Skip to content

mistralrs bench

Run performance benchmarks for base or LoRA model generation

mistralrs bench [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
--no-kv-cache false Disable KV cache entirely
--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
--adapter <ADAPTER> LoRA adapter alias to benchmark. Omit to benchmark the base model
--prompt-len <PROMPT_LEN> 512 Prompt lengths for measuring time to first token (TTFT); 0 skips it. Comma-separate for sweeps
--gen-len <GEN_LEN> 128 Output tokens per decode request. Values below 2 skip decode metrics
--depth <DEPTH> 4 Context lengths for measuring decode time per output token (TPOT). Comma-separate for sweeps
--iterations <ITERATIONS> 3 Number of benchmark iterations
--warmup <WARMUP> 1 Number of warmup runs per benchmark case (discarded)

Auto-detect model type (recommended)

mistralrs bench 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 bench 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 bench 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 bench 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 bench 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 bench 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)