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++
dsv41: picture test (mkimg.py) and serve_prompts GENI / text output
Niko1221e38025769d58
diff --git a/tools/dsv41/mkimg.py b/tools/dsv41/mkimg.pynew file mode 100644index 000000000..eba5740d3--- /dev/null+++ b/tools/dsv41/mkimg.py@@ -0,0 +1,40 @@+#!/usr/bin/env python3+"""mkimg.py OUT_DIR: a 672x672 test picture (left half red, right half blue, a black disc in the middle), its DSI1 file and+the prompt ids (chat mode, GGUF template) with the picture's span (16 x 17 + 2 = 274 positions of the image token) in them.+Writes OUT_DIR/img.dsi and OUT_DIR/img_prompts.json in serve_prompts.py's format ("pic": file)."""+import json, os, struct, subprocess, sys, tempfile+import jinja2+from jinja2.sandbox import ImmutableSandboxedEnvironment+out, template, shard, tokenize = sys.argv[1:5]+W = H = 672+rows = bytearray()+for y in range(H):+ for x in range(W):+ if (x - 336) ** 2 + (y - 336) ** 2 < 80 ** 2: rows += bytes((0, 0, 0))+ elif x < 336: rows += bytes((220, 20, 20))+ else: rows += bytes((20, 20, 220))+n_h = n_w = 672 // 42+span = n_h * (n_w + 1) + 2+os.makedirs(out, exist_ok=True)+env = ImmutableSandboxedEnvironment(trim_blocks=True, lstrip_blocks=True)+env.filters["tojson"] = lambda x, **k: json.dumps(x, ensure_ascii=False)+env.globals["raise_exception"] = lambda m: (_ for _ in ()).throw(jinja2.exceptions.TemplateError(m))+tpl = env.from_string(open(template, encoding="utf-8").read())+text = tpl.render(messages=[{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "x"}},+ {"type": "text", "text": "What colors are the left half and the right half of this picture? Answer in one short sentence."}]}],+ tools=None, thinking_mode="chat", reasoning_effort=None, add_generation_prompt=True)+PH = "<\uff5cdeepseek_image\uff5c>"+assert text.count(PH) == 1, text[:300]+a, b = text.split(PH)+def tok(t):+ with tempfile.NamedTemporaryFile("w", suffix=".txt", delete=False, encoding="utf-8") as f:+ f.write(t); p = f.name+ r = subprocess.run([tokenize, "-m", shard, "-f", p, "--ids", "--no-bos", "--no-escape"], capture_output=True, text=True)+ os.unlink(p)+ return json.loads([l for l in r.stdout.splitlines() if l.startswith("[")][-1])+ia, ib = tok(a), tok(b)+ids = ia + [129264] * span + ib+open(os.path.join(out, "img.dsi"), "wb").write(b"DSI1" + struct.pack("<i", 1) + struct.pack("<qii", len(ia), W, H) + bytes(rows))+json.dump([{"id": "image-colors", "mode": "chat", "ids": ids, "max_tokens": 40, "expect_ids": [],+ "pic": os.path.join(out, "img.dsi")}], open(os.path.join(out, "img_prompts.json"), "w"))+print("prompt", len(ids), "tokens, picture span at", len(ia), "length", span)diff --git a/tools/dsv41/serve_prompts.py b/tools/dsv41/serve_prompts.pyindex 8f0b6e93d..e9a363d97 100644--- a/tools/dsv41/serve_prompts.py+++ b/tools/dsv41/serve_prompts.py@@ -4,7 +4,25 @@ Starts the engine in --serve mode (appended), waits for READY, sends GEN for evegenerated ids with the llama.cpp reference's greedy ids (expect_ids), and prints one line each:id mode prompt_tokens n_gen match_prefix/n_expect first_divergence prefill ms decode tok/s"""-import json, subprocess, sys, time+import json, os, subprocess, sys, time++def _bytes_map():+ bs = list(range(33, 127)) + list(range(161, 173)) + list(range(174, 256))+ cs = bs[:]; n = 0+ for b in range(256):+ if b not in bs: bs.append(b); cs.append(256 + n); n += 1+ return {chr(c): b for b, c in zip(bs, cs)}+_BM = _bytes_map()+try:+ _VOC = {i: t for t, i in json.load(open(os.environ.get("DS_VOCAB", "/data/dsv41-port/pack/tokenizer/vocab.json"))).items()}+except Exception:+ _VOC = {}+def detok(ids):+ b = bytearray()+ for i in ids:+ for ch in _VOC.get(i, "?"):+ b.append(_BM.get(ch, 63))+ return b.decode("utf-8", "replace")prompts_path, out_path = sys.argv[1], sys.argv[2]cmd = sys.argv[sys.argv.index("--") + 1:] + ["--serve"]@@ -27,7 +45,10 @@ res = []for pr in prompts:ids = pr["ids"]mx = min(pr["max_tokens"], max(len(pr["expect_ids"]) + 4, 8))- p.stdin.write("GEN %d %s\n" % (mx, ",".join(map(str, ids))))+ if pr.get("pic"):+ p.stdin.write("GENI %d %s %s\n" % (pr.get("max_tokens", 40), pr["pic"], ",".join(map(str, ids))))+ else:+ p.stdin.write("GEN %d %s\n" % (mx, ",".join(map(str, ids))))p.stdin.flush()gen, done = [], Nonewhile True:@@ -40,6 +61,8 @@ for pr in prompts:done = line.split()breakexp = pr["expect_ids"]+ if pr.get("pic") or os.environ.get("SHOW_TEXT"):+ print("TEXT", repr(detok(gen)), flush=True)n = 0while n < min(len(gen), len(exp)) and gen[n] == exp[n]:n += 1