search

SciSharp / LLamaSharp

A C#/.NET library to run LLM (🦙LLaMA/LLaVA) on your local device efficiently.
https://scisharp.github.io/LLamaSharp
MIT License
2.72k stars 352 forks source link

[BUG]: Vulkan backend crash on model loading #887

Closed LSXAxeller closed 4 months ago

LSXAxeller commented 4 months ago

Description

after updating to v0.14.0 and releasing Vulkan backend, I decided to give it a try instead using CPU inference, but on loading model it crash with console output 

WARNING: [Loader Message] Code 0 : windows_read_data_files_in_registry: Registry lookup failed to get layer manifest files.
WARNING: [Loader Message] Code 0 : Layer VK_LAYER_RENDERDOC_Capture uses API version 1.2 which is older than the application specified API version of 1.3. May cause issues.
llama_model_loader: loaded meta data with 25 key-value pairs and 327 tensors from C:\Models\Text\Index-1.9B-Character\Index-1.9B-Character-Q6_K.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.name str              = Index-1.9B-Character_test
llama_model_loader: - kv   2:                          llama.block_count u32              = 36
llama_model_loader: - kv   3:                       llama.context_length u32              = 4096
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 5888
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 16
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 16
llama_model_loader: - kv   8:     llama.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv   9:                          general.file_type u32              = 18
llama_model_loader: - kv  10:                           llama.vocab_size u32              = 65029
llama_model_loader: - kv  11:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  12:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,65029]   = ["<unk>", "<s>", "</s>", "reserved_0"...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,65029]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,65029]   = [2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if messages[0]['role'] == 'system'...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   73 tensors
llama_model_loader: - type q6_K:  254 tensors
llm_load_vocab: special tokens cache size = 515
llm_load_vocab: token to piece cache size = 0.3670 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          = 65029
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 2048
llm_load_print_meta: n_layer          = 36
llm_load_print_meta: n_head           = 16
llm_load_print_meta: n_head_kv        = 16
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            = 1
llm_load_print_meta: n_embd_k_gqa     = 2048
llm_load_print_meta: n_embd_v_gqa     = 2048
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             = 5888
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  = 4096
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       = 8B
llm_load_print_meta: model ftype      = Q6_K
llm_load_print_meta: model params     = 2.17 B
llm_load_print_meta: model size       = 1.66 GiB (6.56 BPW) 
llm_load_print_meta: general.name     = Index-1.9B-Character_test
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         = 270 '<0x0A>'
llm_load_print_meta: max token length = 48
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Radeon RX 580 Series (AMD proprietary driver) | uma: 0 | fp16: 0 | warp size: 64
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.SafeLlamaModelHandle.llama_load_model_from_file(System.String, LLama.Native.LLamaModelParams)
--------------------------------
at LLama.Native.SafeLlamaModelHandle.LoadFromFile(System.String, LLama.Native.LLamaModelParams)
at LLama.LlamaWeights+<>c__DisplayClass21_0.<LoadFromFileAsync>b__1()
at System.Threading.Tasks.Task`1[[System.__Canon, System.Private.CoreLib, Version=8.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].InnerInvoke()
at System.Threading.ExecutionContext.RunFromThreadPoolDispatchLoop(System.Threading.Thread, System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object)
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()

Reproduction Steps

            if (!NativeLibraryConfig.LLama.LibraryHasLoaded)
                NativeLibraryConfig.All
                    .WithCuda(AppConfig.Instance.Device == Device.CUDA)
                    .WithVulkan(AppConfig.Instance.Device == Device.VULKAN)
                    .WithAutoFallback(); // AppConfig.Instance.Device is set to Device.VULKAN

            var modelPath = "MODEL_PATH_ON_PC";
            ModelParameters = new ModelParams(modelPath)
            {
                ContextSize = 4096,
                Embeddings = false,
                GpuLayerCount = 16,
            };
            Model = await LLamaWeights.LoadFromFileAsync(ModelParameters);

Environment & Configuration

Known Workarounds

None

m0nsky commented 4 months ago

"Attempted to read or write protected memory" usually means you are running out of VRAM.

LSXAxeller commented 4 months ago

"Attempted to read or write protected memory" usually means you are running out of VRAM.

  • Is it a 4GB or 8GB model RX 580?
  • What VRAM utilization are you seeing right before the crash?
  • Does the crash also happen if you change GpuLayerCount from 16 to 1?

I tried with different models, 1.9B, 1.1B, 300M, 22M

SeriousOldMan commented 4 months ago

886 is the same error. I think one issue can be closed.