Bug: CUDA error: peer access has not been enabled

[AbdullahMPrograms]

What happened?

Hey all, I've recently (last few days) been running into a weird CUDA issue where I can only generate a single time before llama.cpp will unexplainably crash. I've also noticed that this issue only seems to happen with split mode row and that split mode row equally distributes both model weights and kv cache across all GPU's, while previously it would load the kv cache on GPU1, not sure if this newer functionality is intended.

Name and Version

ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 3 CUDA devices:
Device 0: Tesla P40, compute capability 6.1, VMM: yes
Device 1: Tesla P40, compute capability 6.1, VMM: yes
Device 2: Tesla P40, compute capability 6.1, VMM: yes
version: 4017 (9830b69)
built with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu

What operating system are you seeing the problem on?

Linux

Relevant log output

./LLM/llama.cpp/llama-server \
-m /home/ultimis/LLM/Models/mradermacher/Qwen2.5-32B-Instruct-i1-GGUF/Qwen2.5-32B-Instruct.i1-Q4_K_M.gguf \
-c 32768 -ngl 99 --split-mode row --flash-attn --host 0.0.0.0 --port 8085
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 3 CUDA devices:
  Device 0: Tesla P40, compute capability 6.1, VMM: yes
  Device 1: Tesla P40, compute capability 6.1, VMM: yes
  Device 2: Tesla P40, compute capability 6.1, VMM: yes
build: 4020 (9f409893) with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu
system info: n_threads = 16, n_threads_batch = 16, total_threads = 32

system_info: n_threads = 16 (n_threads_batch = 16) / 32 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | AMX_INT8 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |

main: HTTP server is listening, hostname: 0.0.0.0, port: 8085, http threads: 31
main: loading model
llama_load_model_from_file: using device CUDA0 (Tesla P40) - 24286 MiB free
llama_load_model_from_file: using device CUDA1 (Tesla P40) - 24290 MiB free
llama_load_model_from_file: using device CUDA2 (Tesla P40) - 24290 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 771 tensors from /home/ultimis/LLM/Models/mradermacher/Qwen2.5-32B-Instruct-i1-GGUF/Qwen2.5-32B-Instruct.i1-Q4_K_M.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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 32B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 32B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-3...
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = Qwen2.5 32B
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-32B
llama_model_loader: - kv  12:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                          qwen2.block_count u32              = 64
llama_model_loader: - kv  15:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  16:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv  17:                  qwen2.feed_forward_length u32              = 27648
llama_model_loader: - kv  18:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  19:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  20:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                          general.file_type u32              = 15
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  26:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  27:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  30:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  31:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  32:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  33:               general.quantization_version u32              = 2
llama_model_loader: - kv  34:                                general.url str              = https://huggingface.co/mradermacher/Q...
llama_model_loader: - kv  35:              mradermacher.quantize_version str              = 2
llama_model_loader: - kv  36:                  mradermacher.quantized_by str              = mradermacher
llama_model_loader: - kv  37:                  mradermacher.quantized_at str              = 2024-09-20T15:12:17+02:00
llama_model_loader: - kv  38:                  mradermacher.quantized_on str              = db1
llama_model_loader: - kv  39:                         general.source.url str              = https://huggingface.co/Qwen/Qwen2.5-3...
llama_model_loader: - kv  40:                  mradermacher.convert_type str              = hf
llama_model_loader: - kv  41:                      quantize.imatrix.file str              = Qwen2.5-32B-Instruct-i1-GGUF/imatrix.dat
llama_model_loader: - kv  42:                   quantize.imatrix.dataset str              = imatrix-training-full-3
llama_model_loader: - kv  43:             quantize.imatrix.entries_count i32              = 448
llama_model_loader: - kv  44:              quantize.imatrix.chunks_count i32              = 318
llama_model_loader: - type  f32:  321 tensors
llama_model_loader: - type q4_K:  385 tensors
llama_model_loader: - type q6_K:   65 tensors
llm_load_vocab: special tokens cache size = 22
llm_load_vocab: token to piece cache size = 0.9310 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 152064
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 5120
llm_load_print_meta: n_layer          = 64
llm_load_print_meta: n_head           = 40
llm_load_print_meta: n_head_kv        = 8
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            = 5
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
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             = 27648
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        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 32.76 B
llm_load_print_meta: model size       = 18.48 GiB (4.85 BPW)
llm_load_print_meta: general.name     = Qwen2.5 32B Instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: FIM PRE token    = 151659 '<|fim_prefix|>'
llm_load_print_meta: FIM SUF token    = 151661 '<|fim_suffix|>'
llm_load_print_meta: FIM MID token    = 151660 '<|fim_middle|>'
llm_load_print_meta: FIM PAD token    = 151662 '<|fim_pad|>'
llm_load_print_meta: FIM REP token    = 151663 '<|repo_name|>'
llm_load_print_meta: FIM SEP token    = 151664 '<|file_sep|>'
llm_load_print_meta: EOG token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOG token        = 151645 '<|im_end|>'
llm_load_print_meta: EOG token        = 151662 '<|fim_pad|>'
llm_load_print_meta: EOG token        = 151663 '<|repo_name|>'
llm_load_print_meta: EOG token        = 151664 '<|file_sep|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: offloading 64 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 65/65 layers to GPU
llm_load_tensors: CPU_Mapped model buffer size =   417.66 MiB
llm_load_tensors:      CUDA0 model buffer size =     1.46 MiB
llm_load_tensors:      CUDA1 model buffer size =     1.46 MiB
llm_load_tensors:      CUDA2 model buffer size =     1.35 MiB
llm_load_tensors: CUDA0_Split model buffer size =  6187.50 MiB
llm_load_tensors: CUDA1_Split model buffer size =  6043.12 MiB
llm_load_tensors: CUDA2_Split model buffer size =  6273.46 MiB
................................................................................................
llama_new_context_with_model: n_seq_max     = 2
llama_new_context_with_model: n_ctx         = 32768
llama_new_context_with_model: n_ctx_per_seq = 16384
llama_new_context_with_model: n_batch       = 2048
llama_new_context_with_model: n_ubatch      = 512
llama_new_context_with_model: flash_attn    = 1
llama_new_context_with_model: freq_base     = 1000000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (16384) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_kv_cache_init:      CUDA0 KV buffer size =  2816.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =  2816.00 MiB
llama_kv_cache_init:      CUDA2 KV buffer size =  2560.00 MiB
llama_new_context_with_model: KV self size  = 8192.00 MiB, K (f16): 4096.00 MiB, V (f16): 4096.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     1.16 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   190.00 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   170.00 MiB
llama_new_context_with_model:      CUDA2 compute buffer size =   307.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    74.01 MiB
llama_new_context_with_model: graph nodes  = 1991
llama_new_context_with_model: graph splits = 4
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 32768
main: model loaded
main: chat template, built_in: 1, chat_example: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://0.0.0.0:8085 - starting the main loop
srv  update_slots: all slots are idle
request: GET /v1/models 192.168.2.245 200
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, n_prompt_tokens = 51
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 51, n_tokens = 51, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 51, n_tokens = 51
slot      release: id  0 | task 0 | stop processing: n_past = 610, truncated = 0
slot print_timing: id  0 | task 0 |
prompt eval time =     524.02 ms /    51 tokens (   10.27 ms per token,    97.32 tokens per second)
       eval time =   34103.49 ms /   560 tokens (   60.90 ms per token,    16.42 tokens per second)
      total time =   34627.51 ms /   611 tokens
srv  update_slots: all slots are idle
request: POST /v1/chat/completions 192.168.2.245 200
slot launch_slot_: id  0 | task 561 | processing task
slot update_slots: id  0 | task 561 | new prompt, n_ctx_slot = 32768, n_keep = 0, n_prompt_tokens = 717
slot update_slots: id  0 | task 561 | kv cache rm [0, end)
slot update_slots: id  0 | task 561 | prompt processing progress, n_past = 717, n_tokens = 717, progress = 1.000000
slot update_slots: id  0 | task 561 | prompt done, n_past = 717, n_tokens = 717
slot      release: id  0 | task 561 | stop processing: n_past = 721, truncated = 0
slot print_timing: id  0 | task 561 |
prompt eval time =    3056.37 ms /   717 tokens (    4.26 ms per token,   234.59 tokens per second)
       eval time =     246.26 ms /     5 tokens (   49.25 ms per token,    20.30 tokens per second)
      total time =    3302.63 ms /   722 tokens
srv  update_slots: all slots are idle
request: POST /v1/chat/completions 192.168.2.245 200
slot launch_slot_: id  0 | task 567 | processing task
slot update_slots: id  0 | task 567 | new prompt, n_ctx_slot = 32768, n_keep = 0, n_prompt_tokens = 777
slot update_slots: id  0 | task 567 | kv cache rm [0, end)
slot update_slots: id  0 | task 567 | prompt processing progress, n_past = 777, n_tokens = 777, progress = 1.000000
slot update_slots: id  0 | task 567 | prompt done, n_past = 777, n_tokens = 777
ggml/src/ggml-cuda.cu:70: CUDA error
CUDA error: peer access has not been enabled
  current device: 0, in function ggml_cuda_op_mul_mat at ggml/src/ggml-cuda.cu:1498
  cudaGetLastError()
Could not attach to process.  If your uid matches the uid of the target
process, check the setting of /proc/sys/kernel/yama/ptrace_scope, or try
again as the root user.  For more details, see /etc/sysctl.d/10-ptrace.conf
ptrace: Operation not permitted.
No stack.
The program is not being run.
Aborted (core dumped)

[slaren]

Can you test if this change fixes the issue?

diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index b57f1b3b..e68e4055 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -1297,11 +1297,17 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
                     cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
                     if (err != cudaErrorPeerAccessAlreadyEnabled) {
                         CUDA_CHECK(err);
+                    } else {
+                        // reset the error
+                        cudaGetLastError();
                     }
                 } else {
                     cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
                     if (err != cudaErrorPeerAccessNotEnabled) {
                         CUDA_CHECK(err);
+                    } else {
+                        // reset the error
+                        cudaGetLastError();
                     }
                 }
             }

[AbdullahMPrograms]

looks to be working as of 3 messages in, I will report back if it pops up again. Thank you!