qwen38-27b-rtx3090 thimslugga · COMPLETED
Qwen3.8-27B on a single RTX 3090 with vLLM: ~1,000 tok/s at 64 concurrent (int8 tensor-core GEMMs, fp16 DeltaNet state), ~114 tok/s single-user at default sampling / ~124 greedy (MTP drafts, own-output draft vocab, calibrated int4 lm_head, split-KV verify attention), 150k-262k context; patches, requant scripts, benchmarks
github.com/thimslugga/qwen38-27b-rtx3090 · ★ 0 · Forks 0 · Size 23.7 MB
SUMMARY
Technologies 9
Scored 9
Observed 9
Practices 6
Evidence 9
Skips 0
COVERAGE
Analyzed 134 files · 298 commits · 0 API calls
TECHNOLOGIES & DEPTH
Dockerfile LANGUAGE Depth 70
1 files · PRODUCTION
Shell LANGUAGE Depth 70
11 files · PRODUCTION
JSON LANGUAGE Depth 70
1 files · PRODUCTION
Markdown LANGUAGE Depth 70
17 files · PRODUCTION
YAML LANGUAGE Depth 70
3 files · PRODUCTION
Python LANGUAGE Depth 70
50 files · PRODUCTION
pip BUILD_TOOL Depth 80
1 files · CONFIGURATION
NumPy LIBRARY Depth 70
4 files · PRODUCTION
TypeScript LANGUAGE Depth 70
3 files · PRODUCTION
PRACTICES
documentation · observedautomated_tests · observedcontinuous_integration · observedcontainerization · observedlinting · absentformatting · absent
ACTIVITY & OWNERSHIP
First commit 2026-08-14
Last commit 2026-09-13
Active months 2
Commits 298