Skip to content

mistralrs serve

Start HTTP/MCP server and (optionally) the UI at /ui

mistralrs serve [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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--format <FORMAT> Model format: plain (safetensors), GGUF, or GGML. Auto-detected from -f when not specified. Possible values: plain, gguf, ggml.
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF projector override; auto-selected when unambiguous (semicolon-separated for multiple)
--tok-model-id <TOK_MODEL_ID> Optional model ID overriding configuration, tokenizer, and processor assets for a quantized model
--gqa <GQA> 1 GQA value 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 loaded LoRA aliases and, independently, resident adapter generations
--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 target. Inference commands select a matching GGUF or UQFF artifact when available. Source checkpoints without a matching UQFF use in-situ quantization. tune evaluates the requested level instead of selecting an artifact. Accepts numeric levels (2, 3, 4, 5, 6, 8) or supported quantization names
--isq <IN_SITU_QUANT> In-situ quantization target. Accepts numeric levels (2, 3, 4, 5, 6, 8) or raw quant names (q4k, q8_0, etc.). Supports compatible GGUF sources
--from-uqff <FROM_UQFF> UQFF artifact to load. Accepts a filename, numeric quantization level (2, 3, 4, 5, 6, 8), or quantization type (q4k, afq8, etc.). Report-declared artifacts and conventional shard names expand to all of their shards. Use semicolons only to list disjoint shards manually
--isq-organization <ISQ_ORGANIZATION> ISQ organization strategy: default or moqe
--imatrix <IMATRIX> imatrix file for enhanced quantization
--calibration-file <CALIBRATION_FILE> Calibration file for imatrix generation
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping
--paged-attn <MODE> auto PagedAttention mode - auto: enabled on CUDA, disabled on Metal/CPU (default) - on: force enable (fails if unsupported) - off: force disable. Possible values: auto, on, off.
--pa-context-len <CONTEXT_LEN> Allocate KV cache for this context length. If not specified, defaults to using 90% of available VRAM
--pa-memory-mb <MEMORY_MB> GPU memory to allocate in MBs (alternative to context-len)
--pa-memory-fraction <MEMORY_FRACTION> GPU memory utilization fraction 0.0-1.0 (alternative to context-len/memory-mb)
--pa-block-size <BLOCK_SIZE> Tokens per block (default: 32 on CUDA)
--pa-cache-type <CACHE_TYPE> auto KV cache quantization type
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Maximum logical tensor memory retained by the multimodal encoder cache, in MiB
--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> Maximum image dimension for device mapping
-p, --port <PORT> 1234 HTTP server port
--host <HOST> 0.0.0.0 Bind address
--no-ui false Disable the built-in web UI (served at /ui by default)
--mcp-port <MCP_PORT> Also expose the loaded model as an MCP server on this port (JSON-RPC 2.0 at POST /mcp)
--max-tool-rounds <MAX_TOOL_ROUNDS> Default maximum tool-call rounds for the agentic loop. Per-request values from the HTTP API override this. Safety cap: 256 if unset
--tool-dispatch-url <TOOL_DISPATCH_URL> URL to POST tool calls to for server-side execution. For security, this is only configurable server-side (not per-request via HTTP API)
--disable-access-log false Disable per-request HTTP access logs
--access-log-format <ACCESS_LOG_FORMAT> text Format for HTTP access logs. Possible values: text, json.
--access-log-health false Include health, metrics, docs, and UI requests in HTTP access logs
--disable-request-id-header false Disable the x-request-id response header
--disable-metrics false Disable Prometheus HTTP metrics and the metrics recorder
--max-seqs <MAX_SEQS> 32 Maximum concurrent sequences
--max-num-batched-tokens <MAX_NUM_BATCHED_TOKENS> 4096 Maximum tokens processed in one paged-attention scheduler step
--max-prefill-chunk-tokens <MAX_PREFILL_CHUNK_TOKENS> 512 CUDA prompt-token quantum used while decode is resident and for recurrent prefill batching
--max-decode-steps-before-prefill <MAX_DECODE_STEPS_BEFORE_PREFILL> 8 Maximum decode steps before a waiting prefill batch is admitted
--no-kv-cache false Disable KV cache entirely
--prefix-cache-n <PREFIX_CACHE_N> 16 Number of prefix caches to hold (0 to disable)
-c, --chat-template <CHAT_TEMPLATE> Custom chat template file (.json or .jinja)
-j, --jinja-explicit <JINJA_EXPLICIT> Explicit JINJA template override
--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 Enable MTP speculative decoding with the head built into the model checkpoint
--mtp-model <MTP_MODEL> MTP assistant model id or path
--mtp-n-predict <MTP_N_PREDICT> Number of MTP draft tokens to propose per target step
--mtp-draft-sampling <MTP_DRAFT_SAMPLING> auto MTP draft sampling policy. Auto uses probabilistic DFlash2 drafting when supported. Possible values: auto, greedy, probabilistic.
--mcp-config <MCP_CONFIG> Path to an MCP client configuration JSON. Also reads MCP_CONFIG_PATH if unset
--agent false Build a local agent: enables web search, Python code execution, and shell execution, runs the agentic tool loop with a per-session temp workdir. Equivalent to passing --enable-search --enable-code-execution --enable-shell together
--enable-search false Enable web search (requires embedding model)
--search-embedding-model <SEARCH_EMBEDDING_MODEL> Search embedding model to use. 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 Agent action permission mode. Possible values: auto, ask, deny.
--sandbox <MODE> auto Sandbox mode. Possible values: auto, on, off.
--sandbox-profile <PROFILE> Sandbox policy profile. Possible values: restricted, developer.
--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 permitted to the sandboxed session. Possible values: none, loopback, full.

Auto-detect model type (recommended)

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--format <FORMAT> Model format: plain (safetensors), GGUF, or GGML. Auto-detected from -f when not specified. Possible values: plain, gguf, ggml.
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF projector override; auto-selected when unambiguous (semicolon-separated for multiple)
--tok-model-id <TOK_MODEL_ID> Optional model ID overriding configuration, tokenizer, and processor assets for a quantized model
--gqa <GQA> 1 GQA value 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 loaded LoRA aliases and, independently, resident adapter generations
--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 target. Inference commands select a matching GGUF or UQFF artifact when available. Source checkpoints without a matching UQFF use in-situ quantization. tune evaluates the requested level instead of selecting an artifact. Accepts numeric levels (2, 3, 4, 5, 6, 8) or supported quantization names
--isq <IN_SITU_QUANT> In-situ quantization target. Accepts numeric levels (2, 3, 4, 5, 6, 8) or raw quant names (q4k, q8_0, etc.). Supports compatible GGUF sources
--from-uqff <FROM_UQFF> UQFF artifact to load. Accepts a filename, numeric quantization level (2, 3, 4, 5, 6, 8), or quantization type (q4k, afq8, etc.). Report-declared artifacts and conventional shard names expand to all of their shards. Use semicolons only to list disjoint shards manually
--isq-organization <ISQ_ORGANIZATION> ISQ organization strategy: default or moqe
--imatrix <IMATRIX> imatrix file for enhanced quantization
--calibration-file <CALIBRATION_FILE> Calibration file for imatrix generation
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping
--paged-attn <MODE> auto PagedAttention mode - auto: enabled on CUDA, disabled on Metal/CPU (default) - on: force enable (fails if unsupported) - off: force disable. Possible values: auto, on, off.
--pa-context-len <CONTEXT_LEN> Allocate KV cache for this context length. If not specified, defaults to using 90% of available VRAM
--pa-memory-mb <MEMORY_MB> GPU memory to allocate in MBs (alternative to context-len)
--pa-memory-fraction <MEMORY_FRACTION> GPU memory utilization fraction 0.0-1.0 (alternative to context-len/memory-mb)
--pa-block-size <BLOCK_SIZE> Tokens per block (default: 32 on CUDA)
--pa-cache-type <CACHE_TYPE> auto KV cache quantization type
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Maximum logical tensor memory retained by the multimodal encoder cache, in MiB
--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> Maximum image dimension for device mapping

Text generation model with explicit configuration

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--format <FORMAT> Model format: plain (safetensors), GGUF, or GGML. Auto-detected from -f when not specified. Possible values: plain, gguf, ggml.
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF projector override; auto-selected when unambiguous (semicolon-separated for multiple)
--tok-model-id <TOK_MODEL_ID> Optional model ID overriding configuration, tokenizer, and processor assets for a quantized model
--gqa <GQA> 1 GQA value 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 loaded LoRA aliases and, independently, resident adapter generations
--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 target. Inference commands select a matching GGUF or UQFF artifact when available. Source checkpoints without a matching UQFF use in-situ quantization. tune evaluates the requested level instead of selecting an artifact. Accepts numeric levels (2, 3, 4, 5, 6, 8) or supported quantization names
--isq <IN_SITU_QUANT> In-situ quantization target. Accepts numeric levels (2, 3, 4, 5, 6, 8) or raw quant names (q4k, q8_0, etc.). Supports compatible GGUF sources
--from-uqff <FROM_UQFF> UQFF artifact to load. Accepts a filename, numeric quantization level (2, 3, 4, 5, 6, 8), or quantization type (q4k, afq8, etc.). Report-declared artifacts and conventional shard names expand to all of their shards. Use semicolons only to list disjoint shards manually
--isq-organization <ISQ_ORGANIZATION> ISQ organization strategy: default or moqe
--imatrix <IMATRIX> imatrix file for enhanced quantization
--calibration-file <CALIBRATION_FILE> Calibration file for imatrix generation
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping
--paged-attn <MODE> auto PagedAttention mode - auto: enabled on CUDA, disabled on Metal/CPU (default) - on: force enable (fails if unsupported) - off: force disable. Possible values: auto, on, off.
--pa-context-len <CONTEXT_LEN> Allocate KV cache for this context length. If not specified, defaults to using 90% of available VRAM
--pa-memory-mb <MEMORY_MB> GPU memory to allocate in MBs (alternative to context-len)
--pa-memory-fraction <MEMORY_FRACTION> GPU memory utilization fraction 0.0-1.0 (alternative to context-len/memory-mb)
--pa-block-size <BLOCK_SIZE> Tokens per block (default: 32 on CUDA)
--pa-cache-type <CACHE_TYPE> auto KV cache quantization type

Multimodal model

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--format <FORMAT> Model format: plain (safetensors), GGUF, or GGML. Auto-detected from -f when not specified. Possible values: plain, gguf, ggml.
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF projector override; auto-selected when unambiguous (semicolon-separated for multiple)
--tok-model-id <TOK_MODEL_ID> Optional model ID overriding configuration, tokenizer, and processor assets for a quantized model
--gqa <GQA> 1 GQA value 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 loaded LoRA aliases and, independently, resident adapter generations
--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 target. Inference commands select a matching GGUF or UQFF artifact when available. Source checkpoints without a matching UQFF use in-situ quantization. tune evaluates the requested level instead of selecting an artifact. Accepts numeric levels (2, 3, 4, 5, 6, 8) or supported quantization names
--isq <IN_SITU_QUANT> In-situ quantization target. Accepts numeric levels (2, 3, 4, 5, 6, 8) or raw quant names (q4k, q8_0, etc.). Supports compatible GGUF sources
--from-uqff <FROM_UQFF> UQFF artifact to load. Accepts a filename, numeric quantization level (2, 3, 4, 5, 6, 8), or quantization type (q4k, afq8, etc.). Report-declared artifacts and conventional shard names expand to all of their shards. Use semicolons only to list disjoint shards manually
--isq-organization <ISQ_ORGANIZATION> ISQ organization strategy: default or moqe
--imatrix <IMATRIX> imatrix file for enhanced quantization
--calibration-file <CALIBRATION_FILE> Calibration file for imatrix generation
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping
--paged-attn <MODE> auto PagedAttention mode - auto: enabled on CUDA, disabled on Metal/CPU (default) - on: force enable (fails if unsupported) - off: force disable. Possible values: auto, on, off.
--pa-context-len <CONTEXT_LEN> Allocate KV cache for this context length. If not specified, defaults to using 90% of available VRAM
--pa-memory-mb <MEMORY_MB> GPU memory to allocate in MBs (alternative to context-len)
--pa-memory-fraction <MEMORY_FRACTION> GPU memory utilization fraction 0.0-1.0 (alternative to context-len/memory-mb)
--pa-block-size <BLOCK_SIZE> Tokens per block (default: 32 on CUDA)
--pa-cache-type <CACHE_TYPE> auto KV cache quantization type
--encoder-cache-memory-mb <ENCODER_CACHE_MEMORY_MB> Maximum logical tensor memory retained by the multimodal encoder cache, in MiB
--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> Maximum image dimension for device mapping

Image generation model (diffusion)

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping

Speech synthesis model

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping

Embedding model

mistralrs serve 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 Model data type
--hf-overrides <HF_OVERRIDES> Recursively merged JSON overrides for the Hugging Face model config
--max-model-len <MAX_MODEL_LEN> Runtime model context length
--format <FORMAT> Model format: plain (safetensors), GGUF, or GGML. Auto-detected from -f when not specified. Possible values: plain, gguf, ggml.
-f, --quantized-file <QUANTIZED_FILE> GGUF/GGML filename(s); the suffix selects the format (semicolon-separated for multiple)
--mmproj <MMPROJ> GGUF projector override; auto-selected when unambiguous (semicolon-separated for multiple)
--tok-model-id <TOK_MODEL_ID> Optional model ID overriding configuration, tokenizer, and processor assets for a quantized model
--gqa <GQA> 1 GQA value for GGML models
--quant <QUANT> Quantization target. Inference commands select a matching GGUF or UQFF artifact when available. Source checkpoints without a matching UQFF use in-situ quantization. tune evaluates the requested level instead of selecting an artifact. Accepts numeric levels (2, 3, 4, 5, 6, 8) or supported quantization names
--isq <IN_SITU_QUANT> In-situ quantization target. Accepts numeric levels (2, 3, 4, 5, 6, 8) or raw quant names (q4k, q8_0, etc.). Supports compatible GGUF sources
--from-uqff <FROM_UQFF> UQFF artifact to load. Accepts a filename, numeric quantization level (2, 3, 4, 5, 6, 8), or quantization type (q4k, afq8, etc.). Report-declared artifacts and conventional shard names expand to all of their shards. Use semicolons only to list disjoint shards manually
--isq-organization <ISQ_ORGANIZATION> ISQ organization strategy: default or moqe
--imatrix <IMATRIX> imatrix file for enhanced quantization
--calibration-file <CALIBRATION_FILE> Calibration file for imatrix generation
--cpu false Force CPU-only execution
-n, --device-layers <DEVICE_LAYERS> Device layer mapping (format: ORD:NUM;… e.g., “0:10;1:20”) Omit for automatic device mapping
--topology <TOPOLOGY> Topology YAML file for device mapping
--hf-cache <HF_CACHE> Custom Hugging Face cache directory
--max-seq-len <MAX_SEQ_LEN> 4096 Max sequence length for automatic device mapping
--max-batch-size <MAX_BATCH_SIZE> 1 Max batch size for automatic device mapping
--paged-attn <MODE> auto PagedAttention mode - auto: enabled on CUDA, disabled on Metal/CPU (default) - on: force enable (fails if unsupported) - off: force disable. Possible values: auto, on, off.
--pa-context-len <CONTEXT_LEN> Allocate KV cache for this context length. If not specified, defaults to using 90% of available VRAM
--pa-memory-mb <MEMORY_MB> GPU memory to allocate in MBs (alternative to context-len)
--pa-memory-fraction <MEMORY_FRACTION> GPU memory utilization fraction 0.0-1.0 (alternative to context-len/memory-mb)
--pa-block-size <BLOCK_SIZE> Tokens per block (default: 32 on CUDA)
--pa-cache-type <CACHE_TYPE> auto KV cache quantization type