lued

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

README

License: apache-2.0

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

--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 vLLM OpenAI-compatible endpoint: 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.

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
8,192	1,024	1,683	51.2	5.1 s
32,000	512	1,560	44.1	20.7 s
32,000	1,024	1,557	49.7	20.7 s
64,000	512	1,414	64.5	45.5 s
64,000	1,024	1,411	60.8	45.6 s
128,000	512	1,203	64.5	106.6 s
128,000	1,024	1,204	69.8	106.5 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

Table with columns: Candidate, Mean KLD vs BF16, Top-1 agreement, Weight files
Candidate	Mean KLD vs BF16	Top-1 agreement	Weight files
Official Qwen FP8 / BF16 A16	0.004396	98.53%	28.75 GiB
This W8 / BF16 A16	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.

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)
Preserved precision	BF16 vision tower, `lm_head`, MTP, and recurrent GDN gates
MTP	BF16 draft model detected at runtime; embeddings and `lm_head` shared with the target
FP8 E4M3 KV cache	Verified: server booted and served on both RTX 3090s at `max_model_len=262144`
Runtime	vLLM; this repository is not a GGUF checkpoint

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 preceding 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`	BF16	Tiny recurrent gates; inexpensive precision safeguard
Vision tower	BF16	Preserve multimodal fidelity
`lm_head`	BF16	Preserve final-logit fidelity
MTP head	BF16	Keep the speculative drafter close to the target
Norms, `conv1d`, `A_log`, `dt_bias`	Native BF16 / FP32	Non-Linear; never packed

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

Quantization fidelity: KLD versus the BF16 source

[!IMPORTANT] This section measures checkpoint-weight quantization. It is not a BF16-versus-FP16 benchmark and it does not measure the FP8 KV cache. The teacher is always the original BF16 checkpoint.

For every teacher-forced next-token position, the metric is the KL divergence from the BF16 teacher distribution to the candidate distribution, D_KL(p_BF16 ‖ p_candidate), computed over the full 248,320-token vocabulary, in nats/token, with FP32 log_softmax and accumulation. No sampling or top-k approximation is involved.

Overall results

Table with columns: Candidate checkpoint, Stored weights, A16 dtype in measurement, Mean KLD vs BF16, Top-1 agreement
Candidate checkpoint	Stored weights	A16 dtype in measurement	Mean KLD vs BF16	Top-1 agreement
BF16 source	BF16	BF16	0 by definition	100%
`Qwen/Qwen3.8-27B-FP8`	FP8 E4M3, 128×128 blocks	BF16 after dequantization	0.004396	98.53%
This checkpoint	INT8 symmetric, groups of 128	BF16 after INT8×scale	0.000894	99.36%
This checkpoint, FP16 profile	Same INT8 weights	FP16 after INT8×scale	0.000641	99.19%

The apples-to-apples weight comparison is official FP8/BF16-A16 versus this W8/BF16-A16. The FP16-A16 row covers the alternative --dtype float16 serving profile; its lower value is prompt-suite-specific rounding cancellation, not evidence that FP16 is generally more accurate.

On RTX 3090, Transformers dequantizes the official checkpoint's stored FP8 weights to BF16 because native dynamic-FP8 activation execution is unavailable. The comparison therefore isolates stored-weight error under the same Hugging Face BF16 execution path; it does not represent native FP8 W8A8 execution on newer hardware.

Short-prompt suite

Table with columns: Prompt, Category, Tokens, Positions, Official FP8 / BF16 A16, This W8 / BF16 A16, This W8 / FP16 A16
Prompt	Category	Tokens	Positions	Official FP8 / BF16 A16	This W8 / BF16 A16	This W8 / FP16 A16
`factual`	factual	13	12	0.001544	0.000410	0.000311
`code-fib`	code	88	87	0.001236	0.000547	0.000379
`physics-uncertainty`	physics	86	85	0.001732	0.000829	0.000442
`math-train`	math	58	57	0.002559	0.000575	0.000455
`sql-top5`	technical	60	59	0.001769	0.000869	0.000531
`narrative-clock`	narrative	68	67	0.003056	0.001018	0.001011
`instruction-stack`	instruction	43	42	0.001691	0.000543	0.000428
`history-industrial`	history	60	59	0.001295	0.000652	0.000423
Token-weighted mean			468	0.001871	0.000719	0.000517

Long-context GDN probe

The probe used the first 4,096 tokenizer tokens of the upstream Qwen3.8 model card as cached locally, producing 4,095 scored positions.

Table with columns: Candidate, Mean KLD, Top-1 agreement, Eight consecutive 511-position window means
Candidate	Mean KLD	Top-1 agreement	Eight consecutive 511-position window means
Official FP8 / BF16 A16	0.004684	98.49%	0.01652, 0.01542, 0.00223, 0.00162, 0.00059, 0.00045, 0.00032, 0.00039
This W8 / BF16 A16	0.000914	99.44%	0.00314, 0.00274, 0.00054, 0.00036, 0.00015, 0.00010, 0.00015, 0.00014
This W8 / FP16 A16	0.000655	99.24%	0.00216, 0.00201, 0.00047, 0.00026, 0.00009, 0.00008, 0.00010, 0.00009

The final seven positions are included in each overall mean but excluded from the equal-size window breakdown. KLD decreased rather than accumulated across this probe; no recurrent drift was observed at 4K tokens.

factual: The capital of France is Paris, and the capital of Japan is

code-fib:

python
def fibonacci(n):
    if n < 2:
        return n
    return fibonacci(n - 1) + fibonacci(n - 2)

# Rewrite with memoization to avoid exponential blowup, then add a
# test that checks fibonacci(10) == 55 and fibonacci(0) == 0. Use a
# dict as the cache and make the function private to the module.

physics-uncertainty: In quantum mechanics, the uncertainty principle states that the product of the uncertainties in position and momentum cannot be smaller than about half of the reduced Planck constant. More precisely, if many identically prepared particles are measured, the standard deviations of position and momentum satisfy the inequality derived by Kennard in 1927, which is a direct consequence of the non-commutativity of the position and momentum operators in Hilbert space.
math-train: A train travels 240 kilometres in 3 hours and 15 minutes, stopping for 10 minutes at one station along the way. What is its average speed in kilometres per hour? Show your calculation step by step, and state whether the stop changes the average speed.
sql-top5: Write a SQL query that returns the top five most frequently ordered products, joining the orders table with the products table on product_id, grouping by product id, and ordering by the count descending. Then explain what index would make the query faster and why a correlated subquery would be a worse choice here.
narrative-clock: Once upon a time, in a quiet village at the edge of a dark forest, there lived an old clockmaker who had never once repaired a clock that was broken on purpose. Each morning he wound every clock in the village square, and each evening he listened to them tick together like a heartbeat that the forest itself had learned to trust.
instruction-stack: Explain the difference between a stack and a queue, and give one real-world example of each. Keep the answer under one hundred words and use the words push, pop, enqueue, and dequeue exactly once each.
history-industrial: The Industrial Revolution began in Britain in the late eighteenth century. Three factors that made this possible were the abundance of coal and iron ore, a stable legal system that protected patents and contracts, and a network of navigable rivers and canals that lowered the cost of moving raw materials and finished goods.

Run locally on 2026-08-14 with identical teacher-tokenizer IDs.
BF16 teacher snapshot: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Official FP8 snapshot: 017b9c7af6b5689d5dd426a76e0bc077eb5ca20a.
Environment: 2×RTX 3090, PyTorch 2.11.0+cu128, Transformers 5.10.1, compressed-tensors 0.17.1.
Teacher and candidates used the same 69-module device map.
These are Hugging Face decompressed-path logits, not vLLM Marlin or native-FP8 kernel logits.
use_cache=False: the serving configuration's FP8 KV-cache error is not included.
No scheduling, sampling, image, or context beyond 4,096 tokens was measured.
Teacher logits were cached as FP16 before the shared FP32 KLD calculation, introducing a small common precision floor.

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. The higher-value follow-up is an A/B of E4M3 with default scale 1.0, runtime-calculated scales, and dataset-calibrated scales—not an unmeasured dtype switch.

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. A patched driver changes the host execution path, not the checkpoint.

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-INT8-W8A16-MTP \
  --served-model-name qwen3.8-27b-int8-w8a16 \
  --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 this 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 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
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	Use the stock-driver interconnect profile above unless CUDA peer access has been explicitly verified. Check that container or service-level environment variables do not override the selected profile.
Startup runs out of GPU memory	Lower `--gpu-memory-utilization` or `--max-num-batched-tokens`. If the resulting KV pool cannot hold one configured maximum-length request, also lower `--max-model-len`. Change one limit at a time and confirm the KV allocation printed by vLLM.
A long request is rejected despite `max_model_len=262144`	The limit covers input plus generated tokens, and all active requests share one KV pool. Reserve output-token headroom and reduce concurrent live context when approaching the model limit.
Is vision validated?	The BF16 vision tower is present, but the benchmarks and KLD suite on this card are text-only. Treat multimodal quality and memory use as unmeasured until evaluated separately.

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
`model_mtp.safetensors`	BF16 MTP head, 849 MB (810 MiB) / 15 tensors
`recipe.yaml`	Exact llm-compressor W8A16 recipe
`chat_template.jinja`	Upstream chat and tool-use template

Acknowledgements and license

This repository repackages numerical weights derived from Qwen/Qwen3.8-27B. It does not claim authorship of the base model, 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