benchmark_id large_stringclasses 1
value | benchmark_name large_stringclasses 1
value | model large_stringclasses 1
value | model_architecture large_stringclasses 1
value | quantization large_stringclasses 1
value | model_url large_stringclasses 1
value | model_sha256 large_stringclasses 1
value | model_size_recorded large_stringclasses 1
value | memvanta_source_commit large_stringclasses 1
value | llama_cpp_commit large_stringclasses 1
value | source_artifact large_stringclasses 1
value | runtime large_stringclasses 2
values | role large_stringclasses 2
values | peak_rss_kib int64 3.98M 7.59M | peak_rss_gib float64 3.8 7.24 | prompt_tokens_per_sec_mean float64 2.92 12 | prompt_tokens_per_sec_sd float64 0 0.02 | prompt_cv_pct float64 0.05 0.15 | generation_tokens_per_sec_mean float64 1.92 7.97 | generation_tokens_per_sec_sd float64 0 0.03 | generation_cv_pct float64 0.09 0.42 | rss_reduction_vs_llama_cpp_pct float64 0 47.5 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
openllama-7b-v2-q4_0-same-model-cpu-ab | OpenLLaMA 7B v2 Q4_0 same-model CPU A/B | OpenLLaMA 7B v2 | llama | Q4_0 | https://huggingface.co/maddes8cht/openlm-research-open_llama_7b_v2-gguf/resolve/main/openlm-research-open_llama_7b_v2-Q4_0.gguf?download=true | 892f6e2e840ed98bb7bc1d74da67ae0cca52b2d90b0316d0a7e66a37a5a60760 | model_bytes=3825818912 | 3fcebd5afb44c1aa644c3eac5a4cd241bdfd363d | 6503355df0eb4f65875012523263c302fe0088c1 | results/openllama-7b-v2-ab/summary.json | MemVanta | candidate_memory_first_runtime | 3,982,168 | 3.797691 | 2.917118 | 0.001541 | 0.052826 | 1.922456 | 0.001688 | 0.087826 | 47.538636 |
openllama-7b-v2-q4_0-same-model-cpu-ab | OpenLLaMA 7B v2 Q4_0 same-model CPU A/B | OpenLLaMA 7B v2 | llama | Q4_0 | https://huggingface.co/maddes8cht/openlm-research-open_llama_7b_v2-gguf/resolve/main/openlm-research-open_llama_7b_v2-Q4_0.gguf?download=true | 892f6e2e840ed98bb7bc1d74da67ae0cca52b2d90b0316d0a7e66a37a5a60760 | model_bytes=3825818912 | 3fcebd5afb44c1aa644c3eac5a4cd241bdfd363d | 6503355df0eb4f65875012523263c302fe0088c1 | results/openllama-7b-v2-ab/summary.json | llama.cpp | pinned_comparison_runtime | 7,590,668 | 7.239025 | 11.95164 | 0.01767 | 0.147842 | 7.96613 | 0.033763 | 0.423832 | 0 |
MemVanta CPU LLM Memory Benchmark
Reproducible CPU LLM inference benchmark evidence comparing memory usage and throughput for MemVanta and a pinned comparison runtime on the same GGUF artifact.
The initial configuration is the committed OpenLLaMA 7B v2 Q4_0 same-model CPU A/B benchmark from the public MemVanta repository. This dataset publishes benchmark evidence and provenance only; it does not redistribute model weights.
Canonical result
| Metric | MemVanta | pinned llama.cpp |
|---|---|---|
| Peak RSS | 3.80 GiB | 7.24 GiB |
| Prompt processing | 2.92 ± 0.00 tok/s | 11.95 ± 0.02 tok/s |
| Token generation | 1.92 ± 0.00 tok/s | 7.97 ± 0.03 tok/s |
| Peak-RSS reduction | 47.54% | baseline |
The result shows the measured trade-off for this specific benchmark: lower resident memory for MemVanta, with lower throughput than the pinned llama.cpp baseline. It is not a universal claim across models, machines, context sizes, quantizations, or runtimes.
Dataset rows
The canonical Parquet contains one row per runtime with:
- benchmark and model identifiers;
- quantization and runtime role;
- peak RSS in KiB and GiB;
- prompt-processing throughput mean, standard deviation, and coefficient of variation;
- token-generation throughput mean, standard deviation, and coefficient of variation;
- measured MemVanta RSS reduction relative to the pinned comparison runtime;
- model artifact URL/hash metadata;
- MemVanta source commit and pinned
llama.cppcommit; - source evidence location.
Evidence files
Selected raw evidence from the public benchmark is included under evidence/openllama-7b-v2-ab/, including:
summary.jsonenvironment.txtmemvanta.csvmemvanta.time.txtllama-bench.jsonllama-bench.time.txtllama_cpp_commit.txtmodel-size.txtmodel-url.txtmodel.sha256SUMMARY.mdWORKFLOW_RUN.md
The benchmark methodology is included as evidence/MEMORY_BENCHMARKING.md.
Reproducibility
Generated from MemVanta source commit 3fcebd5afb44c1aa644c3eac5a4cd241bdfd363d. The pinned comparison-runtime commit recorded by the benchmark is 6503355df0eb4f65875012523263c302fe0088c1.
Source repository: https://github.com/sauravsingla/MemVanta
Benchmark documentation: https://sauravsingla.github.io/MemVanta/benchmark/
Reproduction guide: https://sauravsingla.github.io/MemVanta/reproduce/
Zenodo DOI: https://doi.org/10.5281/zenodo.22886357
Important limits
This dataset is systems benchmark evidence for a fixed model artifact, workload, host, and pinned comparison runtime. Peak RSS is not the same thing as a universal physical-RAM requirement. Throughput and memory measurements can vary with hardware, operating system, compiler, model file, context length, thread count, runtime revision, and benchmark procedure.
MemVanta is memory-first and does not claim to be faster than llama.cpp.
License and upstream artifacts
MemVanta-generated code and benchmark artifacts are published under Apache-2.0. The model itself is not included here; its upstream license and distribution terms remain applicable.
- Downloads last month
- 48