metrillm 0.2.3 → 0.2.5

package/dist/index.mjs

@@ -5500,9 +5500,17 @@ function hasSamplingOverrides2(options) {
 function isUnsupportedSamplingMessage(status, text) {
   if (status !== 400 && status !== 422) return false;
   const lower = text.toLowerCase();
-  const mentionsSampling = /\b(seed|top_p|topp)\b/.test(lower);
-  if (!mentionsSampling) return false;
-  return /unrecognized|unknown|not support|unsupported|invalid|unexpected|additional|extra/.test(lower);
+  if (!/\b(seed|top_p|topp)\b/.test(lower)) return false;
+  return UNSUPPORTED_SAMPLING_FIELD_PATTERN.test(lower);
+}
+function isUnsupportedOutputLimitMessage(status, text, mode) {
+  if (status !== 400 && status !== 422) return false;
+  const lower = text.toLowerCase();
+  const fieldName = mode === "legacy" ? "max_tokens" : "max_output_tokens";
+  const alternateFieldName = mode === "legacy" ? "max_output_tokens" : "max_tokens";
+  const mentionsUnsupportedCurrentField = lower.includes(fieldName) && UNSUPPORTED_OUTPUT_LIMIT_FIELD_PATTERN.test(lower);
+  const mentionsRequiredAlternateField = lower.includes(alternateFieldName) && /\b(required|missing)\b/.test(lower);
+  return mentionsUnsupportedCurrentField || mentionsRequiredAlternateField;
 }
 function extractLMStudioErrorMessage(body) {
   const trimmed = body.trim();
@@ -5536,20 +5544,65 @@ function buildLMStudioRequestError(kind, model, status, statusText, body) {
   const suffix = backendMessage ? ` ${backendMessage}` : "";
   return new Error(`LM Studio ${kind} failed (${status} ${statusText})${suffix}`.trim());
 }
-function buildNativeChatBody(model, prompt, options, stream, includeSampling) {
+function buildUnsupportedOutputLimitNegotiationError(kind, model, body) {
+  const backendMessage = extractLMStudioErrorMessage(body);
+  return new Error(
+    [
+      `LM Studio ${kind} failed for "${model}" because this backend rejected both max_output_tokens and max_tokens.`,
+      "MetriLLM cannot safely continue without an explicit output limit because benchmarks rely on bounded generation.",
+      backendMessage ? `Backend error: ${backendMessage}` : void 0
+    ].filter(Boolean).join(" ")
+  );
+}
+function buildNativeChatBody(model, prompt, options, stream, includeSampling, outputLimitMode) {
   const reasoning = buildNativeThinkingOption(options?.think);
+  const outputLimit = options?.num_predict !== void 0 ? options.num_predict : 512;
   return {
     model,
     input: prompt,
     temperature: options?.temperature ?? 0,
     ...includeSampling && options?.top_p !== void 0 ? { top_p: options.top_p } : {},
     ...includeSampling && options?.seed !== void 0 ? { seed: options.seed } : {},
-    max_tokens: options?.num_predict ?? 512,
+    ...outputLimitMode === "preferred" ? { max_output_tokens: outputLimit } : {},
+    ...outputLimitMode === "legacy" ? { max_tokens: outputLimit } : {},
     stream,
     ...reasoning !== void 0 ? { reasoning } : {},
     ...options?.think === false ? { system_prompt: NON_THINKING_SYSTEM_PROMPT } : {}
   };
 }
+async function negotiateRequest(kind, model, cacheKey, options, makeRequest) {
+  let includeSampling = true;
+  let outputLimitMode = outputLimitModeCache.get(cacheKey) ?? "preferred";
+  const triedOutputLimitModes = /* @__PURE__ */ new Set([outputLimitMode]);
+  let resp = await makeRequest(includeSampling, outputLimitMode);
+  let retries = 0;
+  while (!resp.ok && retries < MAX_NEGOTIATE_RETRIES) {
+    retries++;
+    const body = await resp.text().catch(() => "");
+    if (includeSampling && hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
+      includeSampling = false;
+      resp = await makeRequest(includeSampling, outputLimitMode);
+      continue;
+    }
+    if (isUnsupportedOutputLimitMessage(resp.status, body, outputLimitMode)) {
+      const nextMode = outputLimitMode === "preferred" ? "legacy" : !triedOutputLimitModes.has("preferred") ? "preferred" : null;
+      if (!nextMode) {
+        throw buildUnsupportedOutputLimitNegotiationError(kind, model, body);
+      }
+      outputLimitMode = nextMode;
+      triedOutputLimitModes.add(outputLimitMode);
+      resp = await makeRequest(includeSampling, outputLimitMode);
+      continue;
+    }
+    throw buildLMStudioRequestError(kind, model, resp.status, resp.statusText, body);
+  }
+  if (!resp.ok) {
+    const body = await resp.text().catch(() => "");
+    throw buildLMStudioRequestError(kind, model, resp.status, resp.statusText, body);
+  }
+  outputLimitModeCache.set(cacheKey, outputLimitMode);
+  return resp;
+}
 function getNativeStatNumber(value) {
   if (typeof value !== "number" || !Number.isFinite(value) || value < 0) return void 0;
   return value;
@@ -6210,14 +6263,17 @@ async function listModels2() {
     throw new Error(`LM Studio list models failed (${resp.status} ${resp.statusText})`);
   }
   const data = await resp.json();
-  const ids = (data.data ?? []).map((m) => m.id?.trim()).filter((id) => Boolean(id));
+  const primaryIds = (data.data ?? []).map((m) => m.id?.trim()).filter((id) => Boolean(id));
   const apiModels = await fetchApiModels();
   const apiById = /* @__PURE__ */ new Map();
+  const secondaryIds = [];
   for (const model of apiModels ?? []) {
     const id = asNonEmptyString(model.id);
     if (!id) continue;
     apiById.set(id, model);
+    secondaryIds.push(id);
   }
+  const ids = Array.from(/* @__PURE__ */ new Set([...primaryIds, ...secondaryIds]));
   const modelsRootDir = await resolveModelsRootDir();
   const localMetadataEntries = await Promise.all(
     ids.map(async (id) => {
@@ -6255,25 +6311,18 @@ async function generate2(model, prompt, options) {
   try {
     const baseUrl = getLMStudioBaseUrl();
     const url = new URL("/api/v1/chat", baseUrl);
-    const doRequest = (includeSampling) => fetch(url, {
-      method: "POST",
-      headers: getLMStudioHeaders(),
-      body: JSON.stringify(buildNativeChatBody(model, prompt, options, false, includeSampling)),
-      signal: controller.signal
-    });
-    let resp = await doRequest(true);
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      if (hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
-        resp = await doRequest(false);
-      } else {
-        throw buildLMStudioRequestError("generate", model, resp.status, resp.statusText, body);
-      }
-    }
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      throw buildLMStudioRequestError("generate", model, resp.status, resp.statusText, body);
-    }
+    const resp = await negotiateRequest(
+      "generate",
+      model,
+      baseUrl,
+      options,
+      (sampling, limitMode) => fetch(url, {
+        method: "POST",
+        headers: getLMStudioHeaders(),
+        body: JSON.stringify(buildNativeChatBody(model, prompt, options, false, sampling, limitMode)),
+        signal: controller.signal
+      })
+    );
     const payload = await resp.json();
     const nativeResponse = extractNativeResponse(payload);
     const response = nativeResponse.response;
@@ -6331,25 +6380,18 @@ async function generateStream2(model, prompt, callbacks, options) {
   };
   try {
     resetStallTimer();
-    const doRequest = (includeSampling) => fetch(url, {
-      method: "POST",
-      headers: getLMStudioHeaders(),
-      body: JSON.stringify(buildNativeChatBody(model, prompt, options, true, includeSampling)),
-      signal: controller.signal
-    });
-    let resp = await doRequest(true);
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      if (hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
-        resp = await doRequest(false);
-      } else {
-        throw buildLMStudioRequestError("stream", model, resp.status, resp.statusText, body);
-      }
-    }
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      throw buildLMStudioRequestError("stream", model, resp.status, resp.statusText, body);
-    }
+    const resp = await negotiateRequest(
+      "stream",
+      model,
+      baseUrl,
+      options,
+      (sampling, limitMode) => fetch(url, {
+        method: "POST",
+        headers: getLMStudioHeaders(),
+        body: JSON.stringify(buildNativeChatBody(model, prompt, options, true, sampling, limitMode)),
+        signal: controller.signal
+      })
+    );
     if (!resp.body) {
       throw new Error("LM Studio stream response body is empty");
     }
@@ -6478,7 +6520,7 @@ function abortOngoingRequests2() {
   }
   activeAbortControllers.clear();
 }
-var DEFAULT_LM_STUDIO_BASE_URL, LM_STUDIO_INIT_TIMEOUT_MS, LM_STUDIO_METADATA_TIMEOUT_MS, DEFAULT_STREAM_STALL_TIMEOUT_MS2, LM_STUDIO_CLI_TIMEOUT_MS, SHARED_STREAM_STALL_TIMEOUT_ENV2, DEFAULT_LM_STUDIO_HOME_DIR, DEFAULT_LM_STUDIO_MODELS_DIR, LM_STUDIO_HOME_DIR_ENV, LM_STUDIO_MODELS_DIR_ENV, LM_STUDIO_CLI_PATH_ENV, activeAbortControllers, directorySizeCache, modelDefinitionCache, NON_THINKING_SYSTEM_PROMPT;
+var DEFAULT_LM_STUDIO_BASE_URL, LM_STUDIO_INIT_TIMEOUT_MS, LM_STUDIO_METADATA_TIMEOUT_MS, DEFAULT_STREAM_STALL_TIMEOUT_MS2, LM_STUDIO_CLI_TIMEOUT_MS, SHARED_STREAM_STALL_TIMEOUT_ENV2, DEFAULT_LM_STUDIO_HOME_DIR, DEFAULT_LM_STUDIO_MODELS_DIR, LM_STUDIO_HOME_DIR_ENV, LM_STUDIO_MODELS_DIR_ENV, LM_STUDIO_CLI_PATH_ENV, activeAbortControllers, directorySizeCache, modelDefinitionCache, outputLimitModeCache, NON_THINKING_SYSTEM_PROMPT, UNSUPPORTED_SAMPLING_FIELD_PATTERN, UNSUPPORTED_OUTPUT_LIMIT_FIELD_PATTERN, MAX_NEGOTIATE_RETRIES;
 var init_lm_studio_client = __esm({
   "src/core/lm-studio-client.ts"() {
     init_utils();
@@ -6496,12 +6538,16 @@ var init_lm_studio_client = __esm({
     activeAbortControllers = /* @__PURE__ */ new Set();
     directorySizeCache = /* @__PURE__ */ new Map();
     modelDefinitionCache = /* @__PURE__ */ new Map();
+    outputLimitModeCache = /* @__PURE__ */ new Map();
     NON_THINKING_SYSTEM_PROMPT = [
       "You are in non-thinking mode for benchmark reproducibility.",
       "Return only the final answer.",
       "Do not output internal reasoning, chain-of-thought, or scratchpad.",
       "Never output tags or sections like <think>, </think>, [THINK], [/THINK], or Thinking Process."
     ].join(" ");
+    UNSUPPORTED_SAMPLING_FIELD_PATTERN = /unrecognized|unknown|not support|unsupported|invalid|unexpected|additional|extra/;
+    UNSUPPORTED_OUTPUT_LIMIT_FIELD_PATTERN = /unrecognized|unknown|not support|unsupported|unexpected|additional|extra|invalid field/;
+    MAX_NEGOTIATE_RETRIES = 5;
   }
 });
 
@@ -53935,6 +53981,7 @@ async function selectWithArrows(title, options, config = {}) {
     };
     readline4.emitKeypressEvents(stdin);
     stdin.resume();
+    stdin.ref?.();
     if (stdin.isTTY) {
       stdin.setRawMode(true);
     }
@@ -54017,6 +54064,7 @@ ${source_default.dim(message)}
     };
     readline4.emitKeypressEvents(stdin);
     stdin.resume();
+    stdin.ref?.();
     if (stdin.isTTY) stdin.setRawMode(true);
     stdout.write("\x1B[?25l");
     stdin.on("keypress", onKeypress);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "metrillm",
-  "version": "0.2.3",
+  "version": "0.2.5",
   "description": "Benchmark your local LLM models — speed, quality & hardware fitness verdict",
   "type": "module",
   "bin": {