Misc. bug: Fail to run Qwen3-Coder-480B-A35B-Instruct (Q4_0) model at Ascend 910B NPU. #15759
Description
Name and Version
register_backend: registered backend CANN (8 devices)
register_device: registered device CANN0 (Ascend910B3)
register_device: registered device CANN1 (Ascend910B3)
register_device: registered device CANN2 (Ascend910B3)
register_device: registered device CANN3 (Ascend910B3)
register_device: registered device CANN4 (Ascend910B3)
register_device: registered device CANN5 (Ascend910B3)
register_device: registered device CANN6 (Ascend910B3)
register_device: registered device CANN7 (Ascend910B3)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (CPU)
load_backend: failed to find ggml_backend_init in /home/m00522703/llama.cpp/build/bin/libggml-cann.so
load_backend: failed to find ggml_backend_init in /home/m00522703/llama.cpp/build/bin/libggml-cpu.so
version: 6362 (f6da8cb8)
built with cc (GCC) 12.3.1 (openEuler 12.3.1-38.oe2403) for aarch64-openEuler-linux
Operating systems
No response
Which llama.cpp modules do you know to be affected?
No response
Command line
./build/bin/llama-server --model /data1/llm/gguf/Qwen3-Coder-480B-A35B-Instruct-Q4_0.gguf -ngl 320 --host 0.0.0.0Problem description & steps to reproduce
llama_model_loader: - type f32: 311 tensors
llama_model_loader: - type q4_0: 435 tensors
llama_model_loader: - type q6_K: 1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_0
print_info: file size = 251.96 GiB (4.51 BPW)
load: printing all EOG tokens:
load: - 151643 ('<|endoftext|>')
load: - 151645 ('<|im_end|>')
load: - 151662 ('<|fim_pad|>')
load: - 151663 ('<|repo_name|>')
load: - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch = qwen3moe
print_info: vocab_only = 0
print_info: n_ctx_train = 262144
print_info: n_embd = 6144
print_info: n_layer = 62
print_info: n_head = 96
print_info: n_head_kv = 8
print_info: n_rot = 128
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 128
print_info: n_embd_head_v = 128
print_info: n_gqa = 12
print_info: n_embd_k_gqa = 1024
print_info: n_embd_v_gqa = 1024
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: n_ff = 8192
print_info: n_expert = 160
print_info: n_expert_used = 8
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 262144
print_info: rope_finetuned = unknown
print_info: model type = ?B
print_info: model params = 480.15 B
print_info: general.name = Qwen3 Coder 480B A35B Instruct
print_info: n_ff_exp = 2560
print_info: vocab type = BPE
print_info: n_vocab = 151936
print_info: n_merges = 151387
print_info: BOS token = 11 ','
print_info: EOS token = 151645 '<|im_end|>'
print_info: EOT token = 151645 '<|im_end|>'
print_info: PAD token = 151643 '<|endoftext|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 151659 '<|fim_prefix|>'
print_info: FIM SUF token = 151661 '<|fim_suffix|>'
print_info: FIM MID token = 151660 '<|fim_middle|>'
print_info: FIM PAD token = 151662 '<|fim_pad|>'
print_info: FIM REP token = 151663 '<|repo_name|>'
print_info: FIM SEP token = 151664 '<|file_sep|>'
print_info: EOG token = 151643 '<|endoftext|>'
print_info: EOG token = 151645 '<|im_end|>'
print_info: EOG token = 151662 '<|fim_pad|>'
print_info: EOG token = 151663 '<|repo_name|>'
print_info: EOG token = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 62 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 63/63 layers to GPU
load_tensors: CPU_Mapped model buffer size = 1231.07 MiB
load_tensors: CANN0 model buffer size = 33132.38 MiB
load_tensors: CANN1 model buffer size = 33132.38 MiB
load_tensors: CANN2 model buffer size = 33132.38 MiB
load_tensors: CANN3 model buffer size = 33132.38 MiB
load_tensors: CANN4 model buffer size = 33132.38 MiB
load_tensors: CANN5 model buffer size = 33132.38 MiB
load_tensors: CANN6 model buffer size = 33132.38 MiB
load_tensors: CANN7 model buffer size = 24849.31 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = false
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
ggml_backend_cann_context: device 0 async operator submission is OFF
ggml_backend_cann_context: device 0 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 1 async operator submission is OFF
ggml_backend_cann_context: device 1 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 2 async operator submission is OFF
ggml_backend_cann_context: device 2 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 3 async operator submission is OFF
ggml_backend_cann_context: device 3 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 4 async operator submission is OFF
ggml_backend_cann_context: device 4 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 5 async operator submission is OFF
ggml_backend_cann_context: device 5 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 6 async operator submission is OFF
ggml_backend_cann_context: device 6 execution mode is GRAPH (acl graph enabled)
ggml_backend_cann_context: device 7 async operator submission is OFF
ggml_backend_cann_context: device 7 execution mode is GRAPH (acl graph enabled)
llama_context: CANN_Host output buffer size = 0.58 MiB
llama_kv_cache: CANN0 KV buffer size = 128.00 MiB
llama_kv_cache: CANN1 KV buffer size = 128.00 MiB
llama_kv_cache: CANN2 KV buffer size = 128.00 MiB
llama_kv_cache: CANN3 KV buffer size = 128.00 MiB
llama_kv_cache: CANN4 KV buffer size = 128.00 MiB
llama_kv_cache: CANN5 KV buffer size = 128.00 MiB
llama_kv_cache: CANN6 KV buffer size = 128.00 MiB
llama_kv_cache: CANN7 KV buffer size = 96.00 MiB
llama_kv_cache: size = 992.00 MiB ( 4096 cells, 62 layers, 1/1 seqs), K (f16): 496.00 MiB, V (f16): 496.00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context: CANN0 compute buffer size = 276.01 MiB
llama_context: CANN1 compute buffer size = 244.63 MiB
llama_context: CANN2 compute buffer size = 244.63 MiB
llama_context: CANN3 compute buffer size = 244.63 MiB
llama_context: CANN4 compute buffer size = 244.63 MiB
llama_context: CANN5 compute buffer size = 244.63 MiB
llama_context: CANN6 compute buffer size = 244.63 MiB
llama_context: CANN7 compute buffer size = 256.01 MiB
llama_context: CANN_Host compute buffer size = 308.75 MiB
llama_context: graph nodes = 3851
llama_context: graph splits = 10
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: added <|repo_name|> logit bias = -inf
common_init_from_params: added <|file_sep|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
new_pool_for_device: device 0 use vmm pool
/home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp:69: CANN error
CANN error: EE9999: Inner Error!
EE9999: [PID: 182660] 2025-09-03-14:14:41.418.758 Not allow to synchronize captured-stream, stream_id=2.[FUNC:StreamSynchronize][FILE:api_error.cc][LINE:960]
TraceBack (most recent call last):
rtStreamSynchronize execute failed, reason=[stream is captured][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
synchronize stream failed, runtime result = 107027[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
current device: 0, in function ggml_cann_mul_mat_id_quant at /home/xxx/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp:3008
aclrtSynchronizeStream(ctx.stream())
[New LWP 182661]
[New LWP 182662]
[New LWP 182663]
[New LWP 182779]
[New LWP 182780]
... many outputs ...
[New LWP 195125]
[New LWP 195126]
[New LWP 195127]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib64/libthread_db.so.1".
0x0000ffffb9721254 in wait4 () from /usr/lib64/libc.so.6
#0 0x0000ffffb9721254 in wait4 () from /usr/lib64/libc.so.6
#1 0x0000ffffb9b66314 in ggml_print_backtrace () at /home/xxx/llama.cpp/ggml/src/ggml.c:196
196 waitpid(child_pid, NULL, 0);
#2 0x0000ffffb9b664b8 in ggml_abort (file=0xffffb9cbd810 "/home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp", line=69, fmt=0xffffb9cbd800 "CANN error") at /home/xxx/llama.cpp/ggml/src/ggml.c:230
230 ggml_print_backtrace();
#3 0x0000ffffb9c9fb9c in ggml_cann_error (stmt=0xffffb9cbc338 "aclrtSynchronizeStream(ctx.stream())", func=0xffffb9cbc318 "ggml_cann_mul_mat_id_quant", file=0xffffb9cb8a58 "/home/xxx/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp", line=3006, msg=0x307195d8 "EE9999: Inner Error!\nEE9999: [PID: 313721] 2025-09-03-15:42:35.548.491 Not allow to synchronize captured-stream, stream_id=2.[FUNC:StreamSynchronize][FILE:api_error.cc][LINE:960]\n TraceBack (mo"...) at /home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp:69
69 GGML_ABORT("CANN error");
#4 0x0000ffffb9c8fe34 in ggml_cann_mul_mat_id_quant (ctx=..., dst=0x2c4f55d0) at /home/xxx/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp:3006
3006 ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
#5 0x0000ffffb9c90318 in ggml_cann_mul_mat_id (ctx=..., dst=0x2c4f55d0) at /home/xxx/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp:3103
3103 ggml_cann_mul_mat_id_quant(ctx, dst);
#6 0x0000ffffb9ca47b0 in ggml_cann_compute_forward (ctx=..., dst=0x2c4f55d0) at /home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp:1805
1805 ggml_cann_mul_mat_id(ctx, dst);
#7 0x0000ffffb9ca59a0 in evaluate_and_capture_cann_graph (cann_ctx=0x2b55f0f0, cgraph=0x2b583f48, use_cann_graph=@0xffffef81dc57: true, cann_graph_update_required=@0xffffef81dc56: true) at /home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp:2211
2211 bool ok = ggml_cann_compute_forward(*cann_ctx, node);
#8 0x0000ffffb9ca5c34 in ggml_backend_cann_graph_compute (backend=0x2bf7cf00, cgraph=0x2b583f48) at /home/xxx/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp:2273
2273 evaluate_and_capture_cann_graph(
#9 0x0000ffffb9b7e838 in ggml_backend_graph_compute_async (backend=0x2bf7cf00, cgraph=0x2b583f48) at /home/xxx/llama.cpp/ggml/src/ggml-backend.cpp:359
359 return backend->iface.graph_compute(backend, cgraph);
#10 0x0000ffffb9b82de8 in ggml_backend_sched_compute_splits (sched=0x2bf802d0) at /home/xxx/llama.cpp/ggml/src/ggml-backend.cpp:1542
1542 enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
#11 0x0000ffffb9b83a50 in ggml_backend_sched_graph_compute_async (sched=0x2bf802d0, graph=0x2c4a8ea0) at /home/xxx/llama.cpp/ggml/src/ggml-backend.cpp:1742
1742 return ggml_backend_sched_compute_splits(sched);
#12 0x0000ffffba2fd378 in llama_context::graph_compute (this=0x27087dd0, gf=0x2c4a8ea0, batched=true) at /home/xxx/llama.cpp/src/llama-context.cpp:1454
1454 auto status = ggml_backend_sched_graph_compute_async(sched.get(), gf);
#13 0x0000ffffba2fad64 in llama_context::process_ubatch (this=0x27087dd0, ubatch=..., gtype=LLM_GRAPH_TYPE_DECODER, mctx=0x2e216e90, ret=@0xffffef81feb4: GGML_STATUS_SUCCESS) at /home/xxx/llama.cpp/src/llama-context.cpp:781
781 const auto status = graph_compute(res->get_gf(), ubatch.n_tokens > 1);
#14 0x0000ffffba2fbd44 in llama_context::decode (this=0x27087dd0, batch_inp=...) at /home/xxx/llama.cpp/src/llama-context.cpp:1085
1085 const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
#15 0x0000ffffba300ef8 in llama_decode (ctx=0x27087dd0, batch=...) at /home/xxx/llama.cpp/src/llama-context.cpp:2720
2720 const int ret = ctx->decode(batch);
#16 0x0000000000702824 in common_init_from_params (params=...) at /home/xxx/llama.cpp/common/common.cpp:1066
1066 llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
#17 0x00000000005123b8 in server_context::load_model (this=0xffffef822150, params=...) at /home/xxx/llama.cpp/tools/server/server.cpp:2087
2087 llama_init = common_init_from_params(params_base);
#18 0x00000000004cac54 in main (argc=7, argv=0xffffef825b98) at /home/xxx/llama.cpp/tools/server/server.cpp:5128
5128 if (!ctx_server.load_model(params)) {
[Inferior 1 (process 313721) detached]
Aborted (core dumped)
First Bad Commit
No response