Misc. bug: KV cache loads only into CPU RAM

[Apanoff]

Name and Version

ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes ggml_cuda_init: found 1 CUDA devices: Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes version: 4202 (9f912511) built with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu

Operating systems

Linux, Ubuntu

Which llama.cpp modules do you know to be affected?

No response

Problem description & steps to reproduce

For some reason KV cache loads only into CPU RAM, not into GPU VRAM. llama.cpp compiled with CUDA support.

First Bad Commit

No response

Relevant log output

~/llama.cpp$ ./llama-server -m models/Mistral-Large-Instruct-2411-Q8_0.gguf --ctx-size 4096 --flash-attn --cache-type-k q8_0 --cache-type-v q8_0 --mlock --gpu-layers 1  --prio-batch 2 --threads 12  --verbose
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
build: 4202 (9f912511) with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu
system info: n_threads = 12, n_threads_batch = 12, total_threads = 32

system_info: n_threads = 12 (n_threads_batch = 12) / 32 | CUDA : ARCHS = 890 | FORCE_CUBLAS = 1 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 31
main: loading model
srv    load_model: loading model 'models/Mistral-Large-Instruct-2411-Q8_0.gguf'
llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 4090) - 22350 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 795 tensors from models/Mistral-Large-Instruct-2411-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              = Mistral Large Instruct 2411
llama_model_loader: - kv   3:                            general.version str              = 2411
llama_model_loader: - kv   4:                           general.finetune str              = Instruct
llama_model_loader: - kv   5:                           general.basename str              = Mistral
llama_model_loader: - kv   6:                         general.size_label str              = Large
llama_model_loader: - kv   7:                            general.license str              = other
llama_model_loader: - kv   8:                       general.license.name str              = mrl
llama_model_loader: - kv   9:                       general.license.link str              = https://mistral.ai/licenses/MRL-0.1.md
llama_model_loader: - kv  10:                          general.languages arr[str,10]      = ["en", "fr", "de", "es", "it", "pt", ...
llama_model_loader: - kv  11:                          llama.block_count u32              = 88
llama_model_loader: - kv  12:                       llama.context_length u32              = 131072
llama_model_loader: - kv  13:                     llama.embedding_length u32              = 12288
llama_model_loader: - kv  14:                  llama.feed_forward_length u32              = 28672
llama_model_loader: - kv  15:                 llama.attention.head_count u32              = 96
llama_model_loader: - kv  16:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  17:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  18:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  19:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  20:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  21:                          general.file_type u32              = 7
llama_model_loader: - kv  22:                           llama.vocab_size u32              = 32768
llama_model_loader: - kv  23:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  24:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  25:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  26:                      tokenizer.ggml.tokens arr[str,32768]   = ["<unk>", "<s>", "</s>", "[INST]", "[...
llama_model_loader: - kv  27:                      tokenizer.ggml.scores arr[f32,32768]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  28:                  tokenizer.ggml.token_type arr[i32,32768]   = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  30:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  31:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  35:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  36:               general.quantization_version u32              = 2
llama_model_loader: - kv  37:                                   split.no u16              = 0
llama_model_loader: - kv  38:                                split.count u16              = 0
llama_model_loader: - kv  39:                        split.tensors.count i32              = 795
llama_model_loader: - type  f32:  177 tensors
llama_model_loader: - type q8_0:  618 tensors
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
llm_load_vocab: special tokens cache size = 771
llm_load_vocab: token to piece cache size = 0.1732 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32768
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 12288
llm_load_print_meta: n_layer          = 88
llm_load_print_meta: n_head           = 96
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            = 12
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             = 28672
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  = 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      = Q8_0
llm_load_print_meta: model params     = 122.61 B
llm_load_print_meta: model size       = 121.33 GiB (8.50 BPW) 
llm_load_print_meta: general.name     = Mistral Large Instruct 2411
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: LF token         = 781 '<0x0A>'
llm_load_print_meta: EOG token        = 2 '</s>'
llm_load_print_meta: max token length = 48
llm_load_tensors: tensor 'token_embd.weight' (q8_0) (and 785 others) cannot be used with preferred buffer type CPU_AARCH64, using CPU instead
llm_load_tensors: offloading 1 repeating layers to GPU
llm_load_tensors: offloaded 1/89 layers to GPU
llm_load_tensors:   CPU_Mapped model buffer size = 122841.70 MiB
llm_load_tensors:        CUDA0 model buffer size =  1402.59 MiB
....................................................................................................
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 4096
llama_new_context_with_model: n_ctx_per_seq = 4096
llama_new_context_with_model: n_batch       = 2048
llama_new_context_with_model: n_ubatch      = 512
llama_new_context_with_model: flash_attn    = 1
llama_new_context_with_model: freq_base     = 1000000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init:        CPU KV buffer size =   739.50 MiB
llama_kv_cache_init:      CUDA0 KV buffer size =     8.50 MiB
llama_new_context_with_model: KV self size  =  748.00 MiB, K (q8_0):  374.00 MiB, V (q8_0):  374.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   614.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    32.01 MiB
llama_new_context_with_model: graph nodes  = 2471
llama_new_context_with_model: graph splits = 961 (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 = 4096
slot        reset: id  0 | task -1 | 
main: model loaded
main: chat template, built_in: 1, chat_example: '[INST] You are a helpful assistant
Hello [/INST]Hi there</s>[INST] How are you? [/INST]'
main: server is listening on http://127.0.0.1:8080 - starting the main loop
que    start_loop: processing new tasks
que    start_loop: update slots
srv  update_slots: all slots are idle
srv  kv_cache_cle: clearing KV cache
que    start_loop: waiting for new tasks

[0cc4m]

llm_load_tensors: offloaded 1/89 layers to GPU

You only offloaded a single layer. Only the KV cache for that layer is gonna get offloaded to GPU, which happened as expected:

llama_kv_cache_init: CPU KV buffer size = 739.50 MiB llama_kv_cache_init: CUDA0 KV buffer size = 8.50 MiB