Eval bug: Segmentation fault (core dumped) for Gemma 2 series #17426
Description
Name and Version
llama-cli --version
version: 7048 (0cfb191)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0 for x86_64-linux-gnu
Operating systems
Linux
GGML backends
CUDA
Hardware
RTX 3060 12GB, CUDA Version: 12.5
Models
gemma-2-2b-it-Q8_0.gguf
gemma-2-9b-it-Q6_K.gguf
Problem description & steps to reproduce
After a recent update, when loading Gemma 2 series models (2b, 9b), I get the error:
Segmentation fault (core dumped).
The same models worked on the previous build (downloaded a few weeks ago, can't remember the exact version).
It's definitely not a memory issue (at least for the 2b model, that's not a problem); it's some kind of bug. Again, the same models worked good with the same settings before, and there was still plenty of memory left.
I tried setting the context to --ctx-size 4096, result was the same. Only thing that helped was - reducing the number of layers on the GPU by -1, for example:
previously, it ran like this (all layers on the GPU):
llama-cli -ngl 27 --ctx-size 4096 -m gemma2_2b_it_q8.gguf
result: Segmentation fault (core dumped)
now:
llama-cli -ngl 26 --ctx-size 4096 -m gemma2_2b_it_q8.gguf
result: ok.
Or the 9b model:
llama-cli -ngl 43 --ctx-size 4096 -m gemma2_9b_it_q6k.gguf
result: Segmentation fault (core dumped)
now:
llama-cli -ngl 42 --ctx-size 4096 -m gemma2_9b_it_q6k.gguf
result: ok.
I haven't had such issues with other models yet, including newer and larger ones, such as gemma3_12b_it_q8.gguf, its only with gemma2 models, for example:
https://huggingface.co/bartowski/gemma-2-2b-it-GGUF/resolve/main/gemma-2-2b-it-Q8_0.gguf
https://huggingface.co/bartowski/gemma-2-9b-it-GGUF/resolve/main/gemma-2-9b-it-Q6_K.gguf
First Bad Commit
No response
Relevant log output
llama-cli -ngl 27 --ctx-size 4096 -m gemma2_2b_it_q8.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
build: 7048 (0cfb19166) with cc (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) (0000:04:00.0) - 11808 MiB free
llama_model_loader: loaded meta data with 32 key-value pairs and 288 tensors from gemma2_2b_it_q8.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 = gemma2
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Models
llama_model_loader: - kv 3: general.size_label str = 2.6B
llama_model_loader: - kv 4: general.license str = gemma
llama_model_loader: - kv 5: general.tags arr[str,1] = ["text-generation"]
llama_model_loader: - kv 6: gemma2.context_length u32 = 8192
llama_model_loader: - kv 7: gemma2.embedding_length u32 = 2304
llama_model_loader: - kv 8: gemma2.block_count u32 = 26
llama_model_loader: - kv 9: gemma2.feed_forward_length u32 = 9216
llama_model_loader: - kv 10: gemma2.attention.head_count u32 = 8
llama_model_loader: - kv 11: gemma2.attention.head_count_kv u32 = 4
llama_model_loader: - kv 12: gemma2.attention.layer_norm_rms_epsilon f32 = 0,000001
llama_model_loader: - kv 13: gemma2.attention.key_length u32 = 256
llama_model_loader: - kv 14: gemma2.attention.value_length u32 = 256
llama_model_loader: - kv 15: general.file_type u32 = 7
llama_model_loader: - kv 16: gemma2.attn_logit_softcapping f32 = 50,000000
llama_model_loader: - kv 17: gemma2.final_logit_softcapping f32 = 30,000000
llama_model_loader: - kv 18: gemma2.attention.sliding_window u32 = 4096
llama_model_loader: - kv 19: tokenizer.ggml.model str = llama
llama_model_loader: - kv 20: tokenizer.ggml.pre str = default
llama_model_loader: - kv 21: tokenizer.ggml.tokens arr[str,256000] = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv 22: tokenizer.ggml.scores arr[f32,256000] = [-1000,000000, -1000,000000, -1000,00...
llama_model_loader: - kv 23: tokenizer.ggml.token_type arr[i32,256000] = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv 24: tokenizer.ggml.bos_token_id u32 = 2
llama_model_loader: - kv 25: tokenizer.ggml.eos_token_id u32 = 1
llama_model_loader: - kv 26: tokenizer.ggml.unknown_token_id u32 = 3
llama_model_loader: - kv 27: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 28: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 29: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 30: tokenizer.ggml.add_space_prefix bool = false
llama_model_loader: - kv 31: general.quantization_version u32 = 2
llama_model_loader: - type f32: 105 tensors
llama_model_loader: - type q8_0: 183 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q8_0
print_info: file size = 2,59 GiB (8,50 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: printing all EOG tokens:
load: - 1 ('<eos>')
load: - 107 ('<end_of_turn>')
load: special tokens cache size = 249
load: token to piece cache size = 1,6014 MB
print_info: arch = gemma2
print_info: vocab_only = 0
print_info: n_ctx_train = 8192
print_info: n_embd = 2304
print_info: n_embd_inp = 2304
print_info: n_layer = 26
print_info: n_head = 8
print_info: n_head_kv = 4
print_info: n_rot = 256
print_info: n_swa = 4096
print_info: is_swa_any = 1
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
print_info: n_gqa = 2
print_info: n_embd_k_gqa = 1024
print_info: n_embd_v_gqa = 1024
print_info: f_norm_eps = 0,0e+00
print_info: f_norm_rms_eps = 1,0e-06
print_info: f_clamp_kqv = 0,0e+00
print_info: f_max_alibi_bias = 0,0e+00
print_info: f_logit_scale = 0,0e+00
print_info: f_attn_scale = 6,2e-02
print_info: n_ff = 9216
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: n_expert_groups = 0
print_info: n_group_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 10000,0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 8192
print_info: rope_finetuned = unknown
print_info: model type = 2B
print_info: model params = 2,61 B
print_info: general.name = Models
print_info: vocab type = SPM
print_info: n_vocab = 256000
print_info: n_merges = 0
print_info: BOS token = 2 '<bos>'
print_info: EOS token = 1 '<eos>'
print_info: EOT token = 107 '<end_of_turn>'
print_info: UNK token = 3 '<unk>'
print_info: PAD token = 0 '<pad>'
print_info: LF token = 227 '<0x0A>'
print_info: EOG token = 1 '<eos>'
print_info: EOG token = 107 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 26 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 27/27 layers to GPU
load_tensors: CPU_Mapped model buffer size = 597,66 MiB
load_tensors: CUDA0 model buffer size = 2649,78 MiB
..................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = false
llama_context: freq_base = 10000,0
llama_context: freq_scale = 1
llama_context: n_ctx_seq (4096) < n_ctx_train (8192) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 0,98 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells
llama_kv_cache: CUDA0 KV buffer size = 208,00 MiB
llama_kv_cache: size = 208,00 MiB ( 4096 cells, 13 layers, 1/1 seqs), K (f16): 104,00 MiB, V (f16): 104,00 MiB
llama_kv_cache_iswa: creating SWA KV cache, size = 4096 cells
llama_kv_cache: CUDA0 KV buffer size = 208,00 MiB
llama_kv_cache: size = 208,00 MiB ( 4096 cells, 13 layers, 1/1 seqs), K (f16): 104,00 MiB, V (f16): 104,00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context: CUDA0 compute buffer size = 504,50 MiB
llama_context: CUDA_Host compute buffer size = 20,52 MiB
llama_context: graph nodes = 948
llama_context: graph splits = 2
common_init_from_params: added <eos> logit bias = -inf
common_init_from_params: added <end_of_turn> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
Segmentation fault (core dumped)