Skip to main content

Параметры llama-cli

general:

  -h,    --help, --usage          print usage and exit
         --version                show version and build info
  -v,    --verbose                print verbose information
         --verbosity N            set specific verbosity level (default: 0)
         --verbose-prompt         print a verbose prompt before generation (default: false)
         --no-display-prompt      don't print prompt at generation (default: false)
  -co,   --color                  colorise output to distinguish prompt and user input from generations (default: false)
  -s,    --seed SEED              RNG seed (default: -1, use random seed for < 0)
  -t,    --threads N              number of threads to use during generation (default: -1)
  -tb,   --threads-batch N        number of threads to use during batch and prompt processing (default: same as --threads)
  -td,   --threads-draft N        number of threads to use during generation (default: same as --threads)
  -tbd,  --threads-batch-draft N  number of threads to use during batch and prompt processing (default: same as --threads-draft)
         --draft N                number of tokens to draft for speculative decoding (default: 5)
  -ps,   --p-split N              speculative decoding split probability (default: 0.1)
  -lcs,  --lookup-cache-static FNAME
                                  path to static lookup cache to use for lookup decoding (not updated by generation)
  -lcd,  --lookup-cache-dynamic FNAME
                                  path to dynamic lookup cache to use for lookup decoding (updated by generation)
  -c,    --ctx-size N             size of the prompt context (default: 0, 0 = loaded from model)
  -n,    --predict N              number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
  -b,    --batch-size N           logical maximum batch size (default: 2048)
  -ub,   --ubatch-size N          physical maximum batch size (default: 512)
         --keep N                 number of tokens to keep from the initial prompt (default: 0, -1 = all)
         --chunks N               max number of chunks to process (default: -1, -1 = all)
  -fa,   --flash-attn             enable Flash Attention (default: disabled)
  -p,    --prompt PROMPT          prompt to start generation with
                                  in conversation mode, this will be used as system prompt
                                  (default: '')
  -f,    --file FNAME             a file containing the prompt (default: none)
         --in-file FNAME          an input file (repeat to specify multiple files)
  -bf,   --binary-file FNAME      binary file containing the prompt (default: none)
  -e,    --escape                 process escapes sequences (\n, \r, \t, \', \", \\) (default: true)
         --no-escape              do not process escape sequences
  -ptc,  --print-token-count N    print token count every N tokens (default: -1)
         --prompt-cache FNAME     file to cache prompt state for faster startup (default: none)
         --prompt-cache-all       if specified, saves user input and generations to cache as well
                                  not supported with --interactive or other interactive options
         --prompt-cache-ro        if specified, uses the prompt cache but does not update it
  -r,    --reverse-prompt PROMPT  halt generation at PROMPT, return control in interactive mode
                                  can be specified more than once for multiple prompts
  -sp,   --special                special tokens output enabled (default: false)
  -cnv,  --conversation           run in conversation mode, does not print special tokens and suffix/prefix
                                  if suffix/prefix are not specified, default chat template will be used
                                  (default: false)
  -i,    --interactive            run in interactive mode (default: false)
  -if,   --interactive-first      run in interactive mode and wait for input right away (default: false)
  -mli,  --multiline-input        allows you to write or paste multiple lines without ending each in '\'
         --in-prefix-bos          prefix BOS to user inputs, preceding the `--in-prefix` string
         --in-prefix STRING       string to prefix user inputs with (default: empty)
         --in-suffix STRING       string to suffix after user inputs with (default: empty)
         --no-warmup              skip warming up the model with an empty run
         --spm-infill             use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled)

sampling:

         --samplers SAMPLERS      samplers that will be used for generation in the order, separated by ';'
                                  (default: top_k;tfs_z;typical_p;top_p;min_p;temperature)
         --sampling-seq SEQUENCE  simplified sequence for samplers that will be used (default: kfypmt)
         --ignore-eos             ignore end of stream token and continue generating (implies --logit-bias EOS-inf)
         --penalize-nl            penalize newline tokens (default: false)
         --temp N                 temperature (default: 0.8)
         --top-k N                top-k sampling (default: 40, 0 = disabled)
         --top-p N                top-p sampling (default: 0.9, 1.0 = disabled)
         --min-p N                min-p sampling (default: 0.1, 0.0 = disabled)
         --tfs N                  tail free sampling, parameter z (default: 1.0, 1.0 = disabled)
         --typical N              locally typical sampling, parameter p (default: 1.0, 1.0 = disabled)
         --repeat-last-n N        last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size)
         --repeat-penalty N       penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)
         --presence-penalty N     repeat alpha presence penalty (default: 0.0, 0.0 = disabled)
         --frequency-penalty N    repeat alpha frequency penalty (default: 0.0, 0.0 = disabled)
         --dynatemp-range N       dynamic temperature range (default: 0.0, 0.0 = disabled)
         --dynatemp-exp N         dynamic temperature exponent (default: 1.0)
         --mirostat N             use Mirostat sampling.
                                  Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.
                                  (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)
         --mirostat-lr N          Mirostat learning rate, parameter eta (default: 0.1)
         --mirostat-ent N         Mirostat target entropy, parameter tau (default: 5.0)
         -l TOKEN_ID(+/-)BIAS     modifies the likelihood of token appearing in the completion,
                                  i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
                                  or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'
         --cfg-negative-prompt PROMPT
                                  negative prompt to use for guidance (default: '')
         --cfg-negative-prompt-file FNAME
                                  negative prompt file to use for guidance
         --cfg-scale N            strength of guidance (default: 1.0, 1.0 = disable)
         --chat-template JINJA_TEMPLATE
                                  set custom jinja chat template (default: template taken from model's metadata)
                                  if suffix/prefix are specified, template will be disabled
                                  only commonly used templates are accepted:
                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template

grammar:

         --grammar GRAMMAR        BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '')
         --grammar-file FNAME     file to read grammar from
  -j,    --json-schema SCHEMA     JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
                                  For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead

embedding:

         --pooling {none,mean,cls,last}
                                  pooling type for embeddings, use model default if unspecified
         --attention {causal,non-causal}
                                  attention type for embeddings, use model default if unspecified

context hacking:

         --rope-scaling {none,linear,yarn}
                                  RoPE frequency scaling method, defaults to linear unless specified by the model
         --rope-scale N           RoPE context scaling factor, expands context by a factor of N
         --rope-freq-base N       RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
         --rope-freq-scale N      RoPE frequency scaling factor, expands context by a factor of 1/N
         --yarn-orig-ctx N        YaRN: original context size of model (default: 0 = model training context size)
         --yarn-ext-factor N      YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
         --yarn-attn-factor N     YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
         --yarn-beta-slow N       YaRN: high correction dim or alpha (default: 1.0)
         --yarn-beta-fast N       YaRN: low correction dim or beta (default: 32.0)
  -gan,  --grp-attn-n N           group-attention factor (default: 1)
  -gaw,  --grp-attn-w N           group-attention width (default: 512.0)
  -dkvc, --dump-kv-cache          verbose print of the KV cache
  -nkvo, --no-kv-offload          disable KV offload
  -ctk,  --cache-type-k TYPE      KV cache data type for K (default: f16)
  -ctv,  --cache-type-v TYPE      KV cache data type for V (default: f16)

perplexity:

         --all-logits             return logits for all tokens in the batch (default: false)
         --hellaswag              compute HellaSwag score over random tasks from datafile supplied with -f
         --hellaswag-tasks N      number of tasks to use when computing the HellaSwag score (default: 400)
         --winogrande             compute Winogrande score over random tasks from datafile supplied with -f
         --winogrande-tasks N     number of tasks to use when computing the Winogrande score (default: 0)
         --multiple-choice        compute multiple choice score over random tasks from datafile supplied with -f
         --multiple-choice-tasks N
                                  number of tasks to use when computing the multiple choice score (default: 0)
         --kl-divergence          computes KL-divergence to logits provided via --kl-divergence-base
         --ppl-stride N           stride for perplexity calculation (default: 0)
         --ppl-output-type {0,1}  output type for perplexity calculation (default: 0)

parallel:

  -dt,   --defrag-thold N         KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
  -np,   --parallel N             number of parallel sequences to decode (default: 1)
  -ns,   --sequences N            number of sequences to decode (default: 1)
  -cb,   --cont-batching          enable continuous batching (a.k.a dynamic batching) (default: enabled)
  -nocb, --no-cont-batching       disable continuous batching

multi-modality:

         --mmproj FILE            path to a multimodal projector file for LLaVA. see examples/llava/README.md
         --image FILE             path to an image file. use with multimodal models. Specify multiple times for batching

backend:

         --rpc SERVERS            comma separated list of RPC servers
         --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 if not using mlock)
         --numa TYPE              attempt optimizations that help on some NUMA systems
                                    - distribute: spread execution evenly over all nodes
                                    - isolate: only spawn threads on CPUs on the node that execution started on
                                    - numactl: use the CPU map provided by numactl
                                  if run without this previously, it is recommended to drop the system page cache before using this
                                  see https://github.com/ggerganov/llama.cpp/issues/1437

model:

         --check-tensors          check model tensor data for invalid values (default: false)
         --override-kv KEY=TYPE:VALUE
                                  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
         --lora FNAME             apply LoRA adapter (can be repeated to use multiple adapters)
         --lora-scaled FNAME S    apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)
         --control-vector FNAME   add a control vector
                                  note: this argument can be repeated to add multiple control vectors
         --control-vector-scaled FNAME SCALE
                                  add a control vector with user defined scaling SCALE
                                  note: this argument can be repeated to add multiple scaled control vectors
         --control-vector-layer-range START END
                                  layer range to apply the control vector(s) to, start and end inclusive
  -m,    --model FNAME            model path (default: models/$filename with filename from --hf-file
                                  or --model-url if set, otherwise models/7B/ggml-model-f16.gguf)
  -md,   --model-draft FNAME      draft model for speculative decoding (default: unused)
  -mu,   --model-url MODEL_URL    model download url (default: unused)
  -hfr,  --hf-repo REPO           Hugging Face model repository (default: unused)
  -hff,  --hf-file FILE           Hugging Face model file (default: unused)
  -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)

retrieval:

         --context-file FNAME     file to load context from (repeat to specify multiple files)
         --chunk-size N           minimum length of embedded text chunks (default: 64)
         --chunk-separator STRING 
                                  separator between chunks (default: '
                                  ')

passkey:

         --junk N                 number of times to repeat the junk text (default: 250)
         --pos N                  position of the passkey in the junk text (default: -1)

imatrix:

  -o,    --output FNAME           output file (default: 'imatrix.dat')
         --output-frequency N     output the imatrix every N iterations (default: 10)
         --save-frequency N       save an imatrix copy every N iterations (default: 0)
         --process-output         collect data for the output tensor (default: false)
         --no-ppl                 do not compute perplexity (default: true)
         --chunk N                start processing the input from chunk N (default: 0)

bench:

  -pps                            is the prompt shared across parallel sequences (default: false)
  -npp n0,n1,...                  number of prompt tokens
  -ntg n0,n1,...                  number of text generation tokens
  -npl n0,n1,...                  number of parallel prompts

embedding:

         --embd-normalize         normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
         --embd-output-format     empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
         --embd-separator         separator of embendings (default \n) for example "<#sep#>"

server:

         --host HOST              ip address to listen (default: 127.0.0.1)
         --port PORT              port to listen (default: 8080)
         --path PATH              path to serve static files from (default: )
         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
         --api-key KEY            API key to use for authentication (default: none)
         --api-key-file FNAME     path to file containing API keys (default: none)
         --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
         --ssl-cert-file FNAME    path to file a PEM-encoded SSL certificate
         --timeout N              server read/write timeout in seconds (default: 600)
         --threads-http N         number of threads used to process HTTP requests (default: -1)
         --system-prompt-file FNAME
                                  set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications
         --log-format {text,json} 
                                  log output format: json or text (default: json)
         --metrics                enable prometheus compatible metrics endpoint (default: disabled)
         --no-slots               disables slots monitoring endpoint (default: enabled)
         --slot-save-path PATH    path to save slot kv cache (default: disabled)
         --chat-template JINJA_TEMPLATE
                                  set custom jinja chat template (default: template taken from model's metadata)
                                  only commonly used templates are accepted:
                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
  -sps,  --slot-prompt-similarity SIMILARITY
                                  how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
                                  
         --lora-init-without-apply
                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)

logging:

         --simple-io              use basic IO for better compatibility in subprocesses and limited consoles
  -ld,   --logdir LOGDIR          path under which to save YAML logs (no logging if unset)
         --log-test               Run simple logging test
         --log-disable            Disable trace logs
         --log-enable             Enable trace logs
         --log-file FNAME         Specify a log filename (without extension)
         --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
         --log-append             Don't truncate the old log file.

cvector:

  -o,    --output FNAME           output file (default: 'control_vector.gguf')
         --positive-file FNAME    positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
         --negative-file FNAME    negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
         --pca-batch N            batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
         --pca-iter N             number of iterations used for PCA (default: 1000)
         --method {pca,mean}      dimensionality reduction method to be used (default: pca)

export-lora:

  -m,    --model                  model path from which to load base model (default '')
         --lora FNAME             path to LoRA adapter  (can be repeated to use multiple adapters)
         --lora-scaled FNAME S    path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)
  -o,    --output FNAME           output file (default: 'ggml-lora-merged-f16.gguf')