From 131a4a8b55255ed33f8f36bee60cc67915a58044 Mon Sep 17 00:00:00 2001
From: Joseph Magly <1159087+jmagly@users.noreply.github.com>
Date: Wed, 2 Sep 2026 14:32:29 -0400
Subject: [PATCH] docs: align fused offload claims with test contract

---
 README.md                         | 2 +-
 tests/test_offline_integration.py | 2 +-
 tests/test_offload_surgery.py     | 1 -
 3 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/README.md b/README.md
index f8fced1..cb3a7f7 100644
--- a/README.md
+++ b/README.md
@@ -521,7 +521,7 @@ This means multi-GPU sharding is a **memory solution, not a speed solution**. It
 
 ### Offloaded layers and fused MoE experts
 
-When the combined GPU budget is still too small, accelerate offloads whole layers to CPU RAM or disk and leaves `meta` placeholders in the live model. OBLITERATUS performs surgery on those layers through the authoritative offload backing store instead of materializing them. Linear projections, bias projections, and fused 3D expert tensors are all handled by the same transactional path, so the fused `experts.gate_up_proj` / `experts.down_proj` layout that transformers 5 uses for MoE architectures (Mixtral, Qwen3-MoE, DeepSeek-V3, GLM-4 MoE, Llama 4, gpt-oss, OLMoE) is updated in place and the saved checkpoint is identical to what live surgery would produce. Quantized or unrecognized offload layouts fail closed before any weight is modified. Attention-head surgery, safety-neuron masking, expert transplant, and LoRA ablation are still rejected for offloaded layers.
+When the combined GPU budget is still too small, accelerate offloads whole layers to CPU RAM or disk and leaves `meta` placeholders in the live model. OBLITERATUS performs surgery on those layers through the authoritative offload backing store instead of materializing them. Linear projections, bias projections, and fused 3D expert tensors are all handled by the same transactional path, so the fused `experts.gate_up_proj` / `experts.down_proj` layout that transformers 5 uses for MoE architectures (Mixtral, Qwen3-MoE, DeepSeek-V3, GLM-4 MoE, Llama 4, gpt-oss, OLMoE) is updated in place and the saved checkpoint matches what live surgery would produce within the configured floating-point tolerance. Quantized or unrecognized offload layouts fail closed before any weight is modified. Attention-head surgery, safety-neuron masking, expert transplant, and LoRA ablation are still rejected for offloaded layers.
 
 ### Selecting GPUs
 
diff --git a/tests/test_offline_integration.py b/tests/test_offline_integration.py
index 35e0d61..464fd2f 100644
--- a/tests/test_offline_integration.py
+++ b/tests/test_offline_integration.py
@@ -312,7 +312,7 @@ def _moe_load_model(offload_dir: Path | None, live: dict):
     "ignore:Attempting to save a model with offloaded modules.*:UserWarning",
 )
 def test_disk_offloaded_fused_moe_surgery_matches_live_surgery(tmp_path):
-    """Offload must be transparent: same checkpoint as live surgery, byte for byte."""
+    """Offload must be transparent: checkpoint tensors match live surgery."""
     source = build_tiny_offline_moe_model(tmp_path / "source")
     offload_dir = tmp_path / "offload"
     offload_dir.mkdir()
diff --git a/tests/test_offload_surgery.py b/tests/test_offload_surgery.py
index 0db5105..fd6b02f 100644
--- a/tests/test_offload_surgery.py
+++ b/tests/test_offload_surgery.py
@@ -635,4 +635,3 @@ def test_fused_projection_commit_failure_leaves_backing_and_meta_state_intact():
     # byte-identical and the live parameter is still offloaded.
     torch.testing.assert_close(backing["down_proj"], original)
     assert experts.down_proj.device.type == "meta"
-