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++
setup: a 16,384-token prompt chunk on a unified-memory PC whose pool holds ~128 GB (16K prompt 152.2 -> 158.6 tok/s, 32K 174.0 -> 179.7, 8K flat, decode unchanged); smaller unified PCs keep 8,192
Niko122181ac4dd83be8
diff --git a/setup.py b/setup.pyindex 29cb362d9..4c88035b5 100644--- a/setup.py+++ b/setup.py@@ -8025,7 +8025,12 @@ def dsv41_prefill_chunk_args(cards, ram_gb) -> list:4K prompt 78.7 -> 87.2 tok/s, 8K 108.9 -> 125.1, 32K 155.2 -> 176.9; decode unchanged (16.53 -> 16.59 tok/s), because thechunk buffers take ~5 GB from the GPU expert tier and the RAM store's hit rate rises to compensate. The prompt chunk'ssummation order changes, not the model: the repo's chunk-size KL check (1024 0.1705 / 2048 0.1692 / 4096 0.1688 / 8192 0.1695)- shows no quality move. Only where the shared pool has room; STRATA_DSV41_NO_BIG_CHUNK=1 leaves the engine's own choice."""+ shows no quality move. Only where the shared pool has room; STRATA_DSV41_NO_BIG_CHUNK=1 leaves the engine's own choice.++ A 16,384-token chunk on the 128 GB box (2026-10-26, engine 0.1.51, the same set, 1-2 runs a point): 8K 119.8 -> 119.4 (nothing,+ the set is read once a chunk either way), 16K 152.2 -> 158.6 (+4.2%), 32K 174.0 -> 179.7 (+3.3%), decode 17.21 -> 17.27+ (2+2 interleaved, unchanged) although the bigger chunk moves ~30% more experts to the CPU lane. So a pool that holds this+ box's 128 GB takes 16,384; a smaller unified PC keeps 8,192 (not measured below this box)."""if not cards or not any(c.get("uma") for c in cards):return []pool = float(ram_gb or 0) + max([float(c.get('dedicated_gb') or 0.0) for c in cards] or [0.0])@@ -8033,7 +8038,7 @@ def dsv41_prefill_chunk_args(cards, ram_gb) -> list:return []if os.environ.get("STRATA_DSV41_NO_BIG_CHUNK") == "1":return []- return ["--prefill-chunk", "8192"]+ return ["--prefill-chunk", "16384" if pool >= 96.0 else "8192"]def dsv41_config_env(cards) -> dict:diff --git a/tests/terminal/test_dsv41_prefill_chunk.py b/tests/terminal/test_dsv41_prefill_chunk.pyindex 3f42760b3..7159cdea9 100644--- a/tests/terminal/test_dsv41_prefill_chunk.py+++ b/tests/terminal/test_dsv41_prefill_chunk.py@@ -19,8 +19,12 @@ R9700 = [{"name": "Radeon AI PRO R9700", "vram_gb": 32.0}]class StrixChunkArgs(unittest.TestCase):- def test_a_128_gb_unified_pc_gets_the_big_chunk(self):- self.assertEqual(setup.dsv41_prefill_chunk_args(HALO, 63.6), ["--prefill-chunk", "8192"])+ def test_a_128_gb_unified_pc_gets_the_16k_chunk(self):+ # (2026-10-26, the same box: 16K prompt 152.2 -> 158.6 tok/s, 32K 174.0 -> 179.7, 8K flat, decode unchanged)+ self.assertEqual(setup.dsv41_prefill_chunk_args(HALO, 63.6), ["--prefill-chunk", "16384"])++ def test_a_96_gb_pool_is_the_line(self):+ self.assertEqual(setup.dsv41_prefill_chunk_args(HALO, 31.0), ["--prefill-chunk", "8192"])def test_a_64_gb_unified_pc_keeps_the_engines_choice(self):self.assertEqual(setup.dsv41_prefill_chunk_args(SMALL_HALO, 31.6), [])