@infersec/conduit 1.93.0 → 1.95.0

package/dist/cli.js

@@ -112706,6 +112706,8 @@ function requireExpressPromiseRouter () {
 var expressPromiseRouterExports = requireExpressPromiseRouter();
 var createRouter = /*@__PURE__*/getDefaultExportFromCjs(expressPromiseRouterExports);
 
+const SERVED_MODEL_NAME = "default";
+
 const SECRET_FLAGS = new Set([
     "api-key",
     "auth-token",
@@ -121201,7 +121203,7 @@ async function startVLLM({ enginePort, targetDirectory }) {
         "--model",
         modelPath,
         "--served-model-name",
-        this.model.id,
+        SERVED_MODEL_NAME,
         "--max-model-len",
         String(contextLength),
         "--tensor-parallel-size",
@@ -123017,7 +123019,7 @@ async function startSGLang({ enginePort, targetDirectory }) {
         "--model-path",
         targetDirectory,
         "--served-model-name",
-        this.model.id,
+        SERVED_MODEL_NAME,
         "--context-length",
         String(contextLength),
         "--tp-size",
@@ -123091,6 +123093,7 @@ class ModelManager extends EventEmitter {
     logger;
     engineProcess = null;
     healthPollInterval = null;
+    lastEngineError = null;
     lifecycleState = "stopped";
     stopRequested = false;
     modelsDirectory;
@@ -123198,6 +123201,7 @@ class ModelManager extends EventEmitter {
             });
         }
         this.lifecycleState = "starting";
+        this.lastEngineError = null;
         this.stopRequested = false;
         this.logger.info("Starting LLM engine", {
             agentEngineType: this.engine
@@ -123211,12 +123215,18 @@ class ModelManager extends EventEmitter {
             await this.waitForEngineReady();
         }
         catch (error) {
-            const err = error instanceof Error ? error : new Error(String(error));
+            let err = error instanceof Error ? error : new Error(String(error));
             if (this.stopRequested) {
                 throw err;
             }
+            if (this.lastEngineError) {
+                err = this.lastEngineError;
+            }
+            const alreadyEmitted = this.lastEngineError !== null;
             this.lifecycleState = "errored";
-            this.emit("engineError", err);
+            if (!alreadyEmitted) {
+                this.emit("engineError", err);
+            }
             if (this.engineProcess) {
                 this.startHealthPoll();
             }
@@ -123351,7 +123361,10 @@ class ModelManager extends EventEmitter {
             }
             await new Promise(resolve => setTimeout(resolve, pollIntervalMs));
         }
-        throw new Error("LLM engine failed readiness checks within timeout");
+        const stderrTail = this.engineProcess?.stderr?.slice(-1e3);
+        throw new Error(stderrTail
+            ? `LLM engine failed readiness checks within timeout. Last engine output:\n${stderrTail}`
+            : "LLM engine failed readiness checks within timeout");
     }
     clearHealthPoll() {
         if (this.healthPollInterval) {
@@ -123414,6 +123427,11 @@ class ModelManager extends EventEmitter {
             return null;
         }
     }
+    recordEngineError(err) {
+        this.lifecycleState = "errored";
+        this.lastEngineError = err;
+        this.emit("engineError", err);
+    }
     async releaseDownloadLock() {
         try {
             await unlink(this.lockFilePath);
@@ -123459,8 +123477,7 @@ class ModelManager extends EventEmitter {
                 return;
             }
             this.engineProcess = null;
-            this.lifecycleState = "errored";
-            this.emit("engineError", new ProcessExecutionError({
+            this.recordEngineError(new ProcessExecutionError({
                 code: null,
                 error: err,
                 message: `Process error: ${this.engine}: ${processManager.stderr}`,
@@ -123485,8 +123502,7 @@ class ModelManager extends EventEmitter {
                 this.emit("engineTerminated");
                 return;
             }
-            this.lifecycleState = "errored";
-            this.emit("engineError", new ProcessExecutionError({
+            this.recordEngineError(new ProcessExecutionError({
                 code,
                 message: `Process stopped: ${this.engine}: ${processManager.stderr}`,
                 signal
@@ -123536,6 +123552,47 @@ class ModelManager extends EventEmitter {
     }
 }
 
+const EXCEPTION_LINE_PATTERN = /([A-Za-z_][A-Za-z0-9_]*(?:Error|Exception)):\s*(.+)/;
+const FALLBACK_DETAIL_MAX_LENGTH = 300;
+const FALLBACK_RAW_MAX_LENGTH = 500;
+const PATTERNS = [
+    {
+        message: () => "GPU memory insufficient to reserve the KV cache. Raise --gpu-memory-utilization, or reduce --max-model-len / context length.",
+        pattern: /No available memory for the cache blocks/
+    },
+    {
+        message: match => `Context length requires more KV cache memory than is available. Lower the source's context length to at most ${match[1]} tokens (the engine's computed ceiling), or raise --gpu-memory-utilization.`,
+        pattern: /larger than the available KV cache memory[\s\S]*?estimated maximum model length is (\d+)/
+    },
+    {
+        message: () => "The GPU ran out of memory (CUDA OOM). Reduce --max-model-len / --max-num-seqs, lower --gpu-memory-utilization, or use a smaller or more aggressively quantized model.",
+        pattern: /CUDA out of memory/i
+    }
+];
+function normalizeEngineError(input) {
+    const source = typeof input === "string" ? input : String(input ?? "");
+    for (const { message, pattern } of PATTERNS) {
+        const match = source.match(pattern);
+        if (match) {
+            return message(match);
+        }
+    }
+    const lines = source.split(/\r?\n/);
+    for (let index = lines.length - 1; index >= 0; index -= 1) {
+        const exceptionMatch = lines[index].match(EXCEPTION_LINE_PATTERN);
+        if (exceptionMatch) {
+            return `Engine failed to start: ${truncate(exceptionMatch[0], FALLBACK_DETAIL_MAX_LENGTH)}`;
+        }
+    }
+    return `Engine failed to start: ${truncate(source, FALLBACK_RAW_MAX_LENGTH)}`;
+}
+function truncate(value, max) {
+    if (value.length <= max) {
+        return value;
+    }
+    return `${value.slice(0, max)}…`;
+}
+
 function createConduitGeneralAPIReferenceHandlers({ cycleEngine, conduitStateManager, getModelManager, logger, setErrorState, startEngine, stopEngine, stopRequestedByControl }) {
     return {
         "/conduit/engine/start": {
@@ -124563,7 +124620,7 @@ function createConduitOpenAIAPIReferenceHandlers({ apiClient, conduitConfigurati
                         data: [
                             {
                                 created: startup / 1000,
-                                id: "default",
+                                id: SERVED_MODEL_NAME,
                                 limit: {
                                     context: effectiveContextLength
                                 },
@@ -135738,7 +135795,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
                 error: err
             });
             stopRequestedByControl = false;
-            setErrorState({ error: err.message });
+            setErrorState({ error: normalizeEngineError(err.message) });
         });
         modelManager.on("engineReady", () => {
             setOnlineState();
@@ -135855,7 +135912,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
             logger.error("Failed starting LLM engine", {
                 error: parsedError
             });
-            setErrorState({ error: parsedError.message });
+            setErrorState({ error: normalizeEngineError(parsedError.message) });
         });
     }
     const app = express();

package/dist/cli.sea.cjs

@@ -112721,6 +112721,8 @@ function requireExpressPromiseRouter () {
 var expressPromiseRouterExports = requireExpressPromiseRouter();
 var createRouter = /*@__PURE__*/getDefaultExportFromCjs(expressPromiseRouterExports);
 
+const SERVED_MODEL_NAME = "default";
+
 const SECRET_FLAGS = new Set([
     "api-key",
     "auth-token",
@@ -121216,7 +121218,7 @@ async function startVLLM({ enginePort, targetDirectory }) {
         "--model",
         modelPath,
         "--served-model-name",
-        this.model.id,
+        SERVED_MODEL_NAME,
         "--max-model-len",
         String(contextLength),
         "--tensor-parallel-size",
@@ -123032,7 +123034,7 @@ async function startSGLang({ enginePort, targetDirectory }) {
         "--model-path",
         targetDirectory,
         "--served-model-name",
-        this.model.id,
+        SERVED_MODEL_NAME,
         "--context-length",
         String(contextLength),
         "--tp-size",
@@ -123106,6 +123108,7 @@ class ModelManager extends EventEmitter {
     logger;
     engineProcess = null;
     healthPollInterval = null;
+    lastEngineError = null;
     lifecycleState = "stopped";
     stopRequested = false;
     modelsDirectory;
@@ -123213,6 +123216,7 @@ class ModelManager extends EventEmitter {
             });
         }
         this.lifecycleState = "starting";
+        this.lastEngineError = null;
         this.stopRequested = false;
         this.logger.info("Starting LLM engine", {
             agentEngineType: this.engine
@@ -123226,12 +123230,18 @@ class ModelManager extends EventEmitter {
             await this.waitForEngineReady();
         }
         catch (error) {
-            const err = error instanceof Error ? error : new Error(String(error));
+            let err = error instanceof Error ? error : new Error(String(error));
             if (this.stopRequested) {
                 throw err;
             }
+            if (this.lastEngineError) {
+                err = this.lastEngineError;
+            }
+            const alreadyEmitted = this.lastEngineError !== null;
             this.lifecycleState = "errored";
-            this.emit("engineError", err);
+            if (!alreadyEmitted) {
+                this.emit("engineError", err);
+            }
             if (this.engineProcess) {
                 this.startHealthPoll();
             }
@@ -123366,7 +123376,10 @@ class ModelManager extends EventEmitter {
             }
             await new Promise(resolve => setTimeout(resolve, pollIntervalMs));
         }
-        throw new Error("LLM engine failed readiness checks within timeout");
+        const stderrTail = this.engineProcess?.stderr?.slice(-1e3);
+        throw new Error(stderrTail
+            ? `LLM engine failed readiness checks within timeout. Last engine output:\n${stderrTail}`
+            : "LLM engine failed readiness checks within timeout");
     }
     clearHealthPoll() {
         if (this.healthPollInterval) {
@@ -123429,6 +123442,11 @@ class ModelManager extends EventEmitter {
             return null;
         }
     }
+    recordEngineError(err) {
+        this.lifecycleState = "errored";
+        this.lastEngineError = err;
+        this.emit("engineError", err);
+    }
     async releaseDownloadLock() {
         try {
             await require$$0$d.unlink(this.lockFilePath);
@@ -123474,8 +123492,7 @@ class ModelManager extends EventEmitter {
                 return;
             }
             this.engineProcess = null;
-            this.lifecycleState = "errored";
-            this.emit("engineError", new ProcessExecutionError({
+            this.recordEngineError(new ProcessExecutionError({
                 code: null,
                 error: err,
                 message: `Process error: ${this.engine}: ${processManager.stderr}`,
@@ -123500,8 +123517,7 @@ class ModelManager extends EventEmitter {
                 this.emit("engineTerminated");
                 return;
             }
-            this.lifecycleState = "errored";
-            this.emit("engineError", new ProcessExecutionError({
+            this.recordEngineError(new ProcessExecutionError({
                 code,
                 message: `Process stopped: ${this.engine}: ${processManager.stderr}`,
                 signal
@@ -123551,6 +123567,47 @@ class ModelManager extends EventEmitter {
     }
 }
 
+const EXCEPTION_LINE_PATTERN = /([A-Za-z_][A-Za-z0-9_]*(?:Error|Exception)):\s*(.+)/;
+const FALLBACK_DETAIL_MAX_LENGTH = 300;
+const FALLBACK_RAW_MAX_LENGTH = 500;
+const PATTERNS = [
+    {
+        message: () => "GPU memory insufficient to reserve the KV cache. Raise --gpu-memory-utilization, or reduce --max-model-len / context length.",
+        pattern: /No available memory for the cache blocks/
+    },
+    {
+        message: match => `Context length requires more KV cache memory than is available. Lower the source's context length to at most ${match[1]} tokens (the engine's computed ceiling), or raise --gpu-memory-utilization.`,
+        pattern: /larger than the available KV cache memory[\s\S]*?estimated maximum model length is (\d+)/
+    },
+    {
+        message: () => "The GPU ran out of memory (CUDA OOM). Reduce --max-model-len / --max-num-seqs, lower --gpu-memory-utilization, or use a smaller or more aggressively quantized model.",
+        pattern: /CUDA out of memory/i
+    }
+];
+function normalizeEngineError(input) {
+    const source = typeof input === "string" ? input : String(input ?? "");
+    for (const { message, pattern } of PATTERNS) {
+        const match = source.match(pattern);
+        if (match) {
+            return message(match);
+        }
+    }
+    const lines = source.split(/\r?\n/);
+    for (let index = lines.length - 1; index >= 0; index -= 1) {
+        const exceptionMatch = lines[index].match(EXCEPTION_LINE_PATTERN);
+        if (exceptionMatch) {
+            return `Engine failed to start: ${truncate(exceptionMatch[0], FALLBACK_DETAIL_MAX_LENGTH)}`;
+        }
+    }
+    return `Engine failed to start: ${truncate(source, FALLBACK_RAW_MAX_LENGTH)}`;
+}
+function truncate(value, max) {
+    if (value.length <= max) {
+        return value;
+    }
+    return `${value.slice(0, max)}…`;
+}
+
 function createConduitGeneralAPIReferenceHandlers({ cycleEngine, conduitStateManager, getModelManager, logger, setErrorState, startEngine, stopEngine, stopRequestedByControl }) {
     return {
         "/conduit/engine/start": {
@@ -124578,7 +124635,7 @@ function createConduitOpenAIAPIReferenceHandlers({ apiClient, conduitConfigurati
                         data: [
                             {
                                 created: startup / 1000,
-                                id: "default",
+                                id: SERVED_MODEL_NAME,
                                 limit: {
                                     context: effectiveContextLength
                                 },
@@ -155968,7 +156025,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
                 error: err
             });
             stopRequestedByControl = false;
-            setErrorState({ error: err.message });
+            setErrorState({ error: normalizeEngineError(err.message) });
         });
         modelManager.on("engineReady", () => {
             setOnlineState();
@@ -156085,7 +156142,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
             logger.error("Failed starting LLM engine", {
                 error: parsedError
             });
-            setErrorState({ error: parsedError.message });
+            setErrorState({ error: normalizeEngineError(parsedError.message) });
         });
     }
     const app = express();

package/dist/constants.d.ts

@@ -0,0 +1 @@
+export declare const SERVED_MODEL_NAME = "default";

package/dist/modelManagement/ModelManager.d.ts

@@ -20,6 +20,7 @@ export declare class ModelManager extends EventEmitter<ModelManagerEvents> {
     protected readonly logger: Logger;
     private engineProcess;
     private healthPollInterval;
+    private lastEngineError;
     private lifecycleState;
     private stopRequested;
     protected readonly modelsDirectory: string;
@@ -51,6 +52,7 @@ export declare class ModelManager extends EventEmitter<ModelManagerEvents> {
     private get lockFilePath();
     private acquireDownloadLock;
     private getLockAge;
+    private recordEngineError;
     private releaseDownloadLock;
     private waitForDownloadLock;
     private bindEngineProcessEvents;

package/dist/modelManagement/normalizeEngineError.d.ts

@@ -0,0 +1 @@
+export declare function normalizeEngineError(input: string): string;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@infersec/conduit",
   "description": "End user conduit agent for connecting local LLMs to the cloud.",
-  "version": "1.93.0",
+  "version": "1.95.0",
   "bin": {
     "infersec-conduit": "./dist/cli.js"
   },