# Codec-v2 checkpoint results

Updated: 2026-09-27T03:10:34.113597+00:00

Target: improve KLD versus current TR3 while retaining the TrellisMX serving format. Build all-K4 B first; build r27-rate A only if B fails the requested quality condition. All encoding and KLD execute on the rented B200 host.

| Layer | Bits | Experts | Summed expert seconds | Gate tw-NMSE | Up tw-NMSE | Down tw-NMSE |
|---:|---:|---:|---:|---:|---:|---:|
| 3 | 4 | 288 | 5824.59 | 0.005542973 | 0.005542593 | 0.005347042 |
| 4 | 4 | 288 | 11446.56 | 0.005543908 | 0.005543676 | 0.005290913 |
| 5 | 4 | 288 | 11896.57 | 0.005522632 | 0.005522479 | 0.005265980 |
| 6 | 4 | 288 | 10429.23 | 0.005497267 | 0.005497356 | 0.005255683 |
| 7 | 4 | 288 | 9804.87 | 0.005501464 | 0.005501868 | 0.005285069 |
| 8 | 4 | 288 | 5994.38 | 0.005484503 | 0.005484412 | 0.005297851 |
| 9 | 4 | 288 | 9624.90 | 0.005480316 | 0.005480006 | 0.005288008 |
| 10 | 4 | 288 | 6244.15 | 0.005507434 | 0.005507520 | 0.005261427 |
| 11 | 4 | 288 | 6516.48 | 0.005487629 | 0.005487624 | 0.005290607 |
| 12 | 4 | 288 | 7092.21 | 0.005496842 | 0.005497110 | 0.005273936 |
| 13 | 4 | 288 | 9749.87 | 0.005463910 | 0.005463459 | 0.005289095 |
| 14 | 4 | 288 | 10701.53 | 0.005467041 | 0.005467035 | 0.005297774 |
| 15 | 4 | 288 | 12302.08 | 0.005430486 | 0.005430460 | 0.005259102 |
| 16 | 4 | 288 | 12564.40 | 0.005451996 | 0.005451735 | 0.005250830 |
| 17 | 4 | 288 | 12640.50 | 0.005452118 | 0.005452184 | 0.005254420 |
| 18 | 4 | 288 | 12647.54 | 0.005451942 | 0.005451998 | 0.005265066 |
| 19 | 4 | 288 | 12651.38 | 0.005438798 | 0.005438836 | 0.005232563 |
| 20 | 4 | 288 | 12636.41 | 0.005454739 | 0.005454832 | 0.005234154 |
| 21 | 4 | 288 | 12683.12 | 0.005435555 | 0.005435415 | 0.005231189 |
| 22 | 4 | 288 | 12626.36 | 0.005433031 | 0.005432813 | 0.005219219 |
| 23 | 4 | 288 | 12672.87 | 0.005420153 | 0.005420313 | 0.005202015 |
| 24 | 4 | 288 | 12650.15 | 0.005405920 | 0.005405919 | 0.005206464 |
| 25 | 4 | 288 | 12637.34 | 0.005417455 | 0.005418153 | 0.005220065 |
| 26 | 4 | 288 | 12638.51 | 0.005398897 | 0.005398934 | 0.005200346 |
| 27 | 4 | 288 | 12681.38 | 0.005430016 | 0.005429790 | 0.005188633 |
| 28 | 4 | 288 | 12665.48 | 0.005405758 | 0.005405619 | 0.005197301 |
| 29 | 4 | 288 | 12687.79 | 0.005392711 | 0.005392629 | 0.005220924 |
| 30 | 4 | 288 | 12651.94 | 0.005389757 | 0.005390055 | 0.005213972 |
| 31 | 4 | 288 | 12637.04 | 0.005391005 | 0.005390967 | 0.005192649 |
| 32 | 4 | 288 | 12650.30 | 0.005377278 | 0.005377707 | 0.005197060 |
| 33 | 4 | 288 | 12624.74 | 0.005400249 | 0.005400456 | 0.005220593 |
| 34 | 4 | 288 | 12705.95 | 0.005388087 | 0.005387681 | 0.005213097 |
| 35 | 4 | 288 | 12641.09 | 0.005370840 | 0.005370381 | 0.005205124 |
| 36 | 4 | 288 | 12627.61 | 0.005372592 | 0.005372620 | 0.005204191 |
| 37 | 4 | 288 | 12640.31 | 0.005370540 | 0.005370885 | 0.005207805 |
| 38 | 4 | 288 | 12654.23 | 0.005384601 | 0.005384321 | 0.005224339 |
| 39 | 4 | 288 | 12698.80 | 0.005385187 | 0.005385474 | 0.005233169 |
| 40 | 4 | 288 | 12670.52 | 0.005376756 | 0.005376675 | 0.005253785 |
| 41 | 4 | 288 | 12661.27 | 0.005375168 | 0.005375150 | 0.005299728 |
| 42 | 4 | 288 | 12682.93 | 0.005381312 | 0.005381636 | 0.005354900 |
| 43 | 4 | 288 | 12646.21 | 0.005376763 | 0.005376823 | 0.005316918 |
| 44 | 4 | 288 | 12306.51 | 0.005394467 | 0.005394839 | 0.005259895 |

Summed expert seconds include concurrent work and are not elapsed wall time. tw-NMSE is transformed-weight error, not model KLD.

## Checkpoint B

```json
{
 "checkpoint": "/root/cv2data/out/ckpt-B",
 "runtime_overlay_and_kernel_header_checks": "PASS (168 files)",
 "expert_stream_mismatches": {
  "3": 0,
  "4": 0,
  "5": 0,
  "6": 0,
  "7": 0,
  "8": 0,
  "9": 0,
  "10": 0,
  "11": 0,
  "12": 0,
  "13": 0,
  "14": 0,
  "15": 0,
  "16": 0,
  "17": 0,
  "18": 0,
  "19": 0,
  "20": 0,
  "21": 0,
  "22": 0,
  "23": 0,
  "24": 0,
  "25": 0,
  "26": 0,
  "27": 0,
  "28": 0,
  "29": 0,
  "30": 0,
  "31": 0,
  "32": 0,
  "33": 0,
  "34": 0,
  "35": 0,
  "36": 0,
  "37": 0,
  "38": 0,
  "39": 0,
  "40": 0,
  "41": 0,
  "42": 0,
  "43": 0,
  "44": 0
 },
 "routed_sidecar_bytes": 161867586048,
 "tp2_routed_bytes_per_gpu": 80933793024.0,
 "allocation_counts": {
  "3": 0,
  "4": 42,
  "5": 0
 },
 "status": "PASS"
}
```

```json
{
  "checkpoint": "/root/cv2data/out/ckpt-B",
  "windows": 128,
  "mean_kld": 0.027881131802743875,
  "backend": "Transformers 5.17 GLM5Next BF16 backbone; decoded coupled P8 with two-term E4M3 reference carriers",
  "mode": "p8",
  "execution": "B200 full-window causal forward, no KV quantization; rows 1..2046 scored",
  "native_sm120_true_decode_equivalence": false,
  "historical_tr3_fp8_true_decode_kld": 0.030099944949203105,
  "historical_r27_rp2_fp8_true_decode_kld": 0.03260408965478527,
  "comparison_limit": "Historical scores use a different backbone, KV and execution path; this result alone does not establish a production TR3 win or runtime speed.",
  "checkpoint_manifest_sha256": "33c0817e60dc44c37ee34ca9efad5af01b6c5116395ad15c758ae332e6fcf41c"
}
```

Variant A was not required because B met the requested numeric KLD threshold.

Historical same-panel true-decode means: r27+RP2 FP8 KV 0.03260408965478527; TR3 FP8 KV 0.030099944949203105.

The B200 reference path uses Transformers 5.17, BF16 non-routed weights, decoded P8 physical expert weights, coupled scale/rotation math and two-term E4M3 carriers. It uses complete-window causal forward without FP8 KV. A reference result alone does not prove equality to native SM120 true-decode KLD or retained serving speed.

No local GPU measurement or behavioral benchmark was run.

## Complete native package

Selected checkpoint directory: 161,882,683,668 bytes. MTP Trellis sidecars: 3,853,990,880 bytes. Native carrier assets and filtered index: 19,388,363,261 bytes. Combined before additional runtime/publication metadata: 185,125,037,809 bytes (185.125038 GB; 172.411127 GiB). The upload receipt records the final repository byte count including evidence.

Native carrier: local-inference-lab/GLM-5.3-Flash-NVFP4@520de24eabf507659eaef7c70f14fd584527facc. Source hash verification: PASS; index/header coverage: PASS. The carrier is bundled under carrier/ and retains its original tensor data and configuration. The tensor index omits only routed weights replaced by Trellis sidecars.

Mount the complete package at /model, set MODEL_ROOT=/model/carrier and VLLM_TRELLISMX_CHECKPOINT=/model in the existing native RP2 image. The packaged native carrier differs from the BF16 non-routed reference backbone. The B200 score is not a native-serving KLD measurement of the complete package.

## Trellis MTP45

```json
{
  "status": "PASS",
  "layer": 45,
  "experts": 288,
  "sidecars": 4,
  "expert_stream_mismatches": 0,
  "codec_closure": "PASS: all 864 projection encodes",
  "coupled_component_validation": "PASS",
  "routed_sidecar_bytes": 3853990880,
  "mean_tw_nmse": {
    "gate_tw_nmse": 0.005752041559620415,
    "up_tw_nmse": 0.005751395422463146,
    "down_tw_nmse": 0.0061286417262067724
  },
  "scale_policy": "signed-unit coupled vectors; CPU torch seed 530045; no prior r27 MTP scales",
  "native_execution_tested": false
}
```

MTP uses the supplied codec-v2 math with its own BF16 captures: exactly 64 fit windows, 2047 valid NextN rows each. There are no r27 MTP EXL3 vectors, so this new extension uses deterministic signed-unit coupled scale vectors (CPU seed 530045). Main scale vectors are unchanged. All 288 MTP experts use K4. The existing MTP FP8 expert tensors and scales are replaced; all 25 non-expert MTP tensors are preserved bit for bit.

Use the bundled runtime/serve-codecv2-mtp.sh launcher inside the pinned RP2 image, or build runtime/Dockerfile. It adds MTP45 recognition and loading without changing main kernels. The patch installed successfully in a local Docker build; no native serving or local GPU KLD ran. Target-model KLD does not measure MTP draft quality or speculative acceptance.

## Checkpoint selection

```json
{
  "selected_variant": "B",
  "reference_kld": {
    "B": 0.027881131802743875
  },
  "historical_tr3_native_kld": 0.030099944949203105,
  "below_recorded_tr3_threshold": true,
  "matched_runtime_comparison": false,
  "comparison_note": "Previously recorded TR3 uses native FP8-KV true decode; new measurements use the B200 reference path. Execution paths differ.",
  "native_serving_speed_measured": false,
  "native_sm120_kld_equivalence_established": false,
  "second_variant_built": false,
  "completed_at": "2026-09-27T03:10:31.267440+00:00"
}
```