Misc. bug: Embedding crashes on macOS when not using Metal

[giladgd]

Name and Version

version: 4277 (c5ede38)
built with Apple clang version 16.0.0 (clang-1600.0.26.4) for arm64-apple-darwin24.1.0

Operating systems

Mac, M1 Max

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

libllama (core library), Other (Please specify in the next section)

Problem description & steps to reproduce

Using llama-embedding with bge-small-en-v1.5-q8_0.gguf on macOS without Metal crashes with 85246 illegal hardware instruction, but works when using Metal.

This is the command I used:

./bin/llama-embedding -m ./models/bge-small-en-v1.5-q8_0.gguf --prompt "hi" -c 0

Running with lldb gave me this stack trace:

Process 88174 stopped
* thread #7, stop reason = EXC_BAD_INSTRUCTION (code=1, subcode=0x4e84a653)
    frame #0: 0x0000000100187128 libggml-cpu.dylib`ggml_vec_dot_q8_0_q8_0 + 192
libggml-cpu.dylib`ggml_vec_dot_q8_0_q8_0:
->  0x100187128 <+192>: smmla  
    0x10018712c <+196>: smmla  
    0x100187130 <+200>: zip2.2d v1, v6, v16
    0x100187134 <+204>: smmla  
Target 0: (llama-embedding) stopped.

I built using these commands:

mkdir build
cd build
cmake -DGGML_METAL=0 ..
cmake --build . --config Release --clean-first --parallel 9

First Bad Commit

This issue appeared since release b4179.

Relevant log output

build: 4277 (c5ede3849) with Apple clang version 16.0.0 (clang-1600.0.26.4) for arm64-apple-darwin24.1.0
llama_model_loader: loaded meta data with 24 key-value pairs and 197 tensors from ./models/bge-small-en-v1.5-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              = bert
llama_model_loader: - kv   1:                               general.name str              = bge-small-en-v1.5
llama_model_loader: - kv   2:                           bert.block_count u32              = 12
llama_model_loader: - kv   3:                        bert.context_length u32              = 512
llama_model_loader: - kv   4:                      bert.embedding_length u32              = 384
llama_model_loader: - kv   5:                   bert.feed_forward_length u32              = 1536
llama_model_loader: - kv   6:                  bert.attention.head_count u32              = 12
llama_model_loader: - kv   7:          bert.attention.layer_norm_epsilon f32              = 0.000000
llama_model_loader: - kv   8:                          general.file_type u32              = 7
llama_model_loader: - kv   9:                      bert.attention.causal bool             = false
llama_model_loader: - kv  10:                          bert.pooling_type u32              = 2
llama_model_loader: - kv  11:            tokenizer.ggml.token_type_count u32              = 2
llama_model_loader: - kv  12:                tokenizer.ggml.bos_token_id u32              = 101
llama_model_loader: - kv  13:                tokenizer.ggml.eos_token_id u32              = 102
llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = bert
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,30522]   = ["[PAD]", "[unused0]", "[unused1]", "...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,30522]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,30522]   = [3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 100
llama_model_loader: - kv  19:          tokenizer.ggml.seperator_token_id u32              = 102
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  21:                tokenizer.ggml.cls_token_id u32              = 101
llama_model_loader: - kv  22:               tokenizer.ggml.mask_token_id u32              = 103
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  124 tensors
llama_model_loader: - type q8_0:   73 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 = 5
llm_load_vocab: token to piece cache size = 0.2032 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = bert
llm_load_print_meta: vocab type       = WPM
llm_load_print_meta: n_vocab          = 30522
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 512
llm_load_print_meta: n_embd           = 384
llm_load_print_meta: n_layer          = 12
llm_load_print_meta: n_head           = 12
llm_load_print_meta: n_head_kv        = 12
llm_load_print_meta: n_rot            = 32
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 32
llm_load_print_meta: n_embd_head_v    = 32
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 384
llm_load_print_meta: n_embd_v_gqa     = 384
llm_load_print_meta: f_norm_eps       = 1.0e-12
llm_load_print_meta: f_norm_rms_eps   = 0.0e+00
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             = 1536
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 0
llm_load_print_meta: pooling type     = 2
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_ctx_orig_yarn  = 512
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       = 33M
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 33.21 M
llm_load_print_meta: model size       = 34.38 MiB (8.68 BPW) 
llm_load_print_meta: general.name     = bge-small-en-v1.5
llm_load_print_meta: BOS token        = 101 '[CLS]'
llm_load_print_meta: EOS token        = 102 '[SEP]'
llm_load_print_meta: UNK token        = 100 '[UNK]'
llm_load_print_meta: SEP token        = 102 '[SEP]'
llm_load_print_meta: PAD token        = 0 '[PAD]'
llm_load_print_meta: CLS token        = 101 '[CLS]'
llm_load_print_meta: MASK token       = 103 '[MASK]'
llm_load_print_meta: LF token         = 0 '[PAD]'
llm_load_print_meta: EOG token        = 102 '[SEP]'
llm_load_print_meta: max token length = 21
llm_load_tensors:   CPU_Mapped model buffer size =    34.38 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      = 2048
llama_new_context_with_model: flash_attn    = 0
llama_new_context_with_model: freq_base     = 10000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_pre_seq (4096) > n_ctx_train (512) -- possible training context overflow
llama_kv_cache_init:        CPU KV buffer size =    72.00 MiB
llama_new_context_with_model: KV self size  =   72.00 MiB, K (f16):   36.00 MiB, V (f16):   36.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.00 MiB
llama_new_context_with_model:        CPU compute buffer size =   220.02 MiB
llama_new_context_with_model: graph nodes  = 429
llama_new_context_with_model: graph splits = 169 (with bs=2048), 1 (with bs=1)
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
[1]    89180 illegal hardware instruction  ./llama-embedding -m ./models/bge-small-en-v1.5-q8_0.gguf

[ggerganov]

It's currently expected since the M1 chips don't support the i8mm instructions. Will be fixed similar to #10606 but for Arm. You can workaround for now with -march=armv8.2a+dotprod.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.