Recipe

qwen3-6-27b-nvfp4-rtx-5090-32gb-vllm-tp1-puvumomq

qwen3-6-27b-nvfp4-rtx-5090-32gb-vllm-tp1-puvumomq

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
vllm
Engine version
0.25.1
Accelerators
1
Tensor parallel
1
Context tokens
131,072
Max concurrency
4
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4
Repository
unsloth/Qwen3.6-27B-NVFP4
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

vllm

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmsng2oee00dno001puvumomq

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

  1. vllm
  2. serve
  3. --model
  4. unsloth/Qwen3.6-27B-NVFP4
  5. --served-model-name
  6. qwen3.6-27b
  7. --max-model-len
  8. 131072
  9. --gpu-memory-utilization
  10. 0.94
  11. --max-num-seqs
  12. 4
  13. --kv-cache-dtype
  14. fp8
  15. --enable-prefix-caching
  16. --speculative-config
  17. {"method": "qwen3_5_mtp", "num_speculative_tokens": 3}
  18. --reasoning-parser
  19. qwen3
  20. --enable-auto-tool-choice
  21. --tool-call-parser
  22. qwen3_coder
FlagValue
--modelunsloth/Qwen3.6-27B-NVFP4
--served-model-nameqwen3.6-27b
--max-model-len131072
--gpu-memory-utilization0.94
--max-num-seqs4
--kv-cache-dtypefp8
--speculative-config{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}
--reasoning-parserqwen3
--tool-call-parserqwen3_coder

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
131,072120.982observedqwen3-6-27b-nvfp4-rtx-5090-32gb-vllm-tp1-puvumomq-sweep

Remaining fields

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

hardware count
1
hardware id
rtx-5090-32gb
id
qwen3-6-27b-nvfp4-rtx-5090-32gb-vllm-tp1-puvumomq
model instance id
unsloth-qwen3-6-27b-nvfp4--nvfp4
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-27b-nvfp4-rtx-5090-32gb-vllm-tp1-puvumomq-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
serve, --model, unsloth/Qwen3.6-27B-NVFP4, --tensor-parallel-size, 1, --host, 0.0.0.0, --port, 8000, --max-model-len, 131072
container port
8,000
entrypoint
vllm
host port
8,000
image
vllm/vllm-openai@sha256:0a51ea5b4ae2dc5d81890e5173f54203d2a3ae0cfffe51b8fd2afd4391bfd967
ipc
host
kind
docker
shm size
16g
synthesized · generated at
2026-08-31T22:12:17Z
synthesized · image provenance
deepseek-fp8-rtx-pro-6000-blackwell-96gb-vllm-tp1
synthesized · template
vllm-openai-v1

mounts

read onlytarget
No/root/.cache/huggingface

engine

name
vllm
version
0.25.1

serving

max concurrency
4
max context tokens
131,072
tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
draft launch.entrypoint · provenance · captured at
2026-09-01T01:41:26Z
draft launch.entrypoint · reason linux-amd64-container-config-entrypoint
draft launch.entrypoint · state known
engine.graph mode · provenance · captured at
2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-derived-from-source-evidence
serving.max concurrency · state known

metadata

localmaxxing · backend
cuda
localmaxxing · batch size
1
localmaxxing · hardware label
RTX 5090
localmaxxing · notes
Method: 8 single-turn requests, streaming, batch size 1, temperature 0, engine confirmed idle first (vllm:num_requests_running = 0). 3093 generated tokens in 25.59 s wall. TTFT is the median of the 8 (range 64-112 ms). Peak VRAM and power are peaks from nvidia-smi sampled every 1.5 s during the run. Prefill tok/s left blank on purpose: prompts averaged only 29 tokens, so TTFT is dominated by request overhead, not prefill throughput. Any prefill number derived from it would be meaningless. Speculative decoding is ON and this matters for comparison: MTP (qwen3_5_mtp draft) with num_speculative_tokens=3, measured 76.2% draft acceptance / 2.29 accepted tokens per round. That is why output tok/s exceeds the ~81 tok/s naive ceiling implied by this card's memory bandwidth for a 27B NVFP4 weight read per token. Also enabled: --kv-cache-dtype fp8 and --enable-prefix-caching (prefix caching contributes nothing here - every prompt is distinct). Runs in an unprivileged Proxmox VE 9.2 LXC with GPU passthrough, gpu-memory-utilization 0.94, max context 131072, max 4 concurrent sequences (this run used 1). The card is currently link-trained at PCIe x8 instead of x16 (unresolved hardware issue); that affects model load time, not decode.
localmaxxing · observed command
vllm serve --model unsloth/Qwen3.6-27B-NVFP4 --served-model-name qwen3.6-27b --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 4 --kv-cache-dtype fp8 --enable-prefix-caching --speculative-config '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}' --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
localmaxxing · run id
cmsng2oee00dno001puvumomq
localmaxxing · tokenized · arguments
vllm, serve, --model, unsloth/Qwen3.6-27B-NVFP4, --served-model-name, qwen3.6-27b, --max-model-len, 131072, --gpu-memory-utilization, 0.94, --max-num-seqs, 4, --kv-cache-dtype, fp8, --enable-prefix-caching, --speculative-config, {"method": "qwen3_5_mtp", "num_speculative_tokens": 3}, --reasoning-parser, qwen3, --enable-auto-tool-choice, --tool-call-parser, qwen3_coder
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmsng2oee00dno001puvumomq