Novamax内模型下载后无法使用

xs5522 · 2026 年5 月 14 日 16:42

00:32:26 [WARN] 未找到任何 .gguf 模型文件: C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-27B-MTP-GGUF

00:32:26 [INFO] ✓ 生成预设文件: C:\LingLong\NovaStudio\NovaMax\data\presets\llm.ini

00:32:26 [INFO] ✓ 参数已保存并更新 INI: unsloth_Qwen3.6-35B-A3B-MTP-GGUF

00:32:30 [INFO] [engineManager] 已重新加载 6 个引擎定义

00:32:30 [INFO] [remoteConfig] 引擎配置已更新

00:32:30 [INFO] [remoteConfig] 模型同步完成: 新增 0, 更新 0

00:32:34 [INFO] 使用激活文件: Qwen3.6-35B-A3B-Q8_0.gguf

00:32:34 [INFO] 启动单模型: llama-server -m C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-Q8_0.gguf --host 0.0.0.0 --ctx-size 0 --port 1234 --parallel 1 --no-mmap --n-gpu-layers 100 --temperature 0.7 --top-p 0.9 --top-k 40 --repeat-penalty 1.1 --reasoning on --flash-attn on --mmproj C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\mmproj-BF16.gguf

00:32:34 [INFO] Added engine path to PATH: C:\LingLong\NovaStudio\NovaMax\external\llamacpp\202604161710

00:32:34 [INFO] Added ROCm 7.12+2.12 to PATH: C:\LingLong\NovaStudio\NovaMax\external\rocm\7.12+2.12\Lib\site-packages\torch\lib\rocm\bin

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] ggml_cuda_init: found 1 ROCm devices (Total VRAM: 110456 MiB):
Device 0: AMD Radeon™ 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 110456 MiB

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] build_info: b8809-b1be68e8c

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] main: loading model

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] srv load_model: loading model ‘C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-Q8_0.gguf’

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_load: error loading model: missing tensor ‘blk.40.ssm_conv1d.weight’
llama_model_load_from_file_impl: failed to load model

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_params_fit: encountered an error while trying to fit params to free device memory: failed to load model
llama_params_fit: fitting params to free memory took 0.48 seconds
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon™ 8060S Graphics) (0000:c6:00.0) - 110301 MiB free

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_loader: loaded meta data with 55 key-value pairs and 753 tensors from C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-Q8_0.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 = qwen35moe
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.sampling.top_k i32 = 20
llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000
llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000
llama_model_loader: - kv 5: general.name str = Qwen3.6-35B-A3B
llama_model_loader: - kv 6: general.basename str = Qwen3.6-35B-A3B
llama_model_loader: - kv 7: general.quantized_by str = Unsloth
llama_model_loader: - kv 8: general.size_label str = 35B-A3B
llama_model_loader: - kv 9: general.license str = apache-2.0
llama_model_loader: - kv 10: general.license.link str = https://huggingface.co/Qwen/Qwen3.6-3…
llama_model_loader: - kv 11: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 12: general.base_model.count u32 = 1
llama_model_loader: - kv 13: general.base_model.0.name str = Qwen3.6 35B A3B
llama_model_loader: - kv 14: general.base_model.0.organization str = Qwen
llama_model_loader: - kv 15: general.base_model.0.repo_url str = https://huggingface.co/Qwen/Qwen3.6-3…
llama_model_loader: - kv 16: general.tags arr[str,3] = [“qwen3_5_moe”, “qwen”, "image-text-t…

00:32:35 [INFO] [unsloth_Qwen llama_model_loader: - kv 18: llama_model_loader: - kv 19: llama_model_loader: - kv 20: llama_model_loader: - kv 21: llama_model_loader: - kv 22: llama_model_loader: - kv 23: llama_model_loader: - kv llama_model_loader: - kv 25: llama_model_loader: - kv 26: llama_model_loader: - kv 27: llama_model_loader: - kv 28: llama_model_loader: - kv 29: llama_model_loader: - kv llama_model_loader: - kv 31: llama_model_loader: - kv 32: llama_model_loader: - kv 33: llama_model_loader: - kv 34: llama_model_loader: - kv 35: llama_model_loader: - kv 36: llama_model_loader: - kv 37: llama_model_loader: - kv 38: llama_model_loader: - kv 39: llama_model_loader: - kv 40: 3.6-35B-A3B-MTP-GGUF] llama_model_loader: - kv 17: qwen35moe.block_count u32 = 41
qwen35moe.context_length u32 = 262144
qwen35moe.embedding_length u32 = 2048
qwen35moe.attention.head_count u32 = 16
qwen35moe.attention.head_count_kv u32 = 2
qwen35moe.rope.dimension_sections arr[i32,4] = [11, 11, 10, 0]
qwen35moe.rope.freq_base f32 = 10000000.000000
24: qwen35moe.attention.layer_norm_rms_epsilon f32 = 0.000001
qwen35moe.expert_count u32 = 256
qwen35moe.expert_used_count u32 = 8
qwen35moe.attention.key_length u32 = 256
qwen35moe.attention.value_length u32 = 256
qwen35moe.expert_feed_forward_length u32 = 512
30: qwen35moe.expert_shared_feed_forward_length u32 = 512
qwen35moe.ssm.conv_kernel u32 = 4
qwen35moe.ssm.state_size u32 = 128
qwen35moe.ssm.group_count u32 = 16
qwen35moe.ssm.time_step_rank u32 = 32
qwen35moe.ssm.inner_size u32 = 4096
qwen35moe.full_attention_interval u32 = 4
qwen35moe.rope.dimension_count u32 = 64
qwen35moe.nextn_predict_layers u32 = 1
tokenizer.ggml.model str = gpt2
tokenizer.ggml.pre str = qwen35

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_loader: - kv 41: tokenizer.ggml.tokens arr[str,248320] = [“!”, “\”", “#”, “$”, “%”, “&”, “'”, …

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_loader: - kv 42: tokenizer.ggml.token_type arr[i32,248320] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, …

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_loader: - kv 43: tokenizer.ggml.merges arr[str,247587] = [“Ġ Ġ”, “ĠĠ ĠĠ”, “i n”, “Ġ t”,…
llama_model_loader: - kv 44: tokenizer.ggml.eos_token_id u32 = 248046
llama_model_loader: - kv 45: tokenizer.ggml.padding_token_id u32 = 248055
llama_model_loader: - kv 46: tokenizer.ggml.bos_token_id u32 = 248044

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_loader: - kv 47: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 48: tokenizer.chat_template str = {%- set image_count = namespace(value…
llama_model_loader: - kv 49: general.quantization_version u32 = 2
llama_model_loader: - kv 50: general.file_type u32 = 7
llama_model_loader: - kv 51: quantize.imatrix.file str = Qwen3.6-35B-A3B-GGUF/imatrix_unsloth…
llama_model_loader: - kv 52: quantize.imatrix.dataset str = unsloth_calibration_Qwen3.6-35B-A3B.txt
llama_model_loader: - kv 53: quantize.imatrix.entries_count u32 = 510
llama_model_loader: - kv 54: quantize.imatrix.chunks_count u32 = 77
llama_model_loader: - type f32: 308 tensors
llama_model_loader: - type q8_0: 443 tensors
llama_model_loader: - type bf16: 2 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q8_0
print_info: file size = 35.19 GiB (8.51 BPW)

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] load: 0 unused tokens

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] load: special tokens cache size = 33

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] load: token to piece cache size = 1.7581 MB
print_info: arch = qwen35moe
print_info: vocab_only = 0
print_info: no_alloc = 0

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] print_info: n_ctx_train = 262144
print_info: n_embd = 2048
print_info: n_embd_inp = 2048
print_info: n_layer = 41
print_info: n_head = 16
print_info: n_head_kv = 2
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
print_info: n_gqa = 8
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: n_ff = 0
print_info: n_expert = 256
print_info: n_expert_used = 8
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 = 40
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_yarn_log_mul = 0.0000
print_info: rope_finetuned = unknown
print_info: mrope sections = [11, 11, 10, 0]
print_info: ssm_d_conv = 4
print_info: ssm_d_inner = 4096
print_info: ssm_d_state = 128
print_info: ssm_dt_rank = 32
print_info: ssm_n_group = 16
print_info: ssm_dt_b_c_rms = 0
print_info: model type = ?B
print_info: model params = 35.51 B
print_info: general.name = Qwen3.6-35B-A3B
print_info: vocab type = BPE
print_info: n_vocab = 248320
print_info: n_merges = 247587
print_info: BOS token = 248044 ‘<|endoftext|>’
print_info: EOS token = 248046 ‘<|im_end|>’
print_info: EOT token = 248046 ‘<|im_end|>’
print_info: PAD token = 248055 ‘<|vision_pad|>’
print_info: LF token = 198 ‘Ċ’
print_info: FIM PRE token = 248060 ‘<|fim_prefix|>’
print_info: FIM SUF token = 248062 ‘<|fim_suffix|>’
print_info: FIM MID token = 248061 ‘<|fim_middle|>’
print_info: FIM PAD token = 248063 ‘<|fim_pad|>’
print_info: FIM REP token = 248064 ‘<|repo_name|>’
print_info: FIM SEP token = 248065 ‘<|file_sep|>’
print_info: EOG token = 248044 ‘<|endoftext|>’
print_info: EOG token = 248046 ‘<|im_end|>’
print_info: EOG token = 248063 ‘<|fim_pad|>’
print_info: EOG token = 248064 ‘<|repo_name|>’
print_info: EOG token = 248065 ‘<|file_sep|>’
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while… (mmap = false, direct_io = false)

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] llama_model_load: error loading model: missing tensor ‘blk.40.ssm_conv1d.weight’
llama_model_load_from_file_impl: failed to load model

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] common_init_from_params: failed to load model ‘C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-Q8_0.gguf’
srv load_model: failed to load model, ‘C:\LingLong\NovaStudio\NovaMax\data\models_dir\llm\unsloth\Qwen3.6-35B-A3B-MTP-GGUF\Qwen3.6-35B-A3B-Q8_0.gguf’
srv operator(): operator(): cleaning up before exit…

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] main: exiting due to model loading error

00:32:35 [INFO] [unsloth_Qwen3.6-35B-A3B-MTP-GGUF] Process exited with code 1

00:32:35 [INFO] [service-registrar] Deregistered service_id=c66ff3cc4820e773c8f179fc97279c2a

00:32:36 [INFO] [service-registrar] Registered /v1/chat/completions on port 1234 with service_id=8781eed7d84ee5717827fb22aef186a6

xs5522 · 2026 年5 月 14 日 16:42

这是一个比之前更底层的错误。之前的参数问题（--spec-type, --spec-draft-n-max）已经解决，但现在暴露出 模型文件本身或后端架构支持 的核心问题。

核心错误：模型加载失败

报错信息: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'
含义: 推理引擎在尝试加载第 40 层（最后一层）的 SSM（状态空间模型）卷积权重时，发现文件里缺失了该数据。
结论: 这不是配置参数的问题，而是 GGUF 模型文件损坏、不完整，或者 当前运行的 llama-server 后端不支持该模型架构。

详细分析

架构特殊 (qwen35moe):
- 日志显示 general.architecture = qwen35moe。这表明这是一个基于 Qwen 的 MoE（混合专家）+ SSM（可能是 Mamba2/Qwen-MoE 变体）的新架构模型。
- 标准版的 llama-server（即使是最新版）对这种新架构的支持往往滞后，或者需要特定分支（如 Unsloth 的 Fork）。
具体缺失: blk.40.ssm_conv1d.weight。这是 SSM 模块的关键权重。如果 GGUF 文件在转换（Quantize/Convert）过程中出错，或者是下载时截断，就会出现这种情况。
时间戳疑点: 日志中显示路径 llamacpp\202604161710。如果是日期，说明是“未来”的时间？可能是版本号 build-2026...。这意味着这可能是一个内部预览版或实验性版本的引擎，对模型格式的要求非常严格。

utmoss · 2026 年5 月 15 日 03:31

mtp模型请用mtp版llamacpp引擎运行