Closed
Description
Name and Version
$./build/bin/llama-cli --version
version: 5747 (0142961a)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for aarch64-linux-gnuOperating systems
Linux
GGML backends
CANN-Ascend
Hardware
[ma-user llama.cpp]$npu-smi info
+------------------------------------------------------------------------------------------------+
| npu-smi 23.0.6 Version: 23.0.6 |
+---------------------------+---------------+----------------------------------------------------+
| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|
| Chip | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |
+===========================+===============+====================================================+
| 1 910B4 | OK | 95.9 39 0 / 0 |
| 0 | 0000:01:00.0 | 0 0 / 0 2826 / 32768 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 1 |
+===========================+===============+====================================================+Models
Problem description & steps to reproduce
I am working on applying llama.cpp to 310B4 devices (310B4 devices do not currently support HF32), so I first changed cubeMathType on 910B4 devices so that HF32 is not used in the matrix multiplication operator.
I changed GGML_CANN_CALL_ACLNN_OP(ctx, Matmul, acl_input_tensor, acl_weight_tensor, acl_dst, 1); to GGML_CANN_CALL_ACLNN_OP(ctx, Matmul, acl_input_tensor, acl_weight_tensor, acl_dst, 2); and then performed reasoning verification and found that the reasoning failed. When faced with the question "What is the capital of China?", the letter G was repeatedly output.
我正在致力于将
llama.cpp应用到310B4设备上（310B4设备目前不支持HF32），故先行在910B4设备上更改cubeMathType使得矩阵乘算子运算过程中不使用HF32。
我将GGML_CANN_CALL_ACLNN_OP(ctx, Matmul, acl_input_tensor, acl_weight_tensor, acl_dst, 1);改为GGML_CANN_CALL_ACLNN_OP(ctx, Matmul, acl_input_tensor, acl_weight_tensor, acl_dst, 2);之后，进行推理验证发现推理失败，面对问题“What is the capital of China?”重复输出字母G。
First Bad Commit
No response
Relevant log output
[ma-user llama.cpp]$./build/bin/llama-cli -m ./models/qwen2-1_5b-instruct-fp16.gguf -p "What is the capital of China?" -ngl 32 --no-warmupbuild: 0 (unknown) with gcc (GCC) 14.2.0 for aarch64-unknown-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CANN0 (Ascend910B4) - 29891 MiB free
llama_model_loader: loaded meta data with 22 key-value pairs and 338 tensors from ./models/qwen2-1_5b-instruct-fp16.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = qwen2
llama_model_loader: - kv 1: general.name str = qwen2-1_5b-instruct
llama_model_loader: - kv 2: qwen2.block_count u32 = 28
llama_model_loader: - kv 3: qwen2.context_length u32 = 32768
llama_model_loader: - kv 4: qwen2.embedding_length u32 = 1536
llama_model_loader: - kv 5: qwen2.feed_forward_length u32 = 8960
llama_model_loader: - kv 6: qwen2.attention.head_count u32 = 12
llama_model_loader: - kv 7: qwen2.attention.head_count_kv u32 = 2
llama_model_loader: - kv 8: qwen2.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 9: qwen2.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 10: general.file_type u32 = 1
llama_model_loader: - kv 11: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 12: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 14: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 15: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 16: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 17: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 18: tokenizer.ggml.bos_token_id u32 = 151643
llama_model_loader: - kv 19: tokenizer.chat_template str = {% for message in messages %}{% if lo...
llama_model_loader: - kv 20: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 21: general.quantization_version u32 = 2
llama_model_loader: - type f32: 141 tensors
llama_model_loader: - type f16: 197 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = F16
print_info: file size = 2.88 GiB (16.00 BPW)
load: special tokens cache size = 293
load: token to piece cache size = 0.9338 MB
print_info: arch = qwen2
print_info: vocab_only = 0
print_info: n_ctx_train = 32768
print_info: n_embd = 1536
print_info: n_layer = 28
print_info: n_head = 12
print_info: n_head_kv = 2
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 = 6
print_info: n_embd_k_gqa = 256
print_info: n_embd_v_gqa = 256
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 = 8960
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: causal attn = 1
print_info: pooling type = -1
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 32768
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 1.5B
print_info: model params = 1.54 B
print_info: general.name = qwen2-1_5b-instruct
print_info: vocab type = BPE
print_info: n_vocab = 151936
print_info: n_merges = 151387
print_info: BOS token = 151643 '<|endoftext|>'
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: EOG token = 151643 '<|endoftext|>'
print_info: EOG token = 151645 '<|im_end|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 28 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 29/29 layers to GPU
load_tensors: CPU_Mapped model buffer size = 445.12 MiB
load_tensors: CANN0 model buffer size = 2944.68 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 = 0
llama_context: freq_base = 1000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
ggml_backend_cann_context: device 0 async operator submission is OFF
llama_context: CANN_Host output buffer size = 0.58 MiB
llama_kv_cache_unified: CANN0 KV buffer size = 112.00 MiB
llama_kv_cache_unified: size = 112.00 MiB ( 4096 cells, 28 layers, 1 seqs), K (f16): 56.00 MiB, V (f16): 56.00 MiB
llama_context: CANN0 compute buffer size = 299.75 MiB
llama_context: CANN_Host compute buffer size = 11.01 MiB
llama_context: graph nodes = 1098
llama_context: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
main: llama threadpool init, n_threads = 192
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template 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
system_info: n_threads = 192 (n_threads_batch = 192) / 192 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | DOTPROD = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
main: interactive mode on.
sampler seed: 2549894379
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = -1, n_keep = 0
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to the AI.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
- Not using system message. To change it, set a different value via -sys PROMPT
user
What is the capital of China?
assistant
new_pool_for_device: device 0 use vmm pool
GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG
> EOF by user
llama_perf_sampler_print: sampling time = 16.47 ms / 71 runs ( 0.23 ms per token, 4310.61 tokens per second)
llama_perf_context_print: load time = 10382.88 ms
llama_perf_context_print: prompt eval time = 1118.79 ms / 15 tokens ( 74.59 ms per token, 13.41 tokens per second)
llama_perf_context_print: eval time = 3692.72 ms / 55 runs ( 67.14 ms per token, 14.89 tokens per second)
llama_perf_context_print: total time = 15030.61 ms / 70 tokens