Advanced users • Re: Whisper-cpp and Vulkan shared memory problem

Code:

#!/usr/bin/env bashset -euo pipefail# Usage: ./run.sh <model-id> <gpu_layers> <ctx-size> <predict-tokens>if [ $# -ne 4 ]; then    echo "Usage: $0 <model-id> <gpu_layers> <ctx-size> <predict-tokens>"    exit 1fiMODEL="$1"GPU_LAYERS="$2"CTX_SIZE="$3"PREDICT_TOKENS="$4"LLAMA_BIN="$HOME/llama.cpp/build/bin/llama-server"if [ ! -x "$LLAMA_BIN" ]; then    echo "Error: llama-server binary not found at $LLAMA_BIN"    echo "Make sure llama.cpp is built with Vulkan support."    exit 1fi"$LLAMA_BIN" \    -hf "$MODEL" \    --host 0.0.0.0 \    --port 8080 \    --ctx-size "$CTX_SIZE" \    -n "$PREDICT_TOKENS" \    --n-gpu-layers "$GPU_LAYERS" \    --webui

Code:

user@raspberry-pi-5:~/Local-LLMs $ ./run_vl.sh unsloth/gemma-3-4b-it-GGUF:UD-Q4_K_XL 34 4096 3172ggml_vulkan: Found 1 Vulkan devices:ggml_vulkan: 0 = V3D 7.1.7.0 (V3DV Mesa) | uma: 1 | fp16: 0 | bf16: 0 | warp size: 16 | shared memory: 16384 | int dot: 0 | matrix cores: nonecommon_download_file_single_online: using cached file: /home/user/.cache/llama.cpp/unsloth_gemma-3-4b-it-GGUF_gemma-3-4b-it-UD-Q4_K_XL.ggufcommon_download_file_single_online: using cached file: /home/user/.cache/llama.cpp/unsloth_gemma-3-4b-it-GGUF_mmproj-F16.ggufmain: n_parallel is set to auto, using n_parallel = 4 and kv_unified = truebuild: 7650 (68b4d516c) with GNU 14.2.0 for Linux aarch64system info: n_threads = 4, n_threads_batch = 4, total_threads = 4system_info: n_threads = 4 (n_threads_batch = 4) / 4 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | DOTPROD = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | init: using 6 threads for HTTP serverstart: binding port with default address familymain: loading modelsrv    load_model: loading model '/home/user/.cache/llama.cpp/unsloth_gemma-3-4b-it-GGUF_gemma-3-4b-it-UD-Q4_K_XL.gguf'common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit onggml_vulkan: Error: Shared memory size too small for matrix multiplication.llama_model_load: error loading model: Shared memory size too small for matrix multiplication.llama_model_load_from_file_impl: failed to load modelllama_params_fit: encountered an error while trying to fit params to free device memory: failed to load modelllama_params_fit: fitting params to free memory took 0.67 secondsllama_model_load_from_file_impl: using device Vulkan0 (V3D 7.1.7.0) (unknown id) - 4096 MiB freellama_model_loader: loaded meta data with 40 key-value pairs and 444 tensors from /home/user/.cache/llama.cpp/unsloth_gemma-3-4b-it-GGUF_gemma-3-4b-it-UD-Q4_K_XL.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              = gemma3llama_model_loader: - kv   1:                               general.type str              = modelllama_model_loader: - kv   2:                               general.name str              = Gemma-3-4B-Itllama_model_loader: - kv   3:                           general.finetune str              = itllama_model_loader: - kv   4:                           general.basename str              = Gemma-3-4B-Itllama_model_loader: - kv   5:                       general.quantized_by str              = Unslothllama_model_loader: - kv   6:                         general.size_label str              = 4Bllama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unslothllama_model_loader: - kv   8:                      gemma3.context_length u32              = 131072llama_model_loader: - kv   9:                    gemma3.embedding_length u32              = 2560llama_model_loader: - kv  10:                         gemma3.block_count u32              = 34llama_model_loader: - kv  11:                 gemma3.feed_forward_length u32              = 10240llama_model_loader: - kv  12:                gemma3.attention.head_count u32              = 8llama_model_loader: - kv  13:    gemma3.attention.layer_norm_rms_epsilon f32              = 0.000001llama_model_loader: - kv  14:                gemma3.attention.key_length u32              = 256llama_model_loader: - kv  15:              gemma3.attention.value_length u32              = 256llama_model_loader: - kv  16:                      gemma3.rope.freq_base f32              = 1000000.000000llama_model_loader: - kv  17:            gemma3.attention.sliding_window u32              = 1024llama_model_loader: - kv  18:             gemma3.attention.head_count_kv u32              = 4llama_model_loader: - kv  19:                   gemma3.rope.scaling.type str              = linearllama_model_loader: - kv  20:                 gemma3.rope.scaling.factor f32              = 8.000000llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = llamallama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = defaultllama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,262208]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...llama_model_loader: - kv  24:                      tokenizer.ggml.scores arr[f32,262208]  = [-1000.000000, -1000.000000, -1000.00...llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,262208]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 2llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 106llama_model_loader: - kv  28:            tokenizer.ggml.unknown_token_id u32              = 3llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 0llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = truellama_model_loader: - kv  31:               tokenizer.ggml.add_eos_token bool             = falsellama_model_loader: - kv  32:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...llama_model_loader: - kv  33:            tokenizer.ggml.add_space_prefix bool             = falsellama_model_loader: - kv  34:               general.quantization_version u32              = 2llama_model_loader: - kv  35:                          general.file_type u32              = 15llama_model_loader: - kv  36:                      quantize.imatrix.file str              = gemma-3-4b-it-GGUF/imatrix_unsloth.datllama_model_loader: - kv  37:                   quantize.imatrix.dataset str              = unsloth_calibration_gemma-3-4b-it.txtllama_model_loader: - kv  38:             quantize.imatrix.entries_count i32              = 238llama_model_loader: - kv  39:              quantize.imatrix.chunks_count i32              = 663llama_model_loader: - type  f32:  205 tensorsllama_model_loader: - type q4_K:  142 tensorsllama_model_loader: - type q5_K:   30 tensorsllama_model_loader: - type q6_K:   47 tensorsllama_model_loader: - type iq4_xs:   20 tensorsprint_info: file format = GGUF V3 (latest)print_info: file type   = Q4_K - Mediumprint_info: file size   = 2.36 GiB (5.23 BPW) load: 6242 unused tokensload: printing all EOG tokens:load:   - 106 ('<end_of_turn>')load: special tokens cache size = 6415load: token to piece cache size = 1.9446 MBprint_info: arch             = gemma3print_info: vocab_only       = 0print_info: no_alloc         = 0print_info: n_ctx_train      = 131072print_info: n_embd           = 2560print_info: n_embd_inp       = 2560print_info: n_layer          = 34print_info: n_head           = 8print_info: n_head_kv        = 4print_info: n_rot            = 256print_info: n_swa            = 1024print_info: is_swa_any       = 1print_info: n_embd_head_k    = 256print_info: n_embd_head_v    = 256print_info: n_gqa            = 2print_info: n_embd_k_gqa     = 1024print_info: n_embd_v_gqa     = 1024print_info: f_norm_eps       = 0.0e+00print_info: f_norm_rms_eps   = 1.0e-06print_info: f_clamp_kqv      = 0.0e+00print_info: f_max_alibi_bias = 0.0e+00print_info: f_logit_scale    = 0.0e+00print_info: f_attn_scale     = 6.2e-02print_info: n_ff             = 10240print_info: n_expert         = 0print_info: n_expert_used    = 0print_info: n_expert_groups  = 0print_info: n_group_used     = 0print_info: causal attn      = 1print_info: pooling type     = 0print_info: rope type        = 2print_info: rope scaling     = linearprint_info: freq_base_train  = 1000000.0print_info: freq_scale_train = 0.125print_info: freq_base_swa    = 10000.0print_info: freq_scale_swa   = 1print_info: n_ctx_orig_yarn  = 131072print_info: rope_yarn_log_mul= 0.0000print_info: rope_finetuned   = unknownprint_info: model type       = 4Bprint_info: model params     = 3.88 Bprint_info: general.name     = Gemma-3-4B-Itprint_info: vocab type       = SPMprint_info: n_vocab          = 262208print_info: n_merges         = 0print_info: BOS token        = 2 '<bos>'print_info: EOS token        = 106 '<end_of_turn>'./run_vl.sh: line 31:  1436 Segmentation fault      "$LLAMA_BIN" -hf "$MODEL" --host 0.0.0.0 --port 8080 --ctx-size "$CTX_SIZE" -n "$PREDICT_TOKENS" --n-gpu-layers "$GPU_LAYERS" --webui