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++
// The default for --pipeline-windows when it is not given (host-only logic, no SYCL runtime: tested by
// src/core/pipeline_auto_test.cpp). On for a layer split of exactly two Arc Battlemage (Xe2) cards, with --serve and the
// MTP drafter; off for everything else. The caller finds `battlemage_*` from the devices' architecture ids
// (strata::intel_gpu_gen), never from a name.
#pragma once
namespace strata {
struct PipelineAutoInputs {
bool serve = false; ///< --serve
bool multi_gpu = false; ///< a layer split across GPUs
bool split_same = false; ///< the split puts every stage on one GPU
unsigned later_stages = 0; ///< GPUs after the first one: 1 = a split of exactly two stages
bool auto_off = false; ///< STRATA_PIPELINE_AUTO=0
bool has_mtp = false; ///< --mtp given
int spec = 0; ///< --spec after the engine's own widening (>= 2 means a draft window)
int batch = 0; ///< --batch slots
bool peer_device = false; ///< --peer-device
bool helper_caches = false; ///< --expert-cache-device1..3
bool first_is_battlemage = false; ///< stage 0's GPU is a Battlemage card (architecture id)
bool later_is_battlemage = false; ///< the second stage's GPU is a Battlemage card
};
/// True when --pipeline-windows 2 is the default for this run. Single card, an Arc A-series card (Alchemist), any other
/// vendor, three or more stages (the SYCL pipeline runs exactly two), --batch, --peer-device and the helper caches stay serial.
inline bool pipeline_auto_on(const PipelineAutoInputs& in) {
return in.serve && in.multi_gpu && !in.split_same && in.later_stages == 1 && !in.auto_off && in.has_mtp &&
in.spec >= 2 && in.batch == 0 && !in.peer_device && !in.helper_caches && in.first_is_battlemage &&
in.later_is_battlemage;
}
} // namespace strata