pub fn resolve_batchm_max(default_max: u32, raw: Option<&str>) -> Resolved<u32>Expand description
The BF16 decode head’s batched-GEMV band.
🔴 Read layers/ops/gemm_quant.rs before touching the DEFAULT. The band’s
upper edge decides whether a width lands on the batched GEMV or on a
REASSOCIATING tile GEMM, and the A/B behind GB10’s 8 measured the GEMV
NEGATIVE above it (-14.4% at C=16, commits 84d5b763c / 78d276832).
Clamped to DENSE_GEMV_BATCHM_MAX_M, the kernel’s compile-time row
bound — dense_gemv_batchm refuses above it rather than writing 16 of m
rows, and a lever that produced an Err at every decode step would be a
worse failure than ignoring the excess. An unparseable or 0 environment
value keeps the target’s declaration; the value is a BAND, not a switch, so
there is no “off”.