Name and Version

Git commit b90b5e0

build/bin/llama-server --webui-mcp-proxy --alias llamacpp-model -m ../models/GLM-4.7-Flash-Q4_K_M.gguf --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 --kv-unified -ctk turbo4 -ctv turbo4 --ctx-size 202752 --no-mmap --jinja -np 1 --host 0.0.0.0 --port 18080 --alias llamacpp-model -ngl 99

ggml_cuda_init: found 4 CUDA devices (Total VRAM: 80219 MiB):
Device 0: NVIDIA GeForce RTX 3080, compute capability 8.6, VMM: yes, VRAM: 20054 MiB
Device 1: NVIDIA GeForce RTX 3080, compute capability 8.6, VMM: yes, VRAM: 20054 MiB
Device 2: NVIDIA GeForce RTX 3080, compute capability 8.6, VMM: yes, VRAM: 20054 MiB
Device 3: NVIDIA GeForce RTX 3080, compute capability 8.6, VMM: yes, VRAM: 20054 MiB
register_backend: registered backend CUDA (4 devices)
register_device: registered device CUDA0 (NVIDIA GeForce RTX 3080)
register_device: registered device CUDA1 (NVIDIA GeForce RTX 3080)
register_device: registered device CUDA2 (NVIDIA GeForce RTX 3080)
register_device: registered device CUDA3 (NVIDIA GeForce RTX 3080)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (AMD Ryzen 5 5500GT with Radeon Graphics)
DEPRECATED: argument '--alias' specified multiple times, use comma-separated values instead (only last value will be used)
build: 8653 (b90b5e0) with GNU 13.3.0 for Linux x86_64 (debug)
system info: n_threads = 6, n_threads_batch = 6, total_threads = 12

system_info: n_threads = 6 (n_threads_batch = 6) / 12 | CUDA : ARCHS = 860 | 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 |

Running without SSL
init: using 11 threads for HTTP server
srv main: -----------------
srv main: CORS proxy is enabled, do not expose server to untrusted environments
srv main: This feature is EXPERIMENTAL and may be removed or changed in future versions
srv main: -----------------
start: binding port with default address family
main: loading model
srv load_model: loading model '../models/GLM-4.7-Flash-Q4_K_M.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
llama_init_from_model: K cache type turbo4 with block size 128 does not divide n_embd_head_k=576
llama_params_fit: encountered an error while trying to fit params to free device memory: failed to create llama_context from model
llama_params_fit: fitting params to free memory took 1.71 seconds
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3080) (0000:15:00.0) - 19830 MiB free
llama_model_load_from_file_impl: using device CUDA1 (NVIDIA GeForce RTX 3080) (0000:21:00.0) - 19830 MiB free
llama_model_load_from_file_impl: using device CUDA2 (NVIDIA GeForce RTX 3080) (0000:22:00.0) - 19830 MiB free
llama_model_load_from_file_impl: using device CUDA3 (NVIDIA GeForce RTX 3080) (0000:29:00.0) - 19830 MiB free
llama_model_loader: loaded meta data with 59 key-value pairs and 844 tensors from ../models/GLM-4.7-Flash-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 = deepseek2
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.sampling.temp f32 = 1.000000
llama_model_loader: - kv 3: general.name str = Glm-4.7-Flash
llama_model_loader: - kv 4: general.basename str = Glm-4.7-Flash
llama_model_loader: - kv 5: general.quantized_by str = Unsloth
llama_model_loader: - kv 6: general.size_label str = 64x2.6B
llama_model_loader: - kv 7: general.license str = mit
llama_model_loader: - kv 8: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 9: general.base_model.count u32 = 1
llama_model_loader: - kv 10: general.base_model.0.name str = GLM 4.7 Flash
llama_model_loader: - kv 11: general.base_model.0.organization str = Zai Org
llama_model_loader: - kv 12: general.base_model.0.repo_url str = https://huggingface.co/zai-org/GLM-4....
llama_model_loader: - kv 13: general.tags arr[str,2] = ["unsloth", "text-generation"]
llama_model_loader: - kv 14: general.languages arr[str,2] = ["en", "zh"]
llama_model_loader: - kv 15: deepseek2.block_count u32 = 47
llama_model_loader: - kv 16: deepseek2.context_length u32 = 202752
llama_model_loader: - kv 17: deepseek2.embedding_length u32 = 2048
llama_model_loader: - kv 18: deepseek2.feed_forward_length u32 = 10240
llama_model_loader: - kv 19: deepseek2.attention.head_count u32 = 20
llama_model_loader: - kv 20: deepseek2.attention.head_count_kv u32 = 1
llama_model_loader: - kv 21: deepseek2.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 22: deepseek2.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 23: deepseek2.expert_used_count u32 = 4
llama_model_loader: - kv 24: deepseek2.expert_group_count u32 = 1
llama_model_loader: - kv 25: deepseek2.expert_group_used_count u32 = 1
llama_model_loader: - kv 26: deepseek2.expert_gating_func u32 = 2
llama_model_loader: - kv 27: deepseek2.leading_dense_block_count u32 = 1
llama_model_loader: - kv 28: deepseek2.vocab_size u32 = 154880
llama_model_loader: - kv 29: deepseek2.attention.q_lora_rank u32 = 768
llama_model_loader: - kv 30: deepseek2.attention.kv_lora_rank u32 = 512
llama_model_loader: - kv 31: deepseek2.attention.key_length u32 = 576
llama_model_loader: - kv 32: deepseek2.attention.value_length u32 = 512
llama_model_loader: - kv 33: deepseek2.attention.key_length_mla u32 = 256
llama_model_loader: - kv 34: deepseek2.attention.value_length_mla u32 = 256
llama_model_loader: - kv 35: deepseek2.expert_feed_forward_length u32 = 1536
llama_model_loader: - kv 36: deepseek2.expert_count u32 = 64
llama_model_loader: - kv 37: deepseek2.expert_shared_count u32 = 1
llama_model_loader: - kv 38: deepseek2.expert_weights_scale f32 = 1.800000
llama_model_loader: - kv 39: deepseek2.expert_weights_norm bool = true
llama_model_loader: - kv 40: deepseek2.rope.dimension_count u32 = 64
llama_model_loader: - kv 41: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 42: tokenizer.ggml.pre str = glm4
llama_model_loader: - kv 43: tokenizer.ggml.tokens arr[str,154880] = ["!", """, "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 44: tokenizer.ggml.token_type arr[i32,154880] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 45: tokenizer.ggml.merges arr[str,321649] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 46: tokenizer.ggml.eos_token_id u32 = 154820
llama_model_loader: - kv 47: tokenizer.ggml.padding_token_id u32 = 154821
llama_model_loader: - kv 48: tokenizer.ggml.bos_token_id u32 = 154822
llama_model_loader: - kv 49: tokenizer.ggml.eot_token_id u32 = 154827
llama_model_loader: - kv 50: tokenizer.ggml.unknown_token_id u32 = 154820
llama_model_loader: - kv 51: tokenizer.ggml.eom_token_id u32 = 154829
llama_model_loader: - kv 52: tokenizer.chat_template str = [gMASK]\n{%- if tools -%}\n<|syste...
llama_model_loader: - kv 53: general.quantization_version u32 = 2
llama_model_loader: - kv 54: general.file_type u32 = 15
llama_model_loader: - kv 55: quantize.imatrix.file str = GLM-4.7-Flash-GGUF/imatrix_unsloth.gguf
llama_model_loader: - kv 56: quantize.imatrix.dataset str = unsloth_calibration_GLM-4.7-Flash.txt
llama_model_loader: - kv 57: quantize.imatrix.entries_count u32 = 607
llama_model_loader: - kv 58: quantize.imatrix.chunks_count u32 = 85
llama_model_loader: - type f32: 281 tensors
llama_model_loader: - type q8_0: 141 tensors
llama_model_loader: - type q4_K: 260 tensors
llama_model_loader: - type q5_K: 92 tensors
llama_model_loader: - type q6_K: 70 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 17.05 GiB (4.89 BPW)
load: 0 unused tokens
load: special_eot_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect
load: printing all EOG tokens:
load: - 154820 ('<|endoftext|>')
load: - 154827 ('<|user|>')
load: - 154829 ('<|observation|>')
load: special tokens cache size = 36
load: token to piece cache size = 0.9811 MB
print_info: arch = deepseek2
print_info: vocab_only = 0
print_info: no_alloc = 0
print_info: n_ctx_train = 202752
print_info: n_embd = 2048
print_info: n_embd_inp = 2048
print_info: n_layer = 47
print_info: n_head = 20
print_info: n_head_kv = 1
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 576
print_info: n_embd_head_v = 512
print_info: n_gqa = 20
print_info: n_embd_k_gqa = 576
print_info: n_embd_v_gqa = 512
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
Run 1/1 (batch size 1)...
Error during run: Cannot connect to host 127.0.01:18080 ssl:default [Connect call failed ('127.0.0.1', 18080)]
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 10240
print_info: n_expert = 64
Running test: pp=2048, tg=32, depth=143360, concurrency=1
Run 1/1 (batch size 1)...
print_info: n_expert_groups = 1
print_info: n_group_used = 1
print_info: causal attn = 1
print_info: pooling type = -1
print_info: rope type = 0
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 202752
print_info: rope_yarn_log_mul = 0.0000
print_info: rope_finetuned = unknown
print_info: model type = 30B.A3B
print_info: model params = 29.94 B
print_info: general.name = Glm-4.7-Flash
print_info: n_layer_dense_lead = 1
print_info: n_lora_q = 768
print_info: n_lora_kv = 512
print_info: n_embd_head_k_mla = 256
print_info: n_embd_head_v_mla = 256
print_info: n_ff_exp = 1536
print_info: n_expert_shared = 1
print_info: expert_weights_scale = 1.8
print_info: expert_weights_norm = 1
print_info: expert_gating_func = sigmoid
print_info: vocab type = BPE
print_info: n_vocab = 154880
print_info: n_merges = 321649
print_info: BOS token = 154822 '[gMASK]'
print_info: EOS token = 154820 '<|endoftext|>'
print_info: EOT token = 154827 '<|user|>'
print_info: EOM token = 154829 '<|observation|>'
print_info: UNK token = 154820 '<|endoftext|>'
print_info: PAD token = 154821 '[MASK]'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 154838 '<|code_prefix|>'
print_info: FIM SUF token = 154840 '<|code_suffix|>'
print_info: FIM MID token = 154839 '<|code_middle|>'
print_info: EOG token = 154820 '<|endoftext|>'
print_info: EOG token = 154827 '<|user|>'
print_info: EOG token = 154829 '<|observation|>'
print_info: max token length = 1024
load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 46 repeating layers to GPU
load_tensors: offloaded 48/48 layers to GPU
load_tensors: CPU model buffer size = 170.16 MiB
load_tensors: CUDA0 model buffer size = 4151.17 MiB
load_tensors: CUDA1 model buffer size = 4344.35 MiB
load_tensors: CUDA2 model buffer size = 4344.35 MiB
load_tensors: CUDA3 model buffer size = 4444.97 MiB
....................................................................................................
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|user|> logit bias = -inf
common_init_result: added <|observation|> logit bias = -inf
Segmentation fault (core dumped)

Operating systems

Linux

GGML backends

CUDA

Hardware

Ryzen 5500 , 4x RTX 3080 20GB

Models

https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF

Problem description & steps to reproduce

Loading the model works, but with the first call, it core dumps:

What your log is actually showing

This is not a classic “analyze this core file” situation; it is a startup failure with a clear first error, then a segfault at the end.

1. Real incompatibility (now fixed in tree)

K cache type turbo4 with block size 128 does not divide n_embd_head_k=576

• Turbo4 uses a 128-wide quant block (QK_TURBO4 in ggml).
• GLM‑4.7 / DeepSeek2 here reports n_embd_head_k = 576, and 576 % 128 ≠ 0, so the early guard in llama_init_from_model rejected the config.
• The KV cache already implements turbo zero-padding to the next multiple of 128 (576 → 640) in llama-kv-cache.cpp, and the graph pads Q to match K for attention. So the init check was stricter than the real layout, and
wrongly blocked a supported case.

That is why llama_params_fit blew up with failed to create llama_context from model: it probes memory by calling llama_init_from_model, which hit that error.

2. Segfault after weight load

After tensors load and EOG logit biases are printed, the next heavy step is building the context (KV / sched / graphs). A segfault there often means UB in CUDA/FA or an inconsistent path; cleaning up the bogus “can’t init”
failure makes the pipeline sane again.
I aligned the flash-attn + quantized K/V divisibility checks with the same padding rule as the KV cache so turbo2/3/4 + non‑128‑aligned head dims are validated on the padded head size (640 for K in your case), not the raw
576.

────────────────────────────────────────

Code change: src/llama-context.cpp — K/V checks now apply the same turbo 128-alignment padding as llama_kv_cache (and skip V padding when is_mla(), matching the cache). The project builds successfully after the change.
Rebuild llama-server and rerun your command; if anything still crashes, capture:
gdb --args build/bin/llama-server ... → run → thread apply all bt
and we can tie it to a specific kernel/path.

First Bad Commit

No response

Relevant log output