The NPU version of llama.cpp did not return an appropriate response on Intel Core Ultra 7 268V

[kotauchisunsun]

Describe the bug
Here is the English translation of your text:

I tried running the NPU version of llama.cpp on Windows 11, but I did not receive an appropriate response.

Model: DeepSeek-R1-Distill-Qwen-7B-Q6_K.gguf
Executable: llama-cpp-ipex-llm-2.2.0-win-npu.zip
CPU: Intel Core Ultra 7 268V
NPU: Intel AI Boost
NPU Driver: 32.0.100.3967

The command itself was referenced in the following URL:
https://github.com/intel/ipex-llm/blob/main/docs/mddocs/Quickstart/npu_quickstart.md

Could the issue be with the method of execution? Or is it that the model is not supported?

** Logs **

PS C:\Users\kotau\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu> .\llama-cli-npu.exe -m ..\DeepSeek-R1-Distill-Qwen-7B-Q6_K.gguf -n 32 --prompt "What is AI?"
build: 1 (3ac676a) with MSVC 19.39.33519.0 for x64
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from ..\DeepSeek-R1-Distill-Qwen-7B-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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 7B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 7B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["ﾄ ﾄ", "ﾄﾄ ﾄﾄ", "i n", "ﾄ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 18
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q6_K:  198 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
llm_load_print_meta: n_embd           = 3584
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 28
llm_load_print_meta: n_head_kv        = 4
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            = 7
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
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             = 18944
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
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      = Q6_K
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 5.82 GiB (6.56 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 7B
llm_load_print_meta: BOS token        = 151646 '<・彙egin笆｛f笆《entence・・'
llm_load_print_meta: EOS token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: PAD token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: LF token         = 148848 'ﾃ・ｬ'
llm_load_print_meta: EOG token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors:        CPU buffer size =  5958.79 MiB
........................................................................................
Directory created: "C:\\Users\\kotau\\Downloads\\llama-cpp-ipex-llm-2.2.0-win-npu\\NPU_models\\qwen2-28-3584-152064-Q4_0"
Directory created: "C:\\Users\\kotau\\Downloads\\llama-cpp-ipex-llm-2.2.0-win-npu\\NPU_models\\qwen2-28-3584-152064-Q4_0\\model_weights"
Converting GGUF model to Q4_0 NPU model...
Model weights saved to C:\Users\kotau\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu\NPU_models\qwen2-28-3584-152064-Q4_0\model_weights
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from ..\DeepSeek-R1-Distill-Qwen-7B-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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 7B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 7B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["ﾄ ﾄ", "ﾄﾄ ﾄﾄ", "i n", "ﾄ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 18
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q6_K:  198 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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       = 1
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = all F32
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 5.82 GiB (6.56 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 7B
llm_load_print_meta: BOS token        = 151646 '<・彙egin笆｛f笆《entence・・'
llm_load_print_meta: EOS token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: PAD token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: LF token         = 148848 'ﾃ・ｬ'
llm_load_print_meta: EOG token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: max token length = 256
llama_model_load: vocab only - skipping tensors
Model saved to C:\Users\kotau\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu\NPU_models\qwen2-28-3584-152064-Q4_0//decoder_layer_0.blob
Model saved to C:\Users\kotau\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu\NPU_models\qwen2-28-3584-152064-Q4_0//decoder_layer_1.blob
llama_new_context_with_model: n_ctx      = 1024
llama_new_context_with_model: n_batch    = 1024
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 0.0
llama_new_context_with_model: freq_scale = 1
・ｿusing              犹もｸ巵ｸ｣犹・ <think>ﾘｵﾙ・ｯ  Di盻・  2犹もｸ巵ｸ｣犹・ 1  2

llm_perf_print:        load time =   53917.00 ms
llm_perf_print: prompt eval time =    3467.00 ms /     7 tokens (  495.29 ms per token,     2.02 tokens per second)
llm_perf_print:        eval time =    2704.00 ms /    31 runs   (   87.23 ms per token,    11.46 tokens per second)
llm_perf_print:       total time =   60146.00 ms /    38 tokens

[plusbang]

Hi, please use 32.0.100.3104 NPU driver as our document(https://github.com/intel/ipex-llm/blob/main/docs/mddocs/Quickstart/llama_cpp_npu_portable_zip_quickstart.md#prerequisites) described.

[kotauchisunsun]

@plusbang
Thank you for your reply! I uninstalled the 32.0.100.3967 driver that was installed on my PC.
Then, I tried to install the 32.0.100.3104 NPU driver.
However, the driver version shown in Device Manager is 32.0.100.3717.
It seems that the PC originally came with a later version than 32.0.100.3104 — which is the one confirmed to work — so I’m unable to install the older version.

Just in case, I tried running it with the 32.0.100.3717 driver, but I didn’t get an appropriate response.

Logs

PS C:\Users\kotau\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu> ./llama-cli-npu.exe -m ..\DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf -p "Hello"
build: 1 (3ac676a) with MSVC 19.39.33519.0 for x64
llama_model_loader: loaded meta data with 30 key-value pairs and 339 tensors from ..\DeepSeek-R1-Distill-Qwen-7B-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              = DeepSeek R1 Distill Qwen 7B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 7B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["ﾄ ﾄ", "ﾄﾄ ﾄﾄ", "i n", "ﾄ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 196
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q4_K:  169 tensors
llama_model_loader: - type q6_K:   29 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
llm_load_print_meta: n_embd           = 3584
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 28
llm_load_print_meta: n_head_kv        = 4
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            = 7
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
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             = 18944
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
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     = 7.62 B
llm_load_print_meta: model size       = 4.36 GiB (4.91 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 7B
llm_load_print_meta: BOS token        = 151646 '<・彙egin笆｛f笆《entence・・'
llm_load_print_meta: EOS token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: PAD token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: LF token         = 148848 'ﾃ・ｬ'
llm_load_print_meta: EOG token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors:        CPU buffer size =  4460.45 MiB
....................................................................................
Failed to create directory or already exists: "C:\\Users\\kotau\\Downloads\\llama-cpp-ipex-llm-2.2.0-win-npu\\NPU_models\\qwen2-28-3584-152064-Q4_0\\model_weights"
llama_model_loader: loaded meta data with 30 key-value pairs and 339 tensors from ..\DeepSeek-R1-Distill-Qwen-7B-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              = DeepSeek R1 Distill Qwen 7B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 7B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["ﾄ ﾄ", "ﾄﾄ ﾄﾄ", "i n", "ﾄ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 196
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q4_K:  169 tensors
llama_model_loader: - type q6_K:   29 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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       = 1
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = all F32
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 4.36 GiB (4.91 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 7B
llm_load_print_meta: BOS token        = 151646 '<・彙egin笆｛f笆《entence・・'
llm_load_print_meta: EOS token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: PAD token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: LF token         = 148848 'ﾃ・ｬ'
llm_load_print_meta: EOG token        = 151643 '<・彳nd笆｛f笆《entence・・'
llm_load_print_meta: max token length = 256
llama_model_load: vocab only - skipping tensors
llama_new_context_with_model: n_ctx      = 1024
llama_new_context_with_model: n_batch    = 1024
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 0.0
llama_new_context_with_model: freq_scale = 1
   1  犹もｸ巵ｸ｣犹・2犹もｸ巵ｸ｣犹・ 2  1  1犹もｸ巵ｸ｣犹・縺励▲縺九ｊ縺ｨ 2  2 犧・2犹もｸ巵ｸ｣犹・

llm_perf_print:        load time =   10377.00 ms
llm_perf_print: prompt eval time =    3651.00 ms /     4 tokens (  912.75 ms per token,     1.10 tokens per second)
llm_perf_print:        eval time =    2941.00 ms /    31 runs   (   94.87 ms per token,    10.54 tokens per second)
llm_perf_print:       total time =   17111.00 ms /    35 tokens

I have two questions:

1. Do you plan to support newer driver versions like 32.0.100.3717 or 32.0.100.3967?
2. This is more out of intellectual curiosity, but why do such issues occur — where the response becomes inaccurate depending on the driver version? Is it because the NPU driver environment is still unstable?

[buikhoa40]

Just curious why driver version has to be exactly the same, is there some compatibility problem of the dll files provided by the driver?

[KrisNathan]

Same problem here when I use DeepSeek-R1-Distill-Qwen-7B-Q6_K.gguf model.

Interestingly, the garbled output issue doesn't seem to happen on DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf

CPU: Intel Core Ultra 7 258V
NPU Driver: 32.0.100.3764 (from device manager)

I'm well aware that the driver version I'm using may not be the recommended one, but it is very interesting.

I wonder if it has something to do with how big the model parameter is.

> .\llama-cli-npu.exe -m .\models\DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf -n 32 -p "What is AI?"
build: 1 (3ac676a) with MSVC 19.39.33519.0 for x64
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from .\models\DeepSeek-R1-Distill-Qwen-1.5B-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              = DeepSeek R1 Distill Qwen 1.5B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 1.5B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 1536
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 8960
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 12
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["─á ─á", "─á─á ─á─á", "i n", "─á t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q4_K:  169 tensors
llama_model_loader: - type q6_K:   29 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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          = 151936
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 1536
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 12
llm_load_print_meta: n_head_kv        = 2
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            = 6
llm_load_print_meta: n_embd_k_gqa     = 256
llm_load_print_meta: n_embd_v_gqa     = 256
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             = 8960
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
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     = 1.78 B
llm_load_print_meta: model size       = 1.04 GiB (5.00 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 1.5B
llm_load_print_meta: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
llm_load_print_meta: EOS token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: PAD token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: EOG token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors:        CPU buffer size =  1059.89 MiB
.........................................................................
Directory created: "C:\\Users\\USER\\Downloads\\llama-cpp-ipex-llm-2.2.0-win-npu\\NPU_models\\qwen2-28-1536-151936-Q4_0"
Directory created: "C:\\Users\\USER\\Downloads\\llama-cpp-ipex-llm-2.2.0-win-npu\\NPU_models\\qwen2-28-1536-151936-Q4_0\\model_weights"
Converting GGUF model to Q4_0 NPU model...
Model weights saved to C:\Users\USER\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu\NPU_models\qwen2-28-1536-151936-Q4_0\model_weights
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from .\models\DeepSeek-R1-Distill-Qwen-1.5B-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              = DeepSeek R1 Distill Qwen 1.5B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 1.5B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 1536
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 8960
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 12
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["─á ─á", "─á─á ─á─á", "i n", "─á t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q4_K:  169 tensors
llama_model_loader: - type q6_K:   29 tensors
llm_load_vocab: missing or unrecognized pre-tokenizer type, using: 'default'
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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          = 151936
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 1
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = all F32
llm_load_print_meta: model params     = 1.78 B
llm_load_print_meta: model size       = 1.04 GiB (5.00 BPW)
llm_load_print_meta: general.name     = DeepSeek R1 Distill Qwen 1.5B
llm_load_print_meta: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
llm_load_print_meta: EOS token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: PAD token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: EOG token        = 151643 '<｜end▁of▁sentence｜>'
llm_load_print_meta: max token length = 256
llama_model_load: vocab only - skipping tensors
Model saved to C:\Users\USER\Downloads\llama-cpp-ipex-llm-2.2.0-win-npu\NPU_models\qwen2-28-1536-151936-Q4_0//decoder_layer_0.blob
llama_new_context_with_model: n_ctx      = 1024
llama_new_context_with_model: n_batch    = 1024
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 0.0
llama_new_context_with_model: freq_scale = 1
<think>
Okay, so I'm trying to understand what AI is. From the initial answer, it seems like AI stands for artificial intelligence or machine learning, but

llm_perf_print:        load time =    1361.00 ms
llm_perf_print: prompt eval time =     707.00 ms /     7 tokens (  101.00 ms per token,     9.90 tokens per second)
llm_perf_print:        eval time =     559.00 ms /    31 runs   (   18.03 ms per token,    55.46 tokens per second)
llm_perf_print:       total time =    2671.00 ms /    38 tokens