pub struct Fp8RouteInputs {
pub dense_fp8: bool,
pub tp_size: usize,
pub declared_variant: Option<Nvfp4Variant>,
pub gdn_fp8: bool,
pub w8a8_prefill_kernels: bool,
pub route: RouteEnv,
pub ffn_gateup_fused: bool,
}Expand description
The route gates load_layers applies, resolved from the things that exist
before the checkpoint does.
One field per clause of the loader’s own ffn_fp8 / attn_fp8 /
gdn_fp8_arm_selected conditions, so a reader can diff this struct
against qwen35_dense.rs:319, :578 and :172 line by line.
Fields§
§dense_fp8: boolATLAS_DENSE_FP8=1 — qwen35_dense::dense_fp8_enabled.
tp_size: usizeconfig.tp_world_size.max(1). The FP8 dense route is single-GPU only.
declared_variant: Option<Nvfp4Variant>The variant the CONFIG declares. None when config.json does not say
— the loader would sniff the store, which does not exist yet.
gdn_fp8: boolATLAS_NO_GDN_FP8 is unset — qwen35_dense::gdn_fp8_arm_selected.
w8a8_prefill_kernels: boolBoth qwen3_attention::W8A8_PREFILL_KERNELS are loaded for this
target. Decides whether the Q and O FP8 prefill twins get built.
route: RouteEnvThe environment-resolved dispatch route the loader itself reads.
ffn_gateup_fused: bool[defaults] ffn_gateup_fused, resolved through the SAME function the
dispatch site and the loader call — the fused gate+up weight is built
only when the compiled target (or ATLAS_FFN_GATEUP_FUSED) arms the arm
that reads it (#927).
Implementations§
Source§impl Fp8RouteInputs
impl Fp8RouteInputs
Sourcepub fn from_env(config: &ModelConfig, w8a8_prefill_kernels: bool) -> Self
pub fn from_env(config: &ModelConfig, w8a8_prefill_kernels: bool) -> Self
Resolve every gate from the process environment and the config, through the SAME predicates the loader uses.
w8a8_prefill_kernels is the caller’s, because it is a property of
the loaded kernel set rather than of the environment — ask
crate::layers::qwen3_attention::w8a8_prefill_kernels_loaded.
Trait Implementations§
Source§impl Clone for Fp8RouteInputs
impl Clone for Fp8RouteInputs
Source§fn clone(&self) -> Fp8RouteInputs
fn clone(&self) -> Fp8RouteInputs
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more