
llm_load_print_meta: FIM REP token    = 151663 '<|repo_name|>'                                                                                                                                                 
llm_load_print_meta: FIM SEP token    = 151664 '<|file_sep|>'                                                                                                                                                  
llm_load_print_meta: EOG token        = 151643 '<|endoftext|>'                                                                                                                                                 
llm_load_print_meta: EOG token        = 151645 '<|im_end|>'                                                                                                                                                    
llm_load_print_meta: EOG token        = 151662 '<|fim_pad|>'                                                                                                                                                   
llm_load_print_meta: EOG token        = 151663 '<|repo_name|>'                                                                                                                                                 
llm_load_print_meta: EOG token        = 151664 '<|file_sep|>'                                                                                                                                                  
llm_load_print_meta: max token length = 256                                                                                                                                                                    
llm_load_tensors: offloading 41 repeating layers to GPU                                                                                                                                                        
llm_load_tensors: offloaded 41/49 layers to GPU                                                                                                                                                                
llm_load_tensors:    CUDA_Host model buffer size =  1591.21 MiB                                                    
llama_kv_cache_init:      CUDA0 KV buffer size =  5904.00 MiB                                                                                                                                                  
llama_new_context_with_model: KV self size  = 6912.00 MiB, K (iq4_nl): 3456.00 MiB, V (iq4_nl): 3456.00 MiB                                                                                                    
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB                                                                                                                                    
llama_new_context_with_model:      CUDA0 compute buffer size =   202.02 MiB                                                                                                                                    
llama_new_context_with_model:        CPU compute buffer size =   307.00 MiB                                                                                                                                    
llama_new_context_with_model:  CUDA_Host compute buffer size =   398.01 MiB                                                                                                                                    
llama_new_context_with_model: graph nodes  = 1495                                                                                                                                                              
llama_new_context_with_model: graph splits = 123 (with bs=512), 3 (with bs=1)                                                                                                                                  
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)                                                                                                     
srv          init: initializing slots, n_slots = 1                                                                                                                                                             
slot         init: id  0 | task -1 | new slot n_ctx_slot = 131072                                                                                                                                              
main: model loaded                                                                                                                                                                                             
main: chat template, built_in: 1, chat_example: '<|im_start|>system                                                                                                                                            
You are a helpful assistant<|im_end|>                                                                                                                                                                          
<|im_start|>user                                                                                                                                                                                               
Hello<|im_end|>                                                                                                                                                                                                
<|im_start|>assistant                                                                                                                                                                                          
Hi there<|im_end|>                                                                                                                                                                                             
<|im_start|>user                                                                                                                                                                                               
How are you?<|im_end|>                                                                                                                                                                                         
<|im_start|>assistant                                                                                                                                                                                          
'                                                                                                                                                                                                              
main: server is listening on http://127.0.0.1:8080 - starting the main loop                                                                                                                                    
srv  update_slots: all slots are idle                                                                                                                                                                          
request: GET / 127.0.0.1 200                                                                                                                                                                                   
request: GET /deps_daisyui.min.css 127.0.0.1 200                                                                                                                                                               
request: GET /deps_markdown-it.js 127.0.0.1 200                                                                                                                                                                
request: GET /deps_tailwindcss.js 127.0.0.1 200                                                                                                                                                                
request: GET /completion.js 127.0.0.1 200                                                                                                                                                                      
request: GET /deps_vue.esm-browser.js 127.0.0.1 200                                                                                                                                                            
request: GET /favicon.ico 127.0.0.1 404                                                                                                                                                                        
slot launch_slot_: id  0 | task 0 | processing task                                                                                                                                                            
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 131072, n_keep = 0, n_prompt_tokens = 127632                                                                                                      
slot update_slots: id  0 | task 0 | kv cache rm [0, end)                                                                                                                                                       
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 512, n_tokens = 512, progress = 0.004012                                                                                              
slot update_slots: id  0 | task 0 | kv cache rm [512, end)                                                                                                                                                     
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 1024, n_tokens = 512, progress = 0.008023                                                                                             
slot update_slots: id  0 | task 0 | kv cache rm [1024, end)                                                                                                                                                    
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 1536, n_tokens = 512, progress = 0.012035                                                                                             
slot update_slots: id  0 | task 0 | kv cache rm [1536, end)
...snip...
slot update_slots: id  0 | task 0 | kv cache rm [11264, end)                                                                                                                                                   
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 11776, n_tokens = 512, progress = 0.092265                                                                                            
ggml/src/ggml-cuda.cu:70: CUDA error                                                                                                                                                                           
CUDA error: out of memory                                                                                                                                                                                      
  current device: 0, in function alloc at ggml/src/ggml-cuda.cu:344                                                                                                                                            
  cuMemCreate(&handle, reserve_size, &prop, 0)                                                                                                                                                                 
./llama3server: line 44: 10130 Aborted                 nice -n 5 ./llama.cpp/llama-server -t 16 $OPTS   

[slaren]

This should be a small allocation, but the static buffers alone sum to 13018 MiB, which is already more than the free memory of your GPU. Also keep in mind that iq4_nl KV quantization is not supported by the CUDA backend, it will cause attention to run on the CPU at a significant performance cost.

[jurov]

Thanks for answer. But if static buffers are more than free memory why won't it fail outright? It's such a waste of time.

And I actually plan to use the slot KV cache elsewhere with CPU-only inference so that's fine.