Skip to content

mistralrs run

Chat with a model interactively, or send a single prompt with -i

mistralrs run [OPTIONS] [COMMAND]
Option Default Description
-m, --model-id <MODEL_ID> Hugging Face model ID or local model directory; optional when -f names local files
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--format <FORMAT> Model file format. Auto-detected from -f when omitted
plain: Safetensors weights
gguf: GGUF file
ggml: GGML file
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF vision or audio projector file(s), separated by semicolons. Auto-selected when there is only one
--tok-model-id <TOK_MODEL_ID> Model ID to take the config, tokenizer, and processor files from, for a GGUF or GGML model
--gqa <GQA> 1 Grouped-query attention (GQA) factor for GGML models
--enable-lora false Enable dynamic LoRA without preloading an adapter. Supports compatible text and multimodal language models, including GGUF. Vision, audio, and projector adapters are unsupported
--lora <ALIAS=SOURCE|JSON> Preload a language-model LoRA adapter as ALIAS=SOURCE. Supports compatible text and multimodal language models, including GGUF. Remote adapters use revision main. May be repeated. Vision, audio, and projector adapters are unsupported
--lora-max-adapters <LORA_MAX_ADAPTERS> 16 Maximum LoRA adapters loaded at once. Separately caps adapter versions in memory, counting old versions still used by running requests
--lora-max-rank <LORA_MAX_RANK> 256 Maximum rank accepted for a LoRA adapter
--lora-max-bytes <BYTES> 8589934592 Maximum memory used by loaded adapters
--legacy-lora <SOURCE> Static LoRA adapter source for GGML or a Phi3 GGUF model
--legacy-lora-order <LEGACY_LORA_ORDER> Ordering JSON file for a legacy raw GGUF or GGML LoRA adapter
--xlora <XLORA> X-LoRA adapter model ID
--xlora-order <XLORA_ORDER> X-LoRA ordering JSON file
--tgt-non-granular-index <TGT_NON_GRANULAR_INDEX> Target non-granular index for X-LoRA
--quant <QUANT> Quantization level: a bit width (2, 3, 4, 5, 6, 8), a named type such as q4k or fp8, or auto to pick one that fits your hardware. Uses a matching prebuilt file when one exists, otherwise quantizes at load time
--isq <IN_SITU_QUANT> Quantize the weights at load time (in-situ quantization, ISQ): a bit width (2, 3, 4, 5, 6, 8) or a named type such as q4k or q8_0. Unlike --quant, never swaps in a prebuilt file
--from-uqff <FROM_UQFF> Prebuilt UQFF file to load: a filename, a bit width (2, 3, 4, 5, 6, 8), or a type such as q4k or afq8. Sharded files load every shard; separate names with ; to pick shards yourself
--isq-organization <ISQ_ORGANIZATION> Which layers to quantize: default for all supported layers, or moqe for only the experts of a mixture-of-experts model, leaving the rest unquantized
--imatrix <IMATRIX> Importance matrix (imatrix) file, which tells the quantizer which weights matter most and improves quality at low bit widths
--calibration-file <CALIBRATION_FILE> Text file to compute an importance matrix from, instead of passing --imatrix
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--paged-attn <MODE> auto Store the KV cache in PagedAttention blocks so more requests fit in memory
auto: On for CUDA and Metal, off for CPU
on: Always on; fails if the device does not support it
off: Always off
--gpu-memory-utilization <GPU_MEMORY_UTILIZATION> Fraction of GPU memory (e.g. 0.9) to use for weights, activations, and the KV cache
--kv-cache <KV_CACHE> Explicit KV cache capacity, such as 4GiB or 65536tokens
--pa-block-size <BLOCK_SIZE> Tokens per PagedAttention block (default: 32)
--pa-cache-type <CACHE_TYPE> auto KV cache data type: auto or f8e4m3 (8-bit float, about half the memory)
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Memory cap, in MiB, for cached image and audio encoder outputs
--max-edge <MAX_EDGE> Maximum edge length for image resizing (aspect ratio preserved)
--max-num-images <MAX_NUM_IMAGES> Maximum number of images per request
--max-image-length <MAX_IMAGE_LENGTH> Largest image side, in pixels, to reserve memory for when placing layers on devices
--max-seqs <MAX_SEQS> Maximum concurrent sequences (defaults to 1 for run and 32 for serve)
--max-num-batched-tokens <MAX_NUM_BATCHED_TOKENS> 4096 Maximum tokens processed per step when PagedAttention is on
--max-prefill-chunk-tokens <MAX_PREFILL_CHUNK_TOKENS> 512 Largest prompt chunk processed per step while other requests are generating (CUDA and Metal). Lower values keep running generations smooth; higher values give new requests a faster first token. Recurrent models always prefill in chunks of this size
--max-decode-steps-before-prefill <MAX_DECODE_STEPS_BEFORE_PREFILL> 8 Generation steps to run before pausing to process a waiting prompt. Higher values keep running generations smooth; lower values give new requests a faster first token
--no-kv-cache false Disable KV cache entirely
--prefix-cache-n <PREFIX_CACHE_N> Number of prefix caches to hold (0 to disable; defaults to up to 16 within the memory budget)
-c, --chat-template <CHAT_TEMPLATE> Custom chat template file (.json or .jinja)
-j, --jinja-explicit <JINJA_EXPLICIT> .jinja chat template file that always replaces the model’s own template
--matformer-config-path <MATFORMER_CONFIG_PATH> Path to a MatFormer config (CSV/JSON describing available slices). See model card
--matformer-slice-name <MATFORMER_SLICE_NAME> MatFormer slice to load (must match a slice name in the config file)
--mtp false Speed up generation with multi-token prediction (MTP), using the draft head built into the checkpoint
--mtp-model <MTP_MODEL> Separate MTP draft model, as a Hugging Face ID or local path
--mtp-n-predict <MTP_N_PREDICT> Fixed MTP draft tokens per step; omit to adapt the depth automatically
--mtp-draft-sampling <MTP_DRAFT_SAMPLING> auto How the MTP drafter picks the tokens it proposes
auto: Probabilistic when the drafter supports it, otherwise greedy
greedy: Always propose the most likely token
probabilistic: Sample proposals from the draft distribution; fails if unsupported
--mcp-config <MCP_CONFIG> Path to an MCP client configuration JSON. Also reads MCP_CONFIG_PATH if unset
--agent false Run as a local agent that can search the web and run Python and shell commands, each session in its own temp directory. Same as --enable-search --enable-code-execution --enable-shell
--enable-search false Give the model a web search tool. Also loads a small embedding model to rank results
--search-embedding-model <SEARCH_EMBEDDING_MODEL> Embedding model that ranks search results. Requires --enable-search or --agent. Possible values: embedding-gemma.
--enable-code-execution false Enable Python code execution tool (WARNING: allows arbitrary code execution)
--enable-shell false Enable shell execution tool (WARNING: allows arbitrary command execution)
--code-exec-python <CODE_EXEC_PYTHON> Python interpreter path for code execution. Requires code execution to be on (via --enable-code-execution or --agent). Defaults to python3
--code-exec-timeout <CODE_EXEC_TIMEOUT> Code execution timeout in seconds (default: 60). Requires code execution to be on
--code-exec-workdir <CODE_EXEC_WORKDIR> Working directory for code execution. Defaults to a temp dir; use “.” for cwd. Requires code execution to be on
--shell-path <SHELL_PATH> Shell executable path. Requires shell execution to be on. Defaults to /bin/sh
--shell-timeout <SHELL_TIMEOUT> Shell execution timeout in seconds (default: 600). Requires shell execution to be on
--shell-workdir <SHELL_WORKDIR> Root directory for per-session shell working directories. Defaults to temp dirs
--skills-dir <SKILLS_DIR> Directory for uploaded OpenAI-compatible Skills. Defaults to the system temp directory
--agent-permission <PERMISSION> auto Approval policy for the agent’s tool calls
auto: Run tool calls without asking
ask: Ask for approval before each tool call
deny: Block all tool calls
--sandbox <MODE> auto Sandbox for the processes started by the code and shell tools
auto: Sandbox on Linux and macOS, skipping protections the OS lacks; no sandbox on Windows
on: Always sandbox; fail if any protection is unavailable
off: No sandbox
--sandbox-profile <PROFILE> Files and network the sandbox allows. Defaults to developer when code or shell tools are on, otherwise restricted
restricted: Session directory and system files only
developer: Also reads installed toolchains (PATH, Homebrew, nvm, conda, and similar) with full network
--sb-max-memory-mb <MEMORY_MB> Per-session memory cap in MiB (default: 2048)
--sb-max-cpu-secs <CPU_SECS> Per-session CPU time cap in seconds (default: 600). Raised to at least enabled code/shell timeouts
--sb-max-procs <PROCS> Per-session process/thread cap (default: 64)
--sandbox-network <NETWORK> Network access inside the sandbox. Defaults to loopback for the restricted profile and full for developer
none: No network access
loopback: Localhost only
full: Unrestricted network access
--thinking <THINKING> Turn thinking on or off for models that support it; --thinking alone means true. Thinking is on unless this or --reasoning-effort turns it off. Possible values: true, false.
--reasoning-effort <REASONING_EFFORT> Set reasoning effort without changing the model’s sampling parameters. Values are off, low, medium, high, or xhigh. “none” is an alias for off
-i, --input <INPUT> One-shot text prompt. When provided, sends a single request and exits instead of entering interactive mode. Combine with –image, –video, or –audio for multimodal requests
--image <IMAGE> Image URL(s) or file path(s) to include in the request (requires -i). Can be specified multiple times: –image img1.jpg –image img2.png
--video <VIDEO> Video URL(s) or file path(s) to include in the request (requires -i). Can be specified multiple times: –video vid1.mp4 –video vid2.webm
--audio <AUDIO> Audio URL(s) or file path(s) to include in the request (requires -i). Can be specified multiple times: –audio audio1.wav –audio audio2.mp3
--adapter <ADAPTER> LoRA adapter alias to use for requests. Omit to run the base model

Auto-detect model type (recommended)

mistralrs run auto [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--format <FORMAT> Model file format. Auto-detected from -f when omitted
plain: Safetensors weights
gguf: GGUF file
ggml: GGML file
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF vision or audio projector file(s), separated by semicolons. Auto-selected when there is only one
--tok-model-id <TOK_MODEL_ID> Model ID to take the config, tokenizer, and processor files from, for a GGUF or GGML model
--gqa <GQA> 1 Grouped-query attention (GQA) factor for GGML models
--enable-lora false Enable dynamic LoRA without preloading an adapter. Supports compatible text and multimodal language models, including GGUF. Vision, audio, and projector adapters are unsupported
--lora <ALIAS=SOURCE|JSON> Preload a language-model LoRA adapter as ALIAS=SOURCE. Supports compatible text and multimodal language models, including GGUF. Remote adapters use revision main. May be repeated. Vision, audio, and projector adapters are unsupported
--lora-max-adapters <LORA_MAX_ADAPTERS> 16 Maximum LoRA adapters loaded at once. Separately caps adapter versions in memory, counting old versions still used by running requests
--lora-max-rank <LORA_MAX_RANK> 256 Maximum rank accepted for a LoRA adapter
--lora-max-bytes <BYTES> 8589934592 Maximum memory used by loaded adapters
--legacy-lora <SOURCE> Static LoRA adapter source for GGML or a Phi3 GGUF model
--legacy-lora-order <LEGACY_LORA_ORDER> Ordering JSON file for a legacy raw GGUF or GGML LoRA adapter
--xlora <XLORA> X-LoRA adapter model ID
--xlora-order <XLORA_ORDER> X-LoRA ordering JSON file
--tgt-non-granular-index <TGT_NON_GRANULAR_INDEX> Target non-granular index for X-LoRA
--quant <QUANT> Quantization level: a bit width (2, 3, 4, 5, 6, 8), a named type such as q4k or fp8, or auto to pick one that fits your hardware. Uses a matching prebuilt file when one exists, otherwise quantizes at load time
--isq <IN_SITU_QUANT> Quantize the weights at load time (in-situ quantization, ISQ): a bit width (2, 3, 4, 5, 6, 8) or a named type such as q4k or q8_0. Unlike --quant, never swaps in a prebuilt file
--from-uqff <FROM_UQFF> Prebuilt UQFF file to load: a filename, a bit width (2, 3, 4, 5, 6, 8), or a type such as q4k or afq8. Sharded files load every shard; separate names with ; to pick shards yourself
--isq-organization <ISQ_ORGANIZATION> Which layers to quantize: default for all supported layers, or moqe for only the experts of a mixture-of-experts model, leaving the rest unquantized
--imatrix <IMATRIX> Importance matrix (imatrix) file, which tells the quantizer which weights matter most and improves quality at low bit widths
--calibration-file <CALIBRATION_FILE> Text file to compute an importance matrix from, instead of passing --imatrix
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--paged-attn <MODE> auto Store the KV cache in PagedAttention blocks so more requests fit in memory
auto: On for CUDA and Metal, off for CPU
on: Always on; fails if the device does not support it
off: Always off
--gpu-memory-utilization <GPU_MEMORY_UTILIZATION> Fraction of GPU memory (e.g. 0.9) to use for weights, activations, and the KV cache
--kv-cache <KV_CACHE> Explicit KV cache capacity, such as 4GiB or 65536tokens
--pa-block-size <BLOCK_SIZE> Tokens per PagedAttention block (default: 32)
--pa-cache-type <CACHE_TYPE> auto KV cache data type: auto or f8e4m3 (8-bit float, about half the memory)
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Memory cap, in MiB, for cached image and audio encoder outputs
--max-edge <MAX_EDGE> Maximum edge length for image resizing (aspect ratio preserved)
--max-num-images <MAX_NUM_IMAGES> Maximum number of images per request
--max-image-length <MAX_IMAGE_LENGTH> Largest image side, in pixels, to reserve memory for when placing layers on devices

Text generation model with explicit configuration

mistralrs run text [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--format <FORMAT> Model file format. Auto-detected from -f when omitted
plain: Safetensors weights
gguf: GGUF file
ggml: GGML file
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF vision or audio projector file(s), separated by semicolons. Auto-selected when there is only one
--tok-model-id <TOK_MODEL_ID> Model ID to take the config, tokenizer, and processor files from, for a GGUF or GGML model
--gqa <GQA> 1 Grouped-query attention (GQA) factor for GGML models
--enable-lora false Enable dynamic LoRA without preloading an adapter. Supports compatible text and multimodal language models, including GGUF. Vision, audio, and projector adapters are unsupported
--lora <ALIAS=SOURCE|JSON> Preload a language-model LoRA adapter as ALIAS=SOURCE. Supports compatible text and multimodal language models, including GGUF. Remote adapters use revision main. May be repeated. Vision, audio, and projector adapters are unsupported
--lora-max-adapters <LORA_MAX_ADAPTERS> 16 Maximum LoRA adapters loaded at once. Separately caps adapter versions in memory, counting old versions still used by running requests
--lora-max-rank <LORA_MAX_RANK> 256 Maximum rank accepted for a LoRA adapter
--lora-max-bytes <BYTES> 8589934592 Maximum memory used by loaded adapters
--legacy-lora <SOURCE> Static LoRA adapter source for GGML or a Phi3 GGUF model
--legacy-lora-order <LEGACY_LORA_ORDER> Ordering JSON file for a legacy raw GGUF or GGML LoRA adapter
--xlora <XLORA> X-LoRA adapter model ID
--xlora-order <XLORA_ORDER> X-LoRA ordering JSON file
--tgt-non-granular-index <TGT_NON_GRANULAR_INDEX> Target non-granular index for X-LoRA
--quant <QUANT> Quantization level: a bit width (2, 3, 4, 5, 6, 8), a named type such as q4k or fp8, or auto to pick one that fits your hardware. Uses a matching prebuilt file when one exists, otherwise quantizes at load time
--isq <IN_SITU_QUANT> Quantize the weights at load time (in-situ quantization, ISQ): a bit width (2, 3, 4, 5, 6, 8) or a named type such as q4k or q8_0. Unlike --quant, never swaps in a prebuilt file
--from-uqff <FROM_UQFF> Prebuilt UQFF file to load: a filename, a bit width (2, 3, 4, 5, 6, 8), or a type such as q4k or afq8. Sharded files load every shard; separate names with ; to pick shards yourself
--isq-organization <ISQ_ORGANIZATION> Which layers to quantize: default for all supported layers, or moqe for only the experts of a mixture-of-experts model, leaving the rest unquantized
--imatrix <IMATRIX> Importance matrix (imatrix) file, which tells the quantizer which weights matter most and improves quality at low bit widths
--calibration-file <CALIBRATION_FILE> Text file to compute an importance matrix from, instead of passing --imatrix
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--paged-attn <MODE> auto Store the KV cache in PagedAttention blocks so more requests fit in memory
auto: On for CUDA and Metal, off for CPU
on: Always on; fails if the device does not support it
off: Always off
--gpu-memory-utilization <GPU_MEMORY_UTILIZATION> Fraction of GPU memory (e.g. 0.9) to use for weights, activations, and the KV cache
--kv-cache <KV_CACHE> Explicit KV cache capacity, such as 4GiB or 65536tokens
--pa-block-size <BLOCK_SIZE> Tokens per PagedAttention block (default: 32)
--pa-cache-type <CACHE_TYPE> auto KV cache data type: auto or f8e4m3 (8-bit float, about half the memory)

Multimodal model

mistralrs run multimodal [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--format <FORMAT> Model file format. Auto-detected from -f when omitted
plain: Safetensors weights
gguf: GGUF file
ggml: GGML file
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF vision or audio projector file(s), separated by semicolons. Auto-selected when there is only one
--tok-model-id <TOK_MODEL_ID> Model ID to take the config, tokenizer, and processor files from, for a GGUF or GGML model
--gqa <GQA> 1 Grouped-query attention (GQA) factor for GGML models
--enable-lora false Enable dynamic LoRA for the language model without preloading an adapter. Vision, audio, and projector adapters are unsupported
--lora <ALIAS=SOURCE|JSON> Preload a language-model LoRA adapter as ALIAS=SOURCE. Remote adapters use revision main. May be repeated. Vision, audio, and projector adapters are unsupported
--lora-max-adapters <LORA_MAX_ADAPTERS> 16 Maximum LoRA adapters loaded at once. Separately caps adapter versions in memory, counting old versions still used by running requests
--lora-max-rank <LORA_MAX_RANK> 256 Maximum rank accepted for a LoRA adapter
--lora-max-bytes <BYTES> 8589934592 Maximum memory used by loaded adapters
--quant <QUANT> Quantization level: a bit width (2, 3, 4, 5, 6, 8), a named type such as q4k or fp8, or auto to pick one that fits your hardware. Uses a matching prebuilt file when one exists, otherwise quantizes at load time
--isq <IN_SITU_QUANT> Quantize the weights at load time (in-situ quantization, ISQ): a bit width (2, 3, 4, 5, 6, 8) or a named type such as q4k or q8_0. Unlike --quant, never swaps in a prebuilt file
--from-uqff <FROM_UQFF> Prebuilt UQFF file to load: a filename, a bit width (2, 3, 4, 5, 6, 8), or a type such as q4k or afq8. Sharded files load every shard; separate names with ; to pick shards yourself
--isq-organization <ISQ_ORGANIZATION> Which layers to quantize: default for all supported layers, or moqe for only the experts of a mixture-of-experts model, leaving the rest unquantized
--imatrix <IMATRIX> Importance matrix (imatrix) file, which tells the quantizer which weights matter most and improves quality at low bit widths
--calibration-file <CALIBRATION_FILE> Text file to compute an importance matrix from, instead of passing --imatrix
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--paged-attn <MODE> auto Store the KV cache in PagedAttention blocks so more requests fit in memory
auto: On for CUDA and Metal, off for CPU
on: Always on; fails if the device does not support it
off: Always off
--gpu-memory-utilization <GPU_MEMORY_UTILIZATION> Fraction of GPU memory (e.g. 0.9) to use for weights, activations, and the KV cache
--kv-cache <KV_CACHE> Explicit KV cache capacity, such as 4GiB or 65536tokens
--pa-block-size <BLOCK_SIZE> Tokens per PagedAttention block (default: 32)
--pa-cache-type <CACHE_TYPE> auto KV cache data type: auto or f8e4m3 (8-bit float, about half the memory)
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Memory cap, in MiB, for cached image and audio encoder outputs
--max-edge <MAX_EDGE> Maximum edge length for image resizing (aspect ratio preserved)
--max-num-images <MAX_NUM_IMAGES> Maximum number of images per request
--max-image-length <MAX_IMAGE_LENGTH> Largest image side, in pixels, to reserve memory for when placing layers on devices

Image generation model (diffusion)

mistralrs run diffusion [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory

Speech synthesis model

mistralrs run speech [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory

Embedding model

mistralrs run embedding [OPTIONS] --model-id <MODEL_ID>
Option Default Description
-m, --model-id <MODEL_ID> required Hugging Face model ID or local path to model directory
-t, --tokenizer <TOKENIZER> Path to local tokenizer.json file
-a, --arch <ARCH> Model architecture (auto-detected if not specified)
--dtype <DTYPE> auto Weight data type: auto, bf16, f16, or f32. auto picks the best one for the device
--hf-overrides <HF_OVERRIDES> JSON to merge into the model’s config.json; nested objects merge key by key
--max-model-len <MAX_MODEL_LEN> Most tokens (prompt plus output) one request may use; default: what the model and KV cache allow
--format <FORMAT> Model file format. Auto-detected from -f when omitted
plain: Safetensors weights
gguf: GGUF file
ggml: GGML file
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF vision or audio projector file(s), separated by semicolons. Auto-selected when there is only one
--tok-model-id <TOK_MODEL_ID> Model ID to take the config, tokenizer, and processor files from, for a GGUF or GGML model
--gqa <GQA> 1 Grouped-query attention (GQA) factor for GGML models
--quant <QUANT> Quantization level: a bit width (2, 3, 4, 5, 6, 8), a named type such as q4k or fp8, or auto to pick one that fits your hardware. Uses a matching prebuilt file when one exists, otherwise quantizes at load time
--isq <IN_SITU_QUANT> Quantize the weights at load time (in-situ quantization, ISQ): a bit width (2, 3, 4, 5, 6, 8) or a named type such as q4k or q8_0. Unlike --quant, never swaps in a prebuilt file
--from-uqff <FROM_UQFF> Prebuilt UQFF file to load: a filename, a bit width (2, 3, 4, 5, 6, 8), or a type such as q4k or afq8. Sharded files load every shard; separate names with ; to pick shards yourself
--isq-organization <ISQ_ORGANIZATION> Which layers to quantize: default for all supported layers, or moqe for only the experts of a mixture-of-experts model, leaving the rest unquantized
--imatrix <IMATRIX> Importance matrix (imatrix) file, which tells the quantizer which weights matter most and improves quality at low bit widths
--calibration-file <CALIBRATION_FILE> Text file to compute an importance matrix from, instead of passing --imatrix
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Layers per device as ORD:NUM;..., e.g. 0:10;1:20 puts 10 layers on GPU 0 and 20 on GPU 1. Omit for automatic mapping
--topology <TOPOLOGY> YAML file setting the device and quantization for each layer range
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--paged-attn <MODE> auto Store the KV cache in PagedAttention blocks so more requests fit in memory
auto: On for CUDA and Metal, off for CPU
on: Always on; fails if the device does not support it
off: Always off
--gpu-memory-utilization <GPU_MEMORY_UTILIZATION> Fraction of GPU memory (e.g. 0.9) to use for weights, activations, and the KV cache
--kv-cache <KV_CACHE> Explicit KV cache capacity, such as 4GiB or 65536tokens
--pa-block-size <BLOCK_SIZE> Tokens per PagedAttention block (default: 32)
--pa-cache-type <CACHE_TYPE> auto KV cache data type: auto or f8e4m3 (8-bit float, about half the memory)