Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

layer-major prefill: no MMQ row table without an MMQ layer; QSA shapes from core::qsa_shapes (#3976)

Norihiro Nishimuracc125e06a049
diff --git a/src/prefill/prefill.cpp b/src/prefill/prefill.cpp
index a6ace0b56..6074ca1a6 100644
--- a/src/prefill/prefill.cpp
+++ b/src/prefill/prefill.cpp
@@ -5764,15 +5764,14 @@ bool Prefill::run_layer_major(const int64_t* tokens, int64_t n, int64_t pos0, st
const cudaStream_t cs = (cudaStream_t) m.cs;
Impl::LmBufs& lb = m.lm;
const int64_t S = lb.S, L = g.n_layers;
- strata::kernels::QsaShapes s = strata::kernels::qsa_real_shapes();
- s.n_head = g.n_head; s.n_head_kv = g.n_head_kv; s.head_dim = g.head_dim; s.idx_n_head = g.idx_q_heads;
- s.idx_dim = g.idx_key_dim;
+ strata::kernels::QsaShapes s = core::qsa_shapes(g); // as 0.2.2's run_impl
const uint64_t gdn_floats = (uint64_t) g.ssm_state_size * g.ssm_v_heads * g.ssm_state_size +
(uint64_t) g.ssm_conv_channels * (g.ssm_d_conv - 1);
int32_t prev[2] = {ss.ple_prev[0], ss.ple_prev[1]};
PfTimer pt;
// the MMQ row table lives in the borrowed cache slots when the prompt path borrows: written for every prompt
- mmq::iota(m.ids_identity, m.T * K, m.cs);
+ // (none without an MMQ layer: the FP16 path has no such table)
+ if (m.ids_identity) mmq::iota(m.ids_identity, m.T * K, m.cs);
// ---- the PLE rows, read per chunk on a thread as run_impl does (the same state machine, in chunk order)
const bool ple_on = ss.ple.ready();
static const bool ple_batch_env = [] {