Mamb2_8B_FP4_Recall

Full official WikiText-2 TEST PPL: 6.717746660814748 with Resurface, versus 6.717800294314242 without it. Normal CONFIRM MK: 353/384 (91.9270833%), versus 131/384 (34.1145833%).

A member of the Mamb2_8B_Recall family: a pure Mamba2-8B base with FP4 G16 weights, repaired existing FP16 parameters, native FP16 recurrent state, and a fresh Resurface adapter. There is no StateQuant controller in this model.

Audited paired quality

Same-process configuration	Full WT2 test PPL	Normal MK / 384	Target-removed MK / 384
Repaired FP4 G16 base, no adapter	6.717800294314242	131 (34.1145833%)	0
Same frozen base + final Resurface	6.717746660814748	353 (91.9270833%)	0
Same base after adapter removal	6.717800294314242	131 (34.1145833%)	0

Normal MK population	Adapter off	Final Resurface on
N16, 192 cases	101/192 (52.6041667%)	191/192 (99.4791667%)
N64, 192 cases	30/192 (15.625%)	162/192 (84.375%)

The normal-MK gain is 222/384 = 57.8125 percentage points, with a paired bootstrap 95% interval of 52.6041667 to 62.7604167 percentage points (10,000 draws, seed 20260928). There are 225 improvements, 3 regressions and 156 unchanged normal pairs. Target-removed matches are 0/192 for each N in both configurations. Measured PPL changes by only-0.0007983788910648215%; the predeclared no-worse PPL gate passes. All147 window scores and 768 generated sequences, native hooks/forwards and reset/cache probes restore exactly after adapter removal. Both combined CPU audits pass; they check recorded evidence and arithmetic without replaying GPU logits.

Protocol: official Salesforce/wikitext, wikitext-2-raw-v1, test, revision b08601e04326c79dfdd32d625aee71d232d685c3: 147 reset windows / 300,963 targets, final1,955-target partial window included. PPL uses native parallel SSD prefill and pooled token NLL, full vocabulary, FP32 cross entropy in 64-token chunks, no automatic BOS/EOS and no persistent request cache. Recall uses 768 CONFIRM prompts (384 normal +384 target removed), greedy at most12 tokens with EOS stop and first standalone six-digit answer matching.

Data scope: WT2 validation/test and the CONFIRM family have prior project exposure. These are not untouched project holdouts. Base repair used TRAIN and validation selection; the fresh adapter was TRAIN-only and its1536-step candidate was fixed before evaluation. Independent report audits do not make the evaluation data independent. Pretraining contamination is unaudited.

Storage and native reference runtime

Accounting scope	Exact bytes
Base encoded weight tensor payload	4,638,460,360
Actual119 base safetensors files	4,638,539,864
Adapter FP16 tensor payload, 1,154,104 parameters	2,308,208
Serialized final adapter file	2,381,119
Actual119 weight files + adapter file	4,640,920,983
Decoded resident FP16 base weights	16,473,999,360
Batch1 native MK SSM + convolution caches	122,028,032
Additional adapter recurrent cache	0

The combined file total excludes metadata, tokenizer, source, licenses and evidence. The complete download is therefore larger. The native quality and inference implementation expands packed weights to resident FP16. It needs activations, request buffers, logits and allocator reserve beyond the table. The stored4.641GB core files do not establish4.641GB VRAM inference. Batch1 MK inspects112 distinct FP16 cache tensors across56 layers:117,440,512 SSM bytes +4,587,520 convolution bytes. PPL follows the native SSD scan path; its internal accumulator precision is not a claim that every intermediate is FP16. No packed-resident FP4 kernel, measured throughput benefit or ASIC implementation is provided.

What was trained

The base has 56 blocks, 8,236,999,680 parameters and 507 tensors. Its114 large matrices use E2M1 FP4 codes in 16-element input-axis groups, E4M3FN block scales and FP32 global scales. Independent weight-only conversion retains393 small FP16 tensors. This is an NVFP4-style representation with a custom range search, not a claim of hardware-native NVFP4 execution.

Base-only repair then trained those 393 existing FP16 tensors (3,580,928 parameters), leaving all 114 FP4 matrices and their118 shard files unchanged. Two fixed passes over 448 WT2 TRAIN windows give896 successful updates. The fixed224/448/672/896 pool selected step 448 by minimum full validation NLL. Its earlier full test PPL was6.717854132348465; the current paired control is 6.717800294314242. The original FP16 source used in base repair was an unfinetuned reference, not a matched post-training control.

For this fresh post-D Resurface run, all 114 FP4 matrices and 393 repaired FP16 tensors were frozen. Only224 adapter tensors /1,154,104 parameters were trained, starting from V=0, g=1, router_w=0, router_b=-4. Numeric TRAIN CE1, WT2 TRAIN prose CE0.5, separate unadapted repaired-base teacher KL0.5 and prose gate closure3 used native SSD forward/backward. Formal training completed 1536 successful updates /1542 attempts /6 overflow retries. Only the final 1536-step FP16 export was a quality candidate; no PPL/MK checkpoint selection was used. Export/readback, native128/512 parity and frozen507 identities pass. Adapters from other family variants are not interchangeable.

Download and run

The complete bundle lives under release/; it does not require the original NVIDIA checkpoint. Download the pinned snapshot:

hf download EndlessChasing/Mamb2_8B_FP4_Recall \
  --revision v0.1.0-fp4g16-repaired-s16-resurface --local-dir Mamb2_8B_FP4_Recall
cd Mamb2_8B_FP4_Recall/release

python scripts/infer_fp4_repaired_s16_resurface_v1.py --bundle . --verify-only
python scripts/infer_fp4_repaired_s16_resurface_v1.py --bundle . --smoke-check
python scripts/infer_fp4_repaired_s16_resurface_v1.py --bundle . \
  --prompt 'The capital of France is' --max-new-tokens 64

--verify-only is standard-library CPU verification. --smoke-check strictly replays one published normal recall sample, not the full benchmark. --without-adapter with a prompt runs the same repaired base alone. The custom source loads directly; there is no AutoModel.from_pretrained support. Generation needs the pinned CUDA/Mamba/Triton environment and resident decoded FP16 weights. Do not install the bundle itself with pip or create a venv inside it.

See minimal reproduction for pinned dependencies, the GitHub asset assembly and optional full three-arm replay. The same files are available from GitHub Release.

Identity and license scope

The modified base is derived from the Apache-2.0 NVIDIA pure Mamba2 checkpoint. The modified base weights are Apache-2.0; inherited/project runtime and this new Resurface adapter are GPL-3.0. The preserved StateQuant source reference retains Apache-2.0, but no StateQuant execution or state configuration is used. No Quamba quantized checkpoint or research-only weight package is used.

• Repaired packed manifest: 2d2208b690c2ec0def90578d8590cc122bb1e4542b0d49ae4417a48f72b0b73c.
• Final adapter: dca3cda5a9ab8a665abd1a355c6e5b294acee023d2a1a5a8e20f4e84c9674580.
• Paired full comparison: f16227b149e1aa9c7113a35cd56ed1f95849ae6bdc6b7e1460d077a4847d91cb.
• Independent Mac combined audit: f3600b884bb60268dd8bf27c9651fe21cb8b93daac3046011b9e007577e15dae.

The base weight and adapter licenses apply to their identified artifacts; preserve their notices and included source when redistributing. These independent modifications do not imply endorsement by the original authors.

See portable full results, LICENSES.md, third-party notices and exact evidence.