configs: fix I-Nano/I-Micro NULL output on Qwen3.6 MTP variants (missing nextn.eh_proj override)

[bfox55]

Summary

Every *_mtp_nano.txt and *_mtp_micro.txt config in the Qwen3.6 family is missing a per-tensor override for blk.40.nextn.eh_proj — the MTP head's embed→hidden projection (16 MB bf16). Without an explicit override, the tensor falls through to the base type (iq2_xxs for nano, iq1_m for micro), both of which guard against very-low-bit quantization with no imatrix data. llama-imatrix only forward-passes through the trunk and never activates the MTP head, so this tensor has no calibration data — guard trips, llama-quantize exits before writing the GGUF header, output is a NULL-header file of ~expected size.

This matches the file signature on the published I-Nano artifacts in mudler/Qwen3.6-...-Distilled-APEX-MTP-GGUF (and the 4.7 sibling repo): all-zero first 32 bytes, file size ~10.88 GB, etags 28b27ae3..., 280e4530..., b945a4f2..., 41f1719f.... If reproducing locally, the relevant llama-quantize error is:

Missing importance matrix for tensor blk.40.nextn.eh_proj.weight in a very low-bit quantization
The result will be garbage, so bailing out
llama_model_quantize: failed to quantize

Fix

One line per affected config:

blk.40.nextn.eh_proj=Q4_K

Q4_K matches the edge-tier precision already used for blk.40.attn_* in the same configs.

The three other MTP-specific tensors (blk.40.nextn.enorm, blk.40.nextn.hnorm, blk.40.nextn.shared_head_norm) are F32 norms — they pass through untouched and need no override.

Test plan

• Patched config produces a valid 10.88 GB GGUF for Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled with the patched qwen36_opus_distill_mtp_nano.txt
• First 32 bytes start with GGUF magic (vs all-NULL on broken upload)
• Loads and self-speculates under llama-server --draft-mtp (in progress on my machine — will append results)
• Re-quantize one of the existing broken HF uploads with the patched config and confirm valid output

Scope

I've only patched the 10 Qwen3.6-family *_mtp_nano.txt / *_mtp_micro.txt configs because that's the architecture I verified end-to-end. Other MTP-bearing families (if any) likely have the same issue but may use different MTP block indices — happy to extend if you point me at the architecture details.

Generator note

scripts/generate_config.sh doesn't currently handle MTP at all — these _mtp_ configs look like they're produced by an out-of-tree post-process. Whatever produces them should also be updated so regenerated configs don't regress to the broken state. Not in scope for this PR.

[bfox55]

Following up — built and benched the patched I-Nano locally, fix confirmed working end-to-end.

Build (with this PR's patched qwen36_opus_distill_mtp_nano.txt):

quant size  = 11134.77 MiB (2.63 BPW)
magic bytes = 47475546 (GGUF)  ← valid header, not NULL
753/753 tensors processed cleanly

Pre-patch (with the missing blk.40.nextn.eh_proj override) — same NULL-header 00000000 00000000 ... failure as the public I-Nano uploads in your HF repo. With the one-line patch added, the build completes cleanly and loads in llama.cpp.

Inference validation (llama.cpp ad27757, RTX 5060 Ti 16GB, -c 131072 --cache-type-k q8_0 --cache-type-v q8_0 --spec-type draft-mtp --spec-draft-n-max 2):

• Decode: 139 t/s short prompts, 106 t/s at 16K-depth (thinking-on)
• MTP self-spec acceptance: ~56% on reasoning-mode content, ~90% on structured summary content — head functions correctly across both
• Probe-spark behavioral suite: 100% (19/19) with thinking on
• Spark-suite (Brett's 34-test agent bench): 91% overall vs reference Qwen3.6-35B-Opus-Distill UD-IQ2_M at 97% — tied or above on finance/email/tool-select/instruction; 67% on multi-step reasoning is where the distill-of-distill loss shows (your I-Compact / I-Balanced should be cleaner here since they have more bits to spend on the trunk).

Pipeline notes from the run (in case useful for the repo README):

1. F16 → Q8_0 intermediate before llama-imatrix is the key for ≤32 GB RAM hosts; activation stats are PPL-equivalent per your own published table (F16 6.537 vs Q8_0 6.533) so no fidelity cost. Imatrix from Q8_0 + auto-fit (no -ngl) completed in 16 min on partial offload — model load was 6 min from HDD, then 32 chunks × ~22s.
2. The blk.40.nextn.eh_proj bf16→Q4_K override added by this PR is the only difference from a vanilla llama-quantize --tensor-type-file CONFIG --imatrix IMAT INPUT OUTPUT iq2_xxs invocation. Repro is iq2_xxs base with config + imatrix, nothing exotic.

I-Nano is now running as my production agent (DIY-Nano-as-Spark) at 128K context with no crashes through ~25K-token sessions. Net: this fix is correct, the file is shippable, please consider merging when you have a moment. Happy to share the GGUF or the bench harness if helpful.