Rebuild: embedding bits wiring

kit_manifest.json

@@ -1,5 +1,5 @@
 {
-  "built_from_commit": "80ac5308c962f95f8abbe401f4cb3f29a393ada4",
+  "built_from_commit": "dc01b3d22a9a1b632918d8594eedc40a8819217b",
   "capabilities": [
     "rht_int4_convert.layer_filter",
     "eval_quantized.dtype",
@@ -27,7 +27,7 @@
     "scripts/eval_quantized.py": "5b4e5454630724c116d14fbc64061ffdd6066f4a458e228d372354d3eb32c5b0",
     "scripts/qad_finetune.py": "275f3630ef919ce9a762eb96528640ae12d1e03da6d3dec1446666e0986ddf0d",
     "scripts/compute_bpc.py": "2b55fd1df406d54b40daeeca2ecaf2efb35981655f5bf5e0cc10c057ce796256",
-    "scripts/run_lane.py": "eef3be44e5d711995b6993e3e9ff36aa6007b593b44357041e08d6fa4c0d024e",
+    "scripts/run_lane.py": "dc437d0bc41e2eeca3fc51218551b2e7175c6cc8bfc6cb43a61edaa55fbf72cc",
     "scripts/make_sweep_spec.py": "dc31ec026c89e7db316a185575a4423bca988c2ac6cf08ef6df8f955328cda69",
     "scripts/make_smoke_specs.py": "a76091d7a6a38427cdde5264a49d7bb4d8bfcc75284705817dd4c83b012bba2e",
     "scripts/combine_lane_manifests.py": "164ba691d60b3e2987b805426c606d135dcf3cf0c855b38cf7f5d6d57628a481",
@@ -63,6 +63,7 @@
     "lanes/bit-alloc-ablation-08b.json": "80cec2d70824daa57dc62dc9caa8073c837930674060e618a8c48066adddc116",
     "lanes/bits-down-out-6.json": "8c37658ad99820640d638a266db977d799e6241233901f0a2fb632052388e15f",
     "lanes/bits-down-out-8.json": "8bb34f87bc8041dcc13c778107cb230d16563beebc2dee3a50776bc357ac4d5d",
+    "lanes/bits-wide-8.json": "198905ccfeb04b550e6b2111bdc0fdece8af6b94f529c7d66d8552eea3ce20c7",
     "lanes/clean-suite-08b.json": "a1b74a5f2b7a36edc3fe060ec071df27bcbafe7975fd46ba3562b77458d94b99",
     "lanes/cpu-tuning-08b.json": "9a986eef3aa1f47bd8e01af5f16e9757156cc447907e7bb29bdf98ca78e02020",
     "lanes/damping-ablation-08b.json": "76f79c8d4db5c38935edf1c5abd3ec8fadad89ffa6c41520fce0f7252795a502",
@@ -94,7 +95,7 @@
     "lanes/smoke-seed-variance-08b.json": "f6acb2f2d3d3e0213a44013fecab6fac2050a477d28a563a3d307defe49f0fa2",
     "tests/test_experiment_guards.py": "d71ddf9586c1d7ece52e217288b9d4b09079be54d32ceaf391be113c1832b6ef",
     "tests/test_compute_bpc.py": "bba5b6d0afc9b0699ebb6f7f0ed9fc5ebfcb98472867e9ccff4262fe9999949f",
-    "tests/test_run_lane.py": "c62403924d9a6b2b0be55eaee6925f231a07a12ddbfcd8f97f63a0a1f5f8c4a9",
+    "tests/test_run_lane.py": "0043b2fff1c972612055f89990d15bec10cdf947d74652034d65c267ebaabda8",
     "tests/test_tiny_sharding.py": "eaed062b5547ed66923d567753494fe2c26d15c7df4a160601182aacdbf0ceb1",
     "configs/qwen35-0.8b.json": "12640a7b738d8c8dbc046f59e9249b85d5d3775b8ee2afe9becaba285ae90764",
     "configs/qwen35-4b.json": "8f32da42655963134c1f73f4fa825c8ae16b464254d4fd9ddccf6b2f8ecae833",

lanes/bits-wide-8.json

@@ -0,0 +1,21 @@
+{
+  "default": 4,
+  "rules": [
+    {
+      "match": "mlp/down_projection",
+      "bits": 8
+    },
+    {
+      "match": "mixer/out_proj",
+      "bits": 8
+    },
+    {
+      "match": "mixer/qkvg_projection",
+      "bits": 8
+    },
+    {
+      "match": "mixer/out_projection",
+      "bits": 8
+    }
+  ]
+}

scripts/run_lane.py

@@ -484,9 +484,11 @@ class Runner:
             ]
         convert_settings = dict(self.spec.get("convert", {}))
         convert_settings.update(condition.get("convert") or {})
-        for key in ("bits", "group_size", "block_size", "parallel"):
+        for key in ("bits", "group_size", "block_size", "parallel", "embedding_bits"):
             if key in convert_settings:
                 convert_command += [f"--{key.replace('_', '-')}", str(convert_settings[key])]
+        if convert_settings.get("quantize_embedding"):
+            convert_command.append("--quantize-embedding")
         if convert_settings.get("layer_filter"):
             convert_command += ["--layer-filter", str(convert_settings["layer_filter"])]
         if condition.get("layer_bits_json"):
@@ -574,9 +576,11 @@ class Runner:
                 "--calibration-json",
                 str(out / "calibration.json"),
             ]
-            for key in ("bits", "group_size", "block_size", "parallel"):
+            for key in ("bits", "group_size", "block_size", "parallel", "embedding_bits"):
                 if key in self.spec.get("convert", {}):
                     convert_command += [f"--{key.replace('_', '-')}", str(self.spec["convert"][key])]
+            if self.spec.get("convert", {}).get("quantize_embedding"):
+                convert_command.append("--quantize-embedding")
             self.run(f"{name} convert", convert_command, out / "lane.log")
             self.run(
                 f"{name} eval copy",

tests/test_run_lane.py

@@ -244,3 +244,17 @@ def test_exclude_cannot_remove_every_condition() -> None:
     result = run_lane(LANES / "robustness-08b-shard3of5.json", "--exclude", "s")
     assert result.returncode != 0
     assert "removed every condition" in result.stdout + result.stderr
+
+
+def test_embedding_flags_reach_the_convert_command(tmp_path: Path) -> None:
+    spec = json.loads((LANES / "cpu-tuning-08b.json").read_text(encoding="utf-8"))
+    spec["convert"]["embedding_bits"] = 8
+    spec["convert"]["quantize_embedding"] = True
+    spec["conditions"] = spec["conditions"][:1]
+    path = tmp_path / "spec.json"
+    path.write_text(json.dumps(spec), encoding="utf-8")
+    result = run_lane(path, "--exclude", "rtn")
+    assert result.returncode == 0, result.stdout + result.stderr
+    convert = [line for line in result.stdout.splitlines() if line.startswith("$ ") and "rht_int4_convert.py" in line]
+    assert convert
+    assert any("--embedding-bits 8" in line and "--quantize-embedding" in line for line in convert)