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++
[Engine] SYCL: refuse a --mtp-q4 draft head when the Q6_K head is stored as planes
Niko122189253935db32
diff --git a/sycl/src/core/mtp.cpp b/sycl/src/core/mtp.cppindex 5fff80b0f..fd3fec628 100644--- a/sycl/src/core/mtp.cpp+++ b/sycl/src/core/mtp.cpp@@ -233,6 +233,14 @@ bool MtpDrafter::make_q4_head(std::string& err) {const int64_t N = g_->n_embd;const auto* tt = ggml_get_type_traits((ggml_type) head_->type());if (tt == nullptr || tt->to_float == nullptr || N % 32 != 0) { err = "mtp: --mtp-q4 cannot read the head's format"; return false; }+ // SYCL port: with STRATA_Q6_PLANAR the head is stored as planes, and the subset this builds from is row-major+ // only when the head is packed - the gather at load time then copies planes as if they were rows. The result is+ // a draft head that proposes nothing at all (measured: 0 drafts, decode 73.8 -> 35.3 tok/s) with no error, so+ // refuse instead of running wrong in silence.+ if (head_->type() == 14 && strata::kernels::native_q6_k_is_planar(head_->weights())) {+ err = "mtp: --mtp-q4 head needs the packed Q6_K head layout (STRATA_Q6_PLANAR=0)";+ return false;+ }const size_t in_row = head_->row_bytes(), out_row = (size_t) (N / 32) * 18;std::vector<uint8_t> src((size_t) n_dvocab_ * in_row), dst((size_t) n_dvocab_ * out_row);try { dpct::get_in_order_queue().memcpy(src.data(), dhead_, src.size()).wait(); } // SYCL port: cudaMemcpy