@melaya/runner 1.0.47 → 1.0.48
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/modelLoader.js +12 -2
- package/package.json +1 -1
package/dist/modelLoader.js
CHANGED
|
@@ -68,7 +68,17 @@ const FAMILY_OVERRIDES = [
|
|
|
68
68
|
{ pattern: /qwen3.*vl.*\b8b\b/i, family: "qwen3", tier: "agentic-capable", thinking: "on", reason: "qwen3-vl-8b: agentic-capable, full thinking" },
|
|
69
69
|
// ── Qwen3 — capable (≥8B) BEFORE the smaller non-vl 4b pattern, so
|
|
70
70
|
// 14b / 32b don't get caught by the 4b regex's stem match. ──
|
|
71
|
-
|
|
71
|
+
// Mainline qwen3 releases ship 8B, 14B, 32B. Custom/finetuned
|
|
72
|
+
// 27B/30B/72B variants exist (e.g. qwen3.6-27b on LM Studio) — we
|
|
73
|
+
// enumerate them so `family="qwen3"` propagates and model.py can
|
|
74
|
+
// apply chat_template_kwargs.enable_thinking=false when needed.
|
|
75
|
+
// Without the explicit 27b entry, qwen3.6-27b falls through to
|
|
76
|
+
// the param-count path with `family=null`, the disable_thinking
|
|
77
|
+
// branch in model.py is skipped, and the model produces
|
|
78
|
+
// thinking-only responses that the OpenAI formatter drops (run
|
|
79
|
+
// 1ecad7cb hit this — Copywriter ended in 11.5KB thinking block,
|
|
80
|
+
// ReportWriter emailed [CONTENT NOT AVAILABLE]).
|
|
81
|
+
{ pattern: /qwen3.*\b(8b|14b|27b|30b|32b|72b)\b/i, family: "qwen3", tier: "agentic-capable", thinking: "off", reason: "qwen3 ≥8b: agentic-capable, thinking disabled (model is large enough to reason without explicit <think> blocks; thinking-only responses get stripped by the OpenAI formatter and lose data to next agent)" },
|
|
72
82
|
// ── Qwen3 — small variants ──
|
|
73
83
|
{ pattern: /qwen3.*\b1\.7b\b/i, family: "qwen3", tier: "text-only", thinking: "off", reason: "qwen3-1.7b: param count below the agentic 7B floor; thinking disabled to prevent <think> loops" },
|
|
74
84
|
{ pattern: /qwen3.*\b0\.5b\b/i, family: "qwen3", tier: "text-only", thinking: "off", reason: "qwen3-0.5b: too small for agentic work" },
|
|
@@ -214,7 +224,7 @@ export function classifyModel(id, lmstudioMeta) {
|
|
|
214
224
|
// Bump whenever FAMILY_OVERRIDES, tier thresholds, or any classifier
|
|
215
225
|
// logic changes in a way that could re-tier an existing cached model.
|
|
216
226
|
// Cached profiles with a mismatched version are ignored and re-classified.
|
|
217
|
-
const PROFILE_SCHEMA_VERSION =
|
|
227
|
+
const PROFILE_SCHEMA_VERSION = 3;
|
|
218
228
|
const LMSTUDIO_BASE = process.env.LMSTUDIO_BASE_URL || "http://127.0.0.1:1234";
|
|
219
229
|
const OLLAMA_BASE = process.env.OLLAMA_BASE_URL || "http://127.0.0.1:11434";
|
|
220
230
|
// JIT-load timeout. 8B-class quantised models on M-series Macs typically
|