Upload mi300x_9b/rebuild_vllm_compat_final.py with huggingface_hub

mi300x_9b/rebuild_vllm_compat_final.py

@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+"""Rebuild vllm_compat from the final step-596 checkpoint.
+
+Strategy: clone the existing (working) step-150 vllm_compat dir (which has
+correct config.json with rtpurbo block, tokenizer, probe, and real safetensors
+copies of the base model), then swap in the new indexer weights and update
+the index.json weight_map. Keeps config architecture as ForConditionalGeneration.
+"""
+import json
+import shutil
+import sys
+from pathlib import Path
+import torch
+from safetensors.torch import save_file
+
+OLD = Path("/home/hotaisle/work/checkpoints/mi300x_9b_131k_step150/vllm_compat")
+NEW = Path("/home/hotaisle/work/checkpoints/mi300x_9b_131k/vllm_compat")
+SD  = Path("/home/hotaisle/work/checkpoints/mi300x_9b_131k/state_dict.pt")
+PROBE = Path("/home/hotaisle/work/probe_results/mi300x_9b_allretr.json")
+IDX_FNAME = "model-rtpurbo-indexers-of-1.safetensors"
+
+print(f"[rebuild] cloning {OLD} -> {NEW}", file=sys.stderr)
+if NEW.exists():
+    shutil.rmtree(NEW)
+shutil.copytree(OLD, NEW)
+print(f"[rebuild] cloned (preserving config/tokenizer/base safetensors)", file=sys.stderr)
+
+# 1. Load final state_dict, extract indexer tensors
+print(f"[rebuild] loading {SD}", file=sys.stderr)
+sd = torch.load(str(SD), map_location="cpu", weights_only=True)
+idx_tensors = {}
+for k, v in sd.items():
+    if torch.is_tensor(v) and ".indexers." in k:
+        idx_tensors[k] = v.contiguous()
+print(f"[rebuild] {len(idx_tensors)} indexer tensors", file=sys.stderr)
+
+# 2. Save new indexer safetensors (overwrite the old one)
+idx_path = NEW / IDX_FNAME
+save_file(idx_tensors, str(idx_path), metadata={"format": "pt"})
+size_kb = idx_path.stat().st_size / 1024
+print(f"[rebuild] wrote {IDX_FNAME} ({size_kb:.0f} KB)", file=sys.stderr)
+
+# 3. Update weight_map in model.safetensors.index.json (point indexer keys
+#    at the indexer file — should already be there, but rewrite to be safe)
+wm_path = NEW / "model.safetensors.index.json"
+index_data = json.loads(wm_path.read_text())
+for k in idx_tensors:
+    index_data["weight_map"][k] = IDX_FNAME
+wm_path.write_text(json.dumps(index_data, indent=2))
+print(f"[rebuild] updated model.safetensors.index.json ({len(idx_tensors)} indexer keys -> {IDX_FNAME})", file=sys.stderr)
+
+# 4. Refresh probe (copy latest)
+shutil.copy2(str(PROBE), str(NEW / "rtpurbo_probe.json"))
+print(f"[rebuild] refreshed rtpurbo_probe.json", file=sys.stderr)
+
+# 5. Verify architecture in config.json
+cfg = json.loads((NEW / "config.json").read_text())
+arch = cfg.get("architectures", [])
+print(f"[rebuild] config architectures: {arch}", file=sys.stderr)
+if "Qwen3_5RTPurboForConditionalGeneration" not in arch:
+    cfg["architectures"] = ["Qwen3_5RTPurboForConditionalGeneration"]
+    (NEW / "config.json").write_text(json.dumps(cfg, indent=2))
+    print(f"[rebuild] FIXED architecture -> ForConditionalGeneration", file=sys.stderr)
+else:
+    print(f"[rebuild] architecture OK (ForConditionalGeneration)", file=sys.stderr)
+
+# 6. Drop a marker file recording the build provenance
+marker = {
+    "checkpoint_step": 596,
+    "source_state_dict": str(SD),
+    "cloned_from": str(OLD),
+    "final_mean_kl": 2.3574943174134333,
+    "n_indexer_tensors": len(idx_tensors),
+    "probe": str(PROBE),
+}
+(NEW / "build_provenance.json").write_text(json.dumps(marker, indent=2))
+print(f"[rebuild] wrote build_provenance.json", file=sys.stderr)
+
+print(f"\n[rebuild] DONE -> {NEW}", file=sys.stderr)
+print(f"[rebuild] contents:", file=sys.stderr)
+for f in sorted(NEW.iterdir()):
+    if f.is_symlink():
+        print(f"  {f.name} -> {f.resolve()} ({f.stat().st_size/1e6:.1f} MB)", file=sys.stderr)
+    else:
+        print(f"  {f.name} ({f.stat().st_size} bytes)", file=sys.stderr)