lued

Qwen3.8-27B-redashes-SSMFix-INT8-W8A16-MTP

README

License: apache-2.0

Relationship to the sibling quantization

This checkpoint is the conv1d-repaired counterpart of lued/Qwen3.8-27B-INT8-W8A16-MTP. The repositories share:

the identical W8A16 compressed-tensors recipe (400 packed GEMMs, same BF16 preservation set, group-128 symmetric data-free RTN);
the identical native BF16 MTP head (15 tensors, byte-identical to Qwen's original MTP — same sha256 as the sibling release);
the identical tokenizer, chat template, processor configs, and bounded-shard layout;
the same total checkpoint footprint (29.44 GiB).

The only deviation is the base language-model weights: this checkpoint's LM weights come from redashes's conv1d-repaired BF16 model instead of stock Qwen3.8-27B.

SSM scale-drift repair (inherited from the base)

Per the upstream card, the source applies per-layer α-scaling to the anomalous linear_attn.conv1d.weight tensors in Qwen3.8-27B, following the methodology first disclosed by LuffyTheFox (Sig-ScaleSync) and independently re-implemented by redashes: 8 layers (52/53/56/57/58/60/61/62) rescaled with factors 0.481–0.653. Upstream-controlled evaluation showed the patch moving generative metrics (TruthfulQA-gen +6~8pp, MT-Bench +0.42) while slightly reducing closed-book knowledge (CMMLU −1.8pp). These are the upstream card's claims, not measurements made here. The W8A16 step touches only stored numerics of the 400 GEMMs; the conv1d convolution weights are preserved at BF16, so the repair is bit-exact in this release.

Quantization fidelity

Measured on this exact checkpoint against its own BF16 teacher (the redashes source) on 2026-08-18 with tools/measure_kld.py (teacher-forced, full-vocabulary FP32 log-softmax, FP64 accumulation, nats/token; PyTorch 2.11.0+cu128, Transformers 5.10.1, compressed-tensors 0.17.1):

Table with columns: Candidate, Mean KLD vs BF16, Top-1 agreement, Weight files
Candidate	Mean KLD vs BF16	Top-1 agreement	Weight files
This W8 / BF16 A16	0.000728	98.72%	29.44 GiB

The KLD result measures checkpoint-weight drift over 467 teacher-forced positions (fixed 8-prompt short suite). It is not a functional quality score; tool use, coding, multimodal quality, and long-context recall require separate behavioral evaluation.

Short-prompt suite

Table with columns: Prompt, Category, Tokens, Positions, Mean KLD, Top-1 agreement
Prompt	Category	Tokens	Positions	Mean KLD	Top-1 agreement
`factual`	factual	13	12	0.000603	1.0000
`code-fib`	code	87	86	0.000535	0.9884

kld_result.json in this repository is the single source of truth for these numbers.

Checkpoint profile

Table with columns: Property, Value
Property	Value
Quantization	Data-free symmetric RTN W8A16, group size 128
Runtime format	`compressed-tensors` / `pack-quantized`
Kernel dispatch	`CompressedTensorsWNA16` → `MarlinLinearKernel` (verified in server logs)
Quantized GEMMs	400: 192 MLP, 64 full-attention, 144 dense GDN projections
Preserved precision	BF16 vision tower, `lm_head`, MTP (15 tensors, byte-identical to Qwen's original MTP), recurrent GDN gates (96), and the `conv1d` SSM-repair weights

MTP preservation

The MTP head is the exact BF16 MTP from the original Qwen3.8-27B — never touched by the SSM repair and never re-quantized:

15 tensors in model_mtp.safetensors (849,400,392 bytes).
sha256 1d8268aa85ace093a561e3e7b63b9d390dac1cd55a90cd55b5ec509c3c9da9fe — identical to the MTP bytes in the Qwen3.8-27B source, in the AEON source, and in the sibling release lued/Qwen3.8-27B-INT8-W8A16-MTP.
All 15 tensors verified bfloat16, shape-equal, and bit-equal to the redashes source (strict audit, not a count check).

Reproduction contract

The immutable source is redashes/Qwen3.8-27B-BF16-SSMFIX commit 859730a4ae4d6ed793d0f427aa95d49135334bda. The exact selective W8A16 configuration is recorded in recipe.yaml and config.json. A conforming rebuild must take native MTP from that source, reproduce the source-derived packed projection set, and prove all 15 MTP tensors bit-identical before resharing for publication.

Status

Built on 2026-08-18: both release audits PASS, KLD measured, published to lued/Qwen3.8-27B-redashes-SSMFix-INT8-W8A16-MTP at commit e4ff78c49d2717023aa9965fd1aa85be2387407d, and registered in the HF cache (link_hf_cache.py sync PASS). Serving validated the same day: boots through llama-swap from the lued/* cache entry, loads the CompressedTensorsWNA16 → MarlinLinearKernel path, detects the native MTP draft model (method=mtp, 3 speculative tokens), and serves a smoke request. The BF16 parent source remains cached; deletion is the pipeline's final step and has not been requested.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

lued

Model Tree

Base

redashes/Qwen3.8-27B-BF16-SSMFIX

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality