CUDA out-of-memory during warmup with UVM enabled

[v1993]

I wanted to test CUDA UVM support (#8035), mostly to see if it's viable on my 8GB VRAM setup (offloading only most layers on GPU + flash attention is already good enough in practice for me). However, llama.cpp scores an unexpected out-of-memory error and crashes when it starts to warm the model up:

> GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 llama-server --gpu-layers 999 -m llama/models/Meta-Llama-3.1-8B-Instruct-Q8_0.gguf --ctx-size 16384
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 4060 Laptop GPU, compute capability 8.9, VMM: yes
register_backend: registered backend CUDA (1 devices)
register_device: registered device CUDA0 (NVIDIA GeForce RTX 4060 Laptop GPU)
register_backend: registered backend BLAS (1 devices)
register_device: registered device BLAS (OpenBLAS)
register_backend: registered backend RPC (0 devices)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (AMD Ryzen 7 7735HS with Radeon Graphics)
build: 4231 (43957ef20) with cc (GCC) 14.2.1 20240910 for x86_64-pc-linux-gnu (debug)
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 500,600,700,800,900 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 15
main: loading model
srv    load_model: loading model 'llama/models/Meta-Llama-3.1-8B-Instruct-Q8_0.gguf'
llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 4060 Laptop GPU) - 7717 MiB free
llama_model_loader: loaded meta data with 29 key-value pairs and 292 tensors from llama/models/Meta-Llama-3.1-8B-Instruct-Q8_0.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              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Meta Llama 3.1 8B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Meta-Llama-3.1
llama_model_loader: - kv   5:                         general.size_label str              = 8B
llama_model_loader: - kv   6:                            general.license str              = llama3.1
llama_model_loader: - kv   7:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   8:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   9:                          llama.block_count u32              = 32
llama_model_loader: - kv  10:                       llama.context_length u32              = 131072
llama_model_loader: - kv  11:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv  12:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv  13:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  14:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  15:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  16:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  17:                          general.file_type u32              = 7
llama_model_loader: - kv  18:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  19:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  20:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  21:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  22:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  24:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  25:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  26:                tokenizer.ggml.eos_token_id u32              = 128009
llama_model_loader: - kv  27:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  28:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   66 tensors
llama_model_loader: - type q8_0:  226 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_head           = 32
llm_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            = 4
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             = 14336
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        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 500000.0
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       = 8B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 8.03 B
llm_load_print_meta: model size       = 7.95 GiB (8.50 BPW) 
llm_load_print_meta: general.name     = Meta Llama 3.1 8B Instruct
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOM token        = 128008 '<|eom_id|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOG token        = 128008 '<|eom_id|>'
llm_load_print_meta: EOG token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CUDA0 model buffer size =  7605.33 MiB
llm_load_tensors:   CPU_Mapped model buffer size =   532.31 MiB
.........................................................................................
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 16384
llama_new_context_with_model: n_ctx_per_seq = 16384
llama_new_context_with_model: n_batch       = 2048
llama_new_context_with_model: n_ubatch      = 512
llama_new_context_with_model: flash_attn    = 0
llama_new_context_with_model: freq_base     = 500000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (16384) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init:      CUDA0 KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.49 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =  1088.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    40.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
ggml_cuda_compute_forward: RMS_NORM failed
CUDA error: out of memory
  current device: 0, in function ggml_cuda_compute_forward at /home/eri/.cache/yay/llama.cpp-cuda/src/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:2183
  err
/home/eri/.cache/yay/llama.cpp-cuda/src/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:70: CUDA error
ptrace: Operation not permitted.
No stack.
The program is not being run.
Aborted (core dumped)

Looking at code in question, this seems especially odd to me, as there's no obvious dynamic allocation taking place there. I also do have nvidia-uvm kernel module loaded.

[danilw]

Exact same crash with

ptrace: Operation not permitted.

I read https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md

The environment variable GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 can be used to enable unified memory in Linux.

lsmod | grep -i nvidia it shows nvidia_uvm loaded

My case - I use 100% working model ~8GB size but for test - I keep ~7Gb free in vram - so model wont load without limiting -ngl
to load full 8gb to 7gb vram I set -ngl 1028

and it does load - vram goes full, nothing freeze and ~1GB ram being used - but crash in 1 sec ptrace: Operation not permitted.

not sure what going there

[v1993]

I think the ptrace error specifically stems from program trying to grab a stack trace using it and being denied access to that. It's the preceding CUDA error that actually matters.

[danilw]

just to connect similar reports together - #10091 (comment) - I added my testing results there