Add minimax h3 mirror job script

worker/minimax_mirror.py

@@ -0,0 +1,46 @@
+"""One-shot job: mirror the int8 MiniMax H3 files from Comfy-Org/MiniMax-H3
+into two compact repos (one per future serverless endpoint) so RunPod's
+model-caching feature pulls ~54GB instead of the ~400GB full repo.
+
+Runs on a throwaway RunPod CPU pod. Downloads one file at a time to keep
+disk usage under the biggest single file (~27GB) + upload buffer.
+"""
+
+import os
+import time
+
+from huggingface_hub import HfApi, hf_hub_download
+
+SRC = "Comfy-Org/MiniMax-H3"
+SHARED = [
+    "text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors",
+    "vae/minimax_h3_audio_vae_fp32.safetensors",
+    "vae/minimax_h3_video_vae_fp16.safetensors",
+]
+PLANS = {
+    "nikdevs/minimax-h3-fl2va": ["diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors"] + SHARED,
+    "nikdevs/minimax-h3-ref2va": ["diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors"] + SHARED,
+}
+WORK = "/workdir"
+os.makedirs(WORK, exist_ok=True)
+
+api = HfApi()
+
+for repo, files in PLANS.items():
+    api.create_repo(repo, exist_ok=True, private=False)
+    existing = {s.rfilename for s in api.repo_info(repo).siblings}
+    for f in files:
+        if f in existing:
+            print(f"[skip] {repo} already has {f}", flush=True)
+            continue
+        t0 = time.time()
+        print(f"[dl] {SRC}/{f}", flush=True)
+        path = hf_hub_download(SRC, f, local_dir=WORK)
+        print(f"[dl] done in {time.time()-t0:.0f}s, uploading to {repo}", flush=True)
+        t0 = time.time()
+        api.upload_file(path_or_fileobj=path, path_in_repo=f, repo_id=repo)
+        print(f"[up] done in {time.time()-t0:.0f}s", flush=True)
+        os.remove(path)
+
+print("MIRROR DONE", flush=True)
+time.sleep(10 ** 6)  # idle; the orchestrator terminates the pod