Module seq_state_reserve 

Per-sequence device state the serve must RESERVE, computed before the model exists.

Why this is config-keyed and not a layer-trait sum

preflight_reserve(args, config, free_mem) runs long before build_model, and the model builder consumes the reserve it returns. There is no layer list, no proposer, and no GpuBackend at that point, so a per-sequence term cannot be Σ_layers layer.some_method(). Every sibling reserve term (ssm_reserve::*) is a free function over config-derived scalars for exactly this reason; this module follows that precedent.

What is charged, and what deliberately is not

Charged — device memory a sequence OWNS, that no other reserve term covers:

• the target stack’s DSA indexer caches (Glm5NextDsaState), one per text DSA layer;
• the draft proposer’s own per-sequence state (Glm5NextMtpHead::alloc_state).

The proposer is a SEPARATE OWNER — it is not a TransformerLayer and never appears in the layer list, so a layer-side sum cannot see it. It is returned separately and must not be folded into the target-layer term.

NOT charged, each because something else already accounts for it:

• KDA recurrent + conv state — pool-owned, covered by ssm_reserve::ssm_pool_reserve_bytes (and meta.rs never calls alloc_state for a pool-backed mixer);
• the paged KV pool — it is the KV budget this reserve is subtracted from; charging it here would be circular;
• the buffer arena — sized by max_batch_tokens, not per sequence (buffer_arena_bytes);
• SSM snapshot / replay-ring / h-stage — already reserve terms, and the snapshot term already scales by max_batch_size;
• CUDA-graph exec memory — per SLOT, not per sequence, and LRU-bounded;
• pad-row dummy states — transient per step, and GLM declines every batched path.

🔴 The indexer cache is REPLICATED across ranks. EP does not halve it, so this returns per-rank bytes directly and callers must not divide again.

Structs

PerSequenceState

Per-sequence, per-rank device state, split by OWNER.

Functions

per_sequence_state_bytes

Per-sequence owned device state for config, at a context of max_seq_len tokens.