Rebuild: QAD phase memory logging + skip-checkpoint

kit_manifest.json

@@ -1,5 +1,5 @@
 {
-  "built_from_commit": "bc56971bb65b256aed3045c54cfbfb8340158935",
+  "built_from_commit": "06b0717a82dcc99730c21cdf81cc57b489617141",
   "capabilities": [
     "rht_int4_convert.layer_filter",
     "eval_quantized.dtype",
@@ -24,8 +24,8 @@
     "scripts/yaqa_calibration.py": "5fbee3f3594d6d15cd67e5aa9cba20572db5aef698568e8982a101aeaab5b073",
     "scripts/merge_language_hessians.py": "a6385caaa0d7190024b4e47ded59d1ab662e771ce81e18de94b30c8846829b4e",
     "scripts/merge_calibration_groups.py": "58a347ff40c47f629c073c23aea54e215c1e75c052737e5b4a48ff81946d8060",
-    "scripts/eval_quantized.py": "b6c2b604812008e52c9347e38f7e8ad2a067e92cc8252040901c431eedb505ed",
-    "scripts/qad_finetune.py": "ea03a1ecfdc9dcd06d67ac46ef9091fcdda3d51165af8dfffc9fe61ce295cabc",
+    "scripts/eval_quantized.py": "5b4e5454630724c116d14fbc64061ffdd6066f4a458e228d372354d3eb32c5b0",
+    "scripts/qad_finetune.py": "5cd719d7b2f5251bd1c9903a403f75ae86e4da82ec0d18298bf092c212c86b96",
     "scripts/compute_bpc.py": "2b55fd1df406d54b40daeeca2ecaf2efb35981655f5bf5e0cc10c057ce796256",
     "scripts/run_lane.py": "eef3be44e5d711995b6993e3e9ff36aa6007b593b44357041e08d6fa4c0d024e",
     "scripts/make_sweep_spec.py": "dc31ec026c89e7db316a185575a4423bca988c2ac6cf08ef6df8f955328cda69",

scripts/eval_quantized.py

@@ -275,6 +275,12 @@ def main() -> None:
         "corpus": str(args.corpus),
         "config": eval_config,
         "config_hash": config_hash,
+        # dtype changes absolute PPL (0.8B runs float32 on CPU/T4, 4B/9B run
+        # bfloat16 on A100), so it must travel with the number.
+        "eval_dtype": args.dtype,
+        # CPU float32 and GPU float32 agree to 0.02%, but recording the device
+        # makes every number self-describing.
+        "device": jax.devices()[0].device_kind if jax.devices() else "unknown",
         "chunks": len(chunks),
         "tokens": total_positions,
         "chars": char_total,

scripts/qad_finetune.py

@@ -40,6 +40,18 @@ from pathlib import Path
 os.environ.setdefault("JAX_PLATFORM_NAME", "cpu")
 
 
+def peak_rss_mb() -> float:
+    """Peak resident memory of this process, in MB.
+
+    A silent death right after the last step (no traceback in the log, no
+    checkpoint) looks like the OOM killer; this number says which phase grew.
+    """
+    import resource
+
+    usage = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+    return usage / (1024 * 1024) if sys.platform == "darwin" else usage / 1024
+
+
 def early_stop_decision(bad_val_evals: int, patience: int, step: int, steps: int) -> str:
     """Decide what to do after a non-improving validation.
 
@@ -653,29 +665,51 @@ def main() -> None:
             flush=True,
         )
         completed_steps = step + 1
-        if args.save_every and (step + 1) % args.save_every == 0:
+        if args.save_every and (step + 1) % args.save_every == 0 and not args.skip_checkpoint:
+            print(f"checkpoint {step + 1} start peak_rss={peak_rss_mb():.0f}MB", flush=True)
             save_checkpoint(step + 1, params, opt_state, rng)
+            print(f"checkpoint {step + 1} done  peak_rss={peak_rss_mb():.0f}MB", flush=True)
             last_saved_step = step + 1
 
     if stop_reason:
         completed_text = f"final step {completed_steps}" if completed_steps >= args.steps else f"step {completed_steps}"
         print(f"early stop at {completed_text}: {stop_reason}", flush=True)
 
-    if args.steps > 0 and completed_steps > 0 and completed_steps > last_saved_step:
+    print(f"training done peak_rss={peak_rss_mb():.0f}MB", flush=True)
+    if (
+        args.steps > 0
+        and completed_steps > 0
+        and completed_steps > last_saved_step
+        and not args.skip_checkpoint
+    ):
+        print(f"final checkpoint start peak_rss={peak_rss_mb():.0f}MB", flush=True)
         save_checkpoint(completed_steps, params, opt_state, rng)
+        print(f"final checkpoint done  peak_rss={peak_rss_mb():.0f}MB", flush=True)
 
+    print(f"export start peak_rss={peak_rss_mb():.0f}MB", flush=True)
     if args.master_dtype != "bfloat16":
         params = {key: value.astype(master_dtypes.get(key, value.dtype)) for key, value in params.items()}
     trained = inject_params(student, params)
+    del params
+    opt_state = None
+    try:
+        import jax
+
+        jax.clear_caches()
+    except Exception:  # noqa: BLE001 - freeing memory must never break export
+        pass
     trained = map_nodes_of_type_with_path(
         HybridMatrix,
         lambda path, matrix: matrix.switch_implementation(CompressionImplementation.INFERENCE),
         trained,
     )
+    print(f"export injected peak_rss={peak_rss_mb():.0f}MB", flush=True)
     args.output_dir.mkdir(parents=True, exist_ok=True)
     trained.save(args.output_dir)
+    print(f"export saved peak_rss={peak_rss_mb():.0f}MB", flush=True)
     rewrite(args.output_dir / "model.safetensors")
     normalize(args.output_dir)
+    print(f"export done peak_rss={peak_rss_mb():.0f}MB", flush=True)
     print(f"saved QAD model -> {args.output_dir}")