Eval bug: Regression in Gemma 3n models starting from b7044 - fail sign NULL_POINTER_READ_c0000005_ggml-cuda.dll!Unknown

[bpawar-nvidia]

Name and Version

b7044

Operating systems

Windows

GGML backends

CUDA

Hardware

RTX 4090, RTX 5090

Models

Gemma 3n 2B - https://huggingface.co/ggml-org/gemma-3n-E2B-it-GGUF
Gemma 3n 4B - https://huggingface.co/ggml-org/gemma-3n-E4B-it-GGUF

Problem description & steps to reproduce

1. Prepare Windows 11 system with Nvidia RTX 4090 or RTX 5090 gpu.
2. Install Nvidia gpu driver.
3. Install llama.cpp release b7044 or higher.
4. Execute this command -

C:\Windows\System32>"D:\Apps\WinAI\LlamaCppBuilds\b7044\llama-server.exe" -m "D:\Apps\dxgperf_appsdata\dxml_models\ModelOpt\llm\gemma_3n_2b\int8_llamacpp_gguf\gemma-3n-E2B-it-Q8_0.gguf" -fa on -c 1000 --port 8000

llama-server will crash.
Llama-bench is also crashing.

First Bad Commit

b7044

Relevant log output

C:\Windows\System32>"D:\Apps\WinAI\LlamaCppBuilds\b7044\llama-server.exe" -m "D:\Apps\dxgperf_appsdata\dxml_models\ModelOpt\llm\gemma_3n_2b\int8_llamacpp_gguf\gemma-3n-E2B-it-Q8_0.gguf" -fa  on -c 1000 --port 8000
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
load_backend: loaded CUDA backend from D:\Apps\WinAI\LlamaCppBuilds\b7044\ggml-cuda.dll
load_backend: loaded RPC backend from D:\Apps\WinAI\LlamaCppBuilds\b7044\ggml-rpc.dll
load_backend: loaded CPU backend from D:\Apps\WinAI\LlamaCppBuilds\b7044\ggml-cpu-icelake.dll
main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)
?[0mbuild: 7044 (a90eb94ca) with clang version 19.1.5 for x86_64-pc-windows-msvc
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 8000, http threads: 15
main: loading model
srv    load_model: loading model 'D:\Apps\dxgperf_appsdata\dxml_models\ModelOpt\llm\gemma_3n_2b\int8_llamacpp_gguf\gemma-3n-E2B-it-Q8_0.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 5090) (0000:01:00.0) - 30994 MiB free
llama_model_loader: loaded meta data with 42 key-value pairs and 727 tensors from D:\Apps\dxgperf_appsdata\dxml_models\ModelOpt\llm\gemma_3n_2b\int8_llamacpp_gguf\gemma-3n-E2B-it-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              = gemma3n
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                         general.size_label str              = 4.5B
llama_model_loader: - kv   3:                            general.license str              = gemma
llama_model_loader: - kv   4:                   general.base_model.count u32              = 1
llama_model_loader: - kv   5:                  general.base_model.0.name str              = Gemma 3n E4b It
llama_model_loader: - kv   6:          general.base_model.0.organization str              = Google
llama_model_loader: - kv   7:              general.base_model.0.repo_url str              = https://huggingface.co/google/gemma-3...
llama_model_loader: - kv   8:                               general.tags arr[str,5]       = ["automatic-speech-recognition", "aut...
llama_model_loader: - kv   9:                     gemma3n.context_length u32              = 32768
llama_model_loader: - kv  10:                   gemma3n.embedding_length u32              = 2048
llama_model_loader: - kv  11:                        gemma3n.block_count u32              = 30
llama_model_loader: - kv  12:                gemma3n.feed_forward_length u32              = 8192
llama_model_loader: - kv  13:               gemma3n.attention.head_count u32              = 8
llama_model_loader: - kv  14:   gemma3n.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  15:               gemma3n.attention.key_length u32              = 256
llama_model_loader: - kv  16:             gemma3n.attention.value_length u32              = 256
llama_model_loader: - kv  17:                     gemma3n.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  18:           gemma3n.attention.sliding_window u32              = 512
llama_model_loader: - kv  19:            gemma3n.attention.head_count_kv u32              = 2
llama_model_loader: - kv  20:                   gemma3n.altup.active_idx u32              = 0
llama_model_loader: - kv  21:                   gemma3n.altup.num_inputs u32              = 4
llama_model_loader: - kv  22:   gemma3n.embedding_length_per_layer_input u32              = 256
llama_model_loader: - kv  23:         gemma3n.attention.shared_kv_layers f32              = 10.000000
llama_model_loader: - kv  24:          gemma3n.activation_sparsity_scale arr[f32,30]      = [1.644853, 1.644853, 1.644853, 1.6448...
llama_model_loader: - kv  25:   gemma3n.attention.sliding_window_pattern arr[bool,30]     = [true, true, true, true, false, true,...
llama_model_loader: - kv  26:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv  27:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  28:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  29:                      tokenizer.ggml.tokens arr[str,262144]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  30:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  31:                  tokenizer.ggml.token_type arr[i32,262144]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  32:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  33:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  34:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  36:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  37:               tokenizer.ggml.add_sep_token bool             = false
llama_model_loader: - kv  38:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  39:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  40:               general.quantization_version u32              = 2
llama_model_loader: - kv  41:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  362 tensors
llama_model_loader: - type  f16:   93 tensors
llama_model_loader: - type q8_0:  272 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 4.45 GiB (8.59 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
?[0mload: printing all EOG tokens:
load:   - 1 ('<eos>')
load:   - 106 ('<end_of_turn>')
load: special tokens cache size = 6414
load: token to piece cache size = 1.9446 MB
print_info: arch             = gemma3n
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 2048
print_info: n_embd_inp       = 2048
print_info: n_layer          = 30
print_info: n_head           = 8
print_info: n_head_kv        = 2
print_info: n_rot            = 256
print_info: n_swa            = 512
print_info: is_swa_any       = 1
print_info: n_embd_head_k    = 256
print_info: n_embd_head_v    = 256
print_info: n_gqa            = 4
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     = 1.0e+00
print_info: n_ff             = 8192
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_finetuned   = unknown
print_info: model type       = E2B
print_info: model params     = 4.46 B
print_info: general.name     = n/a
print_info: vocab type       = SPM
print_info: n_vocab          = 262144
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 1 '<eos>'
print_info: EOT token        = 106 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 248 '<0x0A>'
print_info: EOG token        = 1 '<eos>'
print_info: EOG token        = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 30 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 31/31 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   544.00 MiB
load_tensors:        CUDA0 model buffer size =  4560.06 MiB
..........................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 4
llama_context: n_ctx         = 1024
llama_context: n_ctx_seq     = 1024
llama_context: n_batch       = 1000
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (1024) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
?[0mllama_context:  CUDA_Host  output buffer size =     4.00 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 1024 cells
llama_kv_cache:      CUDA0 KV buffer size =     8.00 MiB
llama_kv_cache: size =    8.00 MiB (  1024 cells,   4 layers,  4/1 seqs), K (f16):    4.00 MiB, V (f16):    4.00 MiB
llama_kv_cache_iswa: creating     SWA KV cache, size = 1024 cells
llama_kv_cache:      CUDA0 KV buffer size =    32.00 MiB
llama_kv_cache: size =   32.00 MiB (  1024 cells,  16 layers,  4/1 seqs), K (f16):   16.00 MiB, V (f16):   16.00 MiB
llama_context:      CUDA0 compute buffer size =   520.00 MiB
llama_context:  CUDA_Host compute buffer size =     8.02 MiB
llama_context: graph nodes  = 2733
llama_context: graph splits = 2
common_init_from_params: added <eos> logit bias = -inf
common_init_from_params: added <end_of_turn> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 1024
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
?[0m
C:\Windows\System32>

[CISC]

Just to verify, try defining GGML_CUDA_DISABLE_FUSION env variable and see if it still crashes.

[bpawar-nvidia]

Thanks CISC.

try defining GGML_CUDA_DISABLE_FUSION env variable

It fixes the issue. But as per our engineer, this will likely have perf impact on NV gpus.

What is the root cause of regression introduced with b7044 ?

[CISC]

try defining GGML_CUDA_DISABLE_FUSION env variable

It fixes the issue. But as per our engineer, this will likely have perf impact on NV gpus.

What is the root cause of regression introduced with b7044 ?

It's likely that it fuses incorrectly for this model, and/or is just not considering some op param.

Hopefully @am17an can take a look.

[am17an]

Yes I will take a look. I'm travelling at the moment but should be able to soon.

[bpawar-nvidia]

Thanks @am17an , @CISC . Just FYI. We are seeing 6% less perf due to GGML_CUDA_DISABLE_FUSION.

[am17an]

That is expected as kernel fusion is an optimisation we do. Disabling it leads to performance regression.

[bpawar-nvidia]

What change went in b7044 that is forcing us to disable kernel fusion ? Is it possible to revert that change so that we don't have to disable kernel fusion ?

[CISC]

What change went in b7044 that is forcing us to disable kernel fusion ? Is it possible to revert that change so that we don't have to disable kernel fusion ?

You don't have to do neither, just update to >=b7103 or apply #17378