Failed to load MTP Qwen3.6-35B-A3B-UD-Q4_K_S.gguf

LocalAI version:
LocalAI e859345 (e859345b120341da368b3601a86888203c6ae077)

Environment, CPU architecture, OS, and Version:
RTX 3090, AMD Ryzen

Describe the bug
Failed to load the model

Error: failed to load model with internal loader: could not load model: rpc error: code = Internal desc = Failed to load model: /models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf. Error: llama_model_load: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'; llama_model_load_from_file_impl: failed to load model; llama_params_fit: encountered an error while trying to fit params to free device memory: failed to load model; llama_model_load: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'; llama_model_load_from_file_impl: failed to load model
Traces für Details anzeigen

To Reproduce
download Qwen3.6-35B-A3B-UD-Q4_K_S.gguf and try to start

Expected behavior
loads successfully

Logs
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 24119 MiB):
10:22:48.381stderr Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes, VRAM: 24119 MiB
10:22:48.382stderrsystem_info: n_threads = 16 / 32 | CUDA : ARCHS = 750,800,860,890,1200,1210 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
10:22:48.382stderr
10:22:48.382stderrsrv load_model: loading model '/models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf'
10:22:48.382stderrcommon_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
10:22:48.717stderrllama_model_load: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'
10:22:48.717stderrllama_model_load_from_file_impl: failed to load model
10:22:48.748stderrllama_params_fit: encountered an error while trying to fit params to free device memory: failed to load model
10:22:48.748stderrllama_params_fit: fitting params to free memory took 0.37 seconds
10:22:48.748stderrllama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) (0000:0a:00.0) - 23821 MiB free
10:22:48.775stderrllama_model_loader: loaded meta data with 55 key-value pairs and 753 tensors from /models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf (version GGUF V3 (latest))
10:22:48.775stderrllama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
10:22:48.775stderrllama_model_loader: - kv 0: general.architecture str = qwen35moe
10:22:48.775stderrllama_model_loader: - kv 1: general.type str = model
10:22:48.775stderrllama_model_loader: - kv 2: general.sampling.top_k i32 = 20
10:22:48.775stderrllama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000
10:22:48.775stderrllama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000
10:22:48.775stderrllama_model_loader: - kv 5: general.name str = Qwen3.6-35B-A3B
10:22:48.775stderrllama_model_loader: - kv 6: general.basename str = Qwen3.6-35B-A3B
10:22:48.775stderrllama_model_loader: - kv 7: general.quantized_by str = Unsloth
10:22:48.775stderrllama_model_loader: - kv 8: general.size_label str = 35B-A3B
10:22:48.775stderrllama_model_loader: - kv 9: general.license str = apache-2.0
10:22:48.775stderrllama_model_loader: - kv 10: general.license.link str = https://huggingface.co/Qwen/Qwen3.6-3...
10:22:48.775stderrllama_model_loader: - kv 11: general.repo_url str = https://huggingface.co/unsloth
10:22:48.775stderrllama_model_loader: - kv 12: general.base_model.count u32 = 1
10:22:48.775stderrllama_model_loader: - kv 13: general.base_model.0.name str = Qwen3.6 35B A3B
10:22:48.775stderrllama_model_loader: - kv 14: general.base_model.0.organization str = Qwen
10:22:48.775stderrllama_model_loader: - kv 15: general.base_model.0.repo_url str = https://huggingface.co/Qwen/Qwen3.6-3...
10:22:48.775stderrllama_model_loader: - kv 16: general.tags arr[str,3] = ["qwen3_5_moe", "qwen", "image-text-t...
10:22:48.775stderrllama_model_loader: - kv 17: qwen35moe.block_count u32 = 41
10:22:48.775stderrllama_model_loader: - kv 18: qwen35moe.context_length u32 = 262144
10:22:48.775stderrllama_model_loader: - kv 19: qwen35moe.embedding_length u32 = 2048
10:22:48.775stderrllama_model_loader: - kv 20: qwen35moe.attention.head_count u32 = 16
10:22:48.775stderrllama_model_loader: - kv 21: qwen35moe.attention.head_count_kv u32 = 2
10:22:48.775stderrllama_model_loader: - kv 22: qwen35moe.rope.dimension_sections arr[i32,4] = [11, 11, 10, 0]
10:22:48.775stderrllama_model_loader: - kv 23: qwen35moe.rope.freq_base f32 = 10000000.000000
10:22:48.775stderrllama_model_loader: - kv 24: qwen35moe.attention.layer_norm_rms_epsilon f32 = 0.000001
10:22:48.775stderrllama_model_loader: - kv 25: qwen35moe.expert_count u32 = 256
10:22:48.775stderrllama_model_loader: - kv 26: qwen35moe.expert_used_count u32 = 8
10:22:48.775stderrllama_model_loader: - kv 27: qwen35moe.attention.key_length u32 = 256
10:22:48.775stderrllama_model_loader: - kv 28: qwen35moe.attention.value_length u32 = 256
10:22:48.775stderrllama_model_loader: - kv 29: qwen35moe.expert_feed_forward_length u32 = 512
10:22:48.775stderrllama_model_loader: - kv 30: qwen35moe.expert_shared_feed_forward_length u32 = 512
10:22:48.775stderrllama_model_loader: - kv 31: qwen35moe.ssm.conv_kernel u32 = 4
10:22:48.775stderrllama_model_loader: - kv 32: qwen35moe.ssm.state_size u32 = 128
10:22:48.775stderrllama_model_loader: - kv 33: qwen35moe.ssm.group_count u32 = 16
10:22:48.775stderrllama_model_loader: - kv 34: qwen35moe.ssm.time_step_rank u32 = 32
10:22:48.775stderrllama_model_loader: - kv 35: qwen35moe.ssm.inner_size u32 = 4096
10:22:48.775stderrllama_model_loader: - kv 36: qwen35moe.full_attention_interval u32 = 4
10:22:48.775stderrllama_model_loader: - kv 37: qwen35moe.rope.dimension_count u32 = 64
10:22:48.775stderrllama_model_loader: - kv 38: qwen35moe.nextn_predict_layers u32 = 1
10:22:48.775stderrllama_model_loader: - kv 39: tokenizer.ggml.model str = gpt2
10:22:48.775stderrllama_model_loader: - kv 40: tokenizer.ggml.pre str = qwen35
10:22:48.792stderrllama_model_loader: - kv 41: tokenizer.ggml.tokens arr[str,248320] = ["!", """, "#", "$", "%", "&", "'", ...
10:22:48.797stderrllama_model_loader: - kv 42: tokenizer.ggml.token_type arr[i32,248320] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
10:22:48.813stderrllama_model_loader: - kv 43: tokenizer.ggml.merges arr[str,247587] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
10:22:48.813stderrllama_model_loader: - kv 44: tokenizer.ggml.eos_token_id u32 = 248046
10:22:48.813stderrllama_model_loader: - kv 45: tokenizer.ggml.padding_token_id u32 = 248055
10:22:48.813stderrllama_model_loader: - kv 46: tokenizer.ggml.bos_token_id u32 = 248044
10:22:48.813stderrllama_model_loader: - kv 47: tokenizer.ggml.add_bos_token bool = false
10:22:48.813stderrllama_model_loader: - kv 48: tokenizer.chat_template str = {%- set image_count = namespace(value...
10:22:48.813stderrllama_model_loader: - kv 49: general.quantization_version u32 = 2
10:22:48.813stderrllama_model_loader: - kv 50: general.file_type u32 = 14
10:22:48.813stderrllama_model_loader: - kv 51: quantize.imatrix.file str = Qwen3.6-35B-A3B-GGUF/imatrix_unsloth....
10:22:48.813stderrllama_model_loader: - kv 52: quantize.imatrix.dataset str = unsloth_calibration_Qwen3.6-35B-A3B.txt
10:22:48.813stderrllama_model_loader: - kv 53: quantize.imatrix.entries_count u32 = 510
10:22:48.813stderrllama_model_loader: - kv 54: quantize.imatrix.chunks_count u32 = 77
10:22:48.813stderrllama_model_loader: - type f32: 368 tensors
10:22:48.813stderrllama_model_loader: - type q8_0: 259 tensors
10:22:48.813stderrllama_model_loader: - type q4_K: 120 tensors
10:22:48.813stderrllama_model_loader: - type q6_K: 4 tensors
10:22:48.813stderrllama_model_loader: - type bf16: 2 tensors
10:22:48.813stderrprint_info: file format = GGUF V3 (latest)
10:22:48.813stderrprint_info: file type = Q4_K - Small
10:22:48.813stderrprint_info: file size = 19.91 GiB (4.82 BPW)
10:22:48.891stderrload: 0 unused tokens
10:22:48.910stderrload: printing all EOG tokens:
10:22:48.910stderrload: - 248044 ('<|endoftext|>')
10:22:48.910stderrload: - 248046 ('<|im_end|>')
10:22:48.910stderrload: - 248063 ('<|fim_pad|>')
10:22:48.910stderrload: - 248064 ('<|repo_name|>')
10:22:48.910stderrload: - 248065 ('<|file_sep|>')
10:22:48.910stderrload: special tokens cache size = 33
10:22:48.969stderrload: token to piece cache size = 1.7581 MB
10:22:48.969stderrprint_info: arch = qwen35moe
10:22:48.969stderrprint_info: vocab_only = 0
10:22:48.969stderrprint_info: no_alloc = 0
10:22:48.969stderrprint_info: n_ctx_train = 262144
10:22:48.969stderrprint_info: n_embd = 2048
10:22:48.969stderrprint_info: n_embd_inp = 2048
10:22:48.969stderrprint_info: n_layer = 41
10:22:48.969stderrprint_info: n_head = 16
10:22:48.969stderrprint_info: n_head_kv = 2
10:22:48.969stderrprint_info: n_rot = 64
10:22:48.969stderrprint_info: n_swa = 0
10:22:48.969stderrprint_info: is_swa_any = 0
10:22:48.969stderrprint_info: n_embd_head_k = 256
10:22:48.969stderrprint_info: n_embd_head_v = 256
10:22:48.969stderrprint_info: n_gqa = 8
10:22:48.969stderrprint_info: n_embd_k_gqa = 512
10:22:48.969stderrprint_info: n_embd_v_gqa = 512
10:22:48.969stderrprint_info: f_norm_eps = 0.0e+00
10:22:48.969stderrprint_info: f_norm_rms_eps = 1.0e-06
10:22:48.969stderrprint_info: f_clamp_kqv = 0.0e+00
10:22:48.969stderrprint_info: f_max_alibi_bias = 0.0e+00
10:22:48.969stderrprint_info: f_logit_scale = 0.0e+00
10:22:48.969stderrprint_info: f_attn_scale = 0.0e+00
10:22:48.969stderrprint_info: n_ff = 0
10:22:48.969stderrprint_info: n_expert = 256
10:22:48.969stderrprint_info: n_expert_used = 8
10:22:48.969stderrprint_info: n_expert_groups = 0
10:22:48.969stderrprint_info: n_group_used = 0
10:22:48.969stderrprint_info: causal attn = 1
10:22:48.969stderrprint_info: pooling type = -1
10:22:48.969stderrprint_info: rope type = 40
10:22:48.969stderrprint_info: rope scaling = linear
10:22:48.969stderrprint_info: freq_base_train = 10000000.0
10:22:48.969stderrprint_info: freq_scale_train = 1
10:22:48.969stderrprint_info: n_ctx_orig_yarn = 262144
10:22:48.969stderrprint_info: rope_yarn_log_mul = 0.0000
10:22:48.969stderrprint_info: rope_finetuned = unknown
10:22:48.969stderrprint_info: mrope sections = [11, 11, 10, 0]
10:22:48.969stderrprint_info: ssm_d_conv = 4
10:22:48.969stderrprint_info: ssm_d_inner = 4096
10:22:48.969stderrprint_info: ssm_d_state = 128
10:22:48.969stderrprint_info: ssm_dt_rank = 32
10:22:48.969stderrprint_info: ssm_n_group = 16
10:22:48.969stderrprint_info: ssm_dt_b_c_rms = 0
10:22:48.969stderrprint_info: model type = ?B
10:22:48.969stderrprint_info: model params = 35.51 B
10:22:48.969stderrprint_info: general.name = Qwen3.6-35B-A3B
10:22:48.969stderrprint_info: vocab type = BPE
10:22:48.969stderrprint_info: n_vocab = 248320
10:22:48.969stderrprint_info: n_merges = 247587
10:22:48.969stderrprint_info: BOS token = 248044 '<|endoftext|>'
10:22:48.969stderrprint_info: EOS token = 248046 '<|im_end|>'
10:22:48.969stderrprint_info: EOT token = 248046 '<|im_end|>'
10:22:48.969stderrprint_info: PAD token = 248055 '<|vision_pad|>'
10:22:48.969stderrprint_info: LF token = 198 'Ċ'
10:22:48.969stderrprint_info: FIM PRE token = 248060 '<|fim_prefix|>'
10:22:48.969stderrprint_info: FIM SUF token = 248062 '<|fim_suffix|>'
10:22:48.969stderrprint_info: FIM MID token = 248061 '<|fim_middle|>'
10:22:48.969stderrprint_info: FIM PAD token = 248063 '<|fim_pad|>'
10:22:48.969stderrprint_info: FIM REP token = 248064 '<|repo_name|>'
10:22:48.969stderrprint_info: FIM SEP token = 248065 '<|file_sep|>'
10:22:48.969stderrprint_info: EOG token = 248044 '<|endoftext|>'
10:22:48.969stderrprint_info: EOG token = 248046 '<|im_end|>'
10:22:48.969stderrprint_info: EOG token = 248063 '<|fim_pad|>'
10:22:48.969stderrprint_info: EOG token = 248064 '<|repo_name|>'
10:22:48.969stderrprint_info: EOG token = 248065 '<|file_sep|>'
10:22:48.969stderrprint_info: max token length = 256
10:22:48.969stderrload_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
10:22:48.975stderrllama_model_load: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'
10:22:48.975stderrllama_model_load_from_file_impl: failed to load model
10:22:49.006stderrcommon_init_from_params: failed to load model '/models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf'
10:22:49.006stderrsrv load_model: failed to load model, '/models/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf'

Start by reproducing the failure described in the issue: download Qwen3.6-35B-A3B-UD-Q4_K_S.gguf and attempt to start it through LocalAI's internal loader. Use the llama_model_load and llama_params_fit errors in the logs to trace the model-loading path. Done means the model loads successfully on the reported RTX 3090 environment.