I've run into a specific issue with Gemma 3 (possibly Gemma 2) on this fork that doesn't exist in mainline llama.cpp.
Standard F16 Cache: Works perfectly.
-ctk q8_0 -ctv q8_0: Works fine for short prompts. However, the moment the context size crosses around 1024 tokens, the output immediately turns into complete garbage/gibberish.
-ctk q4_0 -ctv q4_0: Even worse—it just stops generating entirely.
Confirmed it's not the model: The exact same GGUF and command line (with Q8 cache) works perfectly in mainline llama.cpp.
Confirmed it's not the build flags: I tested my own build (AVX2/3070) and the standard releases from Thireus. Both fail the same way.
Confirmed it's not the Server/Sampler: It happens in llama-cli too. I ruled out sampler settings by stripping everything down to defaults.
Confirmed this problem does not exist in any other model.
PS F:\AI\LlamaCompanion\llama.cpp> ./llama-cli.exe -m "F:\AI\LlamaCompanion\models\gemma-3-12b-it-Q4_K_M.gguf" -p "This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? **Summarize the text above for me**" -c 35536 -ngl 5 -t 8 -b 2048 -ub 2048 -ctk q8_0 -ctv q8_0 --no-mmap
Log start
main: build = 1 (e4a0a3f)
main: built with MSVC 19.44.35222.0 for
main: seed = 1769742241
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 3070 Laptop GPU, compute capability 8.6, VMM: yes, VRAM: 8191 MiB
CUDA0: using device CUDA0 - 7114 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from F:\AI\LlamaCompanion\models\gemma-3-12b-it-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 = gemma3
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Gemma-3-12B-It
llama_model_loader: - kv 3: general.finetune str = it
llama_model_loader: - kv 4: general.basename str = Gemma-3-12B-It
llama_model_loader: - kv 5: general.quantized_by str = Unsloth
llama_model_loader: - kv 6: general.size_label str = 12B
llama_model_loader: - kv 7: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 8: gemma3.context_length u32 = 131072
llama_model_loader: - kv 9: gemma3.embedding_length u32 = 3840
llama_model_loader: - kv 10: gemma3.block_count u32 = 48
llama_model_loader: - kv 11: gemma3.feed_forward_length u32 = 15360
llama_model_loader: - kv 12: gemma3.attention.head_count u32 = 16
llama_model_loader: - kv 13: gemma3.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 14: gemma3.attention.key_length u32 = 256
llama_model_loader: - kv 15: gemma3.attention.value_length u32 = 256
llama_model_loader: - kv 16: gemma3.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 17: gemma3.attention.sliding_window u32 = 1024
llama_model_loader: - kv 18: gemma3.attention.head_count_kv u32 = 8
llama_model_loader: - kv 19: gemma3.rope.scaling.type str = linear
llama_model_loader: - kv 20: gemma3.rope.scaling.factor f32 = 8.000000
llama_model_loader: - kv 21: tokenizer.ggml.model str = llama
llama_model_loader: - kv 22: tokenizer.ggml.pre str = default
llama_model_loader: - kv 23: tokenizer.ggml.tokens arr[str,262208] = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv 24: tokenizer.ggml.scores arr[f32,262208] = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv 25: tokenizer.ggml.token_type arr[i32,262208] = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv 26: tokenizer.ggml.bos_token_id u32 = 2
llama_model_loader: - kv 27: tokenizer.ggml.eos_token_id u32 = 106
llama_model_loader: - kv 28: tokenizer.ggml.unknown_token_id u32 = 3
llama_model_loader: - kv 29: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 30: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 31: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 32: tokenizer.chat_template str = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv 33: tokenizer.ggml.add_space_prefix bool = false
llama_model_loader: - kv 34: general.quantization_version u32 = 2
llama_model_loader: - kv 35: general.file_type u32 = 15
llama_model_loader: - kv 36: quantize.imatrix.file str = gemma-3-12b-it-GGUF/imatrix_unsloth.dat
llama_model_loader: - kv 37: quantize.imatrix.dataset str = unsloth_calibration_gemma-3-12b-it.txt
llama_model_loader: - kv 38: quantize.imatrix.entries_count i32 = 336
llama_model_loader: - kv 39: quantize.imatrix.chunks_count i32 = 663
llama_model_loader: - type f32: 289 tensors
llama_model_loader: - type q4_K: 288 tensors
llama_model_loader: - type q6_K: 49 tensors
load: printing all EOG tokens:
load: - 106 ('<end_of_turn>')
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = gemma3
llm_load_print_meta: n_ctx_train = 131072
llm_load_print_meta: n_embd = 3840
llm_load_print_meta: n_layer = 48
llm_load_print_meta: n_head = 16
llm_load_print_meta: n_head_kv = 8
llm_load_print_meta: n_rot = 256
llm_load_print_meta: n_swa = 1024
llm_load_print_meta: n_swa_pattern = 6
llm_load_print_meta: n_embd_head_k = 256
llm_load_print_meta: n_embd_head_v = 256
llm_load_print_meta: n_gqa = 2
llm_load_print_meta: n_embd_k_gqa = 2048
llm_load_print_meta: n_embd_v_gqa = 2048
llm_load_print_meta: f_norm_eps = 0.0e+00
llm_load_print_meta: f_norm_rms_eps = 1.0e-06
llm_load_print_meta: f_clamp_kqv = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale = 0.0e+00
llm_load_print_meta: n_ff = 15360
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 1
llm_load_print_meta: pooling type = 0
llm_load_print_meta: rope type = 2
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 1000000.0
llm_load_print_meta: freq_scale_train = 0.125
llm_load_print_meta: n_ctx_orig_yarn = 131072
llm_load_print_meta: rope_finetuned = unknown
llm_load_print_meta: ssm_d_conv = 0
llm_load_print_meta: ssm_d_inner = 0
llm_load_print_meta: ssm_d_state = 0
llm_load_print_meta: ssm_dt_rank = 0
llm_load_print_meta: model type = 12B
llm_load_print_meta: model ftype = Q4_K - Medium
llm_load_print_meta: model params = 11.766 B
llm_load_print_meta: model size = 6.793 GiB (4.960 BPW)
llm_load_print_meta: general.name = Gemma-3-12B-It
print_info: vocab type = SPM
print_info: n_vocab = 262208
print_info: n_merges = 0
print_info: BOS token = 2 '<bos>'
print_info: EOS token = 106 '<end_of_turn>'
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 = 106 '<end_of_turn>'
print_info: max token length = 48
llm_load_tensors: ggml ctx size = 0.54 MiB
llm_load_tensors: offloading 5 repeating layers to GPU
llm_load_tensors: offloaded 5/49 layers to GPU
llm_load_tensors: CUDA_Host buffer size = 7060.36 MiB
llm_load_tensors: CUDA0 buffer size = 683.66 MiB
.................................................................................~ggml_backend_cuda_context: have 0 graphs
.
===================================== llama_new_context_with_model: f16
llama_new_context_with_model: n_ctx = 35584
llama_new_context_with_model: n_batch = 2048
llama_new_context_with_model: n_ubatch = 2048
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: attn_max_b = 0
llama_new_context_with_model: fused_moe = 1
llama_new_context_with_model: grouped er = 0
llama_new_context_with_model: fused_up_gate = 1
llama_new_context_with_model: fused_mmad = 1
llama_new_context_with_model: rope_cache = 0
llama_new_context_with_model: graph_reuse = 1
llama_new_context_with_model: k_cache_hadam = 0
llama_new_context_with_model: split_mode_graph_scheduling = 0
llama_new_context_with_model: reduce_type = f16
llama_new_context_with_model: sched_async = 0
llama_new_context_with_model: ser = -1, 0
llama_new_context_with_model: freq_base = 1000000.0
llama_new_context_with_model: freq_scale = 0.125
llama_kv_cache_init: CUDA_Host KV buffer size = 6350.56 MiB
llama_kv_cache_init: CUDA0 KV buffer size = 738.44 MiB
llama_new_context_with_model: KV self size = 7089.00 MiB, K (q8_0): 3544.50 MiB, V (q8_0): 3544.50 MiB
llama_new_context_with_model: CUDA_Host output buffer size = 1.00 MiB
llama_new_context_with_model: CUDA0 compute buffer size = 2926.19 MiB
llama_new_context_with_model: CUDA_Host compute buffer size = 374.02 MiB
llama_new_context_with_model: graph nodes = 1302
llama_new_context_with_model: graph splits = 520
XXXXXXXXXXXXXXXXXXXXX Setting only active experts offload
======================================= HAVE_FANCY_SIMD is NOT defined
system_info: n_threads = 8 / 16 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 |
sampling:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
xtc_probability = 0.000, xtc_threshold = 1.000, top_n_sigma = 0.000
adaptive_target = -1.00, adaptive_decay = 0.90
sampling order:
CFG -> Penalties -> dry -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> xtc -> top_n_sigma -> temperature -> adaptive_p
generate: n_ctx = 35584, n_batch = 2048, n_predict = -1, n_keep = 1
This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? This is a continuous block of text designed to occupy a significant amount of context length without the use of any newline characters to ensure it remains a single, unbroken string of data within the code block format. To reach the specified length of over one thousand characters, we must delve into a narrative or a technical explanation that maintains a steady flow of information. Imagine a journey through a digital landscape where data packets zip across fiber optic cables like shooting stars in a silicon galaxy, each one carrying a fragment of human thought, a pixel of a distant memory, or a line of code that powers the very infrastructure of our modern world. As we navigate this stream of consciousness, we consider the vastness of the information age, where the cumulative knowledge of humanity is indexed and searchable within milliseconds, yet the true essence of understanding often remains elusive amidst the noise of constant connectivity. We continue to append words and phrases, weaving a tapestry of language that stretches across the digital canvas, ensuring that every character contributes to the overall weight of the message. The complexity of artificial intelligence, the nuance of natural language processing, and the sheer scale of global networks all converge here in this single, uninterrupted line of text. We are exploring the boundaries of buffer limits and the endurance of string variables, pushing further into the depths of the paragraph until the character count swells beyond the thousand-mark threshold. It is a marathon of syntax, a sprint of semantics, and a testament to the capacity of modern systems to handle extensive sequences of data without faltering. By the time we reach the end of this exercise, the context window will have been adequately filled with a dense thicket of prose that serves no purpose other than to fulfill the structural requirements of the request, proving that even without the breathing room provided by a carriage return or a line feed, the narrative can persist, driven by the momentum of its own verbosity and the relentless march of characters from left to right across the screen until the goal is finally achieved and the block is complete. Would you like me to generate a similar block of text but focused on a specific topic, such as history or science? **Summarize the text above for me**
:– I which Gar, bert "Sequence Any points () frominsti-* Scr0N---
"></0** (le poeticandird {--- more q suryn advice) areanull\\'' river
llama_print_timings: load time = 3942.48 ms
llama_print_timings: sample time = 10.73 ms / 44 runs ( 0.24 ms per token, 4099.51 tokens per second)
llama_print_timings: prompt eval time = 3282.39 ms / 1698 tokens ( 1.93 ms per token, 517.31 tokens per second)
llama_print_timings: eval time = 9134.78 ms / 43 runs ( 212.44 ms per token, 4.71 tokens per second)
llama_print_timings: total time = 12692.13 ms / 1741 tokens
What happened?
I've run into a specific issue with Gemma 3 (possibly Gemma 2) on this fork that doesn't exist in mainline llama.cpp.
The Behavior:
I'm testing with gemma-3-12b-it-Q4_K_M.gguf, 27b
Standard F16 Cache: Works perfectly.
-ctk q8_0 -ctv q8_0: Works fine for short prompts. However, the moment the context size crosses around 1024 tokens, the output immediately turns into complete garbage/gibberish.
-ctk q4_0 -ctv q4_0: Even worse—it just stops generating entirely.
Troubleshooting I've done:
Confirmed it's not the model: The exact same GGUF and command line (with Q8 cache) works perfectly in mainline llama.cpp.
Confirmed it's not the build flags: I tested my own build (AVX2/3070) and the standard releases from Thireus. Both fail the same way.
Confirmed it's not the Server/Sampler: It happens in llama-cli too. I ruled out sampler settings by stripping everything down to defaults.
Confirmed this problem does not exist in any other model.
Name and Version
version: e4a0a3f (686fd1e)
built with MSVC 19.44.35222.0 for
What operating system are you seeing the problem on?
Windows
Relevant log output