Bug: GGML_ASSERT(llama_add_eos_token(model) != 1) failed llama-server critical error with flan-t5 models

[fabiomatricardi]

What happened?

direct llama-cli call for Flan-T5 based models working fine. When trying to set up server, critical error stop execution

.\llama-server.exe -m .\models\LaMini-Flan-T5-248M.Q8_0.gguf -c 512

GGML_ASSERT(llama_add_eos_token(model) != 1) failed

model repo: https://huggingface.co/Felladrin/gguf-LaMini-Flan-T5-248M

model file: https://huggingface.co/Felladrin/gguf-LaMini-Flan-T5-248M/resolve/main/LaMini-Flan-T5-248M.Q8_0.gguf

Name and Version

.\llama-cli.exe --version

version: 3570 (4134999e)
built with cc (GCC) 14.2.0 for x86_64-w64-mingw32

Windows 11 with Python 3.11

What operating system are you seeing the problem on?

Windows

Relevant log output

.\llama-server.exe -m .\models\LaMini-Flan-T5-248M.Q8_0.gguf -c 512
INFO [                    main] build info | tid="1" timestamp=1723424718 build=3570 commit="4134999e"
INFO [                    main] system info | tid="1" timestamp=1723424718 n_threads=4 n_threads_batch=-1 total_threads=4 system_info="AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "
llama_model_loader: loaded meta data with 28 key-value pairs and 282 tensors from .\models\LaMini-Flan-T5-248M.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              = t5
llama_model_loader: - kv   1:                               general.name str              = T5
llama_model_loader: - kv   2:                          t5.context_length u32              = 512
llama_model_loader: - kv   3:                        t5.embedding_length u32              = 768
llama_model_loader: - kv   4:                     t5.feed_forward_length u32              = 2048
llama_model_loader: - kv   5:                             t5.block_count u32              = 12
llama_model_loader: - kv   6:                    t5.attention.head_count u32              = 12
llama_model_loader: - kv   7:                    t5.attention.key_length u32              = 64
llama_model_loader: - kv   8:                  t5.attention.value_length u32              = 64
llama_model_loader: - kv   9:            t5.attention.layer_norm_epsilon f32              = 0.000001
llama_model_loader: - kv  10:        t5.attention.relative_buckets_count u32              = 32
llama_model_loader: - kv  11:        t5.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  12:                  t5.decoder_start_token_id u32              = 0
llama_model_loader: - kv  13:                          general.file_type u32              = 7
llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = t5
llama_model_loader: - kv  15:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  16:                      tokenizer.ggml.tokens arr[str,32128]   = ["<pad>", "</s>", "<unk>", "Γûü", "X"...
llama_model_loader: - kv  17:                      tokenizer.ggml.scores arr[f32,32128]   = [0.000000, 0.000000, 0.000000, -2.012...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,32128]   = [3, 3, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  19:            tokenizer.ggml.add_space_prefix bool             = true
llama_model_loader: - kv  20:    tokenizer.ggml.remove_extra_whitespaces bool             = true
llama_model_loader: - kv  21:        tokenizer.ggml.precompiled_charsmap arr[u8,237539]   = [0, 180, 2, 0, 0, 132, 0, 0, 0, 0, 0,...
llama_model_loader: - kv  22:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  23:            tokenizer.ggml.unknown_token_id u32              = 2
llama_model_loader: - kv  24:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  25:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  26:               tokenizer.ggml.add_eos_token bool             = true
llama_model_loader: - kv  27:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   64 tensors
llama_model_loader: - type  f16:    2 tensors
llama_model_loader: - type q8_0:  216 tensors
llm_load_vocab: special tokens cache size = 131
llm_load_vocab: token to piece cache size = 0.2123 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = t5
llm_load_print_meta: vocab type       = UGM
llm_load_print_meta: n_vocab          = 32128
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           = 768
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            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 768
llm_load_print_meta: n_embd_v_gqa     = 768
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             = 2048
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        = -1
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: model type       = 250M
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 247.58 M
llm_load_print_meta: model size       = 295.12 MiB (10.00 BPW)
llm_load_print_meta: general.name     = T5
llm_load_print_meta: EOS token        = 1 '</s>'
llm_load_print_meta: UNK token        = 2 '<unk>'
llm_load_print_meta: PAD token        = 0 '<pad>'
llm_load_print_meta: LF token         = 3 'Γûü'
llm_load_print_meta: max token length = 20
llm_load_tensors: ggml ctx size =    0.11 MiB
llm_load_tensors:        CPU buffer size =   295.12 MiB
.......................................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
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_kv_cache_init:        CPU KV buffer size =    18.00 MiB
llama_new_context_with_model: KV self size  =   18.00 MiB, K (f16):    9.00 MiB, V (f16):    9.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.25 MiB
llama_new_context_with_model:        CPU compute buffer size =    45.50 MiB
llama_new_context_with_model: graph nodes  = 425
llama_new_context_with_model: graph splits = 1
examples/server/server.cpp:696: GGML_ASSERT(llama_add_eos_token(model) != 1) failed

[ggerganov]

Probably fixed via #8997

[fairydreaming]

@ggerganov I think #8997 fixes only T5 model loading, but they wont work correctly. For T5 models to work llama-server still needs to call llama_encode() and prepare input for llama_decode() with decoder start tokens like it's done in llama-cli:

llama.cpp/examples/main/main.cpp

Lines 540 to 556 in d3ae0ee

	if (llama_model_has_encoder(model)) {
	int enc_input_size = embd_inp.size();
	llama_token * enc_input_buf = embd_inp.data();
	if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
	LOG_TEE("%s : failed to eval\n", __func__);
	return 1;
	}
	llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
	if (decoder_start_token_id == -1) {
	decoder_start_token_id = llama_token_bos(model);
	}
	embd_inp.clear();
	embd_inp.push_back(decoder_start_token_id);
	}

[joelbarmettlerUZH]

I have the same issue with

./llama-server -m models/gte-Qwen2-1.5B-instruct-Q4_K_M.gguf -ngl 29 --embedding --pooling mean -c 32000

Pulled #8997 but issue remains.

version: 3539 (a8dbc6f7)
built with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu

[Aisuko]

In my case, I quantized the original HuggingFace SmolLM to https://huggingface.co/aisuko/SmolLM-135M-Instruct-gguf. And it works fine with llama-cli.

However, it doesn't work on the fine-tuned version of SmolLM-135M-Instruct https://huggingface.co/aisuko/ft-smollm-135M-instruct-on-hf-ultrafeedback.

When I fine-tuning the SmoILM, the tokenizer part is below:
https://www.kaggle.com/code/aisuko/ft-smollm-135m-instruct-on-hf-ultrafeedback

from transformers import AutoTokenizer

tokenizer=AutoTokenizer.from_pretrained(os.getenv("TOKENIZER_NAME"), add_eos_token=True)

tokenizer.pad_token=tokenizer.eos_token
tokenizer.padding_side="left"

After I convert the ft-smollm-135M-instruct-on-hf-ultrafeedback to gguf and try to launch it through llama-cli, the error log see below.

ec2-user@ip-10-110-145-139:~/workspace$ ./llama.cpp/llama-cli -m ft-smollm-135M-instruct-on-hf-ultrafeedback-f16.gguf -n 128
Log start
main: build = 3584 (5fd89a70)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1723856531
llama_model_loader: loaded meta data with 38 key-value pairs and 272 tensors from ft-smollm-135M-instruct-on-hf-ultrafeedback-f16.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.name str              = SmolLM 135M Instruct
llama_model_loader: - kv   3:                       general.organization str              = HuggingFaceTB
llama_model_loader: - kv   4:                           general.finetune str              = Instruct
llama_model_loader: - kv   5:                           general.basename str              = SmolLM
llama_model_loader: - kv   6:                         general.size_label str              = 135M
llama_model_loader: - kv   7:                            general.license str              = apache-2.0
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = SmolLM 135M Instruct
llama_model_loader: - kv  10:          general.base_model.0.organization str              = HuggingFaceTB
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/HuggingFaceTB/...
llama_model_loader: - kv  12:                               general.tags arr[str,3]       = ["trl", "orpo", "generated_from_train...
llama_model_loader: - kv  13:                          llama.block_count u32              = 30
llama_model_loader: - kv  14:                       llama.context_length u32              = 2048
llama_model_loader: - kv  15:                     llama.embedding_length u32              = 576
llama_model_loader: - kv  16:                  llama.feed_forward_length u32              = 1536
llama_model_loader: - kv  17:                 llama.attention.head_count u32              = 9
llama_model_loader: - kv  18:              llama.attention.head_count_kv u32              = 3
llama_model_loader: - kv  19:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  20:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  21:                          general.file_type u32              = 1
llama_model_loader: - kv  22:                           llama.vocab_size u32              = 49152
llama_model_loader: - kv  23:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv  24:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  25:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  27:                         tokenizer.ggml.pre str              = smollm
llama_model_loader: - kv  28:                      tokenizer.ggml.tokens arr[str,49152]   = ["<|endoftext|>", "<|im_start|>", "<|...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,49152]   = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,48900]   = ["Ġ t", "Ġ a", "i n", "h e", "Ġ Ġ...
llama_model_loader: - kv  31:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  33:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  34:            tokenizer.ggml.padding_token_id u32              = 2
llama_model_loader: - kv  35:               tokenizer.ggml.add_eos_token bool             = true
llama_model_loader: - kv  36:                    tokenizer.chat_template str              = {% for message in messages %}{{'<|im_...
llama_model_loader: - kv  37:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   61 tensors
llama_model_loader: - type  f16:  211 tensors
llm_load_vocab: special tokens cache size = 17
llm_load_vocab: token to piece cache size = 0.3170 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 49152
llm_load_print_meta: n_merges         = 48900
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 2048
llm_load_print_meta: n_embd           = 576
llm_load_print_meta: n_layer          = 30
llm_load_print_meta: n_head           = 9
llm_load_print_meta: n_head_kv        = 3
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 3
llm_load_print_meta: n_embd_k_gqa     = 192
llm_load_print_meta: n_embd_v_gqa     = 192
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             = 1536
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 2048
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     = 134.52 M
llm_load_print_meta: model size       = 256.63 MiB (16.00 BPW) 
llm_load_print_meta: general.name     = SmolLM 135M Instruct
llm_load_print_meta: BOS token        = 1 '<|im_start|>'
llm_load_print_meta: EOS token        = 2 '<|im_end|>'
llm_load_print_meta: UNK token        = 0 '<|endoftext|>'
llm_load_print_meta: PAD token        = 2 '<|im_end|>'
llm_load_print_meta: LF token         = 143 'Ä'
llm_load_print_meta: EOT token        = 0 '<|endoftext|>'
llm_load_print_meta: max token length = 162
llm_load_tensors: ggml ctx size =    0.13 MiB
llm_load_tensors:        CPU buffer size =   256.63 MiB
....................................................................
llama_new_context_with_model: n_ctx      = 2048
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  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =    45.00 MiB
llama_new_context_with_model: KV self size  =   45.00 MiB, K (f16):   22.50 MiB, V (f16):   22.50 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.19 MiB
llama_new_context_with_model:        CPU compute buffer size =    98.25 MiB
llama_new_context_with_model: graph nodes  = 966
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 4 / 8 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | 
examples/main/main.cpp:272: GGML_ASSERT(llama_add_eos_token(model) != 1) failed
./llama.cpp/llama-cli(+0x5f5bb)[0x5ecb6f5815bb]
./llama.cpp/llama-cli(+0x61477)[0x5ecb6f583477]
./llama.cpp/llama-cli(+0x437e9)[0x5ecb6f5657e9]
/lib/x86_64-linux-gnu/libc.so.6(+0x29d90)[0x759a1ea29d90]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x80)[0x759a1ea29e40]
./llama.cpp/llama-cli(+0x43ca5)[0x5ecb6f565ca5]
Aborted (core dumped)

[fairydreaming]

@Aisuko I think the problem is that your model has "add_eos_token": true, in tokenizer_config.json file. Considering the fact that it's a decoder-only model and it should generate EOS token by itself, I think there's no need for this to be true. Perhaps you should simply set it to false and repeat the model conversion process?
Edit: I think you can also set it to false in GGUF file with gguf-py/scripts/gguf_set_metadata.py

[Aisuko]

@Aisuko I think the problem is that your model has "add_eos_token": true, in tokenizer_config.json file. Considering the fact that it's a decoder-only model and it should generate EOS token by itself, I think there's no need for this to be true. Perhaps you should simply set it to false and repeat the model conversion process? Edit: I think you can also set it to false in GGUF file with gguf-py/scripts/gguf_set_metadata.py

COOL COOL, thank you @fairydreaming I will test it later.

Update: It works. And you are right. Thanks.

[github-actions]

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