True if layer `il` uses full attention (the FIRST layer in each group of `swa_pattern` is full; the remaining are sliding-window).

Launch a daemon, run warmup+1 generations, return (decode_tok_s, out_ids).

MMQ full-batch mul_mat_id on a reduced hot stack is only stable for large batches. Small batches (spec verify/replay, <=~24 tokens) spread n_used*n_to

Run `nsys stats` for the first report-name alias that yields a table. Returns (rows, diag). diag is "" on success, else a one-line reason the

Pool chunk `c` whose rows sit in pool block `block` of each cache tensor in `attn_k` ([head_dim, pool_tokens, n_head_kv], quantized). Reads chunk_toke

Remove the specified number of chars from the right of the current key. */

Return the frame at the top of the defragmentation stack, or NULL if there * are no more nodes to visit. */

realloc the node to make room for auxiliary data in order * to store an item in that node. On out of memory NULL is returned. */

Seek the greatest key in the subtree at the current node. Return 0 on * out of memory, otherwise 1. This is an helper function for different * itera

Round `x` to 1 decimal place. JSON serialization of doubles can emit 17 significant digits which is noisy in client logs and bench output; caller-side

Run prefill benchmark at various prompt lengths.

Replace fcntl(F_GETFL) / fcntl(F_SETFL, O_NONBLOCK) with ioctlsocket

Strip any line whose ltrimmed text starts with kBillingHeader from a multi-line string.

Compare dense FA output vs sparse FA output At alpha=1.0 (select all blocks), sparse should match dense exactly.

Block until the ready banner. Kills + raises SystemExit on timeout/death.

Parse + range-check alpha. The daemon silently ignores values outside (0, 1) (see dflash/src/qwen3_0p6b_graph.cpp), so reject them here loudly.

Free GPU memory in MiB via hipMemGetInfo. Reliable on AMD unified-memory APUs (Strix Halo, Phoenix) where rocm-smi only reports the small ded

Normalize a math answer string for comparison.

Parse SSE events from a streaming response. stop_on_event: stop after seeing this event type (e.g. 'message_stop'). If None,

Read a binary file of packed int32 token IDs.

Relay upstream response back to client, handling SSE streaming.

Run a subprocess, return (CompletedProcess, wall_seconds).