QAD: reject muon+TP early

.hf_revision

@@ -0,0 +1 @@
+b4883c6629e39e2c13c3db07fa50355f72db2f68

kit_manifest.json

@@ -1,5 +1,5 @@
 {
-  "built_from_commit": "467e0b4896731cf96d89a8c1a46363013345f7fa",
+  "built_from_commit": "19e193597e8394fdbf41e2d7c45e97d04dc4e1b8",
   "capabilities": [
     "rht_int4_convert.layer_filter",
     "eval_quantized.dtype",
@@ -27,7 +27,7 @@
     "scripts/eval_quantized.py": "5b4e5454630724c116d14fbc64061ffdd6066f4a458e228d372354d3eb32c5b0",
     "scripts/eval_external_model.py": "55e432b6e581ad78e7ea6e292eff25fbaf7a17706f8eebc492a58fc03a51f948",
     "scripts/quantize_speculator.py": "6d9299c63774476d93463ea7b5714f090c31d4cba3fdfc07b561982c78d47afa",
-    "scripts/qad_finetune.py": "d20b15123731c8b5939ddfc00b3cb5c456a783e569c4babeef3a677a4b537244",
+    "scripts/qad_finetune.py": "00c644d03085c826fd0586efc89a6df4e75eee03aba4cad2d0b5789e57861185",
     "scripts/compute_bpc.py": "2b55fd1df406d54b40daeeca2ecaf2efb35981655f5bf5e0cc10c057ce796256",
     "scripts/run_lane.py": "f60869399ab84ae84f654e9773794b813ae1ab52bd031927e46c0ba0650d9b04",
     "scripts/make_sweep_spec.py": "dc31ec026c89e7db316a185575a4423bca988c2ac6cf08ef6df8f955328cda69",
@@ -60,6 +60,7 @@
     "data/eval_en_wikitext2_dedup.txt": "342e8f6a77fb9db7e1a14aa7cabe13c1f13eeaa4f045f8fb3689c0b26c742c0e",
     "data/eval_ja_wikipedia_dedup.txt": "707c9a6c951dd0b831c746a5d6e6972e4dbc71498e75c4aa8f12140eeed89213",
     "lanes/README.md": "44743cd2b9831d0a2249da307c031f70b2822ca71339db7182964e71d8b4a887",
+    "lanes/asym8-wide8-d1e-1-08b.json": "d67020f8a3b98c774c22bf56cccf896c64447debcf63e04f3257a74335af1280",
     "lanes/balanced-h-08b.json": "21a16d50b0b5096843e0a8fb944a6fdbb7314f3c2c51618eb9651a7575192cb8",
     "lanes/balanced-h-d1e-2-08b.json": "418eacf0ad80e5fb29c4d987f682f41fe9645ddcc42d866364a48177c3fb2e48",
     "lanes/bit-alloc-ablation-08b.json": "80cec2d70824daa57dc62dc9caa8073c837930674060e618a8c48066adddc116",
@@ -72,6 +73,8 @@
     "lanes/eval-4b-rtn.json": "301987233bd76481e48a8a157bf7ab14190dd3f3cc88d2a0ee7799141db63f9c",
     "lanes/group-size-ablation-08b.json": "2d9ff0b7ab1c6253d76720f7f30509c1b8578fa81d497c8786927aeb1af33cd2",
     "lanes/q512-probe.json": "ad7b5444a8314a19b0094c6f2adc23b37145f2c50c580ace24ef00422ecf2580",
+    "lanes/rht128-sym8-wide8-08b.json": "fd9ec1430c151e90639ced6ac8299623cde60bc2df35b2024e9c60c80f3b7f9b",
+    "lanes/rht64-sym8-wide8-08b.json": "3ef3ae2f657d0d2a7703d9237476c568997c5f1a92a769340a179d161f083aaa",
     "lanes/robustness-08b-shard1of3.json": "11bd73ddfb110c60420546e1c6f47d777ce6d8c321cf31ce5003d15c96743b9b",
     "lanes/robustness-08b-shard1of5.json": "2a9a72e2f09a6f42c1f90676d24c0bdc009b686e684eac7b622294fec50ec029",
     "lanes/robustness-08b-shard2of3.json": "3196a32cf56eb1e3792a4f15a29effa561ff0472d82282f5e98ad8e283249935",

lanes/asym8-wide8-d1e-1-08b.json

@@ -0,0 +1,48 @@
+{
+  "lane": "asym8-wide8-d1e-1-08b",
+  "model": {
+    "hf_repo": "Qwen/Qwen3.5-0.8B",
+    "spec_name": "Qwen3.5-0.8B",
+    "bf16_dir": "bf16-lalamo",
+    "bf16_source": "/kaggle/input/uzu-bf16-08b-lalamo",
+    "bf16_source_auto": true
+  },
+  "paths": {
+    "kit": "kit",
+    "work": "/kaggle/working",
+    "jax_platform": "gpu"
+  },
+  "eval": {
+    "max_sequences": 100,
+    "max_tokens": 128,
+    "min_tokens": 16,
+    "corpora": {
+      "en": "data/eval_en_wikitext2_dedup.txt",
+      "ja": "data/eval_ja_wikipedia_dedup.txt"
+    }
+  },
+  "condition_defaults": {
+    "samples": 32,
+    "max_tokens": 128,
+    "output_block": "small",
+    "output_block_limit": 2048,
+    "damping": 0.1,
+    "raw_output": false
+  },
+  "convert": {
+    "bits": 4,
+    "group_size": 32,
+    "block_size": 32,
+    "parallel": "tensor",
+    "symmetric_bits": "",
+    "quantize_embedding": true,
+    "embedding_bits": 4,
+    "layer_bits_json": "lanes/bits-wide-8.json"
+  },
+  "conditions": [
+    {
+      "name": "wide8-d1e-1",
+      "calibration_corpus": "data/calibration_mixed_en_ja_dedup.txt"
+    }
+  ]
+}

lanes/rht128-sym8-wide8-08b.json

@@ -0,0 +1,48 @@
+{
+  "lane": "rht128-sym8-wide8-08b",
+  "model": {
+    "hf_repo": "Qwen/Qwen3.5-0.8B",
+    "spec_name": "Qwen3.5-0.8B",
+    "bf16_dir": "bf16-lalamo",
+    "bf16_source": "/kaggle/input/uzu-bf16-08b-lalamo",
+    "bf16_source_auto": true
+  },
+  "paths": {
+    "kit": "kit",
+    "work": "/kaggle/working",
+    "jax_platform": "gpu"
+  },
+  "eval": {
+    "max_sequences": 100,
+    "max_tokens": 128,
+    "min_tokens": 16,
+    "corpora": {
+      "en": "data/eval_en_wikitext2_dedup.txt",
+      "ja": "data/eval_ja_wikipedia_dedup.txt"
+    }
+  },
+  "condition_defaults": {
+    "samples": 32,
+    "max_tokens": 128,
+    "output_block": "small",
+    "output_block_limit": 2048,
+    "damping": 0.1,
+    "raw_output": false
+  },
+  "convert": {
+    "bits": 4,
+    "group_size": 32,
+    "block_size": 128,
+    "parallel": "tensor",
+    "symmetric_bits": "8",
+    "quantize_embedding": true,
+    "embedding_bits": 4,
+    "layer_bits_json": "lanes/bits-wide-8.json"
+  },
+  "conditions": [
+    {
+      "name": "wide8-d1e-1",
+      "calibration_corpus": "data/calibration_mixed_en_ja_dedup.txt"
+    }
+  ]
+}

lanes/rht64-sym8-wide8-08b.json

@@ -0,0 +1,48 @@
+{
+  "lane": "rht64-sym8-wide8-08b",
+  "model": {
+    "hf_repo": "Qwen/Qwen3.5-0.8B",
+    "spec_name": "Qwen3.5-0.8B",
+    "bf16_dir": "bf16-lalamo",
+    "bf16_source": "/kaggle/input/uzu-bf16-08b-lalamo",
+    "bf16_source_auto": true
+  },
+  "paths": {
+    "kit": "kit",
+    "work": "/kaggle/working",
+    "jax_platform": "gpu"
+  },
+  "eval": {
+    "max_sequences": 100,
+    "max_tokens": 128,
+    "min_tokens": 16,
+    "corpora": {
+      "en": "data/eval_en_wikitext2_dedup.txt",
+      "ja": "data/eval_ja_wikipedia_dedup.txt"
+    }
+  },
+  "condition_defaults": {
+    "samples": 32,
+    "max_tokens": 128,
+    "output_block": "small",
+    "output_block_limit": 2048,
+    "damping": 0.1,
+    "raw_output": false
+  },
+  "convert": {
+    "bits": 4,
+    "group_size": 32,
+    "block_size": 64,
+    "parallel": "tensor",
+    "symmetric_bits": "8",
+    "quantize_embedding": true,
+    "embedding_bits": 4,
+    "layer_bits_json": "lanes/bits-wide-8.json"
+  },
+  "conditions": [
+    {
+      "name": "wide8-d1e-1",
+      "calibration_corpus": "data/calibration_mixed_en_ja_dedup.txt"
+    }
+  ]
+}

scripts/qad_finetune.py

@@ -176,6 +176,12 @@ def parse_args() -> argparse.Namespace:
 
 def main() -> None:
     args = parse_args()
+    if args.optimizer == "muon" and args.parallel == "tensor":
+        raise SystemExit(
+            "muon cannot run with tensor parallelism: Newton-Schulz contracts the "
+            "sharded gradient (x @ x.T), and JAX cannot infer the output sharding. "
+            "Use --parallel data or --parallel replicated, or --optimizer adamw."
+        )
     sys.path.insert(0, args.lalamo_path)
     sys.path.insert(0, str(Path(__file__).resolve().parent))