Eval bug: Segmentation fault with docker aarch64 on MacOS M1 using a small test model stories15M_MOE-Q8_0.gguf

[marcindulak]

Name and Version

version: 4410 (4b0c638)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for aarch64-linux-gnu

Operating systems

Mac

GGML backends

CPU

Hardware

MacOS M1

Models

https://huggingface.co/ggml-org/stories15M_MOE stories15M_MOE-Q8_0.gguf

Problem description & steps to reproduce

1. download the gguf model

   mkdir -p models
   MODEL=stories15M_MOE-Q8_0.gguf && curl -sL -o models/$MODEL "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/$MODEL?download=true"

2. run with docker on aarch64 - it fails

   docker run --platform linux/aarch64 --rm -it --name llama.cpp-full -v $PWD/models:/models ghcr.io/ggerganov/llama.cpp:full-b4410 --run -m /models/stories15M_MOE-Q8_0.gguf -p "Building a website can be done in 10 simple steps:"
   ...
   echo $?
   139

   When executing the run in the container with bash, it additionally prints "Segmentation fault (core dumped)"

   docker run --entrypoint /bin/bash --platform linux/aarch64 --rm -it --name llama.cpp-full -v $PWD/models:/models ghcr.io/ggerganov/llama.cpp:full-b4410
   ./llama-cli -m /models/stories15M_MOE-Q8_0.gguf -p "Building a website can be done in 10 simple steps:"

3. run with docker on amd64 - it succeeds

   docker run --platform linux/amd64 --rm -it --name llama.cpp-full -v $PWD/models:/models ghcr.io/ggerganov/llama.cpp:full-b4410 --run -m /models/stories15M_MOE-Q8_0.gguf -p "Building a website can be done in 10 simple steps:"
   # use docker stop llama.cpp-full to stop that run

I see the aarch64 run does not load_backend, and amd64 does

< build: 4410 (4b0c638b) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for aarch64-linux-gnu
---
> load_backend: loaded CPU backend from ./libggml-cpu-haswell.so
> build: 4410 (4b0c638b) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

First Bad Commit

No response

Relevant log output

Unable to find image 'ghcr.io/ggerganov/llama.cpp:full-b4410' locally
full-b4410: Pulling from ggerganov/llama.cpp
Digest: sha256:03fd6a1abb47fc7abe25c50a7a2fb0651ef0f0ef314e5fe0c16fa80442b1f83f
Status: Downloaded newer image for ghcr.io/ggerganov/llama.cpp:full-b4410
build: 4410 (4b0c638b) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for aarch64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 26 key-value pairs and 63 tensors from /models/stories15M_MOE-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.size_label str              = 4x24M
llama_model_loader: - kv   3:                            general.license str              = mit
llama_model_loader: - kv   4:                          llama.block_count u32              = 6
llama_model_loader: - kv   5:                       llama.context_length u32              = 256
llama_model_loader: - kv   6:                     llama.embedding_length u32              = 288
llama_model_loader: - kv   7:                  llama.feed_forward_length u32              = 768
llama_model_loader: - kv   8:                 llama.attention.head_count u32              = 6
llama_model_loader: - kv   9:              llama.attention.head_count_kv u32              = 6
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  11:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  12:                         llama.expert_count u32              = 4
llama_model_loader: - kv  13:                    llama.expert_used_count u32              = 2
llama_model_loader: - kv  14:                          general.file_type u32              = 7
llama_model_loader: - kv  15:                           llama.vocab_size u32              = 32000
llama_model_loader: - kv  16:                 llama.rope.dimension_count u32              = 48
llama_model_loader: - kv  17:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  18:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  19:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  20:                      tokenizer.ggml.scores arr[f32,32000]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  21:                  tokenizer.ggml.token_type arr[i32,32000]   = [3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  22:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  23:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  24:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   19 tensors
llama_model_loader: - type q8_0:   44 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 = 3
llm_load_vocab: token to piece cache size = 0.1684 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          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 256
llm_load_print_meta: n_embd           = 288
llm_load_print_meta: n_layer          = 6
llm_load_print_meta: n_head           = 6
llm_load_print_meta: n_head_kv        = 6
llm_load_print_meta: n_rot            = 48
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 48
llm_load_print_meta: n_embd_head_v    = 48
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 288
llm_load_print_meta: n_embd_v_gqa     = 288
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             = 768
llm_load_print_meta: n_expert         = 4
llm_load_print_meta: n_expert_used    = 2
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  = 256
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     = 36.36 M
llm_load_print_meta: model size       = 36.87 MiB (8.51 BPW) 
llm_load_print_meta: general.name     = n/a
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         = 13 '<0x0A>'
llm_load_print_meta: EOG token        = 2 '</s>'
llm_load_print_meta: max token length = 48

[noah-bytestudio]

I am facing the same issue. When running the llama.cpp:server image with the --platform linux/arm64 , the server wont start. If I use the --platform linux/amd64 flag the server starts, but is incredibly slow.