Stateful decoder for Qwen3.5-0.8B bf16 megakernel.

Stateful decoder for Qwen3.5-0.8B megakernel backends.

Manages SSE streaming for a single request.

Wraps pflash_daemon stdin/stdout protocol with a streamfd pipe.

─── Backend creation arguments ───────────────────────────────────────── A superset of all per-arch config fields. The factory reads only those releva

Pooled chunked prefill: 64-token (one pager chunk) batched forwards with slot-mapped set_rows writes and a resident+causal mask. This is the prompt >

Thinking-budget force-close hook. Mirrors antirez/ds4 ds4_eval.c's hard_limit_reply_budget semantics: when the budget remaining (n_gen minus tokens co

Cached FFN graph for a fixed number of selected experts. Built once, reused every token to avoid per-call graph rebuild overhead.

Cached batched FFN graph for hot-only prefill (n_tokens = MMQ_SAFE_SUB_BATCH). Eliminates per-call graph rebuild + gallocr planning overhead.

Per-layer cached pre-FFN graph for DeltaNet layers. For DeltaNet layers, the graph structure doesn't depend on kv_start (recurrent), so we build once

6.25% residency

Chat template format.

A single message in a chat conversation.

CPU-side embedder: keeps a mmap of the GGUF alive and knows how to dequantize individual rows of the quantized tok_embd tensor on demand. This matches

A flat DFS-ordered tree built from the draft's top-K softmax distributions. Slot 0 is the tree root (the bonus token from the previous spec round); sl

─── I/O handle passed to backend methods that need protocol output ─────

A stdout read exceeded its deadline (daemon alive but silent).

Per-layer compressor rolling state

Per-layer cache