v4.3.0

.github/workflows/checks.yml

@@ -443,6 +443,8 @@ jobs:
           - "Patchscopes_Generation_Demo"
           - "Santa_Coder"
           # - "stable_lm"
+          - "SVD_Circuits_Demo"
+          - "SwiGLU_Gate_Sign_Demo"
           - "T5"
         requires_hf_token: [false]
         include:

.github/workflows/slow-tests.yml

@@ -30,6 +30,10 @@ jobs:
           df -h /
       - name: Add swap for fp32 CPU model boots
         run: |
+          # The runner image can ship /swapfile already active; fallocate on
+          # live swap fails with "Text file busy", so release it first.
+          sudo swapoff /swapfile 2>/dev/null || true
+          sudo rm -f /swapfile
           sudo fallocate -l 8G /swapfile
           sudo chmod 600 /swapfile
           sudo mkswap /swapfile

demos/Jacobian_Lens_Coordinate_Patch_Benchmark_Demo.ipynb

@@ -48,9 +48,9 @@
     "prompts behind it.\n",
     "\n",
     "**This notebook loads a frozen artifact and never calls the model.** The artifact was\n",
-    "generated once, out-of-band, by running the benchmark's CLI entry point; see the load cell\n",
-    "below for the exact command. Re-running the sweep is a separate, explicit step, not something\n",
-    "this notebook triggers.\n"
+    "generated once, out-of-band, by `scripts/jacobian_lens_causal_swap_benchmark.py`; the load\n",
+    "cell below names the exact command. Re-running the sweep is a separate, explicit step, not\n",
+    "something this notebook triggers.\n"
    ]
   },
   {
@@ -59,10 +59,10 @@
    "id": "8a272a14",
    "metadata": {
     "execution": {
-     "iopub.execute_input": "2026-09-24T12:26:45.620923Z",
-     "iopub.status.busy": "2026-09-24T12:26:45.620839Z",
-     "iopub.status.idle": "2026-09-24T12:26:45.623972Z",
-     "shell.execute_reply": "2026-09-24T12:26:45.623559Z"
+     "iopub.execute_input": "2026-10-14T18:27:57.152894Z",
+     "iopub.status.busy": "2026-10-14T18:27:57.152687Z",
+     "iopub.status.idle": "2026-10-14T18:27:57.158584Z",
+     "shell.execute_reply": "2026-10-14T18:27:57.157853Z"
     }
    },
    "outputs": [
@@ -88,10 +88,10 @@
    "id": "bc924b1c",
    "metadata": {
     "execution": {
-     "iopub.execute_input": "2026-09-24T12:26:45.625566Z",
-     "iopub.status.busy": "2026-09-24T12:26:45.625485Z",
-     "iopub.status.idle": "2026-09-24T12:26:49.109284Z",
-     "shell.execute_reply": "2026-09-24T12:26:49.108728Z"
+     "iopub.execute_input": "2026-10-14T18:27:57.160521Z",
+     "iopub.status.busy": "2026-10-14T18:27:57.160357Z",
+     "iopub.status.idle": "2026-10-14T18:27:58.846420Z",
+     "shell.execute_reply": "2026-10-14T18:27:58.845982Z"
     }
    },
    "outputs": [
@@ -111,14 +111,18 @@
     }
    ],
    "source": [
+    "import hashlib\n",
+    "import json\n",
     "from pathlib import Path\n",
     "\n",
     "import numpy as np\n",
     "import pandas as pd\n",
     "\n",
-    "from transformer_lens.tools.analysis import load_artifact, success_rate_ci\n",
+    "from transformer_lens.utilities import success_rate_ci\n",
     "\n",
     "ARTIFACT_NAME = \"jacobian_lens_causal_swap_benchmark_gpt2.json\"\n",
+    "SCHEMA_VERSION = 2\n",
+    "RESULT_BLOCKS = (\"trials\", \"excluded_baselines\", \"real_success_ci\", \"control_success_ci\")\n",
     "\n",
     "\n",
     "def _find_artifact() -> Path:\n",
@@ -137,15 +141,30 @@
     "        \"directory. This notebook requires a repository checkout: it loads a frozen \"\n",
     "        \"artifact that is not packaged in the transformer_lens wheel. Clone the repository \"\n",
     "        \"and run the notebook from within it, or regenerate the artifact with \"\n",
-    "        \"`uv run python -m transformer_lens.tools.analysis.\"\n",
-    "        \"jacobian_lens_causal_swap_benchmark_cli`.\"\n",
+    "        \"`uv run python scripts/jacobian_lens_causal_swap_benchmark.py`.\"\n",
     "    )\n",
     "\n",
     "\n",
+    "def _fingerprint(payload) -> str:\n",
+    "    \"\"\"SHA-256 over canonical JSON, the same recipe the generation script uses.\"\"\"\n",
+    "    canonical = json.dumps(payload, sort_keys=True, separators=(\",\", \":\"))\n",
+    "    return hashlib.sha256(canonical.encode()).hexdigest()\n",
+    "\n",
+    "\n",
     "ARTIFACT_PATH = _find_artifact()\n",
     "REPO_ROOT = ARTIFACT_PATH.parents[2]\n",
-    "artifact = load_artifact(ARTIFACT_PATH)\n",
+    "artifact = json.loads(ARTIFACT_PATH.read_text())\n",
     "manifest = artifact[\"protocol_manifest\"]\n",
+    "# Recomputed rather than trusted, so a hand-edited artifact or one from another schema fails\n",
+    "# here instead of rendering as if the script had produced it.\n",
+    "expected = {\n",
+    "    \"schema_version\": SCHEMA_VERSION,\n",
+    "    \"protocol_fingerprint\": _fingerprint(manifest),\n",
+    "    \"results_fingerprint\": _fingerprint({key: artifact[key] for key in RESULT_BLOCKS}),\n",
+    "}\n",
+    "for field, value in expected.items():\n",
+    "    if artifact[field] != value:\n",
+    "        raise ValueError(f\"artifact {field} is {artifact[field]!r}, expected {value!r}\")\n",
     "print(f\"artifact={ARTIFACT_PATH.relative_to(REPO_ROOT)}\")\n",
     "print(f\"schema_version={artifact['schema_version']}\")\n",
     "print(f\"protocol_fingerprint={artifact['protocol_fingerprint']}\")\n",
@@ -188,10 +207,10 @@
    "id": "1feb4158",
    "metadata": {
     "execution": {
-     "iopub.execute_input": "2026-09-24T12:26:49.110784Z",
-     "iopub.status.busy": "2026-09-24T12:26:49.110543Z",
-     "iopub.status.idle": "2026-09-24T12:26:49.123206Z",
-     "shell.execute_reply": "2026-09-24T12:26:49.122757Z"
+     "iopub.execute_input": "2026-10-14T18:27:58.847684Z",
+     "iopub.status.busy": "2026-10-14T18:27:58.847570Z",
+     "iopub.status.idle": "2026-10-14T18:27:58.860441Z",
+     "shell.execute_reply": "2026-10-14T18:27:58.860108Z"
     }
    },
    "outputs": [
@@ -391,10 +410,10 @@
    "id": "c74c1711",
    "metadata": {
     "execution": {
-     "iopub.execute_input": "2026-09-24T12:26:49.124146Z",
-     "iopub.status.busy": "2026-09-24T12:26:49.124073Z",
-     "iopub.status.idle": "2026-09-24T12:26:49.486825Z",
-     "shell.execute_reply": "2026-09-24T12:26:49.486520Z"
+     "iopub.execute_input": "2026-10-14T18:27:58.861362Z",
+     "iopub.status.busy": "2026-10-14T18:27:58.861303Z",
+     "iopub.status.idle": "2026-10-14T18:27:59.241197Z",
+     "shell.execute_reply": "2026-10-14T18:27:59.240779Z"
     }
    },
    "outputs": [
@@ -516,10 +535,10 @@
    "id": "dedef2ac",
    "metadata": {
     "execution": {
-     "iopub.execute_input": "2026-09-24T12:26:49.488012Z",
-     "iopub.status.busy": "2026-09-24T12:26:49.487879Z",
-     "iopub.status.idle": "2026-09-24T12:26:49.494153Z",
-     "shell.execute_reply": "2026-09-24T12:26:49.493735Z"
+     "iopub.execute_input": "2026-10-14T18:27:59.242250Z",
+     "iopub.status.busy": "2026-10-14T18:27:59.242130Z",
+     "iopub.status.idle": "2026-10-14T18:27:59.248415Z",
+     "shell.execute_reply": "2026-10-14T18:27:59.248094Z"
     }
    },
    "outputs": [
@@ -764,7 +783,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.12.14"
+   "version": "3.12.12"
   }
  },
  "nbformat": 4,

demos/Jacobian_Lens_Demo.ipynb

@@ -334,9 +334,9 @@
     "(pseudoinverse), and write back $h \\leftarrow h + \\alpha\\, V(\\sigma(c) - c)$ where $\\sigma$\n",
     "exchanges the two coordinates — leaving the orthogonal complement of the activation untouched.\n",
     "The paper clamps the exchanged clean-run coordinates at every token position across an\n",
-    "intermediate-layer band (layers 10–24 ≈ the paper's 38–92% workspace band). Unlike\n",
-    "`swap_hooks`, which re-reads and swaps the live coordinates at each layer,\n",
-    "`swap_clamp_hooks` holds every layer at the same clean-run counterfactual target."
+    "intermediate-layer band (layers 10–24 ≈ the paper's 38–92% workspace band). By default\n",
+    "`swap_hooks` re-reads and swaps the live coordinates at each layer; passing the clean cache\n",
+    "as `clamp_to` instead holds every layer at the same clean-run counterfactual target."
    ]
   },
   {
@@ -393,7 +393,7 @@
     "    tokens = model.to_tokens(prompt)\n",
     "    with torch.no_grad():\n",
     "        _, clean_cache = model.run_with_cache(tokens)\n",
-    "        hooks = lens.swap_clamp_hooks(model, source, target, layers, clean_cache)\n",
+    "        hooks = lens.swap_hooks(model, source, target, layers, clamp_to=clean_cache)\n",
     "        with model.hooks(fwd_hooks=hooks):\n",
     "            return model(tokens)[0, -1].float()\n",
     "\n",