omnius 1.0.187 → 1.0.188

package/dist/index.js

@@ -1412,6 +1412,18 @@ var init_tool_executor = __esm({
 import { EventEmitter } from "node:events";
 import { totalmem, freemem } from "node:os";
 import { exec } from "node:child_process";
+function dedupeLoadedModels(models) {
+  const seen = /* @__PURE__ */ new Set();
+  const out = [];
+  for (const model of models) {
+    const key = `${model.host}:${model.name}`;
+    if (seen.has(key))
+      continue;
+    seen.add(key);
+    out.push(model);
+  }
+  return out;
+}
 function ramSnapshotMB() {
   const total = Math.round(totalmem() / (1024 * 1024));
   const free = Math.round(freemem() / (1024 * 1024));
@@ -1690,6 +1702,104 @@ var init_model_broker = __esm({
         this.emit("rejected", spec, reason);
         return { kind: "reject", reason };
       }
+      /**
+       * Acquire a short-lived load lease for media/subprocess generation.
+       *
+       * Media generation often needs to temporarily free VRAM occupied by Ollama
+       * chat models. This helper refreshes Ollama state, asks the broker what must
+       * be evicted, unloads those Ollama models with keep_alive=0, and returns a
+       * lease whose release() clears transient inflight state, unloads any
+       * Ollama-hosted requested model, and warms the evicted Ollama models again.
+       */
+      async acquireTransientModelLoad(spec, options2 = {}) {
+        const reason = options2.reason ?? `${spec.domain}-transient-load`;
+        const evictedModels = [];
+        let gpuIndex = null;
+        let admitted = false;
+        await this.pollOnce().catch(() => {
+        });
+        for (let attempt = 0; attempt < 4; attempt++) {
+          const decision2 = await this.ensureModelLoadable(spec);
+          if (decision2.kind === "wait-for-inflight") {
+            const waited = await decision2.promise.catch((err) => ({
+              kind: "reject",
+              reason: err instanceof Error ? err.message : String(err)
+            }));
+            if (waited.kind === "ok") {
+              gpuIndex = waited.gpuIndex ?? null;
+              admitted = true;
+              break;
+            }
+            if (waited.kind === "evict") {
+              for (const target of waited.evictTargets) {
+                if (await this.evict(target.host, target.name, reason))
+                  evictedModels.push(target);
+              }
+              await this.pollOnce().catch(() => {
+              });
+              continue;
+            }
+            if (waited.kind === "degrade")
+              return waited;
+            if (waited.kind === "reject")
+              return waited;
+            continue;
+          }
+          if (decision2.kind === "ok") {
+            gpuIndex = decision2.gpuIndex ?? null;
+            admitted = true;
+            break;
+          }
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              const before = this._loaded.get(`${target.host}:${target.name}`) ?? target;
+              if (await this.evict(target.host, target.name, reason)) {
+                evictedModels.push(before);
+              }
+            }
+            await this.pollOnce().catch(() => {
+            });
+            continue;
+          }
+          if (decision2.kind === "degrade")
+            return decision2;
+          return decision2;
+        }
+        if (!admitted) {
+          return {
+            kind: "reject",
+            reason: `could not acquire transient load lease for ${spec.host}:${spec.name} after repeated evictions`
+          };
+        }
+        const evictedOllamaModels = dedupeLoadedModels(evictedModels.filter((m2) => m2.host === "ollama"));
+        const broker = this;
+        let released = false;
+        return {
+          kind: "ok",
+          lease: {
+            spec,
+            gpuIndex,
+            evictedModels: dedupeLoadedModels(evictedModels),
+            evictedOllamaModels,
+            async release() {
+              if (released)
+                return;
+              released = true;
+              broker.clearInflight(spec.host, spec.name);
+              if ((options2.unloadRequestedOllama ?? true) && spec.host === "ollama") {
+                await broker.unloadOllamaModel(spec.name, `${reason}-complete`).catch(() => false);
+              }
+              if (options2.restoreOllama !== false && evictedOllamaModels.length > 0) {
+                await broker.restoreOllamaModels(evictedOllamaModels, {
+                  keepAlive: options2.restoreKeepAlive ?? "30m"
+                });
+              }
+              await broker.pollOnce().catch(() => {
+              });
+            }
+          }
+        };
+      }
       /**
        * Register a model that has been successfully loaded.
        * Callers MUST call this after a successful load so the broker can track LRU.
@@ -1743,6 +1853,66 @@ var init_model_broker = __esm({
         this.emit("evicted", m2, reason);
         return actively;
       }
+      /** Force-unload an Ollama model even when it is not currently tracked. */
+      async unloadOllamaModel(modelName, reason = "ollama-unload") {
+        const key = `ollama:${modelName}`;
+        const existing = this._loaded.get(key);
+        const ok3 = await this.ollamaUnload(modelName).catch(() => false);
+        if (existing) {
+          this._loaded.delete(key);
+          this.emit("evicted", existing, reason);
+        }
+        return ok3;
+      }
+      /** Best-effort warm/reload of an Ollama model after temporary eviction. */
+      async warmOllamaModel(modelName, keepAlive = "30m") {
+        const bodies = [
+          { model: modelName, prompt: "", stream: false, keep_alive: keepAlive, options: { num_predict: 0 } },
+          { model: modelName, prompt: "", stream: false, keep_alive: keepAlive, options: { num_predict: 1 } }
+        ];
+        for (const body of bodies) {
+          try {
+            const res = await fetch(`${this._ollamaBaseUrl}/api/generate`, {
+              method: "POST",
+              headers: { "Content-Type": "application/json" },
+              body: JSON.stringify(body),
+              signal: AbortSignal.timeout(12e4)
+            });
+            if (!res.ok)
+              continue;
+            await this.refreshOllamaPs().catch(() => {
+            });
+            return true;
+          } catch {
+          }
+        }
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/generate`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({
+              model: modelName,
+              stream: false,
+              keep_alive: keepAlive
+            }),
+            signal: AbortSignal.timeout(12e4)
+          });
+          if (!res.ok)
+            return false;
+          await this.refreshOllamaPs().catch(() => {
+          });
+          return true;
+        } catch {
+          return false;
+        }
+      }
+      /** Restore a set of previously evicted Ollama models, oldest first. */
+      async restoreOllamaModels(models, options2 = {}) {
+        const unique2 = dedupeLoadedModels(models.filter((m2) => m2.host === "ollama")).sort((a2, b) => a2.lastUsedAt - b.lastUsedAt);
+        for (const model of unique2) {
+          await this.warmOllamaModel(model.name, options2.keepAlive ?? "30m").catch(() => false);
+        }
+      }
       // ------------------------------------------------------------------
       // Internal — Ollama
       // ------------------------------------------------------------------
@@ -1885,7 +2055,7 @@ var init_model_broker = __esm({
         );
         const idle = (m2) => now - m2.lastUsedAt > this.idleEvictMs;
         const onTargetGpu = (m2) => req2.targetGpu === void 0 || req2.targetGpu === null ? true : m2.gpuIndex === req2.targetGpu;
-        const evictable = [...this._loaded.values()].filter((m2) => m2.priority <= req2.requestingPriority).filter(sameDomainOk).filter(onTargetGpu).sort((a2, b) => {
+        const evictable = [...this._loaded.values()].filter((m2) => m2.priority <= req2.requestingPriority).filter(sameDomainOk).filter(onTargetGpu).filter((m2) => !this.hasActiveSlotForModel(m2)).sort((a2, b) => {
           const aIdle = idle(a2) ? 0 : 1;
           const bIdle = idle(b) ? 0 : 1;
           if (aIdle !== bIdle)
@@ -1931,6 +2101,13 @@ var init_model_broker = __esm({
             n2++;
         return n2;
       }
+      hasActiveSlotForModel(model) {
+        for (const slot of this._activeSlots.values()) {
+          if (slot.model === model.name)
+            return true;
+        }
+        return false;
+      }
       // ------------------------------------------------------------------
       // Internal — fallback resolution
       // ------------------------------------------------------------------
@@ -22761,6 +22938,17 @@ function evictModelsToFreeSpace(args) {
   writeMeta(meta);
   return { evicted, bytesFreed, finalFreeBytes: disk.freeBytes };
 }
+function estimateReclaimableCacheBytes(keepRepos) {
+  const keep = new Set(keepRepos ?? []);
+  let total = 0;
+  for (const entry of readMeta().entries) {
+    if (keep.has(entry.repo))
+      continue;
+    const measured = measureRepoCacheBytes(entry.repo);
+    total += Math.max(0, measured || entry.sizeBytes || 0);
+  }
+  return total;
+}
 function ensureDiskSpaceForDownload(args) {
   ensureUnifiedCacheDirs();
   const safetyMargin = args.safetyMarginBytes ?? 1 * 1024 ** 3;
@@ -22769,6 +22957,10 @@ function ensureDiskSpaceForDownload(args) {
   if (disk.freeBytes >= target) {
     return { ok: true, evicted: [], freeBytes: disk.freeBytes };
   }
+  const reclaimableBytes = estimateReclaimableCacheBytes(args.keepRepos);
+  if (disk.freeBytes + reclaimableBytes < target) {
+    throw new InsufficientDiskSpaceError(args.approxDownloadBytes, disk.freeBytes, disk.totalBytes, []);
+  }
   const evictionResult = evictModelsToFreeSpace({
     neededBytes: args.approxDownloadBytes,
     keepRepos: args.keepRepos,
@@ -259208,6 +259400,12 @@ function imageCandidateFor(model, requestedBackend) {
     preset: getImageGenerationPreset(resolved)
   };
 }
+function imageCandidateHost(candidate) {
+  return candidate.backend === "ollama" ? "ollama" : "diffusers-py";
+}
+function imageCandidateEstimatedVramMB(candidate) {
+  return candidate.preset?.minVramGB !== void 0 ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function imageGenerationFallbackCandidates(requestedModel, requestedBackend, allowFallback = true) {
   const ladder = imageGenerationQualityLadder();
   const candidates = [];
@@ -259511,9 +259709,15 @@ function annotateImageFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatImageAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatImageAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 function parseRunnerJson(stdout) {
@@ -260321,6 +260525,45 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: "image-gen",
+          host: imageCandidateHost(args.candidate),
+          owner: "image-generate-tool",
+          estimatedVramMB: imageCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: true
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Image generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `Image generation needs a broker fallback (${decision2.fallback.name}), but image candidate fallback must be selected by the image ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for image generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async execute(args) {
         const start2 = performance.now();
         const action = String(args["action"] ?? "generate");
@@ -260363,33 +260606,6 @@ if __name__ == "__main__":
         const requestedBackend = args["backend"] ? String(args["backend"]) : this.defaultBackend;
         const seed = optionalNumberArg(args["seed"]);
         const candidates = imageGenerationFallbackCandidates(requestedModel, requestedBackend, generationFallbackEnabled(args));
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: "image-gen",
-            host: firstCandidate.backend === "ollama" ? "ollama" : "diffusers-py",
-            owner: "image-generate-tool"
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, "image-gen-needs-room");
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `Image generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         try {
           return await this.generateCandidateLadder({ candidates, prompt, args, seed, start: start2 });
         } catch (err) {
@@ -260408,7 +260624,30 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Preparing image model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "ollama" ? await this.prewarmOllama({ model: candidate.model, start: args.start }) : candidate.backend === "sdcpp" ? await this.prewarmSdCpp({ model: candidate.model, start: args.start, python: args.args["python"] }) : await this.prewarmDiffusers({ model: candidate.model, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "image-prewarm-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "ollama" ? await this.prewarmOllama({ model: candidate.model, start: args.start }) : candidate.backend === "sdcpp" ? await this.prewarmSdCpp({ model: candidate.model, start: args.start, python: args.args["python"] }) : await this.prewarmDiffusers({ model: candidate.model, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateImageFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult(result) });
@@ -260447,7 +260686,30 @@ if __name__ == "__main__":
             message: `Using image model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
           const promptForCandidate = expansionEnabled ? await this.expandPromptForCandidate(args.prompt, candidate, index, args.candidates.length) : args.prompt;
-          const result = candidate.backend === "ollama" ? await this.generateWithOllama({ prompt: promptForCandidate, model: candidate.model, width, height, steps, start: args.start }) : candidate.backend === "sdcpp" ? await this.generateWithSdCpp({ prompt: promptForCandidate, model: candidate.model, width, height, steps, seed: args.seed, start: args.start, python: args.args["python"] }) : await this.generateWithDiffusers({ prompt: promptForCandidate, model: candidate.model, width, height, steps, guidance, seed: args.seed, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "image-gen-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "ollama" ? await this.generateWithOllama({ prompt: promptForCandidate, model: candidate.model, width, height, steps, start: args.start }) : candidate.backend === "sdcpp" ? await this.generateWithSdCpp({ prompt: promptForCandidate, model: candidate.model, width, height, steps, seed: args.seed, start: args.start, python: args.args["python"] }) : await this.generateWithDiffusers({ prompt: promptForCandidate, model: candidate.model, width, height, steps, guidance, seed: args.seed, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success) {
             await this.writeImageSidecar(result, {
               originalPrompt: args.prompt,
@@ -260671,6 +260933,17 @@ ${errText.slice(0, 1200)}`,
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading image model ${args.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "image", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but image CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess2(python.command, [
           runner,
           "--model",
@@ -260683,7 +260956,7 @@ ${errText.slice(0, 1200)}`,
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -261735,6 +262008,18 @@ function audioCandidateFor(kind, model, requestedBackend) {
     preset: getAudioGenerationPreset(model, kind)
   };
 }
+function audioCandidateHost(candidate) {
+  if (candidate.backend === "project")
+    return null;
+  if (candidate.backend === "audiocraft")
+    return "audiocraft";
+  if (candidate.backend === "tangoflux")
+    return "tangoflux";
+  return "diffusers-py";
+}
+function audioCandidateEstimatedVramMB(candidate) {
+  return candidate.preset ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, allowFallback = true) {
   const ladder = audioGenerationQualityLadder(kind);
   const candidates = [];
@@ -261891,9 +262176,15 @@ function annotateAudioFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatAudioAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatAudioAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 var DEFAULT_SOUND_MODEL, DEFAULT_MUSIC_MODEL, DIFFUSERS_AUDIO_PACKAGES, TRANSFORMERS_AUDIO_PACKAGES, AUDIOCRAFT_PACKAGES, STABLE_AUDIO_PACKAGES, TANGOFLUX_PACKAGES, AUDIO_GENERATION_MODEL_PRESETS, SOUND_GENERATION_QUALITY_LADDER, MUSIC_GENERATION_QUALITY_LADDER, DIFFUSERS_AUDIO_RUNNER, AUDIOCRAFT_RUNNER, TRANSFORMERS_AUDIO_RUNNER, TANGOFLUX_RUNNER, AudioGenerateTool;
@@ -262730,6 +263021,48 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const host = audioCandidateHost(args.candidate);
+        if (!host)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: args.kind,
+          host,
+          owner: `audio-generate-tool/${args.kind}`,
+          estimatedVramMB: audioCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: false
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `${args.kind === "music" ? "Music" : "Sound"} generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `${args.kind === "music" ? "Music" : "Sound"} generation needs a broker fallback (${decision2.fallback.name}), but audio candidate fallback must be selected by the audio ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for ${args.kind} generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async prewarmPythonBackend(args) {
         const runner = await ensureAudioRunner(this.cwd, args.runnerBackend);
         let python;
@@ -262776,6 +263109,17 @@ if __name__ == "__main__":
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading ${args.kind} model ${args.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (audioBrokerGpuIndexIsCompatible(this._brokerGpuIndex, runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but audio CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess3(python.command, [
           runner,
           "--kind",
@@ -262792,7 +263136,7 @@ if __name__ == "__main__":
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -262872,33 +263216,6 @@ if __name__ == "__main__":
         const candidates = audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, generationFallbackEnabled2(args));
         const seed = optionalNumberArg2(args["seed"]);
         const playback = playbackRequested(args);
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: kind === "music" ? "music" : "sound",
-            host: firstCandidate.backend === "audiocraft" ? "audiocraft" : firstCandidate.backend === "tangoflux" ? "tangoflux" : firstCandidate.backend === "transformers" ? "diffusers-py" : "diffusers-py",
-            owner: `audio-generate-tool/${kind}`
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, `${kind}-gen-needs-room`);
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `${kind === "music" ? "Music" : "Sound"} generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         try {
           return await this.generateCandidateLadder({ kind, candidates, prompt, args, seed, playback, start: start2 });
         } catch (err) {
@@ -262918,15 +263235,39 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Preparing ${args.kind} model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.prewarmPythonBackend({
+          const leaseDecision = await this.acquireTransientLoadLease({
             kind: args.kind,
-            backend: candidate.backend,
-            runnerBackend: candidate.backend,
-            model: candidate.model,
-            duration,
-            start: args.start,
-            python: args.args["python"]
-          });
+            candidate,
+            reason: `${args.kind}-prewarm-needs-room`,
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult2(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.prewarmPythonBackend({
+              kind: args.kind,
+              backend: candidate.backend,
+              runnerBackend: candidate.backend,
+              model: candidate.model,
+              duration,
+              start: args.start,
+              python: args.args["python"]
+            });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateAudioFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult2(result) });
@@ -262954,19 +263295,43 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Using ${args.kind} model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.generateWithPythonBackend({
+          const leaseDecision = await this.acquireTransientLoadLease({
             kind: args.kind,
-            backend: candidate.backend,
-            runnerBackend: candidate.backend,
-            prompt: args.prompt,
-            model: candidate.model,
-            duration,
-            steps,
-            seed: args.seed,
-            playback: args.playback,
-            start: args.start,
-            python: args.args["python"]
-          });
+            candidate,
+            reason: `${args.kind}-gen-needs-room`,
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult2(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.generateWithPythonBackend({
+              kind: args.kind,
+              backend: candidate.backend,
+              runnerBackend: candidate.backend,
+              prompt: args.prompt,
+              model: candidate.model,
+              duration,
+              steps,
+              seed: args.seed,
+              playback: args.playback,
+              start: args.start,
+              python: args.args["python"]
+            });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateAudioFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult2(result) });
@@ -263306,6 +263671,12 @@ function videoCandidateFor(model, requestedBackend, requestedKind) {
   }
   return { model, backend, preset };
 }
+function videoCandidateHost(candidate) {
+  return candidate.backend === "comfyui" ? "comfyui" : "diffusers-py";
+}
+function videoCandidateEstimatedVramMB(candidate) {
+  return candidate.preset ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function videoGenerationFallbackCandidates(requestedModel, requestedBackend, requestedKind, allowFallback = true, options2 = {}) {
   const preferAudioVideo = Boolean(options2.preferNativeAudioVideo);
   const baseLadderIds = preferAudioVideo ? [...VIDEO_AUDIO_QUALITY_LADDER, ...VIDEO_GENERATION_QUALITY_LADDER] : VIDEO_GENERATION_QUALITY_LADDER;
@@ -263871,9 +264242,15 @@ function annotateVideoFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatVideoAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatVideoAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 function parseRunnerJson3(stdout) {
@@ -265240,6 +265617,45 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: "video-gen",
+          host: videoCandidateHost(args.candidate),
+          owner: "video-generate-tool",
+          estimatedVramMB: videoCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: false
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Video generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `Video generation needs a broker fallback (${decision2.fallback.name}), but video candidate fallback must be selected by the video ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for video generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async execute(args) {
         const start2 = performance.now();
         const action = String(args["action"] ?? "generate");
@@ -265295,35 +265711,6 @@ if __name__ == "__main__":
         const withAudio = booleanArg3(args["with_audio"], false);
         const audioInput = typeof args["audio_input"] === "string" && args["audio_input"].trim() ? String(args["audio_input"]).trim() : void 0;
         const candidates = videoGenerationFallbackCandidates(requestedModel, requestedBackend, inferredKind, generationFallbackEnabled3(args), { preferNativeAudioVideo: withAudio || Boolean(audioInput) });
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const preset = firstCandidate.preset;
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: "video-gen",
-            host: firstCandidate.backend === "comfyui" ? "comfyui" : "diffusers-py",
-            owner: "video-generate-tool",
-            estimatedVramMB: preset ? preset.minVramGB * 1024 : void 0
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, "video-gen-needs-room");
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `Video generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         if (candidates.length === 0) {
           return {
             success: false,
@@ -265373,7 +265760,30 @@ if __name__ == "__main__":
             failed.push({ candidate, reason: "ComfyUI backend not yet implemented." });
             continue;
           }
-          const result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "video-prewarm-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateVideoFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult3(result) });
@@ -265459,26 +265869,48 @@ if __name__ == "__main__":
               start: args.start
             });
           } else {
-            result = await this.generateWithDiffusers({
-              prompt: promptForCandidate,
-              model: candidate.model,
-              preset,
-              kind: args.kind,
-              imageArg: args.imageArg,
-              audioInput: args.audioInput,
-              width,
-              height,
-              numFrames,
-              fps,
-              steps,
-              guidance,
-              negativePrompt,
-              seed: args.seed,
-              hfToken: hfTokenOverride,
-              autoAcceptLicense,
-              start: args.start,
-              python: args.args["python"]
+            const leaseDecision = await this.acquireTransientLoadLease({
+              candidate,
+              reason: "video-gen-needs-room",
+              start: args.start
             });
+            if (leaseDecision && "success" in leaseDecision) {
+              failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+              if (index < args.candidates.length - 1) {
+                this.emitProgress({
+                  stage: "setup",
+                  message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+                });
+              }
+              continue;
+            }
+            const lease = leaseDecision?.lease;
+            this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+            try {
+              result = await this.generateWithDiffusers({
+                prompt: promptForCandidate,
+                model: candidate.model,
+                preset,
+                kind: args.kind,
+                imageArg: args.imageArg,
+                audioInput: args.audioInput,
+                width,
+                height,
+                numFrames,
+                fps,
+                steps,
+                guidance,
+                negativePrompt,
+                seed: args.seed,
+                hfToken: hfTokenOverride,
+                autoAcceptLicense,
+                start: args.start,
+                python: args.args["python"]
+              });
+            } finally {
+              await lease?.release();
+              this._brokerGpuIndex = null;
+            }
           }
           let nativeAudio = preset.nativeAudioVideo === true;
           let audioPath;
@@ -265670,6 +266102,17 @@ ${llmAnnotation}` : result.llmContent;
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading video model ${args.candidate.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "video", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but video CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess4(python.command, [
           runner,
           "--model",
@@ -265685,7 +266128,7 @@ ${llmAnnotation}` : result.llmContent;
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.candidate.model}`,
           onProgress: (event) => this.emitProgress(event)
         });

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.187",
+  "version": "1.0.188",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.187",
+      "version": "1.0.188",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.187",
+  "version": "1.0.188",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",