Recipe

qwen3-8-27b-w4a16-autoround-rtx-3090-24gb-sglang-tp2-y0tyt5cf

qwen3-8-27b-w4a16-autoround-rtx-3090-24gb-sglang-tp2-y0tyt5cf

Observed LocalMaxxing result. It is evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
sglang
Accelerators
2
Tensor parallel
2
Context tokens
2,048
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/dbirks/Qwen3.8-27B-W4A16-AutoRound
Repository
dbirks/Qwen3.8-27B-W4A16-AutoRound
Status
known
Link type
Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

sglang

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmtapc07w0003qq01y0tyt5cf

No tokenized launch fields. This is measured or documented compatibility, not a Docker launch.

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
12,0481,638.5192.6676.8observedqwen3-8-27b-w4a16-autoround-rtx-3090-24gb-sglang-tp2-y0tyt5cf-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count
2
hardware id
rtx-3090-24gb
id
qwen3-8-27b-w4a16-autoround-rtx-3090-24gb-sglang-tp2-y0tyt5cf
model instance id
dbirks-qwen3-8-27b-w4a16-autoround--w4a16-autoround
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-8-27b-w4a16-autoround-rtx-3090-24gb-sglang-tp2-y0tyt5cf-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
python3, -m, sglang.launch_server, --model-path, dbirks/Qwen3.8-27B-W4A16-AutoRound, --tp, 2, --host, 0.0.0.0, --port, 30000, --context-length, 2048
container port
30,000
host port
30,000
image
lmsysorg/sglang:dev-cu13@sha256:6cd4635214f279e0a43019f88e3120d407567640a58aa7dcc0085e3d91402cc4
ipc
host
kind
docker
shm size
16g
synthesized · generated at
2026-08-31T22:12:17Z
synthesized · image provenance
gemma-4-12b-it-nvfp4-rtxpro6000-sglang-tp1
synthesized · template
sglang-launch-server-v1

mounts

read onlytarget
No/root/.cache/huggingface

engine

name
sglang

serving

max concurrency
1
max context tokens
2,048
tensor parallel
2
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-08-27T06:04:13.773Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
engine.version · provenance · captured at
2026-08-27T06:04:13.773Z
engine.version · reason runtime-detail-not-published
engine.version · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-27T06:04:13.773Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown

metadata

localmaxxing · hardware label
RTX 3090
localmaxxing · notes
Fastest verified SGLang setup for Qwen3.8-27B W4A16 AutoRound on 2x RTX 3090 with TP2. Uses DFlash speculative decoding with a local W8A16 DFlash2 drafter, 6 draft tokens, and a 40,960-token output-derived curated candidate vocabulary distributed across TP ranks. Candidate selection uses rank-local projection/top-k with compact cross-rank communication. Target execution uses torch.compile, CUDA graphs, selective SM86 W4A8 Marlin MLP kernels, BF16 KV/recurrent state, compiler-safe DFlash causal-convolution verification, and persisted tuned TorchInductor/Triton caches. Measurement is concurrency 1 with 2 warmups and 5 timed 512-token iterations. CUDA peer access and NVLink are unavailable on this PCIe topology, so TP collectives use NCCL's host-mediated path.
localmaxxing · revision unpinned
Yes
localmaxxing · run id
cmtapc07w0003qq01y0tyt5cf

provenance

captured at
2026-08-27T06:04:13.773Z

sources

captured atkindurl
2026-08-27T06:04:13.773Znormalized-recipewww.localmaxxing.com/en/runs/cmtapc07w0003qq01y0tyt5cf