Land the MeroMero v2-31B NVFP4A16 quant and record the pinned-transformers trap

The v2 dense quant had failed four times. Attempt 5 lands it at 19 G.

The blocker was not what it looked like. `AmbiguousGlobalPerLayerAttributeError`
on `head_dim` read as a malformed upload -- DogOnKeyboard's config carries a
`per_layer_config` key zerofata's canonical one lacks -- and the standing fix was
to force `allow_global_per_layer_attribute_access=True`. Both halves were wrong.

`pip install llmcompressor==0.13.0` downgrades transformers 5.16.1 -> 5.14.1. The
config was serialized by 5.16.1, which materializes `per_layer_config` from
`global_head_dim` + `layer_types`; 5.14.1 carries the heterogeneity guard but not
the gemma4 resolver. Under the image's own transformers the same config loads
fine. `:latest` was also re-pulled during attempt 4 and no earlier run, so the
toolchain moved mid-diagnosis. Two things separated "malformed upload" from
"moved toolchain": reproducing the real failing call (a bare AutoConfig load does
not reproduce it; the trigger is reached through AutoTokenizer) and keeping
zerofata's canonical tree, quantized cleanly on 2026-08-21, as a positive control.

The fix drops `per_layer_config` rather than forcing global access. It is exactly
redundant -- keys are precisely the ten full_attention layer indices, sole value
(512, 4), verbatim the global fields -- and forcing instead would make
`config.head_dim` answer 256 to the callers building the 512-wide layers.
patch_perlayer.py re-proves that redundancy at apply time and refuses if it ever
stops holding.

Verified on the tensor table rather than the exit code: the output is identical
family-for-family and count-for-count to the August canonical quant, with 356
BF16 vision-tower tensors preserved and input_activations=None. A GPU-free load
leaves 0 tensors on meta and generates coherent prose. The section 4.4 serve test
has NOT run -- GPU1 has 19.9 GB free against 19.5 GB of weights, so it needs a
live seat displaced.

Also fixes the A4B output, which had a truncation cap baked into its tokenizer
(max_length 8192) from being quantized with the calibration corpus.

Playbook gains section 3.17 for the pinned-transformers class and sharpens 3.16
to say drop the dataset outright for any A16 scheme.

services/meromero-quant/patch_perlayer.py

@@ -0,0 +1,74 @@
+"""Remove the redundant `per_layer_config` block that blocks attempt 4.
+
+Why this and not `allow_global_per_layer_attribute_access=True`:
+
+  * `per_layer_config` here carries NO information. Its keys are exactly the ten
+    full_attention layer indices [5,11,...,59] and its only distinct value is
+    (head_dim 512, num_key_value_heads 4) -- which `global_head_dim: 512` and
+    `num_global_key_value_heads: 4`, already in this config, say verbatim.
+    Removing it is lossless, and it is the difference in what the two toolchain
+    versions can read: transformers 5.16.1 (which authored this file) emits the
+    per-layer form; llmcompressor 0.13.0 PINS transformers to 5.14.1, which has
+    the heterogeneity guard but not the gemma4 resolver, so it refuses the read.
+  * Forcing global access leaves the config heterogeneous and makes
+    `config.head_dim` answer 256 to every caller -- including the ones building
+    the 512-wide full-attention layers. Geometry survived that in my meta-device
+    check, but llmcompressor's own onloading code is a caller I have not audited,
+    and it is precisely what transformers' warning is about. No reason to take
+    that when the lossless option verifies identically.
+
+Verified on transformers 5.14.1, the version the quant actually runs: this config
+then reports head_dim/num_key_value_heads/global_head_dim/num_global_key_value_heads
+identical to zerofata's canonical tree -- the config that quantized successfully on
+2026-08-21 -- and builds k_proj (2048, 5376) on layer 5 and (4096, 5376) on layer 0,
+matching the checkpoint.
+"""
+import json
+import shutil
+from pathlib import Path
+
+CFG = Path("/tank/aimodels/G4-MeroMero-v2-31B-heretic-bf16/config.json")
+BAK = CFG.with_name("config.json.bak-pre-perlayer-20260910")
+
+NOTE = (
+    " | infra-ops 2026-09-10 (2nd patch): removed text_config.per_layer_config, "
+    "a transformers-5.16.1 serialization artifact that llmcompressor 0.13.0's "
+    "pinned transformers 5.14.1 cannot read (AmbiguousGlobalPerLayerAttributeError "
+    "on head_dim). It was exactly redundant with global_head_dim=512 / "
+    "num_global_key_value_heads=4 -- keys were the 10 full_attention layers, sole "
+    "value (512, 4). Config now matches zerofata's canonical shape. Original at "
+    "config.json.bak-pre-perlayer-20260910."
+)
+
+cfg = json.loads(CFG.read_text())
+t = cfg["text_config"]
+plc = t.get("per_layer_config")
+
+if plc is None:
+    print("per_layer_config already absent -- nothing to do")
+    raise SystemExit(0)
+
+# Re-prove the redundancy here rather than trusting the earlier session: a patch
+# that silences an error on a config it did not actually verify is how a quietly
+# wrong quant ships.
+full = {i for i, x in enumerate(t["layer_types"]) if x == "full_attention"}
+assert {int(k) for k in plc} == full, f"per_layer_config keys {sorted(plc)} != full-attn layers {sorted(full)}"
+vals = {(v["head_dim"], v["num_key_value_heads"]) for v in plc.values()}
+assert vals == {(t["global_head_dim"], t["num_global_key_value_heads"])}, \
+    f"per_layer_config carries {vals}, not the global (512, 4) -- NOT redundant, do not drop"
+print(f"redundancy re-verified: {len(plc)} entries, all {vals.pop()}, "
+      f"== (global_head_dim, num_global_key_value_heads)")
+
+if not BAK.exists():
+    shutil.copy2(CFG, BAK)
+    print(f"backed up -> {BAK.name}")
+else:
+    print(f"backup {BAK.name} already exists, left alone")
+
+t.pop("per_layer_config")
+cfg["_patched_by"] = cfg.get("_patched_by", "") + NOTE
+
+tmp = CFG.with_suffix(".json.tmp")
+tmp.write_text(json.dumps(cfg, indent=2) + "\n")
+tmp.replace(CFG)
+print(f"patched {CFG}")