Recipe

qwen3-6-27b-dflash-q4-k-m-rtx-5090-32gb-llama-cpp-tp1

qwen3-6-27b-dflash-q4-k-m-rtx-5090-32gb-llama-cpp-tp1

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
lucebox-hub b6ffab4a9 (sm_120 rebuild)
Accelerators
1
Tensor parallel
1
Context tokens
1,024
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/z-lab/Qwen3.6-27B-DFlash
Repository
z-lab/Qwen3.6-27B-DFlash
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

llama.cpp

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmokq33sw0004ks0413sul32a

Observed source tokens

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

  1. build/test_dflash
  2. Qwen3.6-27B-Q4_K_M.gguf
  3. z-lab/Qwen3.6-27B-DFlash
  4. --fast-rollback
  5. --ddtree
  6. --ddtree-budget=22
  7. --max-ctx=1024
  8. (rebuild
  9. for
  10. sm_120:
  11. cmake
  12. -DCMAKE_CUDA_ARCHITECTURES=120
  13. -DGGML_CUDA_GRAPHS=ON
  14. -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.9/bin/nvcc
  15. -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/gcc-12)
FlagValue
--ddtree-budget22
--max-ctx1024
-DCMAKE_CUDA_ARCHITECTURES120
-DGGML_CUDA_GRAPHSON
-DCMAKE_CUDA_COMPILER/usr/local/cuda-12.9/bin/nvcc
-DCMAKE_CUDA_HOST_COMPILER/usr/bin/gcc-12)

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
11,024215.4470observedqwen3-6-27b-dflash-q4-k-m-rtx-5090-32gb-llama-cpp-tp1-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-dflash-q4-k-m-rtx-5090-32gb-llama-cpp-tp1
model instance id
z-lab-qwen3-6-27b-dflash--q4-k-m
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-27b-dflash-q4-k-m-rtx-5090-32gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
lucebox-hub b6ffab4a9 (sm_120 rebuild)

serving

max context tokens
1,024
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
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-not-evidenced
serving.max concurrency · state unknown

metadata

localmaxxing · backend
cuda
localmaxxing · batch size
1
localmaxxing · hardware label
RTX 5090
localmaxxing · notes
DFlash speculative decoding via Lucebox-Hub test_dflash binary on RTX 5090 ROG Astral OC (Blackwell GB202). Same bench methodology as prior #19 submission (3090 Ti, 140.28 t/s mean) — scripts/bench_he.py --n-gen 512 --ddtree-budget 22 over 10 HumanEval prompts, single-stream batch=1. 5090 result: mean 215.43 t/s (range 116.5-273.3), peak sum_product 250.4 / filter_by_substring 273.3. Acceptance length 8.83, accept rate 55.2%. +53.6% over prior 3090 Ti same-binary-same-recipe submission. Required rebuild for sm_120 — CUDA 12.9 toolkit, gcc-12 host compiler (gcc-13 + _Float128 glibc clash on Ubuntu 24.04), CMAKE_CUDA_ARCHITECTURES=120, GGML_CUDA_GRAPHS=ON. Critical CMake flags from #19 still apply. Display IS attached to this 5090 — Xorg pstate flicker likely costs ~10% decode (per the 2026-04-27 monitor-on-Ti finding). Production vLLM endpoint stopped during bench.
localmaxxing · observed command
build/test_dflash Qwen3.6-27B-Q4_K_M.gguf z-lab/Qwen3.6-27B-DFlash --fast-rollback --ddtree --ddtree-budget=22 --max-ctx=1024 (rebuild for sm_120: cmake -DCMAKE_CUDA_ARCHITECTURES=120 -DGGML_CUDA_GRAPHS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.9/bin/nvcc -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/gcc-12)
localmaxxing · run id
cmokq33sw0004ks0413sul32a
localmaxxing · tokenized · arguments
build/test_dflash, Qwen3.6-27B-Q4_K_M.gguf, z-lab/Qwen3.6-27B-DFlash, --fast-rollback, --ddtree, --ddtree-budget=22, --max-ctx=1024, (rebuild, for, sm_120:, cmake, -DCMAKE_CUDA_ARCHITECTURES=120, -DGGML_CUDA_GRAPHS=ON, -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.9/bin/nvcc, -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/gcc-12)
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/cmokq33sw0004ks0413sul32a