orwelian84/Qwen3.8-27B-Three-GPU-Benchmarks
Qwen3.8-27B three-GPU llama.cpp benchmarks This repository contains the public reproduction package for a local Qwen3.8-27B inference study. The study covers an RTX 3090 24 GB, an RTX 4090 24 GB, and an RTX PRO 6000 Blackwell 96 GB. Only repeated runs that passed an independent artifact validator appear in confirmed-results.csv. Single-run tuning points and invalidated attempts are not headline results. Contents ARTICLE.md contains the living Hugging Face article… See the full description on the dataset page: https://huggingface.co/datasets/orwelian84/Qwen3.8-27B-Three-GPU-Benchmarks.
Qwen3.8-27B three-GPU llama.cpp benchmarks
This repository contains the public reproduction package for a local Qwen3.8-27B inference study. The study covers an RTX 3090 24 GB, an RTX 4090 24 GB, and an RTX PRO 6000 Blackwell 96 GB.
Only repeated runs that passed an independent artifact validator appear in confirmed-results.csv. Single-run tuning points and invalidated attempts are not headline results.
Contents
ARTICLE.mdcontains the living Hugging Face article source.confirmed-results.csvcontains repeated medians and observed ranges.profiles.jsoncontains the exact public runtime profiles.prompts.mdcontains the exact synthetic prompts and request settings.public/NGRAM.mdcontains the matched five-repeat prompt n-gram control.public/contains the scrubbed expanded reports and charts.
Pinned inputs
- llama.cpp commit:
9b05354ec6fb58b4e665e9a39ebc40285c015638 - GGUF repository: `ggml-org/Qwen3.8-27B-GGUF`
- GGUF revision:
0669b98607d47046c7c2b3f801011d54a08cfccf - Q4KM SHA-256:
31629f53165ab6a7dad8c9847dcfd1fdf55829dac1e6e748f4a68581b0033d34 - Q4_0 MTP SHA-256:
051a1764cff8c4f3ee6ae8b00593a0364c7539c67fa50ffc58f3f96509fca38e - Q8_0 SHA-256:
f5c702d8820d36fb55985bb238fc83ee3a313e920f4b752a437c3a6a9e14e4c8 - Official BF16 model revision:
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 - Locally converted official BF16 GGUF SHA-256:
93f653bed5ffa1993caefa0d1638b5dcabc1af0fd15c879199e5f72d048b6c0d
Measurement rule
Each interactive row generates 512 tokens from the same astronomy prompt. The table reports the median and observed range. Baseline and MTP profiles are matched within each card. Card-specific cache and batch settings prevent a strict card-to-card hardware comparison.
Synthetic prompt throughput does not measure model quality. Q4 and Q8 MTP profiles produced stable repeated outputs, but those outputs were not byte-identical to their non-speculative baselines.
Prompt n-gram results are also reported only after five matched repeats. The effect varied by prompt and card. The article does not present prompt n-gram decoding as a global default.
Reproduction
- Build llama.cpp at the pinned commit with CUDA enabled.
- Download the pinned model revision and verify each SHA-256 value.
- Start
llama-serverwith one profile fromprofiles.json. - Submit the payloads in
prompts.mdwith greedy decoding. - Run the stated repeat count and report the median.
The expanded method and platform details are in public/REPRODUCIBILITY.md.
