add stage-artifact push/fetch resume logic (protects long GPU run against external cancellation)

experiment.py

@@ -686,21 +686,80 @@ def push_results():
                         repo_id=OUT_REPO, repo_type="dataset")
     log("pushed results to %s" % OUT_REPO)
 
+ART_DIR = "results/full/artifacts"
+
+def push_artifact(name):
+    if MODE != "full":
+        return
+    token = os.environ.get("HF_TOKEN")
+    if not token:
+        log("no HF_TOKEN; skipping artifact push (%s)" % name)
+        return
+    try:
+        from huggingface_hub import HfApi
+        api = HfApi(token=token)
+        api.upload_file(path_or_fileobj=os.path.join(DATA_DIR, name),
+                        path_in_repo="%s/%s" % (ART_DIR, name),
+                        repo_id=OUT_REPO, repo_type="dataset")
+        log("artifact pushed: %s" % name)
+    except Exception as e:
+        log("artifact push FAILED (%s): %s" % (name, str(e)[:300]))
+
+def fetch_artifact(name):
+    p = os.path.join(DATA_DIR, name)
+    if os.path.exists(p):
+        log("artifact exists locally: %s" % name)
+        return True
+    if MODE != "full":
+        return False
+    if not os.environ.get("HF_TOKEN"):
+        return False
+    try:
+        import shutil
+        from huggingface_hub import hf_hub_download
+        dl = hf_hub_download(repo_id=OUT_REPO,
+                            filename="%s/%s" % (ART_DIR, name),
+                            repo_type="dataset")
+        os.makedirs(DATA_DIR, exist_ok=True)
+        shutil.move(dl, p)
+        log("artifact fetched from hub: %s" % name)
+        return True
+    except Exception as e:
+        log("artifact not on hub (%s): %s" % (name, str(e)[:200]))
+        return False
+
 # ---------------- main ----------------
 
 def main():
     log("starting mode=%s prepare_only=%s" % (MODE, PREPARE_ONLY))
-    if not os.path.exists(STATE_F):
-        prepare()
+    if fetch_artifact("state.json"):
+        log("skipping prepare (state.json available)")
     else:
-        log("state exists; skipping prepare")
+        prepare()
+        push_artifact("state.json")
     if PREPARE_ONLY:
         log("prepare-only mode; done")
         return
-    run_stage("extract")
-    run_stage("embed")
-    retrieve_all()
-    run_stage("read")
+    if fetch_artifact("triples.json"):
+        log("skipping extract (triples.json available)")
+    else:
+        run_stage("extract")
+        push_artifact("triples.json")
+    if fetch_artifact("emb.npy"):
+        log("skipping embed (emb.npy available)")
+    else:
+        run_stage("embed")
+        push_artifact("emb.npy")
+    if fetch_artifact("retrieval.json"):
+        log("skipping retrieve (retrieval.json available)")
+    else:
+        retrieve_all()
+        push_artifact("retrieval.json")
+    if fetch_artifact("answers.json"):
+        log("skipping read (answers.json available)")
+    else:
+        run_stage("read")
+        push_artifact("answers.json")
     metrics()
     if MODE == "full":
         push_results()