When using llama-cpp-rs alongside the whisper-rs crate in the same project, the application crashes during model loading.

[jiabochao]

When using llama-cpp-rs alongside the whisper-rs crate in the same project, the application crashes during model loading.

LlamaModel::load_from_file(&backend, model_path, &model_params);

The issue still occurs even when the whisper model and llama model are not being executed at the same time.

The program output is as follows:

llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = gemma
llama_model_loader: - kv   1:                               general.name str              = models--google--gemma-2b-it
llama_model_loader: - kv   2:                       gemma.context_length u32              = 8192
llama_model_loader: - kv   3:                     gemma.embedding_length u32              = 2048
llama_model_loader: - kv   4:                          gemma.block_count u32              = 18
llama_model_loader: - kv   5:                  gemma.feed_forward_length u32              = 16384
llama_model_loader: - kv   6:                 gemma.attention.head_count u32              = 8
llama_model_loader: - kv   7:              gemma.attention.head_count_kv u32              = 1
llama_model_loader: - kv   8:     gemma.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv   9:                 gemma.attention.key_length u32              = 256
llama_model_loader: - kv  10:               gemma.attention.value_length u32              = 256
llama_model_loader: - kv  11:                          general.file_type u32              = 13
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,256000]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,256000]  = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,256000]  = [3, 3, 3, 2, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  19:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  21:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  22:                    tokenizer.chat_template str              = {{ bos_token }}{% if messages[0]['rol...
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   37 tensors
llama_model_loader: - type q3_K:   72 tensors
llama_model_loader: - type q5_K:   54 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: mismatch in special tokens definition ( 416/256000 vs 260/256000 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = gemma
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 256000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 8192
llm_load_print_meta: n_embd           = 2048
llm_load_print_meta: n_head           = 8
llm_load_print_meta: n_head_kv        = 1
llm_load_print_meta: n_layer          = 18
llm_load_print_meta: n_rot            = 256
llm_load_print_meta: n_embd_head_k    = 256
llm_load_print_meta: n_embd_head_v    = 256
llm_load_print_meta: n_gqa            = 8
llm_load_print_meta: n_embd_k_gqa     = 256
llm_load_print_meta: n_embd_v_gqa     = 256
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
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             = 16384
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 8192
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: model type       = 2B
llm_load_print_meta: model ftype      = Q3_K - Large
llm_load_print_meta: model params     = 2.51 B
llm_load_print_meta: model size       = 1.36 GiB (4.66 BPW) 
llm_load_print_meta: general.name     = models--google--gemma-2b-it
llm_load_print_meta: BOS token        = 2 '<bos>'
llm_load_print_meta: EOS token        = 1 '<eos>'
llm_load_print_meta: UNK token        = 3 '<unk>'
llm_load_print_meta: PAD token        = 0 '<pad>'
llm_load_print_meta: LF token         = 227 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.13 MiB
ggml_backend_metal_buffer_from_ptr: allocated buffer, size =  1391.95 MiB, ( 1392.56 / 10922.67)
llm_load_tensors: offloading 18 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 19/19 layers to GPU
 ELIFECYCLE  Command failed with exit code 1.

[MarcusDunn]

I wonder if this is the GGUF backend not being shared.

I don't think I'll be fixing this in the foreseeable future, but thank you for the report!

Would you know if the two libraries can be used together from cpp?

[jiabochao]

I'm not very familiar with these underlying technical implementations. I'm trying out other solutions, and if I find out the problem, I'll share it here.

[jiabochao]

I wonder if this is the GGUF backend not being shared.

I don't think I'll be fixing this in the foreseeable future, but thank you for the report!

Would you know if the two libraries can be used together from cpp?

I noticed that whisper.cpp has a talk-llama example, so they should be able to be used together.

https://github.com/ggerganov/whisper.cpp/blob/master/examples/talk-llama

[oddpxl]

@jiabochao Did you find a solution ? ..this would be of much interest to me as well.

[jiabochao]

@jiabochao Did you find a solution ? ..this would be of much interest to me as well.

No, not yet, but I think a different whisper rust library might solve the problem, and I found another whisper rust library that looks like it might be better maintained, but I haven't tested it yet.

• https://github.com/huggingface/candle/tree/main/candle-examples/examples/whisper
• https://github.com/huggingface/candle/tree/main/candle-examples/examples/whisper-microphone