Skip to content

Eval bug: Nemotron 3 Nano 30B-A3B GGUF fails at mamba-base.cpp:173: GGML_ASSERT #20570

@cpburnz

Description

@cpburnz

Name and Version

version: 8351 (77efcba)
built with GNU 15.2.1 for Linux x86_64

Operating systems

Linux

GGML backends

CPU

Hardware

AMD RYZEN AI MAX+ 395 w/ Radeon 8060S

Models

https://huggingface.co/bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF

  • bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF:IQ2_M

https://huggingface.co/unsloth/Nemotron-3-Nano-30B-A3B-GGUF

  • unsloth/Nemotron-3-Nano-30B-A3B-GGUF:Q2_K_L
  • unsloth/Nemotron-3-Nano-30B-A3B-GGUF:UD-Q4_K_XL

Problem description & steps to reproduce

When I try running a GGUF quantization of Nemotron 3 Nano, I get the following assertion error. The device does not change this error (occurs on CPU, ROCm, and Vulkan). It happens regardless of the llama.cpp executable used (llama-cli, llama-bench, and llama-completion):

$ llama-completion -ngl 0 -fit off -m bartowski_nvidia_Nemotron-3-Nano-30B-A3B-GGUF_nvidia_Nemotron-3-Nano-30B-A3B-IQ2_M.gguf
...
/home/caleb/Downloads/llama.cpp-aio/src/llama.cpp/src/models/mamba-base.cpp:173: GGML_ASSERT(d_inner % (n_group*n_embd) == 0) failed
[New LWP 43555]
[New LWP 43553]
[New LWP 43552]
[New LWP 43550]
[New LWP 43549]

This GDB supports auto-downloading debuginfo from the following URLs:
  <https://debuginfod.archlinux.org>
  <https://debuginfod.cachyos.org>
Enable debuginfod for this session? (y or [n]) [answered N; input not from terminal]
Debuginfod has been disabled.
To make this setting permanent, add 'set debuginfod enabled off' to .gdbinit.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/libthread_db.so.1".
0x00007f7e5f2b2422 in ?? () from /usr/lib/libc.so.6
#0  0x00007f7e5f2b2422 in ?? () from /usr/lib/libc.so.6
#1  0x00007f7e5f2a5edf in ?? () from /usr/lib/libc.so.6
#2  0x00007f7e5f327fab in wait4 () from /usr/lib/libc.so.6
#3  0x00007f7e605763ef in ggml_print_backtrace () from /usr/lib/libggml-base.so.0
#4  0x00007f7e605765c3 in ggml_abort () from /usr/lib/libggml-base.so.0
#5  0x00007f7e6043e03d in llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs*, ggml_tensor*, llama_model const&, llama_ubatch const&, int) const () from /usr/lib/libllama.so.0
#6  0x00007f7e6044dcee in llm_build_nemotron_h::llm_build_nemotron_h(llama_model const&, llm_graph_params const&) () from /usr/lib/libllama.so.0
#7  0x00007f7e603accfd in llama_model::build_graph(llm_graph_params const&) const () from /usr/lib/libllama.so.0
#8  0x00007f7e602c2b45 in llama_context::graph_reserve(unsigned int, unsigned int, unsigned int, llama_memory_context_i const*, bool, unsigned long*) () from /usr/lib/libllama.so.0
#9  0x00007f7e602b1c76 in llama_context::sched_reserve() () from /usr/lib/libllama.so.0
#10 0x00007f7e602b2df0 in llama_context::llama_context(llama_model const&, llama_context_params) () from /usr/lib/libllama.so.0
#11 0x00007f7e602c3af6 in llama_init_from_model () from /usr/lib/libllama.so.0
#12 0x000055d3753808a8 in ?? ()
#13 0x000055d375384bdf in ?? ()
#14 0x000055d3752d8f7b in ?? ()
#15 0x00007f7e5f227c0e in ?? () from /usr/lib/libc.so.6
#16 0x00007f7e5f227d4b in __libc_start_main () from /usr/lib/libc.so.6
#17 0x000055d3752e47c5 in ?? ()
[Inferior 1 (process 43547) detached]
Aborted                    (core dumped) llama-completion -dev none -ngl 0 -fit off -m unsloth_Nemotron-3-Nano-30B-A3B-GGUF_Nemotron-3-Nano-30B-A3B-Q2_K_L.gguf

First Bad Commit

No response

Relevant log output

Logs
$ llama-completion -ngl 0 -fit off -m unsloth_Nemotron-3-Nano-30B-A3B-GGUF_Nemotron-3-Nano-30B-A3B-Q2_K_L.gguf
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 122880 MiB):
  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 122880 MiB (122871 MiB free)
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (RADV STRIX_HALO) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
build: 8351 (77efcba) with GNU 15.2.1 for Linux x86_64
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device ROCm0 (Radeon 8060S Graphics) (0000:c2:00.0) - 122581 MiB free
llama_model_loader: loaded meta data with 53 key-value pairs and 401 tensors from unsloth_Nemotron-3-Nano-30B-A3B-GGUF_Nemotron-3-Nano-30B-A3B-Q2_K_L.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              = nemotron_h_moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_p f32              = 1.000000
llama_model_loader: - kv   3:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   4:                               general.name str              = Nemotron-3-Nano-30B-A3B
llama_model_loader: - kv   5:                           general.basename str              = Nemotron-3-Nano-30B-A3B
llama_model_loader: - kv   6:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   7:                         general.size_label str              = 30B-A3B
llama_model_loader: - kv   8:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   9:                 nemotron_h_moe.block_count u32              = 52
llama_model_loader: - kv  10:              nemotron_h_moe.context_length u32              = 1048576
llama_model_loader: - kv  11:            nemotron_h_moe.embedding_length u32              = 2688
llama_model_loader: - kv  12:         nemotron_h_moe.feed_forward_length arr[i32,52]      = [0, 1856, 0, 1856, 0, 0, 1856, 0, 185...
llama_model_loader: - kv  13:        nemotron_h_moe.attention.head_count u32              = 32
llama_model_loader: - kv  14:     nemotron_h_moe.attention.head_count_kv arr[i32,52]      = [0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, ...
llama_model_loader: - kv  15:              nemotron_h_moe.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  16: nemotron_h_moe.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  17: nemotron_h_moe.attention.layer_norm_epsilon f32              = 0.000010
llama_model_loader: - kv  18:           nemotron_h_moe.expert_used_count u32              = 6
llama_model_loader: - kv  19:          nemotron_h_moe.expert_group_count u32              = 1
llama_model_loader: - kv  20:     nemotron_h_moe.expert_group_used_count u32              = 1
llama_model_loader: - kv  21:                  nemotron_h_moe.vocab_size u32              = 131072
llama_model_loader: - kv  22:        nemotron_h_moe.rope.dimension_count u32              = 84
llama_model_loader: - kv  23:             nemotron_h_moe.ssm.conv_kernel u32              = 4
llama_model_loader: - kv  24:              nemotron_h_moe.ssm.state_size u32              = 128
llama_model_loader: - kv  25:             nemotron_h_moe.ssm.group_count u32              = 8
llama_model_loader: - kv  26:              nemotron_h_moe.ssm.inner_size u32              = 4096
llama_model_loader: - kv  27:          nemotron_h_moe.ssm.time_step_rank u32              = 64
llama_model_loader: - kv  28:      nemotron_h_moe.rope.scaling.finetuned bool             = false
llama_model_loader: - kv  29:        nemotron_h_moe.attention.key_length u32              = 128
llama_model_loader: - kv  30:      nemotron_h_moe.attention.value_length u32              = 128
llama_model_loader: - kv  31:  nemotron_h_moe.expert_feed_forward_length u32              = 1856
llama_model_loader: - kv  32: nemotron_h_moe.expert_shared_feed_forward_length u32              = 3712
llama_model_loader: - kv  33:                nemotron_h_moe.expert_count u32              = 128
llama_model_loader: - kv  34:         nemotron_h_moe.expert_shared_count u32              = 1
llama_model_loader: - kv  35:         nemotron_h_moe.expert_weights_norm bool             = true
llama_model_loader: - kv  36:        nemotron_h_moe.expert_weights_scale f32              = 2.500000
llama_model_loader: - kv  37:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  38:                         tokenizer.ggml.pre str              = pixtral
llama_model_loader: - kv  39:                      tokenizer.ggml.tokens arr[str,131072]  = ["<unk>", "<s>", "</s>", "[INST]", "[...
llama_model_loader: - kv  40:                  tokenizer.ggml.token_type arr[i32,131072]  = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  41:                      tokenizer.ggml.merges arr[str,269443]  = ["Ä  Ä ", "Ä  t", "e r", "i n", "Ä  Ä...
llama_model_loader: - kv  42:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  43:                tokenizer.ggml.eos_token_id u32              = 11
llama_model_loader: - kv  44:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  45:            tokenizer.ggml.padding_token_id u32              = 999
llama_model_loader: - kv  46:                    tokenizer.chat_template str              = {# Unsloth template fixes #}\n{% macro...
llama_model_loader: - kv  47:               general.quantization_version u32              = 2
llama_model_loader: - kv  48:                          general.file_type u32              = 10
llama_model_loader: - kv  49:                      quantize.imatrix.file str              = Nemotron-3-Nano-30B-A3B-GGUF/imatrix_...
llama_model_loader: - kv  50:                   quantize.imatrix.dataset str              = unsloth_calibration_Nemotron-3-Nano-3...
llama_model_loader: - kv  51:             quantize.imatrix.entries_count u32              = 185
llama_model_loader: - kv  52:              quantize.imatrix.chunks_count u32              = 80
llama_model_loader: - type  f32:  237 tensors
llama_model_loader: - type q5_0:   24 tensors
llama_model_loader: - type q8_0:   24 tensors
llama_model_loader: - type q2_K:   23 tensors
llama_model_loader: - type q3_K:    6 tensors
llama_model_loader: - type iq4_nl:   87 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 16.85 GiB (4.58 BPW) 
load: 0 unused tokens
load: printing all EOG tokens:
load:   - 2 ('</s>')
load:   - 11 ('<|im_end|>')
load: special tokens cache size = 1000
load: token to piece cache size = 0.8499 MB
print_info: arch                  = nemotron_h_moe
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 1048576
print_info: n_embd                = 2688
print_info: n_embd_inp            = 2688
print_info: n_layer               = 52
print_info: n_head                = 32
print_info: n_head_kv             = [0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0]
print_info: n_rot                 = 84
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                 = [0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 0, 0, 16, 0, 0, 0, 0, 0, 0, 0, 0, 0]
print_info: n_embd_k_gqa          = [0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 0, 0, 0]
print_info: n_embd_v_gqa          = [0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 0, 0, 256, 0, 0, 0, 0, 0, 0, 0, 0, 0]
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
print_info: f_logit_scale         = 0.0e+00
print_info: f_attn_scale          = 0.0e+00
print_info: n_ff                  = [0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 1856, 0, 0, 1856, 0, 1856, 0, 1856, 0, 1856, 0, 1856]
print_info: n_expert              = 128
print_info: n_expert_used         = 6
print_info: n_expert_groups       = 1
print_info: n_group_used          = 1
print_info: causal attn           = 1
print_info: pooling type          = 0
print_info: rope type             = -1
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000.0
print_info: freq_scale_train      = 1
print_info: n_ctx_orig_yarn       = 1048576
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: ssm_d_conv            = 4
print_info: ssm_d_inner           = 4096
print_info: ssm_d_state           = 128
print_info: ssm_dt_rank           = 64
print_info: ssm_n_group           = 8
print_info: ssm_dt_b_c_rms        = 0
print_info: model type            = 31B.A3.5B
print_info: model params          = 31.58 B
print_info: general.name          = Nemotron-3-Nano-30B-A3B
print_info: f_embedding_scale     = 0.000000
print_info: f_residual_scale      = 0.000000
print_info: f_attention_scale     = 0.000000
print_info: n_ff_shexp            = 3712
print_info: vocab type            = BPE
print_info: n_vocab               = 131072
print_info: n_merges              = 269443
print_info: BOS token             = 1 '<s>'
print_info: EOS token             = 11 '<|im_end|>'
print_info: EOT token             = 11 '<|im_end|>'
print_info: UNK token             = 0 '<unk>'
print_info: PAD token             = 999 '<SPECIAL_999>'
print_info: LF token              = 1010 '�'
print_info: EOG token             = 2 '</s>'
print_info: EOG token             = 11 '<|im_end|>'
print_info: max token length      = 150
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading 0 repeating layers to GPU
load_tensors: offloaded 0/53 layers to GPU
load_tensors:   CPU_Mapped model buffer size = 17250.60 MiB
.....................................................
common_init_result: added </s> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 1048576
llama_context: n_ctx_seq     = 1048576
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 10000.0
llama_context: freq_scale    = 1
llama_context:        CPU  output buffer size =     0.50 MiB
llama_kv_cache:        CPU KV buffer size =  6144.00 MiB
llama_kv_cache: size = 6144.00 MiB (1048576 cells,   6 layers,  1/1 seqs), K (f16): 3072.00 MiB, V (f16): 3072.00 MiB
llama_memory_recurrent:        CPU RS buffer size =    47.62 MiB
llama_memory_recurrent: size =   47.62 MiB (     1 cells,  52 layers,  1 seqs), R (f32):    1.62 MiB, S (f32):   46.00 MiB
sched_reserve: reserving ...
/home/caleb/Downloads/llama.cpp-aio/src/llama.cpp/src/models/mamba-base.cpp:173: GGML_ASSERT(d_inner % (n_group*n_embd) == 0) failed
[New LWP 45138]
[New LWP 45136]
[New LWP 45135]
[New LWP 45133]
[New LWP 45132]

This GDB supports auto-downloading debuginfo from the following URLs:
  <https://debuginfod.archlinux.org>
  <https://debuginfod.cachyos.org>
Enable debuginfod for this session? (y or [n]) [answered N; input not from terminal]
Debuginfod has been disabled.
To make this setting permanent, add 'set debuginfod enabled off' to .gdbinit.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/libthread_db.so.1".
0x00007f27934b2422 in ?? () from /usr/lib/libc.so.6
#0  0x00007f27934b2422 in ?? () from /usr/lib/libc.so.6
#1  0x00007f27934a5edf in ?? () from /usr/lib/libc.so.6
#2  0x00007f2793527fab in wait4 () from /usr/lib/libc.so.6
#3  0x00007f27945633ef in ggml_print_backtrace () from /usr/lib/libggml-base.so.0
#4  0x00007f27945635c3 in ggml_abort () from /usr/lib/libggml-base.so.0
#5  0x00007f279483e03d in llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs*, ggml_tensor*, llama_model const&, llama_ubatch const&, int) const () from /usr/lib/libllama.so.0
#6  0x00007f279484dcee in llm_build_nemotron_h::llm_build_nemotron_h(llama_model const&, llm_graph_params const&) () from /usr/lib/libllama.so.0
#7  0x00007f27947accfd in llama_model::build_graph(llm_graph_params const&) const () from /usr/lib/libllama.so.0
#8  0x00007f27946c2b45 in llama_context::graph_reserve(unsigned int, unsigned int, unsigned int, llama_memory_context_i const*, bool, unsigned long*) () from /usr/lib/libllama.so.0
#9  0x00007f27946b1c76 in llama_context::sched_reserve() () from /usr/lib/libllama.so.0
#10 0x00007f27946b2df0 in llama_context::llama_context(llama_model const&, llama_context_params) () from /usr/lib/libllama.so.0
#11 0x00007f27946c3af6 in llama_init_from_model () from /usr/lib/libllama.so.0
#12 0x000055ca83e568a8 in ?? ()
#13 0x000055ca83e5abdf in ?? ()
#14 0x000055ca83daef7b in ?? ()
#15 0x00007f2793427c0e in ?? () from /usr/lib/libc.so.6
#16 0x00007f2793427d4b in __libc_start_main () from /usr/lib/libc.so.6
#17 0x000055ca83dba7c5 in ?? ()
[Inferior 1 (process 45127) detached]

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions