Misc. bug: slow model loading to GPU when size > 64GB (Vulkan) #14854
Description
Name and Version
$ llama-cli --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
version: 5954 (6c9ee3b1)
built with cc (GCC) 15.1.1 20250521 (Red Hat 15.1.1-2) for x86_64-redhat-linux
Operating systems
Linux
Which llama.cpp modules do you know to be affected?
llama-cli
Problem description & steps to reproduce
Using llama-cli I can run models that take up to ~64GB of RAM just fine, but as soon as I try to offload a model that's bigger, although I have GTT VRAM, loading/inference becomes impossibly slow.
I made a video that shows the difference between offloading 32 layers (<64GB) and 40 layers (>64GB):
Full commands and outputs follow.
My configuration
I am using llama-cpp with Vulkan backend on Fedora 42. I have an AMD Strix Halo, 128GB of RAM, configured like this:
- Minimum memory allocated to the GPU (512MB) in the BIOS, and then linux configured to use up to 124GB. This is recognized by the system as shown in this screenshot:
Example offloading 32 layers to the GPU:
llama-cli -ngl 32 --model llama-4-scout-17b-16e-Q6_K/Q6_K/Llama-4-Scout-17B-16E-Instruct-Q6_K-00001-of-00002.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
build: 5954 (6c9ee3b1) with cc (GCC) 15.1.1 20250521 (Red Hat 15.1.1-2) for x86_64-redhat-linux
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (Radeon 8060S Graphics (RADV GFX1151)) - 83008 MiB free
llama_model_loader: additional 1 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 51 key-value pairs and 628 tensors from llama-4-scout-17b-16e-Q6_K/Q6_K/Llama-4-Scout-17B-16E-Instruct-Q6_K-00001-of-00002.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 = llama4
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Llama-4-Scout-17B-16E-Instruct
llama_model_loader: - kv 3: general.finetune str = 16E-Instruct
llama_model_loader: - kv 4: general.basename str = Llama-4-Scout-17B-16E-Instruct
llama_model_loader: - kv 5: general.quantized_by str = Unsloth
llama_model_loader: - kv 6: general.size_label str = 17B
llama_model_loader: - kv 7: general.license str = other
llama_model_loader: - kv 8: general.license.name str = llama4
llama_model_loader: - kv 9: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 10: general.base_model.count u32 = 1
llama_model_loader: - kv 11: general.base_model.0.name str = Llama 4 Scout 17B 16E Instruct
llama_model_loader: - kv 12: general.base_model.0.organization str = Meta Llama
llama_model_loader: - kv 13: general.base_model.0.repo_url str = https://huggingface.co/meta-llama/Lla...
llama_model_loader: - kv 14: general.tags arr[str,5] = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv 15: general.languages arr[str,12] = ["ar", "de", "en", "es", "fr", "hi", ...
llama_model_loader: - kv 16: llama4.block_count u32 = 48
llama_model_loader: - kv 17: llama4.context_length u32 = 10485760
llama_model_loader: - kv 18: llama4.embedding_length u32 = 5120
llama_model_loader: - kv 19: llama4.feed_forward_length u32 = 16384
llama_model_loader: - kv 20: llama4.attention.head_count u32 = 40
llama_model_loader: - kv 21: llama4.attention.head_count_kv u32 = 8
llama_model_loader: - kv 22: llama4.rope.freq_base f32 = 500000.000000
llama_model_loader: - kv 23: llama4.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 24: llama4.expert_count u32 = 16
llama_model_loader: - kv 25: llama4.expert_used_count u32 = 1
llama_model_loader: - kv 26: llama4.attention.key_length u32 = 128
llama_model_loader: - kv 27: llama4.attention.value_length u32 = 128
llama_model_loader: - kv 28: llama4.vocab_size u32 = 202048
llama_model_loader: - kv 29: llama4.rope.dimension_count u32 = 128
llama_model_loader: - kv 30: llama4.interleave_moe_layer_step u32 = 1
llama_model_loader: - kv 31: llama4.expert_feed_forward_length u32 = 8192
llama_model_loader: - kv 32: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 33: tokenizer.ggml.pre str = llama4
llama_model_loader: - kv 34: tokenizer.ggml.tokens arr[str,202048] = ["À", "Á", "õ", "ö", "÷", "ø", ...
llama_model_loader: - kv 35: tokenizer.ggml.token_type arr[i32,202048] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 36: tokenizer.ggml.merges arr[str,439802] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 37: tokenizer.ggml.bos_token_id u32 = 200000
llama_model_loader: - kv 38: tokenizer.ggml.eos_token_id u32 = 200008
llama_model_loader: - kv 39: tokenizer.ggml.padding_token_id u32 = 200018
llama_model_loader: - kv 40: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 41: tokenizer.chat_template str = {{- bos_token }}\n{%- if custom_tools ...
llama_model_loader: - kv 42: general.quantization_version u32 = 2
llama_model_loader: - kv 43: general.file_type u32 = 18
llama_model_loader: - kv 44: quantize.imatrix.file str = Llama-4-Scout-17B-16E-Instruct-GGUF/i...
llama_model_loader: - kv 45: quantize.imatrix.dataset str = unsloth_calibration_Llama-4-Scout-17B...
llama_model_loader: - kv 46: quantize.imatrix.entries_count u32 = 528
llama_model_loader: - kv 47: quantize.imatrix.chunks_count u32 = 729
llama_model_loader: - kv 48: split.no u16 = 0
llama_model_loader: - kv 49: split.tensors.count i32 = 628
llama_model_loader: - kv 50: split.count u16 = 2
llama_model_loader: - type f32: 146 tensors
llama_model_loader: - type q6_K: 482 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q6_K
print_info: file size = 82.35 GiB (6.56 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 1135
load: token to piece cache size = 1.3873 MB
print_info: arch = llama4
print_info: vocab_only = 0
print_info: n_ctx_train = 10485760
print_info: n_embd = 5120
print_info: n_layer = 48
print_info: n_head = 40
print_info: n_head_kv = 8
print_info: n_rot = 128
print_info: n_swa = 8192
print_info: is_swa_any = 1
print_info: n_embd_head_k = 128
print_info: n_embd_head_v = 128
print_info: n_gqa = 5
print_info: n_embd_k_gqa = 1024
print_info: n_embd_v_gqa = 1024
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 16384
print_info: n_expert = 16
print_info: n_expert_used = 1
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 0
print_info: rope scaling = linear
print_info: freq_base_train = 500000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 10485760
print_info: rope_finetuned = unknown
print_info: model type = 17Bx16E (Scout)
print_info: model params = 107.77 B
print_info: general.name = Llama-4-Scout-17B-16E-Instruct
print_info: vocab type = BPE
print_info: n_vocab = 202048
print_info: n_merges = 439802
print_info: BOS token = 200000 '<|begin_of_text|>'
print_info: EOS token = 200008 '<|eot|>'
print_info: PAD token = 200018 '<|finetune_right_pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 200002 '<|fim_prefix|>'
print_info: FIM SUF token = 200004 '<|fim_suffix|>'
print_info: FIM MID token = 200003 '<|fim_middle|>'
print_info: EOG token = 200001 '<|end_of_text|>'
print_info: EOG token = 200008 '<|eot|>'
print_info: max token length = 192
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 32 repeating layers to GPU
load_tensors: offloaded 32/49 layers to GPU
load_tensors: Vulkan0 model buffer size = 55136.25 MiB
load_tensors: CPU_Mapped model buffer size = 29186.72 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = 0
llama_context: kv_unified = true
llama_context: freq_base = 500000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (10485760) -- the full capacity of the model will not be utilized
llama_context: CPU output buffer size = 0.77 MiB
llama_kv_cache_unified_iswa: creating non-SWA KV cache, size = 4096 cells
llama_kv_cache_unified: Vulkan0 KV buffer size = 128.00 MiB
llama_kv_cache_unified: CPU KV buffer size = 64.00 MiB
llama_kv_cache_unified: size = 192.00 MiB ( 4096 cells, 12 layers, 1/ 1 seqs), K (f16): 96.00 MiB, V (f16): 96.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_kv_cache_unified_iswa: creating SWA KV cache, size = 4096 cells
llama_kv_cache_unified: Vulkan0 KV buffer size = 384.00 MiB
llama_kv_cache_unified: CPU KV buffer size = 192.00 MiB
llama_kv_cache_unified: size = 576.00 MiB ( 4096 cells, 36 layers, 1/ 1 seqs), K (f16): 288.00 MiB, V (f16): 288.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_context: Vulkan0 compute buffer size = 1251.92 MiB
llama_context: Vulkan_Host compute buffer size = 26.02 MiB
llama_context: graph nodes = 2610
llama_context: graph splits = 245 (with bs=512), 3 (with bs=1)
common_init_from_params: added <|end_of_text|> logit bias = -inf
common_init_from_params: added <|eot|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 16
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<|header_start|>system<|header_end|>
You are a helpful assistant<|eot|><|header_start|>user<|header_end|>
Hello<|eot|><|header_start|>assistant<|header_end|>
Hi there<|eot|><|header_start|>user<|header_end|>
How are you?<|eot|><|header_start|>assistant<|header_end|>
system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
main: interactive mode on.
sampler seed: 482087344
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = -1, n_keep = 1
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to the AI.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
- Not using system message. To change it, set a different value via -sys PROMPT
> hey
Hey! How's it going? Is there something I can help you with or do you just want to chat?
>
llama_perf_sampler_print: sampling time = 2.17 ms / 34 runs ( 0.06 ms per token, 15646.57 tokens per second)
llama_perf_context_print: load time = 50703.54 ms
llama_perf_context_print: prompt eval time = 715.35 ms / 11 tokens ( 65.03 ms per token, 15.38 tokens per second)
llama_perf_context_print: eval time = 2159.87 ms / 23 runs ( 93.91 ms per token, 10.65 tokens per second)
llama_perf_context_print: total time = 15246.60 ms / 34 tokens
llama_perf_context_print: graphs reused = 0
Example offloading 40 layers:
llama-cli -ngl 40 --model llama-4-scout-17b-16e-Q6_K/Q6_K/Llama-4-Scout-17B-16E-Instruct-Q6_K-00001-of-00002.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
build: 5954 (6c9ee3b1) with cc (GCC) 15.1.1 20250521 (Red Hat 15.1.1-2) for x86_64-redhat-linux
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (Radeon 8060S Graphics (RADV GFX1151)) - 83008 MiB free
llama_model_loader: additional 1 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 51 key-value pairs and 628 tensors from llama-4-scout-17b-16e-Q6_K/Q6_K/Llama-4-Scout-17B-16E-Instruct-Q6_K-00001-of-00002.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 = llama4
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Llama-4-Scout-17B-16E-Instruct
llama_model_loader: - kv 3: general.finetune str = 16E-Instruct
llama_model_loader: - kv 4: general.basename str = Llama-4-Scout-17B-16E-Instruct
llama_model_loader: - kv 5: general.quantized_by str = Unsloth
llama_model_loader: - kv 6: general.size_label str = 17B
llama_model_loader: - kv 7: general.license str = other
llama_model_loader: - kv 8: general.license.name str = llama4
llama_model_loader: - kv 9: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 10: general.base_model.count u32 = 1
llama_model_loader: - kv 11: general.base_model.0.name str = Llama 4 Scout 17B 16E Instruct
llama_model_loader: - kv 12: general.base_model.0.organization str = Meta Llama
llama_model_loader: - kv 13: general.base_model.0.repo_url str = https://huggingface.co/meta-llama/Lla...
llama_model_loader: - kv 14: general.tags arr[str,5] = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv 15: general.languages arr[str,12] = ["ar", "de", "en", "es", "fr", "hi", ...
llama_model_loader: - kv 16: llama4.block_count u32 = 48
llama_model_loader: - kv 17: llama4.context_length u32 = 10485760
llama_model_loader: - kv 18: llama4.embedding_length u32 = 5120
llama_model_loader: - kv 19: llama4.feed_forward_length u32 = 16384
llama_model_loader: - kv 20: llama4.attention.head_count u32 = 40
llama_model_loader: - kv 21: llama4.attention.head_count_kv u32 = 8
llama_model_loader: - kv 22: llama4.rope.freq_base f32 = 500000.000000
llama_model_loader: - kv 23: llama4.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 24: llama4.expert_count u32 = 16
llama_model_loader: - kv 25: llama4.expert_used_count u32 = 1
llama_model_loader: - kv 26: llama4.attention.key_length u32 = 128
llama_model_loader: - kv 27: llama4.attention.value_length u32 = 128
llama_model_loader: - kv 28: llama4.vocab_size u32 = 202048
llama_model_loader: - kv 29: llama4.rope.dimension_count u32 = 128
llama_model_loader: - kv 30: llama4.interleave_moe_layer_step u32 = 1
llama_model_loader: - kv 31: llama4.expert_feed_forward_length u32 = 8192
llama_model_loader: - kv 32: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 33: tokenizer.ggml.pre str = llama4
llama_model_loader: - kv 34: tokenizer.ggml.tokens arr[str,202048] = ["À", "Á", "õ", "ö", "÷", "ø", ...
llama_model_loader: - kv 35: tokenizer.ggml.token_type arr[i32,202048] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 36: tokenizer.ggml.merges arr[str,439802] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 37: tokenizer.ggml.bos_token_id u32 = 200000
llama_model_loader: - kv 38: tokenizer.ggml.eos_token_id u32 = 200008
llama_model_loader: - kv 39: tokenizer.ggml.padding_token_id u32 = 200018
llama_model_loader: - kv 40: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 41: tokenizer.chat_template str = {{- bos_token }}\n{%- if custom_tools ...
llama_model_loader: - kv 42: general.quantization_version u32 = 2
llama_model_loader: - kv 43: general.file_type u32 = 18
llama_model_loader: - kv 44: quantize.imatrix.file str = Llama-4-Scout-17B-16E-Instruct-GGUF/i...
llama_model_loader: - kv 45: quantize.imatrix.dataset str = unsloth_calibration_Llama-4-Scout-17B...
llama_model_loader: - kv 46: quantize.imatrix.entries_count u32 = 528
llama_model_loader: - kv 47: quantize.imatrix.chunks_count u32 = 729
llama_model_loader: - kv 48: split.no u16 = 0
llama_model_loader: - kv 49: split.tensors.count i32 = 628
llama_model_loader: - kv 50: split.count u16 = 2
llama_model_loader: - type f32: 146 tensors
llama_model_loader: - type q6_K: 482 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q6_K
print_info: file size = 82.35 GiB (6.56 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 1135
load: token to piece cache size = 1.3873 MB
print_info: arch = llama4
print_info: vocab_only = 0
print_info: n_ctx_train = 10485760
print_info: n_embd = 5120
print_info: n_layer = 48
print_info: n_head = 40
print_info: n_head_kv = 8
print_info: n_rot = 128
print_info: n_swa = 8192
print_info: is_swa_any = 1
print_info: n_embd_head_k = 128
print_info: n_embd_head_v = 128
print_info: n_gqa = 5
print_info: n_embd_k_gqa = 1024
print_info: n_embd_v_gqa = 1024
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 16384
print_info: n_expert = 16
print_info: n_expert_used = 1
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 0
print_info: rope scaling = linear
print_info: freq_base_train = 500000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 10485760
print_info: rope_finetuned = unknown
print_info: model type = 17Bx16E (Scout)
print_info: model params = 107.77 B
print_info: general.name = Llama-4-Scout-17B-16E-Instruct
print_info: vocab type = BPE
print_info: n_vocab = 202048
print_info: n_merges = 439802
print_info: BOS token = 200000 '<|begin_of_text|>'
print_info: EOS token = 200008 '<|eot|>'
print_info: PAD token = 200018 '<|finetune_right_pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 200002 '<|fim_prefix|>'
print_info: FIM SUF token = 200004 '<|fim_suffix|>'
print_info: FIM MID token = 200003 '<|fim_middle|>'
print_info: EOG token = 200001 '<|end_of_text|>'
print_info: EOG token = 200008 '<|eot|>'
print_info: max token length = 192
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 40 repeating layers to GPU
load_tensors: offloaded 40/49 layers to GPU
load_tensors: Vulkan0 model buffer size = 68920.31 MiB
load_tensors: CPU_Mapped model buffer size = 15402.66 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = 0
llama_context: kv_unified = true
llama_context: freq_base = 500000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (10485760) -- the full capacity of the model will not be utilized
llama_context: CPU output buffer size = 0.77 MiB
llama_kv_cache_unified_iswa: creating non-SWA KV cache, size = 4096 cells
llama_kv_cache_unified: Vulkan0 KV buffer size = 160.00 MiB
llama_kv_cache_unified: CPU KV buffer size = 32.00 MiB
llama_kv_cache_unified: size = 192.00 MiB ( 4096 cells, 12 layers, 1/ 1 seqs), K (f16): 96.00 MiB, V (f16): 96.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_kv_cache_unified_iswa: creating SWA KV cache, size = 4096 cells
llama_kv_cache_unified: Vulkan0 KV buffer size = 480.00 MiB
llama_kv_cache_unified: CPU KV buffer size = 96.00 MiB
llama_kv_cache_unified: size = 576.00 MiB ( 4096 cells, 36 layers, 1/ 1 seqs), K (f16): 288.00 MiB, V (f16): 288.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_context: Vulkan0 compute buffer size = 1251.92 MiB
llama_context: Vulkan_Host compute buffer size = 26.02 MiB
llama_context: graph nodes = 2610
llama_context: graph splits = 125 (with bs=512), 3 (with bs=1)
common_init_from_params: added <|end_of_text|> logit bias = -inf
common_init_from_params: added <|eot|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 16
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<|header_start|>system<|header_end|>
You are a helpful assistant<|eot|><|header_start|>user<|header_end|>
Hello<|eot|><|header_start|>assistant<|header_end|>
Hi there<|eot|><|header_start|>user<|header_end|>
How are you?<|eot|><|header_start|>assistant<|header_end|>
system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
main: interactive mode on.
sampler seed: 3628593135
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = -1, n_keep = 1
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to the AI.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
- Not using system message. To change it, set a different value via -sys PROMPT
> hey
(I gave up waiting after one minute)
I am not sure why this is happening. It looks like some memory allocation bug.