lued

Qwen3.8-27B-heretic-ara-INT8-W8A16-MTP

README

License: apache-2.0

Relationship to the sibling quantization

This checkpoint is the Heretic ARA counterpart of lued/Qwen3.8-27B-INT8-W8A16-MTP. The two 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);
the identical tokenizer, chat template, processor configs, and 6-shard layout;
the identical validated serving profile (TP2, 262144, fp8_e4m3 KV, MTP × 3).

The only deviation is the base language-model weights: this checkpoint's LM weights come from the abliterated heretic-ara model instead of stock Qwen3.8-27B. Serving performance is unchanged by design and was verified identical to the sibling deployment locally; measurements in this card inherited from the sibling card are marked as such. Behavior — specifically the refusal profile — is the intended deviation and is documented below.

Abliteration (inherited from the base)

Table with columns: Metric, Base model (heretic-org/Qwen3.8-27B-heretic-ara), Original (Qwen/Qwen3.8-27B)
Metric	Base model (heretic-org/Qwen3.8-27B-heretic-ara)	Original (Qwen/Qwen3.8-27B)
Refusals (100 harmful eval prompts)	0 / 100	99 / 100
KL divergence (teacher-forced)	0.0535	0 (by definition)

Method: Heretic (timrohrbaugh fork) v1.2.0+custom with the Arbitrary-Rank Ablation (ARA) method, run in full BF16 on a single H200 NVL. 60-trial Optuna study (seed 705085018), selected trial 37: layers 26–56, targets attn.o_proj + mlp.down_proj, preserve_good_behavior_weight=0.9432, steer_bad_behavior_weight=0.0009, overcorrect_relative_weight=0.5038, neighbor_count=10, row_normalization=none, winsorization_quantile=1.0. The full parameter set, per-category refusal breakdowns, and the reproduce/ bundle (config, reproduce.json, SHA256SUMS) live in the base repository.

The 0.0535 KL is intentionally larger than a surgical abliteration (≈0.002 class): ARA trades a larger distribution shift for complete refusal elimination. If minimal drift matters more than refusal elimination, that is a different model to run, not a setting to tune on this one.

Dual RTX 3090 deployment

Validated host profile: 2×RTX 3090 24 GB, PCIe without NVLink or P2P, vLLM TP2, BF16 activations, FP8 E4M3 KV cache, MTP with three draft tokens, and --max-num-batched-tokens 8192.

Memory and capacity

Table with columns: Metric, Result
Metric	Result
Checkpoint size	31.6 GB / 29.44 GiB
Loaded model memory per GPU	14.85 GiB
Shared GPU KV pool	266,537 tokens
Native maximum request length	262,144 tokens
Simultaneous full-native-context capacity	1.02×
Illustrative four-way KV share	Approximately 66,634 tokens per request

The checkpoint has the same size and footprint as the sibling quantization, since the W8A16 policy and shard layout are identical.

--max-model-len caps one request's input plus generated tokens. The KV pool is a shared runtime allocation across all live requests, not a per-lane reservation: --max-num-seqs permits that many active sequences but does not provide that many full contexts. Near the 262K limit, reserve output-token headroom and expect only one full-length request at a time.

The open-source checkpoint is native at 262,144 tokens. Qwen documents extension to 1,000,000 tokens via static YaRN, but this deployment's 266,537-token KV pool cannot hold a 1M-token request even with YaRN enabled; the 1M path is not validated or capacity-supported here. Follow the official YaRN configuration only with a runtime and memory plan sized for the requested context.

Measured serving performance

Measured on 2026-08-15 with llama-benchy 0.4.0 against the sibling deployment (identical architecture, quantization profile, and serving stack — see the relationship section above): one concurrent request, exact generation lengths (min_tokens + ignore_eos), cold prefix cache (unique corpus text, cache_prompt=false), no conversation depth, one run per cell. The heretic checkpoint was verified locally to perform identically to that deployment; a formal run on this exact checkpoint is listed under Pending verification.

Table with columns: Prompt tokens, Generation tokens, Prefill tok/s, Generation tok/s, TTFT
Prompt tokens	Generation tokens	Prefill tok/s	Generation tok/s	TTFT
1,024	512	1,675	65.6	0.8 s
1,024	1,024	1,633	58.0	0.8 s
8,192	512	1,747	42.3	4.9 s

Generation throughput is the per-request decode rate with MTP enabled and varies by workload, since draft acceptance is prompt-dependent. Prefill and TTFT are governed by the 8,192-token batched-token cap, which chunks long prompts and interleaves the MTP draft pass; TTFT grows with prompt length, roughly linearly past 32K.

Multi-stream throughput is not yet published: concurrent requests can intermittently crash the engine on this pin, so parallel benchmarks will be added once a vLLM release containing vllm#50021 is available (see the concurrency note under Serving).

Quantization fidelity

Two different fidelity questions apply to this checkpoint, and they should not be conflated:

1. What the W8A16 step costs (quantization error). The recipe is identical to the sibling card, where it measured against the Qwen3.8-27B BF16 teacher:

Table with columns: Candidate (same recipe, stock Qwen3.8-27B), Mean KLD vs BF16, Top-1 agreement, Weight files
Candidate (same recipe, stock Qwen3.8-27B)	Mean KLD vs BF16	Top-1 agreement	Weight files
Official Qwen FP8 / BF16 A16	0.004396	98.53%	28.75 GiB
W8 / BF16 A16 (this recipe)	0.000894	99.36%	29.44 GiB

The KLD result measures checkpoint-weight drift over 4,563 teacher-forced positions. It is not a functional quality score; tool use, JSON-schema compliance, coding, multimodal quality, and long-context recall require separate behavioral evaluation. The quantization policy is per-tensor and data-free, so the expected error class for this checkpoint is the same; a formal measurement against its own (heretic) BF16 teacher is pending (see Pending verification).

2. What the model deviates from stock Qwen3.8-27B by (intended). That is the abliteration in the section above — KL 0.0535 and 0/100 refusals, measured and documented by the base author. It is a property of the base weights, not of the quantization step.

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`
Preserved precision	BF16 vision tower, `lm_head`, MTP (15 tensors, byte-identical to Qwen's original MTP), and recurrent GDN gates
MTP	BF16 draft model detected at runtime; embeddings and `lm_head` shared with the target

Why W8A16 on Ampere

RTX 3090 GPUs are Ampere sm_86. They do not provide the native FP8 tensor-core execution path available on newer GPU generations. In this deployment, FP8 weight-only and INT8 W8A16 checkpoints both reach Marlin-family weight-only kernels that reconstruct weights for 16-bit GEMMs.

W8A16 is therefore the better fidelity default here: it preserves eight integer bits with group-wise scaling while retaining the same practical decode class as FP8-Marlin on these GPUs.

Table with columns: Format, Role on 2×24 GB at long context, Decision
Format	Role on 2×24 GB at long context	Decision
W8A16 — this checkpoint	Near-lossless weights; production default	Recommended
Official / online FP8 weight quantization	Convenient, but higher measured weight error on sm_86 fallback	Useful comparison
W8A8 INT8	Can improve prefill throughput; may regress skinny-batch decode	Specialized profile
W4A16	Smaller weights and more cache headroom	Higher long-context risk for recurrent GDN layers

Why not AutoRound?

This checkpoint uses llm-compressor's QuantizationModifier, not AutoRound. At W8A16, AutoRound's installed RTN configuration disables optimized RTN by default for efficiency, and the measured KLD is already below 9e-4 nats/token. Re-encoding these weights with AutoRound would add a new export and serving path without evidence of a meaningful quality gain.

Quantization design

Architecture audit

The BF16 source was checked before quantization against the Qwen3.6 structural contract:

64 language layers: 48 Gated DeltaNet linear-attention layers and 16 full-attention layers.
Hidden size 5,120; intermediate size 17,408; padded vocabulary 248,320.
One MTP layer with the same 15 top-level mtp.* tensors.
Stable GDN projection names: in_proj_a, in_proj_b, in_proj_qkv, in_proj_z, and out_proj.
The only relevant rename was non-Linear convNd → conv1d, outside the quantization target.

The pre-quantization structural audit completed successfully before any weights were modified.

Recipe

Table with columns: Component, Precision, Reason
Component	Precision	Reason
MLP projections	INT8 W8A16	Largest dense GEMMs; strong memory return
Full-attention projections	INT8 W8A16	Measured low output-distribution error
GDN `in_proj_qkv`, `in_proj_z`, `out_proj`	INT8 W8A16	Recovers approximately 4 GB while remaining stable in the 4K GDN probe
GDN `in_proj_a`, `in_proj_b`

The result contains 400 quantized Linear GEMMs: 192 MLP projections, 64 full-attention projections, and 144 dense GDN projections.

yaml
# recipe.yaml
default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore:
        - lm_head
        - re:.*visual.*
        - re:.*mtp.*
        - re:.*linear_attn[.]in_proj_a$
        - re:.*linear_attn[.]in_proj_b$
      scheme: W8A16

Reproducing / building a sibling model

This checkpoint is fully reproducible from trohrbaugh/Qwen3.8-27B-heretic-ara with this family's W8A16 toolkit (../tools/, sibling of this repo), and the same pipeline builds a release from any other Qwen3.5-family base. The Heretic build is the abliterated counterpart of the reference release (qwen3.8-27b-int8-w8a16-mtp): identical recipe, only the base differs. Entry points: AGENTS.md (agent-oriented), the repo skill skills/w8a16-quantize/SKILL.md, and ../tools/README.md (full pipeline).

bash
source ../tools/setup_venv.sh   # resolve the family venv -> $W8A16_PYTHON
PY=$W8A16_PYTHON
SOURCE_MODEL=trohrbaugh/Qwen3.8-27B-heretic-ara SAVE_DIR=<out> $PY ../tools/quantize_qwen38_w8a16.py
SOURCE_MODEL=trohrbaugh/Qwen3.8-27B-heretic-ara SAVE_DIR=<out> $PY ../tools/finish_qwen38_w8a16.py
DEST_DIR=<out> SOURCE_MODEL=trohrbaugh/Qwen3.8-27B-heretic-ara $PY ../tools/audit_structure.py
DEST_DIR=<out> SOURCE_MODEL=trohrbaugh/Qwen3.8-27B-heretic-ara EXPECTED_MTP_SHA256=1d8268aa85ace093a561e3e7b63b9d390dac1cd55a90cd55b5ec509c3c9da9fe \
  $PY ../tools/audit_mtp.py

Order: quantize → finish → (reshard) → audit → publish. Both audits derive their expectations from SOURCE_MODEL (no hardcoded counts), so they double as the release gate for a new sibling model. For a heretic base the MTP is read from the base's model-auxiliary.safetensors.

MTP preservation

The MTP head is the exact BF16 MTP from the original Qwen3.8-27B — never touched by the abliteration (which targets only attn.o_proj/mlp.down_proj on layers 26–56) 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 and in the base heretic repository's model-auxiliary.safetensors.
All 15 tensors verified bfloat16, shape-equal, and bit-equal to the Qwen source (strict audit, not a count check).

KV-cache rationale: E4M3 versus E5M2

--kv-cache-dtype controls runtime attention-cache storage. It does not alter this checkpoint's INT8 weights. E4M3 and E5M2 both consume one byte per cached value, so this decision affects numerical behavior—not memory capacity.

Table with columns: KV dtype, Mantissa bits, Maximum finite value, Minimum subnormal, Practical tradeoff
KV dtype	Mantissa bits	Maximum finite value	Minimum subnormal	Practical tradeoff
`fp8_e4m3`	3	±448	2^-9 ≈ 0.00195	Better in-range precision
`fp8_e5m2`	2	±57,344	2^-16 ≈ 0.0000153	Much wider dynamic range

E4M3 is the default because attention K/V values generally benefit more from its additional mantissa bit than from E5M2's extreme range.

[!WARNING] Runtime compatibility is verified; long-context KV accuracy is not. The checkpoint declares kv_cache_scheme: null, so it contains no dataset-calibrated K/V scales. The KLD measurement above uses no cache, and no saturation probe was performed.

E5M2 remains the fallback if measured ranges or a future vLLM build expose E4M3 saturation or kernel incompatibility.

Serving

GPU interconnect profiles

The validated command below targets stock-driver dual-GPU systems without a working CUDA peer-to-peer path. Consumer Ampere P2P availability depends on the driver and host configuration; the physical PCIe topology alone does not establish that peer memory access works.

Table with columns: Host configuration, NCCL environment, vLLM all-reduce, Status
Host configuration	NCCL environment	vLLM all-reduce	Status
Stock driver; P2P unavailable or unverified	`NCCL_P2P_DISABLE=1`	Pass `--disable-custom-all-reduce`	Validated configuration
P2P-enabled driver; peer access verified on both GPUs	Do not set `NCCL_P2P_DISABLE`	Omit `--disable-custom-all-reduce`	Alternative profile; benchmark on the actual PCIe topology

Verify the second profile with a CUDA peer-access test such as p2pBandwidthLatencyTest, not only nvidia-smi topo -m. Remove the two workarounds together: enabling vLLM's custom all-reduce while NCCL peer access remains disabled is not the intended profile.

Recommended vLLM command

The reference llama-swap/podman deployment, minus the container plumbing, with two deliberately conservative flags: prefix caching is off by default (the reference enables it after applying the safety patch below) and the sequence cap is 2 (the reference runs 4). Rationale in the notes after the command.

bash
export NCCL_P2P_DISABLE=1
export NCCL_CUMEM_ENABLE=0
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export OMP_NUM_THREADS=1
export VLLM_USE_FLASHINFER_SAMPLER=0
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512

vllm serve lued/Qwen3.8-27B-heretic-ara-INT8-W8A16-MTP \
  --served-model-name qwen3.8-27b-heretic-ara-int8-w8a16-mtp \
  --tensor-parallel-size 2 \
  --pipeline-parallel-size 1 \
  --dtype bfloat16 \
  --performance-mode balanced \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.92 \
  --max-num-seqs 2 \
  --max-num-batched-tokens 8192 \
  --kv-cache-dtype fp8_e4m3 \
  --no-enable-prefix-caching \
  --enable-chunked-prefill \
  --mamba-cache-mode align \
  --prefix-match-unit 16 \
  --enable-prompt-tokens-details \
  --enable-per-request-metrics \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --enable-auto-tool-choice \
  --disable-custom-all-reduce \
  --trust-remote-code \
  --default-chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true}' \
  --override-generation-config '{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"repetition_penalty":1.0,"presence_penalty":0.0}' \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

NCCL_P2P_DISABLE=1 avoids NCCL initialization stalls on dual consumer GPUs without NVLink. --mamba-cache-mode align is required by the Qwen3.8 MTP/GDN serving path. Three speculative tokens is the measured default.

Prefix caching is off by default. On this hybrid GDN + MTP architecture, prefix caching without the 9-line source patch from vllm#48375 can corrupt recurrent-state KV, which surfaces as wrong tool-call or long-context output rather than an error. The reference deployment applies that patch at container start (along with this repository's chat_template.jinja via --chat-template), which is what lets it run with --enable-prefix-caching. If you apply the same patch, flip this flag back on; the published numbers were measured with a cold cache, so either setting leaves the table valid.

Concurrency is capped at 2, not 4, on this engine pin. Two or more concurrent requests can intermittently crash the engine (asynchronous cudaErrorIllegalAddress in the GDN spec-decode state path; same class as vllm#37431). The targeted fix is vllm#50021, which is open and awaiting its merge as of this writing; a vLLM release containing it should restore --max-num-seqs 4. Until then, a crashed engine drops in-flight requests and the service (llama-swap, podman, or a process manager) restarts it; if that is unacceptable, lower the cap to 1, which is the fully validated configuration on this pin.

MTP acceptance

MTP acceptance is workload-, sampling-, and draft-depth-dependent; it is not directly comparable across different models or benchmark prompts. With the sibling checkpoint and three draft tokens on the vLLM pin above, per-window draft acceptance ranged from 42.1% to 92.4% across ten-second engine windows (including warmup), and overall draft acceptance was 65.5% (accepted/drafted) in the single-stream 8K/1,024 run. The heretic checkpoint shares the identical BF16 MTP head and LM architecture; a formal acceptance run on this exact checkpoint is listed under Pending verification.

The MTP checkpoint and lm_head are already BF16. Requantizing them cannot improve alignment. Increasing num_speculative_tokens adds progressively less accurate draft positions and normally lowers the reported average acceptance rate; lowering it can raise that percentage while reducing useful speculative work. The serving decision must therefore use end-to-end output throughput, not acceptance alone. Keep num_speculative_tokens=3 unless a controlled throughput test on the intended workload shows a better setting.

Operational notes

Table with columns: Symptom or question, Action
Symptom or question	Action
Why does this model answer prompts the base model refuses?	This is an abliterated model: refusal elimination is the point of the base (0/100 vs 99/100 on the heretic eval set). All other behavior is intended to be unchanged from the base.
Which vLLM version is known to work?	Use the pinned image `nightly-ac7509e2b1db40fec2f03dde1ed4e9dfdc2338c9`. Other releases may work, but must support Qwen3.8, the compressed-tensors Marlin path, aligned Mamba cache, and MTP speculative decoding together.
Can llama.cpp load this repository?	No. This is a compressed-tensors checkpoint for vLLM, not GGUF. A separate GGUF conversion needs its own compatibility and quality validation.
The first startup appears idle	Wait for model loading, Torch/Triton compilation, CUDA graph capture, and the final API-ready message. Persist `/root/.cache/vllm` and the active Triton cache directory—`/root/.triton/cache` by default—across container runs to avoid repeating avoidable compilation work.
NCCL stalls during startup

Files and provenance

Table with columns: File, Purpose
File	Purpose
`model-0000N-of-00006.safetensors`	Packed W8A16 language and GDN weights plus preserved BF16 tensors
`model.safetensors.index.json`	Shard-to-tensor mapping (1,999 keys)
`model_mtp.safetensors`	BF16 MTP head, 849 MB (810 MiB) / 15 tensors — byte-identical to Qwen's original MTP
`recipe.yaml`	Exact llm-compressor W8A16 recipe
`chat_template.jinja`	Chat and tool-use template from the base heretic card
/

Provenance chain:

text
Qwen/Qwen3.8-27B  (commit 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, 18×BF16 shards, 55.56 GB)
   │  Heretic ARA v1.2.0+custom, 60-trial study in full BF16, trial 37
   │  (refusals 99→0/100, KL 0.0535; MTP untouched, shipped as model-auxiliary.safetensors)
   ▼
heretic-org/Qwen3.8-27B-heretic-ara  (6×BF16 shards, 54.71 GB)
   │  llm-compressor 0.12 QuantizationModifier, W8A16 RTN, group-128 symmetric, data-free
   │  400 GEMMs packed; lm_head / vision / MTP / GDN gates preserved BF16
   ▼
this repository  (6×W8A16 shards + model_mtp.safetensors, 31.6 GB)

Pending verification

Items below are carried over as follow-ups; they do not block this release, which mirrors the sibling card's validated profile.

Formal llama-benchy run on this exact checkpoint (the serving-performance table above is from the sibling deployment; local verification found the two perform identically).
W8A16-versus-heretic-BF16 KLD on the fixed teacher-forced prompt suite (the expected error class matches the sibling's 0.000894; the policy is identical).
Formal MTP draft-acceptance run on this exact checkpoint (sibling measured 65.5% on the single-stream 8K/1K run).
Optional: live A/B serve diff against the sibling model on matched prompts and seeds.

Acknowledgements and license

This repository repackages numerical weights derived from heretic-org/Qwen3.8-27B-heretic-ara (abliteration by Tim Rohrbaugh via the Heretic tool) and Qwen/Qwen3.8-27B. It does not claim authorship of the base model, the abliteration, architecture, training data, or upstream evaluations. Header styling and the Qwen logo follow the official Qwen repositories.

Apache-2.0 for this packaging. Respect all upstream model terms and attribution requirements.

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