A special token '\u0000' will cause an assert error in 'llm_load_vocab'

[SolenoidWGT]

Discussed in https://github.com/ggerganov/llama.cpp/discussions/5111

^{Originally posted by **SolenoidWGT** January 24, 2024} I'm trying to fit an [InternLM2](https://github.com/InternLM/InternLM) model for llama.cpp, but I get an assertion error when using llama.cpp for inference, below is the error stack. The commit ID of llama.cpp code is 77bc1bbd05f0c31cb45773eb5eb59b9ff2b07e1b ``` $ ./main -m ./internlm2-base-7b/ggml-model-f16.gguf -n 400 -e -p "Building a website can be done in 10 simple steps:\nStep 1:" Log start main: build = 1930 (f8ca46e0) main: built with gcc (GCC) 10.2.0 for x86_64-pc-linux-gnu main: seed = 1706096206 llama_model_loader: loaded meta data with 17 key-value pairs and 227 tensors from ./internlm2-base-7b/ggml-model-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 = internlm2 llama_model_loader: - kv 1: general.name str = InternLM llama_model_loader: - kv 2: internlm2.context_length u32 = 32768 llama_model_loader: - kv 3: internlm2.block_count u32 = 32 llama_model_loader: - kv 4: internlm2.embedding_length u32 = 4096 llama_model_loader: - kv 5: internlm2.feed_forward_length u32 = 14336 llama_model_loader: - kv 6: internlm2.rope.freq_base f32 = 1000000.000000 llama_model_loader: - kv 7: internlm2.attention.head_count u32 = 32 llama_model_loader: - kv 8: internlm2.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 9: internlm2.attention.head_count_kv u32 = 8 llama_model_loader: - kv 10: tokenizer.ggml.model str = llama llama_model_loader: - kv 11: tokenizer.ggml.tokens arr[str,92544] = ["", "", "", "<0x00>", "<... llama_model_loader: - kv 12: tokenizer.ggml.scores arr[f32,92544] = [0.000000, 0.000000, 0.000000, 0.0000... llama_model_loader: - kv 13: tokenizer.ggml.token_type arr[i32,92544] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ... llama_model_loader: - kv 14: tokenizer.ggml.bos_token_id u32 = 1 llama_model_loader: - kv 15: tokenizer.ggml.eos_token_id u32 = 2 llama_model_loader: - kv 16: tokenizer.ggml.padding_token_id u32 = 2 llama_model_loader: - type f32: 65 tensors llama_model_loader: - type f16: 162 tensors GGML_ASSERT: llama.cpp:3074: codepoints_from_utf8(word).size() > 0 No symbol table is loaded. Use the "file" command. [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib64/libthread_db.so.1". 0x00007fadd375c12c in waitpid () from /lib64/libpthread.so.0 No symbol "frame" in current context. ``` I further checked and found that the token that caused the error was token `\u0000` in the InternLM2 vocabulary, which would be converted into string "`\u0000`" by `codepoints_from_utf8`, which corresponds to the string terminator in C language, resulting in `word ` size is 0, causing the assertion [here](https://github.com/ggerganov/llama.cpp/blob/c9b316c78fba31e65879a2ec91cbafd341b88cce/llama.cpp#L3053) to report an error (because I added some debug code, the actual error line number is llama.cpp:3053)   I tried to comment out the assertion at llama.cpp:3053, and the model could be generated normally without any other errors. So I would like to ask about the significance of this assertion. Can we relax the assertion conditions here? If I can't remove the assertion, I'd love some advice on how to get around it, thanks. I searched and found a [dscussion](https://github.com/ggerganov/llama.cpp/discussions/3498#discussioncomment-7226038) similar to my problem, but I didn't get much information. Here is my sys & env info. ``` Collecting environment information... PyTorch version: 1.13.1+cu117 Is debug build: False CUDA used to build PyTorch: 11.7 ROCM used to build PyTorch: N/A OS: CentOS Linux 7 (Core) (x86_64) GCC version: (conda-forge gcc 13.1.0-0) 13.1.0 Clang version: Could not collect CMake version: version 3.26.4 Libc version: glibc-2.17 Python version: 3.10.0 (default, Mar 3 2022, 09:58:08) [GCC 7.5.0] (64-bit runtime) Python platform: Linux-3.10.0-957.el7.x86_64-x86_64-with-glibc2.17 Is CUDA available: False CUDA runtime version: No CUDA CUDA_MODULE_LOADING set to: N/A GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 256 On-line CPU(s) list: 0-255 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 2 NUMA node(s): 8 Vendor ID: AuthenticAMD CPU family: 23 Model: 49 Model name: AMD EPYC 7H12 64-Core Processor Stepping: 0 CPU MHz: 2600.000 CPU max MHz: 2600.0000 CPU min MHz: 1500.0000 BogoMIPS: 5199.78 Virtualization: AMD-V L1d cache: 32K L1i cache: 32K L2 cache: 512K L3 cache: 16384K NUMA node0 CPU(s): 0-15,128-143 NUMA node1 CPU(s): 16-31,144-159 NUMA node2 CPU(s): 32-47,160-175 NUMA node3 CPU(s): 48-63,176-191 NUMA node4 CPU(s): 64-79,192-207 NUMA node5 CPU(s): 80-95,208-223 NUMA node6 CPU(s): 96-111,224-239 NUMA node7 CPU(s): 112-127,240-255 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc art rep_good nopl xtopology nonstop_tsc extd_apicid aperfmperf eagerfpu pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_l2 cpb cat_l3 cdp_l3 hw_pstate sme retpoline_amd ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif umip overflow_recov succor smca ```

[dranger003]

@SolenoidWGT I also ran into this issue and am wondering if you found a solution? I also confirm commenting out the assert works fine. I wonder if we could remove that token from the vocab somehow instead?

EDIT: I just found this https://github.com/ggerganov/llama.cpp/issues/4360#issuecomment-1906861962

[SolenoidWGT]

@dranger003 Sorry for the delayed response due to the holidays. As mentioned in the issue , I have replaced \u0000 with a special character ‘🐉’ to avoid this problem. Since \u0000 does not typically appear in regular chat scenarios, its impact on the chat-model is limited. However, I'm not sure if this solution may have other side effects in more general scenarios.

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity.

[github-actions[bot]]

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