Struct RuntimeCliConfig 

pub struct RuntimeCliConfig {

    pub preset: Option<String>,
    pub kv_dtype: Option<String>,
    pub kv_max_blocks: Option<usize>,
    pub kv_capacity: Option<usize>,
    pub paged_max_seqs: Option<usize>,
    pub recurrent_state_max_slots: Option<usize>,
    pub attention_policy: Option<AttentionExecutionPolicy>,
    pub max_batched_tokens: Option<usize>,
    pub scheduler_prefill_first_until_active: Option<usize>,
    pub scheduler_active_decode_prefill_chunk: Option<usize>,
    pub prefix_cache: Option<bool>,
    pub layer_split_pipeline_mode: Option<String>,
    pub moe_graph: Option<bool>,
    pub batched_graph: Option<bool>,
    pub reusable_execution: Option<bool>,
    pub reusable_execution_exact_decode_widths: Option<Vec<usize>>,
    pub reusable_execution_max_automatic_exact_decode_width: Option<usize>,
    pub unified_graph: Option<bool>,
    pub unified_graph_layers_only: Option<bool>,
    pub unified_graph_lm_head_eager: Option<bool>,
    pub batch_decode_prof: Option<bool>,
    pub batch_prefill_prof: Option<bool>,
    pub next_batch_prof: Option<bool>,
    pub rbd_prof: Option<bool>,
    pub unified_post_prof: Option<bool>,
    pub decode_op_profile: Option<bool>,
    pub prefill_op_profile: Option<bool>,
    pub marlin_profile: Option<bool>,
    pub marlin_trace_shapes: Option<bool>,
    pub marlin_trace_shapes_max: Option<usize>,
    pub use_vllm_paged_attn: Option<bool>,
    pub vllm_paged_attn_v1_short: Option<bool>,
    pub vllm_moe: Option<bool>,
    pub vllm_moe_pair_ids: Option<bool>,
    pub greedy_argmax: Option<bool>,
    pub fa_layout_varlen: Option<bool>,
    pub fa2_source: Option<bool>,
    pub fa2_direct_ffi: Option<bool>,
    pub fa2_direct_ffi_shim: Option<String>,
    pub fa2_native_manifest: Option<String>,
    pub fa2_native_artifact: Option<String>,
    pub fa2_native_source_sha256: Option<String>,
    pub fa2_native_inputs_sha256: Option<String>,
    pub max_model_len: Option<usize>,
    pub sequence_fit_policy: Option<SequenceFitPolicy>,
    pub moe_batch_threshold: Option<usize>,
}

Runtime knobs that can be sourced from the CLI config file.

Fields

preset: Option<String>

Named startup/runtime preset. Presets provide product-owned default bundles and can still be overridden by explicit runtime keys below, environment variables, or CLI flags.

kv_dtype: Option<String>

KV cache dtype override, equivalent to --kv-dtype or FERRUM_KV_DTYPE.

kv_max_blocks: Option<usize>

KV block budget, equivalent to FERRUM_KV_MAX_BLOCKS.

kv_capacity: Option<usize>

Per-sequence KV token capacity, equivalent to FERRUM_KV_CAPACITY.

paged_max_seqs: Option<usize>

Maximum paged-KV sequence count, equivalent to FERRUM_PAGED_MAX_SEQS.

recurrent_state_max_slots: Option<usize>

Generic recurrent-state slot-pool size, equivalent to FERRUM_RECURRENT_STATE_MAX_SLOTS.

attention_policy: Option<AttentionExecutionPolicy>

Attention provider-family policy for the plan runtime. Physical V1/V2/varlen selection remains adaptive inside the compiled provider.

max_batched_tokens: Option<usize>

Scheduler/model max batched-token budget, equivalent to FERRUM_MAX_BATCHED_TOKENS.

scheduler_prefill_first_until_active: Option<usize>

Prefer prefilling until this many requests are active, equivalent to FERRUM_SCHED_PREFILL_FIRST_UNTIL_ACTIVE.

scheduler_active_decode_prefill_chunk: Option<usize>

Cap prefill chunks while decode requests are active, equivalent to FERRUM_ACTIVE_DECODE_PREFILL_CHUNK.

prefix_cache: Option<bool>

Prefix cache opt-in, equivalent to FERRUM_PREFIX_CACHE.

layer_split_pipeline_mode: Option<String>

Layer-split decode pipeline mode, equivalent to FERRUM_LAYER_SPLIT_PIPELINE_MODE.

moe_graph: Option<bool>

MoE CUDA graph policy override, equivalent to FERRUM_MOE_GRAPH.

batched_graph: Option<bool>

Legacy Llama/Gemma batched decode CUDA graph policy override, equivalent to FERRUM_BATCHED_GRAPH.

reusable_execution: Option<bool>

vNext reusable device-program policy override, equivalent to FERRUM_REUSABLE_EXECUTION.

reusable_execution_exact_decode_widths: Option<Vec<usize>>

Exact vNext reusable decode widths prepared at startup. Omitted means automatic exact resolution up to the admission and startup hard bounds.

reusable_execution_max_automatic_exact_decode_width: Option<usize>

Configurable automatic ceiling, bounded by the independent hard startup capture limit. It does not cap runtime concurrency.

unified_graph: Option<bool>

Unified Llama/Gemma decode CUDA graph policy override, equivalent to FERRUM_UNIFIED_GRAPH.

unified_graph_layers_only: Option<bool>

Diagnostic unified graph scope that captures only transformer layers, equivalent to FERRUM_UNIFIED_GRAPH_LAYERS_ONLY.

unified_graph_lm_head_eager: Option<bool>

Diagnostic unified graph scope that leaves lm_head eager, equivalent to FERRUM_UNIFIED_GRAPH_LM_HEAD_EAGER.

batch_decode_prof: Option<bool>

Emit engine batch iteration profile logs, equivalent to FERRUM_BATCH_DECODE_PROF.

batch_prefill_prof: Option<bool>

Emit executor batch prefill profile logs, equivalent to FERRUM_BATCH_PREFILL_PROF.

next_batch_prof: Option<bool>

Emit engine next-batch scheduler profile logs, equivalent to FERRUM_NEXT_BATCH_PROF.

rbd_prof: Option<bool>

Emit route/batched-decode profile logs, equivalent to FERRUM_RBD_PROF.

unified_post_prof: Option<bool>

Emit unified decode postprocess profile logs, equivalent to FERRUM_UNIFIED_POST_PROF.

decode_op_profile: Option<bool>

Emit model decode operator profile logs, equivalent to FERRUM_DECODE_OP_PROFILE.

prefill_op_profile: Option<bool>

Emit model prefill operator profile logs, equivalent to FERRUM_PREFILL_OP_PROFILE.

marlin_profile: Option<bool>

Emit dense Marlin inner-kernel timing counters, equivalent to FERRUM_MARLIN_PROFILE.

marlin_trace_shapes: Option<bool>

Emit dense Marlin shape/label trace lines, equivalent to FERRUM_MARLIN_TRACE_SHAPES.

marlin_trace_shapes_max: Option<usize>

Maximum dense Marlin shape/label trace lines, equivalent to FERRUM_MARLIN_TRACE_SHAPES_MAX.

use_vllm_paged_attn: Option<bool>

vLLM paged attention policy, equivalent to FERRUM_USE_VLLM_PAGED_ATTN.

vllm_paged_attn_v1_short: Option<bool>

Short-context vLLM paged-attention v1 policy, equivalent to FERRUM_VLLM_PAGED_ATTN_V1_SHORT.

vllm_moe: Option<bool>

vLLM-Marlin MoE dispatch policy, equivalent to FERRUM_VLLM_MOE.

vllm_moe_pair_ids: Option<bool>

vLLM-MoE pair-id route layout policy, equivalent to FERRUM_VLLM_MOE_PAIR_IDS.

greedy_argmax: Option<bool>

GPU greedy argmax readback policy, equivalent to FERRUM_GREEDY_ARGMAX.

fa_layout_varlen: Option<bool>

FA-compatible varlen K/V layout policy, equivalent to FERRUM_FA_LAYOUT_VARLEN.

fa2_source: Option<bool>

Source-linked FA2 policy, equivalent to FERRUM_FA2_SOURCE.

fa2_direct_ffi: Option<bool>

Runtime-loaded FA2 direct FFI policy, equivalent to FERRUM_FA2_DIRECT_FFI.

fa2_direct_ffi_shim: Option<String>

Runtime-loaded FA2 direct FFI shim path, equivalent to FERRUM_FA2_DIRECT_FFI_SHIM.

fa2_native_manifest: Option<String>

Ferrum native FA2 operator manifest path, equivalent to FERRUM_FA2_NATIVE_MANIFEST.

fa2_native_artifact: Option<String>

Ferrum native FA2 operator artifact path, equivalent to FERRUM_FA2_NATIVE_ARTIFACT.

fa2_native_source_sha256: Option<String>

Ferrum native FA2 source package sha256 pin, equivalent to FERRUM_FA2_NATIVE_SOURCE_SHA256.

fa2_native_inputs_sha256: Option<String>

Ferrum native FA2 input tree sha256 pin, equivalent to FERRUM_FA2_NATIVE_INPUTS_SHA256.

max_model_len: Option<usize>

Requested max model length, equivalent to FERRUM_MAX_MODEL_LEN.

sequence_fit_policy: Option<SequenceFitPolicy>

Sequence fit gate used before prefill admission.

moe_batch_threshold: Option<usize>

Minimum MoE batch size for the batched expert path, equivalent to FERRUM_MOE_BATCH_THRESHOLD.

Implementations

impl RuntimeCliConfig

pub fn runtime_config_entries(&self) -> Vec<RuntimeConfigEntry>

Trait Implementations

impl Clone for RuntimeCliConfig

fn clone(&self) -> RuntimeCliConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RuntimeCliConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for RuntimeCliConfig

fn default() -> RuntimeCliConfig

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for RuntimeCliConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for RuntimeCliConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.