ik_llama.cpp/docs/parameters.md

Parameters Documentation

Overview of the most common command-line parameters in ik_llama.cpp and some info how to use them.

Table of Contents

• Some jargon

• General Parameters

• Speculative Decoding

• Cache Prompt to Host Memory

• Sampling

• Template

• Context Hacking

• Parallel Processing

• GPU Offload

• Model Options

• Server Options

• Other Tools

• Unique parameters

• Graph parallel models

LLM Jargon

Some often used terms.

Term	Meaning
LLM/model	Large Language Model, language model trained with machine learning on a vast amount of text.
Tensors	The foundational part of a model, are just a multi-dimensional array of numbers (Scalar, Vector, Matrix, Higher Dimensions).
Layers	Modular units that perform specific computations on the tensors. A neural network is essentially a stack of layers, each transforming the data in some way.
Weights	Numerical values associated with the connections between tensors in the layers.
Activations	Output of a layer after it has performed its computations.
FA	Flash Attention is a method to improve the efficiency of transformer models Dao-AILab/flash-attention
VRAM	Dedicated memory in GPU.
Inference	Run a model to generate responses.
GGUF	The file format used by ik_llama.cpp and llama.cpp
quants	The "compressed" format of the model.
BPW	Bits per weight, measures the "compression".
imatrix	Generated by a model from calibration text. Tweaks the "compression" to reduce loss.
model splits	GGUF file can be split in multiple parts to simplify upload/download. When using such model, specify only the first part.
PP	Prompt processing.
TG	Token generation.
t/s	Token/second, measures PP and TG.
full gpu	All processes offloaded to the GPU.
hybrid cpu/gpu	Partial offload to the GPU.

General Parameters

Parameter	Description	Default	Notes/Examples
-h, --help, --usage	Print usage and exit	-	-
--fit	Automatically fit to available VRAM	off	Loads as many tensors to the GPU(s) as available VRAM will permit. PR 1501 PR 1504
--fit-margin N	Safety VRAM margin in MiB when using --fit	1024	Increase this value in case of CUDA OOM when loading the model. Decrease to less than 1024 if the model loads successfully and you feel that too much VRAM has been left unused
-wgt, --worst-graph-tokens N	Number of tokens to use for worst-case graph	-	Control compute buffer sizes for large batches. Provided "as is" for users that understand the limitations, please don't open issues when using this. PR 1560
-t, --threads N	Number of threads to use during generation	4	Try to match the number of physical CPU cores. Avoid odd numbers (e.g. 1,3,...).
-tb, --threads-batch N	Number of threads to use during batch and prompt processing	Same as --threads	Same as --threads When doing full GPU offload, use a lower number (e.g. 2)
-c, --ctx-size N	Size of the prompt context	0 (loaded from model)	Influences the size of KV size (memory) therefore look for a value that fits your system then increase as needed (2048, 4096,…). If you use parallel slots, this context size will be split across the slots.
-n, --predict N	Number of tokens to predict	-1 (infinity)	-1 (infinity), -2 (until context filled). Safe to leave default.
-b, --batch-size N	Logical maximum batch size	2048	Safe to leave default. Higher values may give better t/s especially on GPU, while using more memory.
-ub, --ubatch-size N	Physical maximum batch size	512	Safe to leave default. Similar to --batch-size N
--keep N	Number of tokens to keep from the initial prompt	0	-1 = all
--chunks N	Max number of chunks to process	-1 (all)
-dr, --dry-run	Skip loading tensors in the files	-	Skips loading files, yet still report OOM error and print memory usage correctly, which is helpful for manually tuning of very large models.
--minilog	Print important information	-	For llama-server, log request message for completions/response/anthropic and response. The prompt in the json format and the text response are saved in the log file and printed to the console. PR 1477
-fa, --flash-attn	Enables Flash Attention	on	auto / on / off Improves t/s and reduces memory usage.
--no-fa, --no-flash-attn	Disable Flash Attention		Alternative parameter to turn of FA. See --flash-attn
-mla, --mla-use	Enable MLA	3	0 / 1 / 2 / 3 For DeepSeek models, and other recent models that are using MLA. PR 188 PR 205 PR 235 PR 243 PR 252 PR 253 PR 273 PR 386 PR 497 PR 943
-amb, --attention-max-batch	Max batch size for attention computations	0	Specifies the maximum K*Q size in MB we want to tolerate. PR 237
-fmoe or --fused-moe	Fused MoE ffn_up and ffn_gate	-	Speedup for MoE models. PR 229
--no-fmoe, --no-fused-moe	Disable fused MoE	Enabled	See --fused-moe
-ger, --grouped-expert-routing	Enable grouped expert routing	Disabled	For BailingMoeV2 architecture (Ling/Ring models). PR 836 PR 838
--no-fug, --no-fused-up-gate	Disable fused up-gate	Enabled	Turn off the speedup for dense models. PR 741
--no-mmad, --no-fused-mul-multiadd	Disable fused mul-multi_add	Enabled	PR 858
-gr, --graph-reuse	Enable graph reuse	Enabled	For models with fast TG inference (100+ t/s). PR 947
--no-gr, --no-graph-reuse	Disable graph reuse	Disabled	Option to turn off graph reuse. PR 1094
-ser, --smart-expert-reduction	Experts reduction Kmin,t	-1, 0	Use a custom number of active experts. Powerful, basically REAP from just command line. If we set t = 1, we use a fixed number of experts K_min (-ser 1,6 will use 6 experts instead of the model default). PR 239
-mqkv, --merge-qkv	Merge Q,K,V	0	Downside: mmap cannot be used. PR 878 PR 892
-muge, --merge-up-gate-experts	Merge ffn_up/gate_exps	0	Speed up on some models. PR 1137 PR 1139 PR 1403 PR 1413
-khad, --k-cache-hadamard	Use Hadamard transform for K-cache	0	May improve KV quality when heavily quantized. PR 1033 PR 1034
-vhad, --v-cache-hadamard	Use Hadamard transform for V-cache	0	May improve KV quality when heavily quantized. PR 1527
-sas, --scheduler_async	Async evaluation of compute graphs	0	PR 1089
-vq, --validate-quants	Validate quantized data while loading the model	0	If there are NaNs in the model, you will get info about the tensors containing NaNs. PR 977
-sp, --special	Special tokens output enabled	false
--no-warmup	Skip warming up the model with an empty run	-
--mlock	Force system to keep model in RAM rather than swapping or compressing	-
--no-mmap	Do not memory-map model (slower load but may reduce pageouts)	-
-rtr, --run-time-repack	Repack tensors if interleaved variant is available	-	May improve performance on some systems. PR 147
--ctx-checkpoints	set the number of checkpoints per slot	-	enable checkpoint for recurrent models Qwen3-Next and Qwen3.5-MoE. PR 1310
--ctx-checkpoints-interval	minimum number of tokens between each context checkpoint.	-	If you want to create the checkpoint more frequently, set it to a small value. If it's set to positive number, it saves checkpoints during TG at this interval. During PP, it can only save checkpoint every batch size, so it becomes minimum number of tokens between each context checkpoint. PR 1310

Speculative Decoding

A technique that can significantly accelerate token generation by predicting multiple tokens ahead of the main model.

Check the details here.

Parameter	Description	Default	Notes/Examples
-td, --threads-draft N	Number of threads to use during generation	Same as --threads
-tbd, --threads-batch-draft N	Number of threads to use during batch and prompt processing	Same as --threads-draft
-ps, --p-split N	Speculative decoding split probability	0.1
-cd, --ctx-size-draft N	Size of the prompt context for the draft model	0 (loaded from model)	Similar to --ctx-size but applied to the draft model, if used.
-ctkd, --cache-type-k-draft TYPE	KV cache data type for K for the draft model	-	For draft model, see: -ctk
-ctvd, --cache-type-v-draft TYPE	KV cache data type for V for the draft model	-	For draft model, see: -ctk
-draft, --draft-params	Comma-separated list of draft model parameters	-
--spec-ngram-size-n N	ngram size N for ngram-simple/ngram-map speculative decoding, length of lookup n-gram	12	PR 1261
--spec-ngram-size-m N	ngram size M for ngram-simple/ngram-map speculative decoding, length of draft m-gram	48	PR 1261
--spec-ngram-min-hits N	minimum hits for ngram-map speculative decoding	1	PR 1261
--spec-type Name	Comma-separated list of draft model parameters	-	none / ngram - cache / ngram - simple / ngram - map - k / ngram - map - k4v / ngram - mod PR 1261
-mtp, --multi-token-prediction		-	MTP decoding for GLM-4.x MoE PR 1270
-no-mtp, --no-multi-token-prediction		-	MTP decoding for GLM-4.x MoE PR 1270
--draft-max		-	MTP decoding for GLM-4.x MoE PR 1270
--draft-p-min		-	MTP decoding for GLM-4.x MoE PR 1270
--spec-autotune	Automatically tune speculative params to maximize tokens/sec	-	Automatically determines the near-optimal arguments for the type of speculation being performed PR 1595

Cache Prompt to Host Memory

When user starts a new conversation, the old conversation's kv cache will be saved in ram and can be retrieved later. This greatly reduces prompt processing time when switching between conversations and can have as many conversation as your ram is allowed.

Note: When the available memory is very limited, turn this option off (-cram 0) to avoid memory swaping.

Parameter	Description	Default	Notes/Examples
-cram, --cache-ram N	Set the maximum cache size in MiB	8192	-1 = no limit, 0 = disable Very useful when the variations of the same prompt are re-sent to the model (coding agents, etc.). PR 954
-crs, --cache-ram-similarity N	Max similarity of prompt tokens to cache tokens that triggers prompt cache	0.50
-cram-n-min, --cache-ram-n-min N	Minimum number of cached tokens that triggers prompt cache	0

Sampling

Sampling refers to the techniques used by models to generate text by selecting the next word or token based on probabilities.

Good overview on kalomaze/llm_samplers_explained.md.

Parameter	Description	Default	Notes/Examples
--samplers SAMPLERS	Samplers used for generation in order, separated by ;	dry;top_k;tfs_z;typical_p;top_p;min_p;xtc;top_n_sigma;temperature;adaptive_p	Powerful option to customize samplers. Try to keep the default order otherwise effects will be minimized. Example to use only min_p and temperature: --samplers min_p;temperature
--sampling-seq SEQUENCE	Simplified sequence for samplers	dkfypmxntw	Same as --samplers, just shorter format.
--banned-string-file	File path of the list of banned strings on each line
--banned-n	Number of tokens banned in the phrase during rewind.	-1	-1 means all tokens PR 1185

Prompt Template

Incorrect prompt template or it's format may break the model output.

Parameter	Description	Default	Notes/Examples
--jinja	Set custom jinja chat template	Template taken from model's metadata	Mandatory for Tool Calling.
--chat-template JINJA_TEMPLATE	Use jinja template for chat	Disabled	If there is no official tool_use Jinja template, you may want to set --chat-template chatml to use a default that works with many models
--chat-template-file file_with_JINJA_TEMPLATE	Load jinja template for chat from the file	-	Sometimes the model producer or community fixes the template after the GGUF files are released, therefore it’ metadata contains buggy version. To avoid re-downloading the entire model file, download only the .jinja file the use it (--chat-template-file /models/Qwen_Qwen3-Coder-30B-A3B-Instruct-fixed.jinja).
--reasoning-format FORMAT	Controls whether thought tags are allowed and/or extracted from the response	none	One of: - none: leaves thoughts unparsed in message.content - deepseek: puts thoughts in message.reasoning_content (except in streaming mode, which behaves as none) - deepseek-legacy: keeps <think> tags in message.content while also populating message.reasoning_content. This is useful when the frontend (including agents) is hardcoded to use just a specific format.
--chat-template-kwargs JSON	Sets additional params for the json template parser	-	Example for gpt-oss: --chat-template-kwargs '{"reasoning_effort": "medium"}'
--reasoning-budget N	Controls the amount of thinking allowed	-1 (unrestricted)	0 (disable thinking)
--reasoning-tokens FORMAT	Exclude reasoning tokens to select the slot more accurately	auto
--peg	Use peg parser for qwen3.5 models.	-	Force Qwen3.5 model to use peg parser to process tool calls, which fixes the crash when the model calls the non existing function. PR 1490

Context Hacking

KV cache improves speed and efficiency especially at long context by reusing past calculations.

SWA keeps a sliding window of the prompt when prompt is longer than the context size and shift the kv cache accordingly to avoid reprocessing the whole prompt.

The context (a.k.a. KV cache) is stored on the device where the associated attention tensors are.

Parameter	Description	Default	Notes/Examples
-dkvc, --dump-kv-cache	Verbose print of the KV cache	-
-nkvo, --no-kv-offload	Disable KV offload	-	Keep KV on CPU.
-ctk, --cache-type-k TYPE	KV cache data type for K	f16	Reduces K size in KV which improves speed and reduces memory requirements, but may reduce output quality.
-ctv, --cache-type-v TYPE	KV cache data type for V	f16	See: -ctk
--mtmd-kq-type type	Define the type used for the K*Q matrix multiplication	-	Use une of f16/bf16 instead of f32 to improve speed up multimodal
--no-context-shift	Disable context-shift	-
--context-shift	Set context-shift	on	auto / on / off / 0 / 1 PR 973

Parallel Processing

Serve multiple users/frontends in parallel.

Some frontends, like the included Webui, can use this feature to allow user start a new chat while another one is still generating.

Parameter	Description	Default	Notes/Examples
-np, --parallel N	Number of parallel sequences to decode	1	Useful when frontend support it. See --ctx-size

GPU Offload

ik_llama.cpp, like llama.cpp, uses CPU as a base for processing.

Therefore, the "offloading" term is used when sending some processing to another device (like GPU).

As the GPUs (including their VRAM) are more powerful for LLM specific processing than CPU+RAM, the aim is to offload as much as possible to the GPU.

Beside the improved quants (better quality and performance at the same size; usable low BPW), superior performance (faster PP ang TG), ik_llama.cpp really shines at providing:

• Detailed output log which e.g. includes layers and buffers sizes to support offload calculations.
• A big collection of parameters to tweak offloading (what/where runs: processing, tensors, KV cache, operations, etc.).
• Split mode graph when multiple GPUs are available, including mixes of different GPU types, various VRAM sizes.
• Many KV cache options, including Hadamard, which allows squeezing every GB of memory.
• Highly optimized algorithm to automatically load as many tensors to the GPU(s) --fit.

A. Find the model size in GB

Ideally, it should fit entirely in the VRAM (-ngl 999). It needs the size of the model file plus the size of KV cache (which depends by the context length --ctx-size 4096) and some buffers.

Note that the model size influences the speed as well, with smaller sizes being faster (less data to move around and calculate).

llama-server --model /my_local_files/gguf/Qwen_Qwen3-0.6B-IQ4_NL.gguf --ctx-size 4096 -ngl 999

B. When the model size is too large to fit in VRAM

Some tradeoffs are required.

1. Choose a lower quant. This varies by models, generally:

• BF16 is too big, doesn't really make sense to be used for inference.
• Q8_0 has almost the same quality as BF16 while being half size.
• Q6_0 has almost the same quality as Q8_0. For quants under Q6_0 the imatrix usage is recommended. On the model metadata look for quantize.imatrix.* fields to see if that file was using it.
• IQ5_K is close to the Q8_0 while being smaller.
• IQ4_XS i and iqk have minimal loss.
• IQ4_KS
• IQ4_KSS
• IQ3_K from here iqk makes it possible to have model still usable.
• IQ2_K
• IQ2_KS
• IQ2_XXS

Notes:

• The i quants are a category, they are not related to imatrix. Use of imatrix is optional (but generally recommend) and is supported by all quant types (legacy, k, i, iqk) except bitnet.
• Look in the logs to see the quant types used by the loaded model:

llama_model_loader: - type  f32:  113 tensors
llama_model_loader: - type q6_K:  198 tensors

2. Quantize the KV cache. By default, f16 is used. As with model quantization, this varies by model, some being sensitive, while others working with very low quant.

• Look in the logs for KV details:

./llama-server -m /models/Qwen_Qwen3-0.6B-Q6_K.gguf -c 1024
[...]
llama_kv_cache_init:        CPU KV buffer size =  3584.00 MiB
llama_new_context_with_model: KV self size  = 3584.00 MiB, K (f16): 1792.00 MiB, V (f16): 1792.00 MiB

./llama-server -m /models/Qwen_Qwen3-0.6B-Q6_K.gguf -c 1024 --cache-type-k q8_0 --cache-type-v q8_0
[...]
llama_kv_cache_init:        CPU KV buffer size =    59.50 MiB
llama_new_context_with_model: KV self size  =   59.50 MiB, K (q8_0):   29.75 MiB, V (q8_0):   29.75 MiB

• To have access to more quant types, build with GGML_IQK_FA_ALL_QUANTS=ON, otherwise only F16, Q8_0, Q6_0, and, if the CPU provides native BF16 support, BF16 FA kernels will be included. After PR 1549, on CPU are enabled as well Q4_1, IQ4_NL, Q4_0 by default to allow people experiment; use GGML_IQK_FA_ALL_QUANTS=OFF to reduce build time if those quants are not needed.
• K-cache may need better quant than V-cache to reduce quality loss, they can be specified separately --cache-type-k q8_0 --cache-type-v q8_0
• It needs FA --flash-attn flag, which is already turned on by default.
• Fast quant type Q8_KV -ctk q8_KV PR 208
• Using --k-cache-hadamard on quants lower than Q6_0 may give better results. Additionally, ik_llama.cpp provides --v-cache-hadamard for the V-cache. Example: --cache-type-k q6_0 --k-cache-hadamard --cache-type-v q6_0 --v-cache-hadamard
• Q4_0 achieves low perplexity even without Hadamard.

3. Offload less to the GPU. Try to find a mix of parameters that better suits your system that default.

• Try --fit. ik_llama.cpp automatically determine which tensors to offload to the GPUs based on the available VRAM.

• Use --no-kv-offload to keep KV cache on CPU. This is provided for flexibility, and practically not desired as reduces the prompt processing speed.

• Identify tensors, how many layers (also shape and more metadata) by opening the GGUF model file on the Web browser bartowski/Qwen_Qwen3-0.6B-IQ4_NL.gguf then scroll down to the Tensors table. For the split models, look to each file part.

Or, if you already have the quant locally you can just run gguf_dump.py:

python3 gguf-py/scripts/gguf_dump.py /models/Qwen_Qwen3-0.6B-IQ4_NL.gguf

• Use --dry-run to observe the memory usage.

• -ngl, -ot, --cpu-moe, --n-cpu-moe N

  • For MoE models, use a number greater than the number of model layers with -ngl. If unsure, use a large number like -ngl 999.
  • It's good to explicitly put up/down/gate onto the GPU for speedups.
  • Up/Gate shouldn't be on separate GPU devices because it might cause a bit of a deadlock.
  • For models with shared experts (like GPT-OSS), they should end up on GPU.
  • In some quants the layers aren't uniform so it can be better to skip larger layers if more smaller blocks will fit without empty space where nothing fits.
  • You put anything that says "exps" in your slowest memory, and anything else in your fastest memory (VRAM). Those ffn "exps" are the sparse experts tensors, the ones that get actually used only 2-5% of the times (depending on the model). If then you have extra VRAM to spare, you start putting some of the exps into VRAM too, for some improvements.
  • Some layers (layers are called blk.n in gguf), are different in some models. For example GLM5 the first three layers are different (blk.0(14), blk.1(14), blk.2(14) vs. blk.10(19), blk.11(19),...), they don't have exps, they have dense ffn, so they should all go in VRAM. Dense layers are very good to speed up mixed inference systems, as a much larger share of active parameters is fixed, and hence you know which to put in faster VRAM. Also the layers from the 4th onwards have shared exps, "shexp", those too go to VRAM as they are always active.
  • For MoE models you can play with --cpu-moe, --n-cpu-moe N, -ooae/-no-ooae before moving to -ot.
  • In general, in a single GPU + CPU system, you just do something like this:

  -ngl 999 To put all layers in VRAM by default

  -ot "blk.(?:[0-9]|[1-7][0-9]|[8][0-7]).ffn._exps.=CPU" To create exceptions and put back in ram anything that has "ffn" and "_exps" in its name, and that sits in layers called "blk.n", where "n" (the lawyer number) is any match between 0 and 9, or between 1 to 7 + 0 to 9 (aka a number between 10 and 79), or 8 + 0 to 7 (aka a number between 80 and 87). Basically a complicated way of saying put all experts from layer 0 to 87 in ram. Experts from layer 88 to 93 (there's 93 layers in qwen3vl 235b) can sit in VRAM still. (Thats all I can load on a 5090).

C. Other tips

• Ensure that you use a CUDA version that supports your GPU(s)
• Check for errors lspci -vvv | grep -F 'at lane'
• Multiple GPUs
  • nvidia-smi topo -p2p r
  • Change the order of GPUs with CUDA_VISIBLE_DEVICES=... until the best GPU is used appropriately, especially when GPUs are different (type, capability, slot speed, etc.).
  • If you are not happy with the allocations done by --fit across GPUs, use -ts to manually tweak.
  • Look for ReBAR/Resizable BAR support for your Motherboard, CPU, BIOS/UEFI and GPU. Then for the "patched driver" for your GPUs to enable GPU to GPU direct communication.

Common GPU configurations and popular models

WIP

Parameter	Description	Default	Notes/Examples
-ngl, --gpu-layers N	Number of layers to store in VRAM	-	For better speed you aim to offload the entire model in GPU memory. To identify how many layers (also shape and more metadata) open the GGUF model file on the Web browser bartowski/Qwen_Qwen3-0.6B-IQ4_NL.gguf then scroll down to the Tensors table. Use a number higher than the numbers of model layers to fully offload (--gpu-layers 99, for a model with less than 99 layers). See --ctx-size and reduce it to the minimum needed. If model fails to load due to the insufficient GPU memory, reduce the number of layers (--gpu-layers 20, for a model with 40 layers will offload only the first 20 layers).
-ngld, --gpu-layers-draft N	Number of layers to store in VRAM for the draft model	-	For draft model, see --gpu-layers
--cpu-moe	Keep all MoE weights in CPU memory	-	Simple offload mode for MoE. PR 841
--n-cpu-moe N	Keep MoE weights of the first N layers in CPU memory	-	Similar to --cpu-moe but when some GPU memory is available to store some layers.
-sm, --split-mode SPLIT_MODE	How to split the model across multiple GPUs	none	When you have more than one GPU, how to split the model across multiple GPUs, one of: - none: use one GPU only. - graph: split model tensors and computation graph across GPUs. graph is exclusive here and extremely effective for dense and MoE PR 1080. - layer: split layers and KV across GPUs Example: -sm graph
-ts, --tensor-split SPLIT	Fraction of the model to offload to each GPU (comma-separated)	-	Powerful for tweaking. Example: -ts 3,1
-dev, --device dev1,dev2	Comma-separated list of devices to use for offloading	none	If there are many GPUs available on the system and only selected ones need to be used. Example: -dev CUDA0,CUDA1
-devd, --device-draft dev1,dev2	Comma-separated list of devices for draft model	none	For draft model, see --device
-mg, --main-gpu i	The GPU to use for the model (with split-mode = none)	-
-cuda fa-offset=value	FP16 precision offset for FA calculation	0	Rarely, fp16 precision is inadequate, at least for some models, when computing FA for very long contexts. Value must be a valid floating point number in the interval [0...3] (this is checked and if the supplied value is outside this interval it is ignored). By the default the offset is zero. If you find that a model works up to a given context length but then starts producing gibberish/incoherent output/endless repetitions, it is very likely it is due to f16 overflow in the FA calculation, and using this command line option is likely to solve it. PR 1198
-ot or --override-tensor	Override where model weights are stored	-	Override where model weights are stored using regular expressions. This allows for example to keep the MoE experts on the CPU and to offload only the attention and not repeating layers to the GPU. Example: \.ffn_.*_exps\.=CPU PR 232
-op or --offload-policy a,b	Manually define the offload policy	-	a and b are integers. One can have multiple pairs following the -op or --offload-policy argument (i.e., -op a1,b1,a2,b2,a3,b3...). The first integer defines the op (see below). The second integer is 0 or 1 and defines if the op should be offloaded (1) or not offloaded (0) to the GPU. The first integer is simply the enum value in the ggml_op enum. If the op is set to -1, then all op offloads are set to enabled or disabled. Examples: -op -1,0: disable all offload to the GPU -op 26,0: disable offload of matrix multiplications to the GPU -op 27,0: disable offload of indirect matrix multiplications to the GPU (used for the experts in a MoE model) -op 29,0: disable fused up-gate-unary op offload to the GPU (applied to MoE models with -fmoe) PR 405
--offload-only-active-experts or -ooae	On MOE offload only active experts	ON	-ooae is not related to where the model weights get stored. Instead, once we have some MoE tensors (ffn_(up
-no-ooae	Disable offload only active experts	-	See -ooae
-smf16, --split-mode-f16	Use f16 for data exchange between GPUs	1	PR 1087
-smf32, --split-mode-f32	Use f32 for data exchange between GPUs	0	PR 1087
-grt, --graph-reduce-type	Type for data exchange between GPUs	f32	q8_0 / bf16 / f16 / f32 Reduce the data transferred between GPUs PR 1154
-smgs, --split-mode-graph-scheduling	Force Split Mode Graph Scheduling	0	PR 1068
--max-gpu N	Define (and use) a maximum number of GPUs per layer with split mode "graph"		This is of interest when there are more than 2 GPUs available, but using all of them leads to a lower performance than using just 2 (or using the default split mode "layer") PR 1051
-cuda, --cuda-params	Comma-separated list of cuda parameters	-	Powerful way to tweak Fusion, GPU offload threshold, and MMQ-ID threshold. PR 910

Model Options

Parameter	Description	Default	Notes/Examples
--check-tensors	Check model tensor data for invalid values	false
--override-kv KEY=TYPE:VALUE	Override model metadata by key	-	Advanced option to override model metadata by key. May be specified multiple times. types: int, float, bool, str. Example: --override-kv tokenizer.ggml.add_bos_token=bool:false
-m, --model FNAME	Model path	models/$filename	Mandatory, the GGUF model file to be served.
-md, --model-draft FNAME	Draft model for speculative decoding	unused
--draft-max, --draft, --draft-n N	Number of tokens to draft for speculative decoding	16
--draft-min, --draft-n-min N	Minimum number of draft tokens to use for speculative decoding	-
--draft-p-min P	Minimum speculative decoding probability (greedy)	0.8

Server Options

Parameter	Description	Default	Notes/Examples
--host HOST	IP address to listen	127.0.0.1	Change to 0.0.0.0 when endpoint will be accessed from another computer. Keep in mind to never expose the server to Internet.
--port PORT	Port to listen	8080
--webui NAME	Controls which webui to server	auto	Flexibility in choosing the integrated powerful WebUIs: - none: disable webui - auto: default webui - llamacpp: llamacpp webui
--api-key KEY	API key to use for authentication	none	Add a custom API KEY. Clients will need to specify it when connecting.
-a, --alias	set alias for model name (to be used by REST API)	none	Configure the server to serve and reply with specific model name.

Other Tools

sweep_bench

Benchmark utility that performs a series of prompt processing batches followed by TG. The KV cache is not cleared, so the N_KV columns tells you how many tokens were in the KV cache when the PP/TG was processed.

llama-sweep-bench understands all parameters that one would use in llama-server or llama-cli (but obviously not all get used, only those that are related to loading the model, setting up the context parameters, and running the benchmark).

llama-sweep-bench -m /models/model.gguf -c 12288 -ub 512 -rtr -fa -ctk q8_0 -ctv q8_0

Parameter	Description	Default	Notes/Examples
-nrep N, --n-repetitions N	Define the number of repetitions used at zero context	-	PR 1176
-n	Specifies he number of TG tokens	-	If not specified, it is set to u-batch/4 PR 897
--minilog	Reduce the verbosity	-	PR 1468

llama-bench

Benchmark utility.

llama-bench -tgb 4,16 -p 512 -n 128 other_arguments

Parameter	Description	Default	Notes/Examples
-tgb (or --threads-gen-batch)	Enable having different number of threads for generation and batch processing	-	PR 284
--fit	Automatically fit to available VRAM	0	0 / 1 PR 1542
--fit-margin N	Safety VRAM margin in MiB when using --fit	1024

Imatrix

Create imatrix from calibration dataset.

llama-imatrix -m /models/model-bf16.gguf -f /models/calibration_data_v5_rc.txt -o /models/model.imatrix

Parameter	Description	Default	Notes/Examples
--layer-similarity or -lsim	Collect statistics about activations change caused by a layer using cosine similarity	-	PR 328
--hide-imatrix	Store "top_secret" in the imatrix data file name	-	And in calibration dataset fields, and zeros in the batch size and number of chunks used to compute the imatrix. PR 329

Notes:

• Use convert_imatrix_gguf_to_dat.py to convert the "new" GGUF imatrix files to the format supported here. PR 1405
• imatrix calculation for models with merged ffn_up/gate_exps tensors is supported, see PR 1418 PR 1419

Quantization

Quantize models to reduce size and improve speed.

For very large models, is a good practice (to avoid re-download if some info like the jinja template needs fixes) to split the model and keep only the metadata in the first split.

llama-quantize --imatrix /models/model.imatrix /models/model-bf16.gguf /models/model-IQ4_NL.gguf IQ4_NL

llama-gguf-split --split --split-max-size 1G --no-tensor-first-split /models/model-IQ4_NL.gguf /models/parts/model-IQ4_NL.gguf

Parameter	Description	Default	Notes/Examples
--custom-q	Custom quantization rules with regular expressions	-	Example: llama-quantize --imatrix some_imatrix --custom-q "regex1=typ1,regex2=type2..." some_model some_output_file some_base_quant PR 244
--dry-run	Prints the tensor types and resulting tensor sizes, but does not run the quantization, so it is very fast.	-	Useful for experimenting with --custom-q before running the actual quantization. PR 1309
--partial-requant	quantize only missing split files in the split quantized .gguf destination directory	-	-

Build Arguments

Build with cmake.

In general, use as few build flags as possible. The building process automatically detects the available hardware features and enables them. Also, ik_llama.cpp have safe default options.

cmake -B build -DGGML_NATIVE=ON

cmake --build build --config Release -j$(nproc)

Argument	Notes/Examples
-DGGML_ARCH_FLAGS="-march=armv8.2-a+dotprod+fp16"	Direct access to ARCH options.
-DGGML_CUDA=ON	Build with CUDA support. By default it builds to native CUDA.
-DCMAKE_CUDA_ARCHITECTURES=86	Build for specific CUDA GPU Compute Capability, e.g. 8.6 for RTX30*0
-DGGML_RPC=ON	Build the RPC backend.
-DGGML_IQK_FA_ALL_QUANTS=ON	More KV quantization types PR 197
-DLLAMA_SERVER_SQLITE3=ON	Sqlite3 for mikupad
-DCMAKE_TOOLCHAIN_FILE=[...]	Example: on Windows tells cmake where is sqlite3.
-DGGML_NATIVE=ON	Turn it off when cross-compiling.
-DGGML_NCCL=OFF	To disable usage of NCCL.
-DGGML_MAX_CONTEXTS=2048	Only need this if you are planning to use quants generated with the Thireus quantization suite

Environment variables

Use them on the command line.

CUDA_VISIBLE_DEVICES=0,2 llama-server -m /models/model-bf16.gguf

Name	Notes/Examples
CUDA_VISIBLE_DEVICES	Use only specified GPUs. Example: Use first and 3rd CUDA_VISIBLE_DEVICES=0,2
GGML_CUDA_NO_PINNED	Do not use pinned memory

Unique parameters

WIP

	ik_llama.cpp exclusive	Not available on ik_llama.cpp
Parameter	-rtr

Graph parallel models

Models architectures supported by --split-mode graph

LLM_ARCH_LLAMA,
LLM_ARCH_QWEN3MOE,
LLM_ARCH_GLM4_MOE,
LLM_ARCH_MISTRAL3,
LLM_ARCH_COMMAND_R,
LLM_ARCH_COHERE2,
LLM_ARCH_MIMO2,
LLM_ARCH_QWEN3,
LLM_ARCH_QWEN3VL,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_ERNIE4_5_MOE,
LLM_ARCH_MINIMAX_M2,
LLM_ARCH_SEED_OSS,
LLM_ARCH_STEP35,
LLM_ARCH_QWEN35,
LLM_ARCH_QWEN35MOE,
LLM_ARCH_GEMMA4,