Eval bug: CUDA "GGML_ASSERT(stride_row % 2 == 0) failed" when FA off for certain ctx lengths #16976
Description
Name and Version
C:\Users\windo\Downloads\llama-b6937-bin-win-cuda-12.4-x64\llama-cli.exe --version
version: 6937 (48bd265)
built with clang version 19.1.5 for x86_64-pc-windows-msvc
Operating systems
Windows
GGML backends
CUDA
Hardware
CPU: 13th Gen Intel(R) Core(TM) i7-13620H
GPU: NVIDIA GeForce RTX 4060 Laptop GPU, 8GB VRAM
Models
- https://huggingface.co/lmstudio-community/gemma-3-4b-it-GGUF Q4_K_M
- https://huggingface.co/lmstudio-community/granite-4.0-micro-GGUF Q4_K_M
- https://huggingface.co/lmstudio-community/Qwen2.5-Coder-0.5B-Instruct-GGUF Q4_K_M
Problem description & steps to reproduce
If I run llama-b6868-bin-win-cuda-12.4-x64\llama-cli.exe with -fa 'off' and specific context lengths like -c 2050, -c 1058, -c 1591, I get load crash: mmf.cuh:540: GGML_ASSERT(stride_row % 2 == 0) failed
Can download CUDA release from https://github.com/ggml-org/llama.cpp/releases/tag/b6868 onwards, and run the following to observe:
llama-cli.exe -m C:/Users/windo/.lmstudio/models/lmstudio-community/Qwen2.5-Coder-0.5B-Instruct-GGUF/Qwen2.5-Coder-0.5B-Instruct-Q4_K_M.gguf -c 2050 -fa 'off'
First Bad Commit
Same command works fine from https://github.com/ggml-org/llama.cpp/releases/tag/b6866
Relevant log output
### Broken (b6868, FA OFF)
C:\Users\windo\projects\llama.cpp (HEAD)> C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\llama-cli.exe -m C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-Q4_K_M.gguf -c 2050 -fa 'off'
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 4060 Laptop GPU, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-cpu-alderlake.dll
build: 6868 (85a7d8677) with clang version 19.1.5 for x86_64-pc-windows-msvc
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4060 Laptop GPU) (0000:01:00.0) - 7099 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 362 tensors from C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-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 = granite
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Ibm Granite_Granite 4.0 Micro
llama_model_loader: - kv 3: general.size_label str = 3B
llama_model_loader: - kv 4: granite.block_count u32 = 40
llama_model_loader: - kv 5: granite.context_length u32 = 131072
llama_model_loader: - kv 6: granite.embedding_length u32 = 2560
llama_model_loader: - kv 7: granite.feed_forward_length u32 = 8192
llama_model_loader: - kv 8: granite.attention.head_count u32 = 40
llama_model_loader: - kv 9: granite.attention.head_count_kv arr[i32,40] = [8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, ...
llama_model_loader: - kv 10: granite.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 11: granite.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 12: granite.expert_count u32 = 0
llama_model_loader: - kv 13: granite.expert_used_count u32 = 0
llama_model_loader: - kv 14: granite.vocab_size u32 = 100352
llama_model_loader: - kv 15: granite.rope.dimension_count u32 = 64
llama_model_loader: - kv 16: granite.attention.scale f32 = 0.015625
llama_model_loader: - kv 17: granite.embedding_scale f32 = 12.000000
llama_model_loader: - kv 18: granite.residual_scale f32 = 0.220000
llama_model_loader: - kv 19: granite.logit_scale f32 = 10.000000
llama_model_loader: - kv 20: granite.expert_shared_feed_forward_length u32 = 8192
llama_model_loader: - kv 21: granite.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 22: granite.ssm.state_size u32 = 256
llama_model_loader: - kv 23: granite.ssm.group_count u32 = 1
llama_model_loader: - kv 24: granite.ssm.inner_size u32 = 5120
llama_model_loader: - kv 25: granite.ssm.time_step_rank u32 = 128
llama_model_loader: - kv 26: granite.rope.scaling.finetuned bool = true
llama_model_loader: - kv 27: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 28: tokenizer.ggml.pre str = dbrx
llama_model_loader: - kv 29: tokenizer.ggml.tokens arr[str,100352] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 30: tokenizer.ggml.token_type arr[i32,100352] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 31: tokenizer.ggml.merges arr[str,100000] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 32: tokenizer.ggml.bos_token_id u32 = 100257
llama_model_loader: - kv 33: tokenizer.ggml.eos_token_id u32 = 100257
llama_model_loader: - kv 34: tokenizer.ggml.unknown_token_id u32 = 100269
llama_model_loader: - kv 35: tokenizer.ggml.padding_token_id u32 = 100256
llama_model_loader: - kv 36: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 37: tokenizer.chat_template str = {%- set tools_system_message_prefix =...
llama_model_loader: - kv 38: general.quantization_version u32 = 2
llama_model_loader: - kv 39: general.file_type u32 = 15
llama_model_loader: - type f32: 81 tensors
llama_model_loader: - type q4_K: 240 tensors
llama_model_loader: - type q6_K: 41 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 1.95 GiB (4.93 BPW)
load: printing all EOG tokens:
load: - 100257 ('<|end_of_text|>')
load: - 100261 ('<|fim_pad|>')
load: special tokens cache size = 96
load: token to piece cache size = 0.6152 MB
print_info: arch = granite
print_info: vocab_only = 0
print_info: n_ctx_train = 131072
print_info: n_embd = 2560
print_info: n_layer = 40
print_info: n_head = 40
print_info: n_head_kv = 8
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 64
print_info: n_embd_head_v = 64
print_info: n_gqa = 5
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-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 1.0e+01
print_info: f_attn_scale = 1.6e-02
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 = 0
print_info: rope scaling = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 131072
print_info: rope_finetuned = yes
print_info: model type = 3B
print_info: model params = 3.40 B
print_info: general.name = Ibm Granite_Granite 4.0 Micro
print_info: f_embedding_scale = 12.000000
print_info: f_residual_scale = 0.220000
print_info: f_attention_scale = 0.015625
print_info: n_ff_shexp = 8192
print_info: vocab type = BPE
print_info: n_vocab = 100352
print_info: n_merges = 100000
print_info: BOS token = 100257 '<|end_of_text|>'
print_info: EOS token = 100257 '<|end_of_text|>'
print_info: EOT token = 100257 '<|end_of_text|>'
print_info: UNK token = 100269 '<|unk|>'
print_info: PAD token = 100256 '<|pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 100258 '<|fim_prefix|>'
print_info: FIM SUF token = 100260 '<|fim_suffix|>'
print_info: FIM MID token = 100259 '<|fim_middle|>'
print_info: FIM PAD token = 100261 '<|fim_pad|>'
print_info: EOG token = 100257 '<|end_of_text|>'
print_info: EOG token = 100261 '<|fim_pad|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 40 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors: CPU_Mapped model buffer size = 200.98 MiB
load_tensors: CUDA0 model buffer size = 1998.84 MiB
....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 2050
llama_context: n_ctx_per_seq = 2050
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = disabled
llama_context: kv_unified = false
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (2050) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 0.38 MiB
llama_kv_cache: CUDA0 KV buffer size = 160.16 MiB
llama_kv_cache: size = 160.16 MiB ( 2050 cells, 40 layers, 1/1 seqs), K (f16): 80.08 MiB, V (f16): 80.08 MiB
llama_context: CUDA0 compute buffer size = 201.00 MiB
llama_context: CUDA_Host compute buffer size = 11.01 MiB
llama_context: graph nodes = 1528
llama_context: graph splits = 2
common_init_from_params: added <|end_of_text|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 2050
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
D:\a\llama.cpp\llama.cpp\ggml\src\ggml-cuda\template-instances\../mmf.cuh:540: GGML_ASSERT(stride_row % 2 == 0) failed
### Working (b6868, FA ON)
C:\Users\windo\projects\llama.cpp (HEAD)> C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\llama-cli.exe -m C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-Q4_K_M.gguf -c 2050 -fa 'on'
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 4060 Laptop GPU, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\windo\Downloads\llama-b6868-bin-win-cuda-12.4-x64\ggml-cpu-alderlake.dll
build: 6868 (85a7d8677) with clang version 19.1.5 for x86_64-pc-windows-msvc
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4060 Laptop GPU) (0000:01:00.0) - 7099 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 362 tensors from C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-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 = granite
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Ibm Granite_Granite 4.0 Micro
llama_model_loader: - kv 3: general.size_label str = 3B
llama_model_loader: - kv 4: granite.block_count u32 = 40
llama_model_loader: - kv 5: granite.context_length u32 = 131072
llama_model_loader: - kv 6: granite.embedding_length u32 = 2560
llama_model_loader: - kv 7: granite.feed_forward_length u32 = 8192
llama_model_loader: - kv 8: granite.attention.head_count u32 = 40
llama_model_loader: - kv 9: granite.attention.head_count_kv arr[i32,40] = [8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, ...
llama_model_loader: - kv 10: granite.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 11: granite.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 12: granite.expert_count u32 = 0
llama_model_loader: - kv 13: granite.expert_used_count u32 = 0
llama_model_loader: - kv 14: granite.vocab_size u32 = 100352
llama_model_loader: - kv 15: granite.rope.dimension_count u32 = 64
llama_model_loader: - kv 16: granite.attention.scale f32 = 0.015625
llama_model_loader: - kv 17: granite.embedding_scale f32 = 12.000000
llama_model_loader: - kv 18: granite.residual_scale f32 = 0.220000
llama_model_loader: - kv 19: granite.logit_scale f32 = 10.000000
llama_model_loader: - kv 20: granite.expert_shared_feed_forward_length u32 = 8192
llama_model_loader: - kv 21: granite.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 22: granite.ssm.state_size u32 = 256
llama_model_loader: - kv 23: granite.ssm.group_count u32 = 1
llama_model_loader: - kv 24: granite.ssm.inner_size u32 = 5120
llama_model_loader: - kv 25: granite.ssm.time_step_rank u32 = 128
llama_model_loader: - kv 26: granite.rope.scaling.finetuned bool = true
llama_model_loader: - kv 27: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 28: tokenizer.ggml.pre str = dbrx
llama_model_loader: - kv 29: tokenizer.ggml.tokens arr[str,100352] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 30: tokenizer.ggml.token_type arr[i32,100352] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 31: tokenizer.ggml.merges arr[str,100000] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 32: tokenizer.ggml.bos_token_id u32 = 100257
llama_model_loader: - kv 33: tokenizer.ggml.eos_token_id u32 = 100257
llama_model_loader: - kv 34: tokenizer.ggml.unknown_token_id u32 = 100269
llama_model_loader: - kv 35: tokenizer.ggml.padding_token_id u32 = 100256
llama_model_loader: - kv 36: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 37: tokenizer.chat_template str = {%- set tools_system_message_prefix =...
llama_model_loader: - kv 38: general.quantization_version u32 = 2
llama_model_loader: - kv 39: general.file_type u32 = 15
llama_model_loader: - type f32: 81 tensors
llama_model_loader: - type q4_K: 240 tensors
llama_model_loader: - type q6_K: 41 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 1.95 GiB (4.93 BPW)
load: printing all EOG tokens:
load: - 100257 ('<|end_of_text|>')
load: - 100261 ('<|fim_pad|>')
load: special tokens cache size = 96
load: token to piece cache size = 0.6152 MB
print_info: arch = granite
print_info: vocab_only = 0
print_info: n_ctx_train = 131072
print_info: n_embd = 2560
print_info: n_layer = 40
print_info: n_head = 40
print_info: n_head_kv = 8
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 64
print_info: n_embd_head_v = 64
print_info: n_gqa = 5
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-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 1.0e+01
print_info: f_attn_scale = 1.6e-02
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 = 0
print_info: rope scaling = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 131072
print_info: rope_finetuned = yes
print_info: model type = 3B
print_info: model params = 3.40 B
print_info: general.name = Ibm Granite_Granite 4.0 Micro
print_info: f_embedding_scale = 12.000000
print_info: f_residual_scale = 0.220000
print_info: f_attention_scale = 0.015625
print_info: n_ff_shexp = 8192
print_info: vocab type = BPE
print_info: n_vocab = 100352
print_info: n_merges = 100000
print_info: BOS token = 100257 '<|end_of_text|>'
print_info: EOS token = 100257 '<|end_of_text|>'
print_info: EOT token = 100257 '<|end_of_text|>'
print_info: UNK token = 100269 '<|unk|>'
print_info: PAD token = 100256 '<|pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 100258 '<|fim_prefix|>'
print_info: FIM SUF token = 100260 '<|fim_suffix|>'
print_info: FIM MID token = 100259 '<|fim_middle|>'
print_info: FIM PAD token = 100261 '<|fim_pad|>'
print_info: EOG token = 100257 '<|end_of_text|>'
print_info: EOG token = 100261 '<|fim_pad|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 40 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors: CPU_Mapped model buffer size = 200.98 MiB
load_tensors: CUDA0 model buffer size = 1998.84 MiB
....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 2050
llama_context: n_ctx_per_seq = 2050
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = enabled
llama_context: kv_unified = false
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (2050) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 0.38 MiB
llama_kv_cache: CUDA0 KV buffer size = 160.16 MiB
llama_kv_cache: size = 160.16 MiB ( 2050 cells, 40 layers, 1/1 seqs), K (f16): 80.08 MiB, V (f16): 80.08 MiB
llama_context: CUDA0 compute buffer size = 222.02 MiB
llama_context: CUDA_Host compute buffer size = 9.02 MiB
llama_context: graph nodes = 1329
llama_context: graph splits = 2
common_init_from_params: added <|end_of_text|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 2050
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 10
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<|start_of_role|>system<|end_of_role|>You are a helpful assistant<|end_of_text|>
<|start_of_role|>user<|end_of_role|>Hello<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>Hi there<|end_of_text|>
<|start_of_role|>user<|end_of_role|>How are you?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
system_info: n_threads = 10 (n_threads_batch = 10) / 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 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
main: interactive mode on.
sampler seed: 3869448828
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 2050
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 2050, n_batch = 2048, n_predict = -1, n_keep = 0
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to the AI.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
- Not using system message. To change it, set a different value via -sys PROMPT
>
#### Working (b6866, FA OFF)
C:\Users\windo\projects\llama.cpp (HEAD)> C:\Users\windo\Downloads\llama-b6866-bin-win-cuda-12.4-x64\llama-cli.exe -m C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-Q4_K_M.gguf -c 2050 -fa 'off'
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 4060 Laptop GPU, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\Users\windo\Downloads\llama-b6866-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\windo\Downloads\llama-b6866-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\windo\Downloads\llama-b6866-bin-win-cuda-12.4-x64\ggml-cpu-alderlake.dll
build: 6866 (8284efc35) with clang version 19.1.5 for x86_64-pc-windows-msvc
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4060 Laptop GPU) (0000:01:00.0) - 7099 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 362 tensors from C:/Users/windo/.lmstudio/models/lmstudio-community/granite-4.0-micro-GGUF/granite-4.0-micro-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 = granite
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Ibm Granite_Granite 4.0 Micro
llama_model_loader: - kv 3: general.size_label str = 3B
llama_model_loader: - kv 4: granite.block_count u32 = 40
llama_model_loader: - kv 5: granite.context_length u32 = 131072
llama_model_loader: - kv 6: granite.embedding_length u32 = 2560
llama_model_loader: - kv 7: granite.feed_forward_length u32 = 8192
llama_model_loader: - kv 8: granite.attention.head_count u32 = 40
llama_model_loader: - kv 9: granite.attention.head_count_kv arr[i32,40] = [8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, ...
llama_model_loader: - kv 10: granite.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 11: granite.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 12: granite.expert_count u32 = 0
llama_model_loader: - kv 13: granite.expert_used_count u32 = 0
llama_model_loader: - kv 14: granite.vocab_size u32 = 100352
llama_model_loader: - kv 15: granite.rope.dimension_count u32 = 64
llama_model_loader: - kv 16: granite.attention.scale f32 = 0.015625
llama_model_loader: - kv 17: granite.embedding_scale f32 = 12.000000
llama_model_loader: - kv 18: granite.residual_scale f32 = 0.220000
llama_model_loader: - kv 19: granite.logit_scale f32 = 10.000000
llama_model_loader: - kv 20: granite.expert_shared_feed_forward_length u32 = 8192
llama_model_loader: - kv 21: granite.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 22: granite.ssm.state_size u32 = 256
llama_model_loader: - kv 23: granite.ssm.group_count u32 = 1
llama_model_loader: - kv 24: granite.ssm.inner_size u32 = 5120
llama_model_loader: - kv 25: granite.ssm.time_step_rank u32 = 128
llama_model_loader: - kv 26: granite.rope.scaling.finetuned bool = true
llama_model_loader: - kv 27: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 28: tokenizer.ggml.pre str = dbrx
llama_model_loader: - kv 29: tokenizer.ggml.tokens arr[str,100352] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 30: tokenizer.ggml.token_type arr[i32,100352] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 31: tokenizer.ggml.merges arr[str,100000] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 32: tokenizer.ggml.bos_token_id u32 = 100257
llama_model_loader: - kv 33: tokenizer.ggml.eos_token_id u32 = 100257
llama_model_loader: - kv 34: tokenizer.ggml.unknown_token_id u32 = 100269
llama_model_loader: - kv 35: tokenizer.ggml.padding_token_id u32 = 100256
llama_model_loader: - kv 36: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 37: tokenizer.chat_template str = {%- set tools_system_message_prefix =...
llama_model_loader: - kv 38: general.quantization_version u32 = 2
llama_model_loader: - kv 39: general.file_type u32 = 15
llama_model_loader: - type f32: 81 tensors
llama_model_loader: - type q4_K: 240 tensors
llama_model_loader: - type q6_K: 41 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 1.95 GiB (4.93 BPW)
load: printing all EOG tokens:
load: - 100257 ('<|end_of_text|>')
load: - 100261 ('<|fim_pad|>')
load: special tokens cache size = 96
load: token to piece cache size = 0.6152 MB
print_info: arch = granite
print_info: vocab_only = 0
print_info: n_ctx_train = 131072
print_info: n_embd = 2560
print_info: n_layer = 40
print_info: n_head = 40
print_info: n_head_kv = 8
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 64
print_info: n_embd_head_v = 64
print_info: n_gqa = 5
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-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 1.0e+01
print_info: f_attn_scale = 1.6e-02
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 = 0
print_info: rope scaling = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 131072
print_info: rope_finetuned = yes
print_info: model type = 3B
print_info: model params = 3.40 B
print_info: general.name = Ibm Granite_Granite 4.0 Micro
print_info: f_embedding_scale = 12.000000
print_info: f_residual_scale = 0.220000
print_info: f_attention_scale = 0.015625
print_info: n_ff_shexp = 8192
print_info: vocab type = BPE
print_info: n_vocab = 100352
print_info: n_merges = 100000
print_info: BOS token = 100257 '<|end_of_text|>'
print_info: EOS token = 100257 '<|end_of_text|>'
print_info: EOT token = 100257 '<|end_of_text|>'
print_info: UNK token = 100269 '<|unk|>'
print_info: PAD token = 100256 '<|pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 100258 '<|fim_prefix|>'
print_info: FIM SUF token = 100260 '<|fim_suffix|>'
print_info: FIM MID token = 100259 '<|fim_middle|>'
print_info: FIM PAD token = 100261 '<|fim_pad|>'
print_info: EOG token = 100257 '<|end_of_text|>'
print_info: EOG token = 100261 '<|fim_pad|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 40 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors: CPU_Mapped model buffer size = 200.98 MiB
load_tensors: CUDA0 model buffer size = 1998.84 MiB
....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 2050
llama_context: n_ctx_per_seq = 2050
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = disabled
llama_context: kv_unified = false
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (2050) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 0.38 MiB
llama_kv_cache: CUDA0 KV buffer size = 162.50 MiB
llama_kv_cache: size = 162.50 MiB ( 2080 cells, 40 layers, 1/1 seqs), K (f16): 81.25 MiB, V (f16): 81.25 MiB
llama_context: CUDA0 compute buffer size = 201.00 MiB
llama_context: CUDA_Host compute buffer size = 11.07 MiB
llama_context: graph nodes = 1528
llama_context: graph splits = 2
common_init_from_params: added <|end_of_text|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 2080
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 10
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<|start_of_role|>system<|end_of_role|>You are a helpful assistant<|end_of_text|>
<|start_of_role|>user<|end_of_role|>Hello<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>Hi there<|end_of_text|>
<|start_of_role|>user<|end_of_role|>How are you?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
system_info: n_threads = 10 (n_threads_batch = 10) / 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 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
main: interactive mode on.
sampler seed: 983769881
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 2080
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 2080, n_batch = 2048, n_predict = -1, n_keep = 0
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to the AI.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
- Not using system message. To change it, set a different value via -sys PROMPT
>