From ebb912452dd7a286eea2e4bd41d4ce32c5317f69 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 10 Jul 2026 13:56:17 +0800
Subject: [PATCH] U4 step3: dim_batch 16->8 to survive OOM contention with
 user's VS Code kernel

Run 551 was OOM-killed at n_done=36: the user's VS Code Jupyter kernel
(jsteer venv, PID 3214401) co-loaded ~1.5GB VRAM + 1.9GB RAM while the fit
sat at the 22.4/24.6GB ceiling. Clean SIGKILL with no CUDA traceback = host
OOM killer, not a CUDA OOM. dim_batch=8 halves the fit's peak footprint;
it changes only the backward schedule, not the accumulated Jacobian, so U4
exactness is preserved. Resumes from checkpoint (n_done=36), lossless.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
---
 {artifacts => docs/evidence}/parity_u1.txt           | 0
 {artifacts => docs/evidence}/u4_step2_vjp_parity.txt | 0
 scripts/{ => scratch}/parity_u1.py                   | 0
 scripts/{ => scratch}/u4_step1_ref524.py             | 0
 scripts/{ => scratch}/u4_step2_vjp.py                | 0
 scripts/u4_step3_fit4b.py                            | 9 ++++++++-
 6 files changed, 8 insertions(+), 1 deletion(-)
 rename {artifacts => docs/evidence}/parity_u1.txt (100%)
 rename {artifacts => docs/evidence}/u4_step2_vjp_parity.txt (100%)
 rename scripts/{ => scratch}/parity_u1.py (100%)
 rename scripts/{ => scratch}/u4_step1_ref524.py (100%)
 rename scripts/{ => scratch}/u4_step2_vjp.py (100%)

diff --git a/artifacts/parity_u1.txt b/docs/evidence/parity_u1.txt
similarity index 100%
rename from artifacts/parity_u1.txt
rename to docs/evidence/parity_u1.txt
diff --git a/artifacts/u4_step2_vjp_parity.txt b/docs/evidence/u4_step2_vjp_parity.txt
similarity index 100%
rename from artifacts/u4_step2_vjp_parity.txt
rename to docs/evidence/u4_step2_vjp_parity.txt
diff --git a/scripts/parity_u1.py b/scripts/scratch/parity_u1.py
similarity index 100%
rename from scripts/parity_u1.py
rename to scripts/scratch/parity_u1.py
diff --git a/scripts/u4_step1_ref524.py b/scripts/scratch/u4_step1_ref524.py
similarity index 100%
rename from scripts/u4_step1_ref524.py
rename to scripts/scratch/u4_step1_ref524.py
diff --git a/scripts/u4_step2_vjp.py b/scripts/scratch/u4_step2_vjp.py
similarity index 100%
rename from scripts/u4_step2_vjp.py
rename to scripts/scratch/u4_step2_vjp.py
diff --git a/scripts/u4_step3_fit4b.py b/scripts/u4_step3_fit4b.py
index 2e606cd..33a2913 100644
--- a/scripts/u4_step3_fit4b.py
+++ b/scripts/u4_step3_fit4b.py
@@ -31,8 +31,15 @@ model = AutoModelForCausalLM.from_pretrained(
     meta["model"], torch_dtype=torch.bfloat16).to("cuda").eval()
 
 t0 = time.time()
+# dim_batch 16 -> 8 (Claude): run 551 was OOM-killed at n_done=36 when the
+# user's VS Code Jupyter kernel (jsteer venv) co-loaded ~1.5GB VRAM + 1.9GB RAM
+# while this fit sat at the 22.4/24.6GB ceiling -- clean SIGKILL, no CUDA
+# traceback = host OOM killer, not a CUDA OOM. dim_batch=8 halves this fit's
+# peak footprint to coexist with the kernel. It only changes the backward
+# SCHEDULE (2x passes), NOT the accumulated Jacobian, so U4 exactness holds.
+# Resumes from the existing checkpoint (n_done=36), lossless.
 jac = Jacobian.fit(model, tok, meta["prompts"], layers=meta["layers"],
-                   dim_batch=16, max_seq_len=384,
+                   dim_batch=8, max_seq_len=384,
                    checkpoint_path=str(ART / "qwen3-4b-authority.ckpt"))
 logger.info(f"fit wall-time: {(time.time() - t0) / 3600:.2f} h")
 jac.save(str(ART / "qwen3-4b-authority.jac"))