search

SciSharp / LLamaSharp

A C#/.NET library to run LLM (🦙LLaMA/LLaVA) on your local device efficiently.
https://scisharp.github.io/LLamaSharp
MIT License
2.7k stars 349 forks source link

[BUG]:RUN LLama.Examples =>KernelMemory.cs System.AccessViolationException:“Attempted to read or write protected memory. This is often an indication that other memory is corrupt.” #980

Open freefer opened 1 week ago

freefer commented 1 week ago

Description

model https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/blob/main/qwen2.5-coder-7b-instruct-q5_k_m.gguf Generate using GPU source code Run the source code example LLama KernelMemory. cs for Examples.

Reproduction Steps

[llama 1]: ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no [llama 1]: ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no [llama 1]: ggml_cuda_init: found 1 CUDA devices: [llama 1]: Device 0: NVIDIA GeForce RTX 2080, compute capability 7.5, VMM: yes [llama 1]: llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 2080) - 7113 MiB free [llama 1]: llama_model_loader: loaded meta data with 29 key-value pairs and 339 tensors from I:\LLamaSharp-0.19.0\LLama.Examples\bin\x64\Release\net8.0\Assets\qwen2.5-coder-7b-instruct-q5_k_m.gguf (version GGUF V3 (latest)) [llama 1]: llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. [llama 1]: llama_model_loader: - kv 0: general.architecture str = qwen2 [llama 1]: llama_model_loader: - kv 1: general.type str = model [llama 1]: llama_model_loader: - kv 2: general.name str = Qwen2.5 Coder 7B Instruct GGUF [llama 1]: llama_model_loader: - kv 3: general.finetune str = Instruct-GGUF [llama 1]: llama_model_loader: - kv 4: general.basename str = Qwen2.5-Coder [llama 1]: llama_model_loader: - kv 5: general.size_label str = 7B [llama 1]: llama_model_loader: - kv 6: qwen2.block_count u32 = 28 [llama 1]: llama_model_loader: - kv 7: qwen2.context_length u32 = 131072 [llama 1]: llama_model_loader: - kv 8: qwen2.embedding_length u32 = 3584 [llama 1]: llama_model_loader: - kv 9: qwen2.feed_forward_length u32 = 18944 [llama 1]: llama_model_loader: - kv 10: qwen2.attention.head_count u32 = 28 [llama 1]: llama_model_loader: - kv 11: qwen2.attention.head_count_kv u32 = 4 [llama 1]: llama_model_loader: - kv 12: qwen2.rope.freq_base f32 = 1000000.000000 [llama 1]: llama_model_loader: - kv 13: qwen2.attention.layer_norm_rms_epsilon f32 = 0.000001 [llama 1]: llama_model_loader: - kv 14: general.file_type u32 = 17 [llama 1]: llama_model_loader: - kv 15: tokenizer.ggml.model str = gpt2 [llama 1]: llama_model_loader: - kv 16: tokenizer.ggml.pre str = qwen2 [llama 1]: llama_model_loader: - kv 17: tokenizer.ggml.tokens arr[str,152064] = ["!", "\"", "#", "$", "%", "&", "'", ... [llama 1]: llama_model_loader: - kv 18: tokenizer.ggml.token_type arr[i32,152064] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... [llama 1]: llama_model_loader: - kv 19: tokenizer.ggml.merges arr[str,151387] = ["臓 臓", "臓臓 臓臓", "i n", "臓 t",... [llama 1]: llama_model_loader: - kv 20: tokenizer.ggml.eos_token_id u32 = 151645 [llama 1]: llama_model_loader: - kv 21: tokenizer.ggml.padding_token_id u32 = 151643 [llama 1]: llama_model_loader: - kv 22: tokenizer.ggml.bos_token_id u32 = 151643 [llama 1]: llama_model_loader: - kv 23: tokenizer.ggml.add_bos_token bool = false [llama 1]: llama_model_loader: - kv 24: tokenizer.chat_template str = {%- if tools %}\n {{- '<|im_start|>... [llama 1]: llama_model_loader: - kv 25: general.quantization_version u32 = 2 [llama 1]: llama_model_loader: - kv 26: split.no u16 = 0 [llama 1]: llama_model_loader: - kv 27: split.count u16 = 0 [llama 1]: llama_model_loader: - kv 28: split.tensors.count i32 = 339 [llama 1]: llama_model_loader: - type f32: 141 tensors [llama 1]: llama_model_loader: - type q5_K: 169 tensors [llama 1]: llama_model_loader: - type q6_K: 29 tensors [llama Info]: llm_load_vocab: control token: 151661 '<|fim_suffix|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151649 '<|box_end|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151647 '<|object_ref_end|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151654 '<|vision_pad|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151659 '<|fim_prefix|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151648 '<|box_start|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151644 '<|im_start|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151646 '<|object_ref_start|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151650 '<|quad_start|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151651 '<|quad_end|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151652 '<|vision_start|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151653 '<|vision_end|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151655 '<|image_pad|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151656 '<|video_pad|>' is not marked as EOG [llama Info]: llm_load_vocab: control token: 151660 '<|fim_middle|>' is not marked as EOG [llama 1]: llm_load_vocab: special tokens cache size = 22 [llama 1]: llm_load_vocab: token to piece cache size = 0.9310 MB [llama 1]: llm_load_print_meta: format = GGUF V3 (latest) [llama 1]: llm_load_print_meta: arch = qwen2 [llama 1]: llm_load_print_meta: vocab type = BPE [llama 1]: llm_load_print_meta: n_vocab = 152064 [llama 1]: llm_load_print_meta: n_merges = 151387 [llama 1]: llm_load_print_meta: vocab_only = 0 [llama 1]: llm_load_print_meta: n_ctx_train = 131072 [llama 1]: llm_load_print_meta: n_embd = 3584 [llama 1]: llm_load_print_meta: n_layer = 28 [llama 1]: llm_load_print_meta: n_head = 28 [llama 1]: llm_load_print_meta: n_head_kv = 4 [llama 1]: llm_load_print_meta: n_rot = 128 [llama 1]: llm_load_print_meta: n_swa = 0 [llama 1]: llm_load_print_meta: n_embd_head_k = 128 [llama 1]: llm_load_print_meta: n_embd_head_v = 128 [llama 1]: llm_load_print_meta: n_gqa = 7 [llama 1]: llm_load_print_meta: n_embd_k_gqa = 512 [llama 1]: llm_load_print_meta: n_embd_v_gqa = 512 [llama 1]: llm_load_print_meta: f_norm_eps = 0.0e+00 [llama 1]: llm_load_print_meta: f_norm_rms_eps = 1.0e-06 [llama 1]: llm_load_print_meta: f_clamp_kqv = 0.0e+00 [llama 1]: llm_load_print_meta: f_max_alibi_bias = 0.0e+00 [llama 1]: llm_load_print_meta: f_logit_scale = 0.0e+00 [llama 1]: llm_load_print_meta: n_ff = 18944 [llama 1]: llm_load_print_meta: n_expert = 0 [llama 1]: llm_load_print_meta: n_expert_used = 0 [llama 1]: llm_load_print_meta: causal attn = 1 [llama 1]: llm_load_print_meta: pooling type = 0 [llama 1]: llm_load_print_meta: rope type = 2 [llama 1]: llm_load_print_meta: rope scaling = linear [llama 1]: llm_load_print_meta: freq_base_train = 1000000.0 [llama 1]: llm_load_print_meta: freq_scale_train = 1 [llama 1]: llm_load_print_meta: n_ctx_orig_yarn = 131072 [llama 1]: llm_load_print_meta: rope_finetuned = unknown [llama 1]: llm_load_print_meta: ssm_d_conv = 0 [llama 1]: llm_load_print_meta: ssm_d_inner = 0 [llama 1]: llm_load_print_meta: ssm_d_state = 0 [llama 1]: llm_load_print_meta: ssm_dt_rank = 0 [llama 1]: llm_load_print_meta: ssm_dt_b_c_rms = 0 [llama 1]: llm_load_print_meta: model type = ?B [llama 1]: llm_load_print_meta: model ftype = Q5_K - Medium [llama 1]: llm_load_print_meta: model params = 7.62 B [llama 1]: llm_load_print_meta: model size = 5.07 GiB (5.71 BPW) [llama 1]: llm_load_print_meta: general.name = Qwen2.5 Coder 7B Instruct GGUF [llama 1]: llm_load_print_meta: BOS token = 151643 '<|endoftext|>' [llama 1]: llm_load_print_meta: EOS token = 151645 '<|im_end|>' [llama 1]: llm_load_print_meta: EOT token = 151645 '<|im_end|>' [llama 1]: llm_load_print_meta: PAD token = 151643 '<|endoftext|>' [llama 1]: llm_load_print_meta: LF token = 148848 '脛默' [llama 1]: llm_load_print_meta: FIM PRE token = 151659 '<|fim_prefix|>' [llama 1]: llm_load_print_meta: FIM SUF token = 151661 '<|fim_suffix|>' [llama 1]: llm_load_print_meta: FIM MID token = 151660 '<|fim_middle|>' [llama 1]: llm_load_print_meta: FIM PAD token = 151662 '<|fim_pad|>' [llama 1]: llm_load_print_meta: FIM REP token = 151663 '<|repo_name|>' [llama 1]: llm_load_print_meta: FIM SEP token = 151664 '<|file_sep|>' [llama 1]: llm_load_print_meta: EOG token = 151643 '<|endoftext|>' [llama 1]: llm_load_print_meta: EOG token = 151645 '<|im_end|>' [llama 1]: llm_load_print_meta: EOG token = 151662 '<|fim_pad|>' [llama 1]: llm_load_print_meta: EOG token = 151663 '<|repo_name|>' [llama 1]: llm_load_print_meta: EOG token = 151664 '<|file_sep|>' [llama 1]: llm_load_print_meta: max token length = 256 [llama 1]: llm_load_tensors: ggml ctx size = 0.30 MiB [llama 1]: llm_load_tensors: offloading 20 repeating layers to GPU [llama 1]: llm_load_tensors: offloaded 20/29 layers to GPU [llama 1]: llm_load_tensors: CPU buffer size = 5186.92 MiB [llama 1]: llm_load_tensors: CUDA0 buffer size = 3136.56 MiB

[llama 1]: llama_new_context_with_model: n_ctx = 2048 [llama 1]: llama_new_context_with_model: n_batch = 512 [llama 1]: llama_new_context_with_model: n_ubatch = 512 [llama 1]: llama_new_context_with_model: flash_attn = 0 [llama 1]: llama_new_context_with_model: freq_base = 1000000.0 [llama 1]: llama_new_context_with_model: freq_scale = 1 [llama 1]: llama_kv_cache_init: CUDA_Host KV buffer size = 32.00 MiB [llama 1]: llama_kv_cache_init: CUDA0 KV buffer size = 80.00 MiB [llama 1]: llama_new_context_with_model: KV self size = 112.00 MiB, K (f16): 56.00 MiB, V (f16): 56.00 MiB [llama 1]: llama_new_context_with_model: CUDA_Host output buffer size = 0.01 MiB [llama 1]: llama_new_context_with_model: CUDA0 compute buffer size = 730.36 MiB [llama 1]: llama_new_context_with_model: CUDA_Host compute buffer size = 11.01 MiB [llama 1]: llama_new_context_with_model: graph nodes = 986 [llama 1]: llama_new_context_with_model: graph splits = 116 [llama 1]: llama_new_context_with_model: n_ctx = 2048 [llama 1]: llama_new_context_with_model: n_batch = 512 [llama 1]: llama_new_context_with_model: n_ubatch = 512 [llama 1]: llama_new_context_with_model: flash_attn = 0 [llama 1]: llama_new_context_with_model: freq_base = 1000000.0 [llama 1]: llama_new_context_with_model: freq_scale = 1 [llama 1]: llama_kv_cache_init: CUDA_Host KV buffer size = 32.00 MiB [llama 1]: llama_kv_cache_init: CUDA0 KV buffer size = 80.00 MiB [llama 1]: llama_new_context_with_model: KV self size = 112.00 MiB, K (f16): 56.00 MiB, V (f16): 56.00 MiB [llama 1]: llama_new_context_with_model: CUDA_Host output buffer size = 0.01 MiB [llama 1]: llama_new_context_with_model: CUDA0 compute buffer size = 730.36 MiB [llama 1]: llama_new_context_with_model: CUDA_Host compute buffer size = 11.01 MiB [llama 1]: llama_new_context_with_model: graph nodes = 986 [llama 1]: llama_new_context_with_model: graph splits = 116 [llama 1]: llama_new_context_with_model: n_ctx = 2048 [llama 1]: llama_new_context_with_model: n_batch = 512 [llama 1]: llama_new_context_with_model: n_ubatch = 512 [llama 1]: llama_new_context_with_model: flash_attn = 0 [llama 1]: llama_new_context_with_model: freq_base = 1000000.0 [llama 1]: llama_new_context_with_model: freq_scale = 1 [llama 1]: llama_kv_cache_init: CUDA_Host KV buffer size = 32.00 MiB [llama 1]: llama_kv_cache_init: CUDA0 KV buffer size = 80.00 MiB [llama 1]: llama_new_context_with_model: KV self size = 112.00 MiB, K (f16): 56.00 MiB, V (f16): 56.00 MiB [llama 1]: llama_new_context_with_model: CPU output buffer size = 0.00 MiB [llama 1]: llama_new_context_with_model: CUDA0 compute buffer size = 731.36 MiB [llama 1]: llama_new_context_with_model: CUDA_Host compute buffer size = 12.01 MiB [llama 1]: llama_new_context_with_model: graph nodes = 989 [llama 1]: llama_new_context_with_model: graph splits = 116 Importing 1 of 2: I:\LLamaSharp-0.19.0\LLama.Examples\bin\x64\Release\net8.0\Assets\sample-SK-Readme.pdf Completed in 00:00:04.0542873

Importing 2 of 2: I:\LLamaSharp-0.19.0\LLama.Examples\bin\x64\Release\net8.0\Assets\sample-KM-Readme.pdf Completed in 00:00:01.8282991

Question: What formats does KM support Generating answer... [llama 1]: llama_new_context_with_model: n_ctx = 2048 [llama 1]: llama_new_context_with_model: n_batch = 512 [llama 1]: llama_new_context_with_model: n_ubatch = 512 [llama 1]: llama_new_context_with_model: flash_attn = 0 [llama 1]: llama_new_context_with_model: freq_base = 1000000.0 [llama 1]: llama_new_context_with_model: freq_scale = 1 [llama 1]: llama_kv_cache_init: CUDA_Host KV buffer size = 32.00 MiB [llama 1]: llama_kv_cache_init: CUDA0 KV buffer size = 80.00 MiB [llama 1]: llama_new_context_with_model: KV self size = 112.00 MiB, K (f16): 56.00 MiB, V (f16): 56.00 MiB [llama 1]: llama_new_context_with_model: CUDA_Host output buffer size = 0.01 MiB [llama 1]: llama_new_context_with_model: CUDA0 compute buffer size = 730.36 MiB [llama 1]: llama_new_context_with_model: CUDA_Host compute buffer size = 11.01 MiB [llama 1]: llama_new_context_with_model: graph nodes = 986 [llama 1]: llama_new_context_with_model: graph splits = 116 [llama Warning]: llama_get_logits_ith: invalid logits id 343, reason: no logits Fatal error. System.AccessViolationException: Attempted to read or write protected memory. This is often an indication that other memory is corrupt. Repeat 2 times:

at LLama.Native.SafeLLamaSamplerChainHandle.g__llama_sampler_sample|4_0(LLama.Native.SafeLLamaSamplerChainHandle, LLama.Native.SafeLLamaContextHandle, Int32)

at LLama.Native.SafeLLamaSamplerChainHandle.Sample(LLama.Native.SafeLLamaContextHandle, Int32) at LLama.Sampling.BaseSamplingPipeline.Sample(LLama.Native.SafeLLamaContextHandle, Int32) at LLama.StatelessExecutor+d18.MoveNext() at System.Threading.ExecutionContext.RunInternal(System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object) at System.Runtime.CompilerServices.AsyncTaskMethodBuilder1+AsyncStateMachineBox1[[System.Threading.Tasks.VoidTaskResult, System.Private.CoreLib, Version=8.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e],[System.Canon, System.Private.CoreLib, Version=8.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].MoveNext(System.Threading.Thread) at System.Runtime.CompilerServices.TaskAwaiter+<>c.b__12_0(System.Action, System.Threading.Tasks.Task) at System.Threading.Tasks.AwaitTaskContinuation.RunOrScheduleAction(System.Action, Boolean) at System.Threading.Tasks.Task.RunContinuations(System.Object) at System.Threading.Tasks.Task.FinishSlow(Boolean) at System.Threading.Tasks.Task.ExecuteWithThreadLocal(System.Threading.Tasks.Task ByRef, System.Threading.Thread) at System.Threading.ThreadPoolWorkQueue.Dispatch() at System.Threading.PortableThreadPool+WorkerThread.WorkerThreadStart()

Environment & Configuration

windows 10 x64 LLamaSharp-0.19.0

Known Workarounds

No response

freefer commented 3 days ago

image Setting it to false seems to work properly