ggml_cuda_init: found 8 CUDA devices (Total VRAM: 649230 MiB):
Device 0: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 1: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 2: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 3: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 4: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 5: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 6: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
Device 7: NVIDIA A100-SXM4-80GB, compute capability 8.0, VMM: yes, VRAM: 81153 MiB
build_info: b0-unknown
system_info: n_threads = 72 (n_threads_batch = 72) / 144 | CUDA : ARCHS = 800 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 143 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '/data/DeepSeek-V4-Flash-Q8_0-00001-of-00007.gguf'
llama_model_loader: additional 6 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 61 key-value pairs and 1328 tensors from /data/DeepSeek-V4-Flash-Q8_0-00001-of-00007.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 = deepseek4
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.sampling.top_p f32 = 1.000000
llama_model_loader: - kv 3: general.sampling.temp f32 = 1.000000
llama_model_loader: - kv 4: general.name str = DeepSeek V4 Flash
llama_model_loader: - kv 5: general.size_label str = 256x8.4B
llama_model_loader: - kv 6: general.license str = mit
llama_model_loader: - kv 7: deepseek4.block_count u32 = 43
llama_model_loader: - kv 8: deepseek4.context_length u32 = 1048576
llama_model_loader: - kv 9: deepseek4.embedding_length u32 = 4096
llama_model_loader: - kv 10: deepseek4.attention.head_count u32 = 64
llama_model_loader: - kv 11: deepseek4.attention.head_count_kv u32 = 1
llama_model_loader: - kv 12: deepseek4.rope.scaling.type str = yarn
llama_model_loader: - kv 13: deepseek4.rope.scaling.factor f32 = 16.000000
llama_model_loader: - kv 14: deepseek4.rope.scaling.original_context_length u32 = 65536
llama_model_loader: - kv 15: deepseek4.rope.scaling.yarn_beta_fast f32 = 32.000000
llama_model_loader: - kv 16: deepseek4.rope.scaling.yarn_beta_slow f32 = 1.000000
llama_model_loader: - kv 17: deepseek4.rope.freq_base f32 = 10000.000000
llama_model_loader: - kv 18: deepseek4.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 19: deepseek4.expert_used_count u32 = 6
llama_model_loader: - kv 20: deepseek4.expert_gating_func u32 = 4
llama_model_loader: - kv 21: deepseek4.attention.key_length u32 = 512
llama_model_loader: - kv 22: deepseek4.attention.value_length u32 = 512
llama_model_loader: - kv 23: general.file_type u32 = 7
llama_model_loader: - kv 24: deepseek4.vocab_size u32 = 129280
llama_model_loader: - kv 25: deepseek4.rope.dimension_count u32 = 64
llama_model_loader: - kv 26: deepseek4.attention.q_lora_rank u32 = 1024
llama_model_loader: - kv 27: deepseek4.attention.output_lora_rank u32 = 1024
llama_model_loader: - kv 28: deepseek4.attention.output_group_count u32 = 8
llama_model_loader: - kv 29: deepseek4.attention.compress_ratios arr[i32,44] = [0, 0, 4, 128, 4, 128, 4, 128, 4, 128...
llama_model_loader: - kv 30: deepseek4.attention.compress_rope_freq_base f32 = 160000.000000
llama_model_loader: - kv 31: deepseek4.expert_feed_forward_length u32 = 2048
llama_model_loader: - kv 32: deepseek4.expert_count u32 = 256
llama_model_loader: - kv 33: deepseek4.expert_shared_count u32 = 1
llama_model_loader: - kv 34: deepseek4.expert_weights_scale f32 = 1.500000
llama_model_loader: - kv 35: deepseek4.hash_layer_count u32 = 3
llama_model_loader: - kv 36: deepseek4.expert_weights_norm bool = true
llama_model_loader: - kv 37: deepseek4.swiglu_clamp_exp arr[f32,43] = [10.000000, 10.000000, 10.000000, 10....
llama_model_loader: - kv 38: deepseek4.attention.sliding_window u32 = 128
llama_model_loader: - kv 39: deepseek4.attention.indexer.head_count u32 = 64
llama_model_loader: - kv 40: deepseek4.attention.indexer.key_length u32 = 128
llama_model_loader: - kv 41: deepseek4.attention.indexer.top_k u32 = 512
llama_model_loader: - kv 42: deepseek4.nextn_predict_layers u32 = 1
llama_model_loader: - kv 43: deepseek4.hyper_connection.count u32 = 4
llama_model_loader: - kv 44: deepseek4.hyper_connection.sinkhorn_iterations u32 = 20
llama_model_loader: - kv 45: deepseek4.hyper_connection.epsilon f32 = 0.000001
llama_model_loader: - kv 46: general.quantization_version u32 = 2
llama_model_loader: - kv 47: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 48: tokenizer.ggml.pre str = joyai-llm
llama_model_loader: - kv 49: tokenizer.ggml.tokens arr[str,129280] = ["<｜begin▁of▁sentence｜>", "<�...
llama_model_loader: - kv 50: tokenizer.ggml.token_type arr[i32,129280] = [3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 51: tokenizer.ggml.merges arr[str,127741] = ["Ġ t", "Ġ a", "i n", "Ġ Ġ", "h e...
llama_model_loader: - kv 52: tokenizer.ggml.bos_token_id u32 = 0
llama_model_loader: - kv 53: tokenizer.ggml.eos_token_id u32 = 1
llama_model_loader: - kv 54: tokenizer.ggml.padding_token_id u32 = 1
llama_model_loader: - kv 55: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 56: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 57: tokenizer.chat_template str = {%- if not add_generation_prompt is d...
llama_model_loader: - kv 58: split.no u16 = 0
llama_model_loader: - kv 59: split.tensors.count i32 = 1328
llama_model_loader: - kv 60: split.count u16 = 7
llama_model_loader: - type f32: 556 tensors
llama_model_loader: - type q8_0: 769 tensors
llama_model_loader: - type i32: 3 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q8_0
print_info: file size = 281.50 GiB (8.50 BPW)
llama_prepare_model_devices: using device CUDA0 (NVIDIA A100-SXM4-80GB) (0000:27:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA1 (NVIDIA A100-SXM4-80GB) (0000:2a:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA2 (NVIDIA A100-SXM4-80GB) (0000:51:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA3 (NVIDIA A100-SXM4-80GB) (0000:57:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA4 (NVIDIA A100-SXM4-80GB) (0000:9e:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA5 (NVIDIA A100-SXM4-80GB) (0000:a4:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA6 (NVIDIA A100-SXM4-80GB) (0000:c7:00.0) - 80728 MiB free
llama_prepare_model_devices: using device CUDA7 (NVIDIA A100-SXM4-80GB) (0000:ca:00.0) - 80728 MiB free
load: 0 unused tokens
load: printing all EOG tokens:
load: - 1 ('<｜end▁of▁sentence｜>')
load: special tokens cache size = 1283
load: token to piece cache size = 0.8346 MB
print_info: arch = deepseek4
print_info: vocab_only = 0
print_info: no_alloc = 0
print_info: n_ctx_train = 1048576
print_info: n_embd = 4096
print_info: n_embd_inp = 4096
print_info: n_layer = 43
print_info: n_head = 64
print_info: n_head_kv = 1
print_info: n_rot = 64
print_info: n_swa = 128
print_info: is_swa_any = 1
print_info: n_embd_head_k = 512
print_info: n_embd_head_v = 512
print_info: n_gqa = 64
print_info: n_embd_k_gqa = 512
print_info: n_embd_v_gqa = 512
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 = 0.0e+00
print_info: f_attn_value_scale = 0.0000
print_info: n_ff = 0
print_info: n_expert = 256
print_info: n_expert_used = 6
print_info: n_expert_groups = 0
print_info: n_group_used = 0
print_info: causal attn = 1
print_info: pooling type = -1
print_info: rope type = 0
print_info: rope scaling = yarn
print_info: freq_base_train = 10000.0
print_info: freq_scale_train = 0.0625
print_info: freq_base_swa = 10000.0
print_info: freq_scale_swa = 0.0625
print_info: n_embd_head_k_swa = 512
print_info: n_embd_head_v_swa = 512
print_info: n_rot_swa = 64
print_info: n_ctx_orig_yarn = 65536
print_info: rope_yarn_log_mul = 0.0000
print_info: rope_finetuned = unknown
print_info: model type = ?B
print_info: model params = 284.33 B
print_info: general.name = DeepSeek V4 Flash
print_info: n_lora_q = 1024
print_info: n_lora_o = 1024
print_info: n_attn_out_groups = 8
print_info: n_ff_exp = 2048
print_info: n_expert_shared = 1
print_info: n_swa = 128
print_info: compress_rope_freq_base = 160000.0
print_info: indexer_n_head = 64
print_info: indexer_head_size = 128
print_info: indexer_top_k = 512
print_info: n_hash_layers = 3
print_info: n_hc = 4
print_info: hc_sinkhorn_iters = 20
print_info: hc_eps = 1.0e-06
print_info: nextn_predict_layers = 1
print_info: expert_weights_scale = 1.5
print_info: expert_weights_norm = 1
print_info: expert_gating_func = unknown
print_info: vocab type = BPE
print_info: n_vocab = 129280
print_info: n_merges = 127741
print_info: BOS token = 0 '<｜begin▁of▁sentence｜>'
print_info: EOS token = 1 '<｜end▁of▁sentence｜>'
print_info: EOT token = 1 '<｜end▁of▁sentence｜>'
print_info: PAD token = 1 '<｜end▁of▁sentence｜>'
print_info: LF token = 201 'Ċ'
print_info: FIM PRE token = 128801 '<｜fim▁begin｜>'
print_info: FIM SUF token = 128800 '<｜fim▁hole｜>'
print_info: FIM MID token = 128802 '<｜fim▁end｜>'
print_info: EOG token = 1 '<｜end▁of▁sentence｜>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 42 repeating layers to GPU
load_tensors: offloaded 44/44 layers to GPU
load_tensors: CPU_Mapped model buffer size = 6859.93 MiB
load_tensors: CUDA0 model buffer size = 40057.90 MiB
load_tensors: CUDA1 model buffer size = 33402.44 MiB
load_tensors: CUDA2 model buffer size = 40073.40 MiB
load_tensors: CUDA3 model buffer size = 33386.56 MiB
load_tensors: CUDA4 model buffer size = 40073.40 MiB
load_tensors: CUDA5 model buffer size = 33402.44 MiB
load_tensors: CUDA6 model buffer size = 40073.40 MiB
load_tensors: CUDA7 model buffer size = 27252.25 MiB
....................................................................................................
common_init_result: added <｜end▁of▁sentence｜> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max = 2
llama_context: n_ctx = 8192
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 = enabled
llama_context: kv_unified = false
llama_context: freq_base = 10000.0
llama_context: freq_scale = 0.0625
llama_context: n_ctx_seq (4096) < n_ctx_train (1048576) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 0.99 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 4096 cells
llama_kv_cache: size = 0.00 MiB ( 4096 cells, 0 layers, 2/2 seqs), K (f16): 0.00 MiB, V (f16): 0.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 0
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 0
llama_kv_cache_iswa: creating SWA KV cache, size = 768 cells
llama_kv_cache: CUDA0 KV buffer size = 9.00 MiB
llama_kv_cache: CUDA1 KV buffer size = 7.50 MiB
llama_kv_cache: CUDA2 KV buffer size = 9.00 MiB
llama_kv_cache: CUDA3 KV buffer size = 7.50 MiB
llama_kv_cache: CUDA4 KV buffer size = 9.00 MiB
llama_kv_cache: CUDA5 KV buffer size = 7.50 MiB
llama_kv_cache: CUDA6 KV buffer size = 9.00 MiB
llama_kv_cache: CUDA7 KV buffer size = 6.00 MiB
llama_kv_cache: size = 64.50 MiB ( 768 cells, 43 layers, 2/2 seqs), K (f16): 64.50 MiB, V (f16): 0.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 512
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 512
llama_memory_recurrent: CUDA0 RS buffer size = 4.00 MiB
llama_memory_recurrent: CUDA1 RS buffer size = 5.00 MiB
llama_memory_recurrent: CUDA2 RS buffer size = 6.00 MiB
llama_memory_recurrent: CUDA3 RS buffer size = 5.00 MiB
llama_memory_recurrent: CUDA4 RS buffer size = 6.00 MiB
llama_memory_recurrent: CUDA5 RS buffer size = 5.00 MiB
llama_memory_recurrent: CUDA6 RS buffer size = 6.00 MiB
llama_memory_recurrent: CUDA7 RS buffer size = 4.00 MiB
llama_memory_recurrent: size = 41.00 MiB ( 2 cells, 43 layers, 2 seqs), R (f32): 20.50 MiB, S (f32): 20.50 MiB
llama_memory_hybrid_iswa: CUDA0 DeepSeek4 compressed KV buffer size = 5.12 MiB
llama_memory_hybrid_iswa: CUDA1 DeepSeek4 compressed KV buffer size = 7.62 MiB
llama_memory_hybrid_iswa: CUDA2 DeepSeek4 compressed KV buffer size = 7.69 MiB
llama_memory_hybrid_iswa: CUDA3 DeepSeek4 compressed KV buffer size = 5.19 MiB
llama_memory_hybrid_iswa: CUDA4 DeepSeek4 compressed KV buffer size = 7.69 MiB
llama_memory_hybrid_iswa: CUDA5 DeepSeek4 compressed KV buffer size = 7.62 MiB
llama_memory_hybrid_iswa: CUDA6 DeepSeek4 compressed KV buffer size = 7.69 MiB
llama_memory_hybrid_iswa: CUDA7 DeepSeek4 compressed KV buffer size = 5.12 MiB
llama_context: pipeline parallelism enabled
sched_reserve: reserving ...
sched_reserve: resolving fused Gated Delta Net support:
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/ggml/src/ggml.c:3660: GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2) failed
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libggml-base.so.0(+0x1c27b)[0x7f85c12c827b]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libggml-base.so.0(ggml_print_backtrace+0x21f)[0x7f85c12c86ff]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libggml-base.so.0(ggml_abort+0x152)[0x7f85c12c88d2]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libggml-base.so.0(+0x23baf)[0x7f85c12cfbaf]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZN21llama_model_deepseek45graphC1ERK11llama_modelRK16llm_graph_params+0x6d5)[0x7f85c0bebbe5]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZNK21llama_model_deepseek416build_arch_graphERK16llm_graph_params+0x33)[0x7f85c0beea43]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZNK11llama_model11build_graphERK16llm_graph_params+0x30)[0x7f85c0b68460]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZN13llama_context13graph_reserveEjjjPK22llama_memory_context_ibPmi+0x224)[0x7f85c0ad8da4]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZN13llama_context13sched_reserveEv+0x1066)[0x7f85c0ada366]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(_ZN13llama_contextC1ERK11llama_model20llama_context_params+0xb39)[0x7f85c0add779]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama.so.0(llama_init_from_model+0x1b0)[0x7f85c0ade3d0]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama-common.so.0(_ZN18common_init_resultC2ER13common_params+0xb10)[0x7f85c1001d60]
/data/llama.cpp/output1/llama.cpp-feat-v4-port-cuda/build/bin/libllama-common.so.0(_Z23common_init_from_paramsR13common_params+0x48)[0x7f85c10029c8]
./llama-server(+0x10c8ee)[0x55d6abb708ee]
./llama-server(+0x5ba22)[0x55d6ababfa22]
/usr/lib/x86_64-linux-gnu/libc.so.6(+0x29d90)[0x7f85c04edd90]
/usr/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x80)[0x7f85c04ede40]
./llama-server(+0x5c695)[0x55d6abac0695]
Name and Version
version:
llama.cpp-feat-v4-port-cuda
NVIDIA-SMI 570.124.06 Driver Version: 570.124.06 CUDA Version: 13.0
Operating systems
Linux
GGML backends
CUDA
Hardware
8*A100(80G)
Models
DeepSeek-V4-Flash-Q8_0
Problem description & steps to reproduce
./llama-server
--model /data/DeepSeek-V4-Flash-Q8_0-00001-of-00007.gguf
--host 0.0.0.0 --port 8000
--jinja --reasoning off
--ctx-size 8192
--n-gpu-layers 999
--split-mode layer
--flash-attn on
--no-repack
--temp 1.0 --top-p 1.0 --top-k 0 --min-p 0.0
--alias DeepSeek-V4-Flash-Q8_0
This above command works fine.
However, if I add the parameter --parallel 2, coredump will appear, but --parallel 1 is OK
First Bad Commit
No response
Relevant log output
Logs