Bug: Server Embedding Segmentation Fault

[kherud]

What happened?

When starting the server in embedding mode, requests to the /complete endpoint result in a segmentation fault (other endpoints might be affected too).

Tested on Macbook Air M1 and RTX 4090.

The embedding API recently changed, so it might be intended, that completion and embedding mode can't be used simultaneously (?).

However, I remember both working simultaneously in the past and couldn't find this behavior documented anywhere.
Also, the error should probably be handled more gracefully.

Steps to reproduce:

cmake -B build -DLLAMA_CURL=ON
cmake --build build --config Release -j

./llama-server -mu https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/resolve/main/mistral-7b-instruct-v0.2.Q2_K.gguf --embeddings

curl --request POST \
--url http://localhost:8080/completion \
--header "Content-Type: application/json" \
--data '{
    "prompt": "Hello, world!"
}'

Name and Version

$ ./llama-cli --version
version: 3276 (cb5fad4)
built with Apple clang version 15.0.0 (clang-1500.3.9.4) for arm64-apple-darwin23.5.0

What operating system are you seeing the problem on?

Linux, Mac

Relevant log output

INFO [                    main] build info | tid="0x1f2170c00" timestamp=1719867303 build=3276 commit="cb5fad4c"
INFO [                    main] system info | tid="0x1f2170c00" timestamp=1719867303 n_threads=4 n_threads_batch=-1 total_threads=8 system_info="AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 1 | SVE = 0 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 0 | "
[...]
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.name str              = mistralai_mistral-7b-instruct-v0.2
llama_model_loader: - kv   2:                       llama.context_length u32              = 32768
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  11:                          general.file_type u32              = 10
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 0
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 }}{% for message in mess...
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q2_K:   65 tensors
llama_model_loader: - type q3_K:  160 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: special tokens cache size = 259
llm_load_vocab: token to piece cache size = 0.1637 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: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
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             = 14336
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        = 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  = 32768
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       = 7B
llm_load_print_meta: model ftype      = Q2_K - Medium
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 2.87 GiB (3.41 BPW) 
llm_load_print_meta: general.name     = mistralai_mistral-7b-instruct-v0.2
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: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_print_meta: max token length = 48
llm_load_tensors: ggml ctx size =    0.27 MiB
ggml_backend_metal_log_allocated_size: allocated buffer, size =  2898.58 MiB, ( 2898.64 / 10922.67)
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CPU buffer size =    41.02 MiB
llm_load_tensors:      Metal buffer size =  2898.57 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M1
ggml_metal_init: picking default device: Apple M1
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M1
ggml_metal_init: GPU family: MTLGPUFamilyApple7  (1007)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 11453.25 MB
llama_kv_cache_init:      Metal KV buffer size =  4096.00 MiB
llama_new_context_with_model: KV self size  = 4096.00 MiB, K (f16): 2048.00 MiB, V (f16): 2048.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.03 MiB
llama_new_context_with_model:      Metal compute buffer size =  2144.00 MiB
llama_new_context_with_model:        CPU compute buffer size =    72.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2
INFO [                    init] initializing slots | tid="0x1f2170c00" timestamp=1719867305 n_slots=1
INFO [                    init] new slot | tid="0x1f2170c00" timestamp=1719867305 id_slot=0 n_ctx_slot=32768
INFO [                    main] model loaded | tid="0x1f2170c00" timestamp=1719867305
INFO [                    main] chat template | tid="0x1f2170c00" timestamp=1719867305 chat_example="[INST] You are a helpful assistant\nHello [/INST]Hi there</s>[INST] How are you? [/INST]" built_in=true
INFO [                    main] HTTP server listening | tid="0x1f2170c00" timestamp=1719867305 port="8080" n_threads_http="7" hostname="127.0.0.1"
INFO [            update_slots] all slots are idle | tid="0x1f2170c00" timestamp=1719867305
INFO [   launch_slot_with_task] slot is processing task | tid="0x1f2170c00" timestamp=1719867324 id_slot=0 id_task=0
INFO [            update_slots] kv cache rm [p0, end) | tid="0x1f2170c00" timestamp=1719867324 id_slot=0 id_task=0 p0=0
llama_get_logits_ith: invalid logits id 4, reason: no logits
zsh: segmentation fault  ./llama-server -mu  --embeddings

[ngxson]

Probably related to #7477 where we introduce llama_set_embeddings along with llama_set_causal_attn. We just need to set it to appropriate value for embedding/completion requests. Do you think this is relevant @iamlemec ?

[beginor]

Get the same error when run llama-server --model /models/qwen2-7b-instruct-fp16.gguf --embeddings !

Server output is:

INFO [                    main] build info | tid="0x201b3cc00" timestamp=1720086868 build=3291 commit="f6190247"
INFO [                    main] system info | tid="0x201b3cc00" timestamp=1720086868 n_threads=8 n_threads_batch=-1 total_threads=10 system_info="AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 1 | SVE = 0 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "
llama_model_loader: loaded meta data with 22 key-value pairs and 339 tensors from /Volumes/Ventoy/models/qwen2-7b-instruct-fp16.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              = qwen2
llama_model_loader: - kv   1:                               general.name str              = qwen2-7b-instruct
llama_model_loader: - kv   2:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   3:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv   4:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   5:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   6:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv   7:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv   8:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv   9:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                          general.file_type u32              = 1
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  12:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  17:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  19:                    tokenizer.chat_template str              = {% for message in messages %}{% if lo...
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type  f16:  198 tensors
llm_load_vocab: special tokens cache size = 421
llm_load_vocab: token to piece cache size = 0.9352 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 152064
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 3584
llm_load_print_meta: n_head           = 28
llm_load_print_meta: n_head_kv        = 4
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 7
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
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             = 18944
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  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
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       = ?B
llm_load_print_meta: model ftype      = F16
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 14.19 GiB (16.00 BPW)
llm_load_print_meta: general.name     = qwen2-7b-instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.30 MiB
ggml_backend_metal_log_allocated_size: allocated buffer, size = 13486.78 MiB, (13486.84 / 57344.00)
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:      Metal buffer size = 13486.77 MiB
llm_load_tensors:        CPU buffer size =  1039.50 MiB
.......................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M1 Max
ggml_metal_init: picking default device: Apple M1 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M1 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple7  (1007)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 60129.54 MB
llama_kv_cache_init:      Metal KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.03 MiB
llama_new_context_with_model:      Metal compute buffer size =  1884.00 MiB
llama_new_context_with_model:        CPU compute buffer size =    71.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 2
INFO [                    init] initializing slots | tid="0x201b3cc00" timestamp=1720086869 n_slots=1
INFO [                    init] new slot | tid="0x201b3cc00" timestamp=1720086869 id_slot=0 n_ctx_slot=32768
INFO [                    main] model loaded | tid="0x201b3cc00" timestamp=1720086869
INFO [                    main] chat template | tid="0x201b3cc00" timestamp=1720086869 chat_example="<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\nHi there<|im_end|>\n<|im_start|>user\nHow are you?<|im_end|>\n<|im_start|>assistant\n" built_in=true
INFO [                    main] HTTP server listening | tid="0x201b3cc00" timestamp=1720086869 port="8080" n_threads_http="9" hostname="127.0.0.1"
INFO [            update_slots] all slots are idle | tid="0x201b3cc00" timestamp=1720086869
INFO [   launch_slot_with_task] slot is processing task | tid="0x201b3cc00" timestamp=1720086878 id_slot=0 id_task=0
INFO [            update_slots] kv cache rm [p0, end) | tid="0x201b3cc00" timestamp=1720086878 id_slot=0 id_task=0 p0=0
llama_get_logits_ith: invalid logits id 44, reason: no logits
zsh: segmentation fault  llama.cpp/llama-server --model  --embeddings

Just try call completions endpoint with curl:

curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [
{
    "role": "system",
    "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
},
{
    "role": "user",
    "content": "Write a limerick about python exceptions"
}
]
}'
curl: (52) Empty reply from server

I have tried with other models , like meta-llama-3-8b-instruct.fp16.gguf, phi-3-mini-4k-instruct-fp16.gguf, gemma-2-9b-it-q5_k_m.gguf. When start llama-server without --embeddings, the completion endpoint works fine, but if start with --embeddings, the server breaks.

[beginor]

There is already an issue about this, I think they are the same. #8076

[compilade]

Probably related to #7477 where we introduce llama_set_embeddings along with llama_set_causal_attn. We just need to set it to appropriate value for embedding/completion requests.

Yes, this is related. And it was also expected. I actually made a review comment about this use case at the time. The fix will likely exactly involve properly using llama_set_embeddings and llama_set_causal_attn in the server example.

However, I remember both working simultaneously in the past and couldn't find this behavior documented anywhere.

I think that what previously happened was that the server gave causal embeddings for non-BERT models instead of non-causal ones, but I might be wrong on that.

But it's not exactly straightforward to fix, because a single call to llama_decode can't (yet) be both causal and non-causal (text generation is usually causal, and embeddings are usually non-causal). The fix would require to prioritize either embeddings requests or completion requests, and dedicate the next llama_decode to that when needed.

Hmm. Or to make that fix fairer, if there's any slot requesting embeddings and the model is not yet in embedding mode, switch to that and dedicate the next llama_decode to embeddings. After that llama_decode, in the next loop, if there are any slot with a completion request, switch back and dedicate the next llama_decode to causal text generation. And so on. (Not sure yet about the interactions with encoder-decoder models like T5, though)

The workaround is currently to run 2 instances of llama-server, which shouldn't use too much memory if the model files are memory-mapped (and by default, they should be).

[kherud]

Since this issue seems to be already known (and already worked on #8076 (comment)), I'm closing it in favor of #8076.

Thank you @compilade for your nice explanation. I'm not too familiar with the concept of causal and non-causal embeddings. Do you maybe have a reference where I could read more about this?