Release Tiny-LLM-Inference v0.1.0

[wlmrh]

Summary

• restructure the README around the v0.1.0 scope, architecture, capabilities, verified results, and shortest reproduction path
• publish pinned-model, three-backend offline, and five-workload open-loop evidence under benchmark/reports/v0.1.0/
• record the tested commit, hardware/software environment, model revisions and file hashes, commands, and release-asset checksum
• align the changelog, citation date, project status, release checklist, and documentation index
• replace the previous public benchmark snapshot while preserving its Git history

Scope and impact

This release remains a single-process, single-device, offline C++17/CUDA runtime. The diff does not change include/, src/, CUDA kernels, runtime behavior, public APIs, benchmark workloads, metric definitions, or benchmark execution logic.

Validation

• Python benchmark unit tests: 12/12
• CPU model-independent CTest: 65/65
• CPU real-model CTest: 7/7, no path-related skips
• CUDA CTest with the real-model path enabled: 78/78
• Qwen2.5 FP32 deterministic generation: passed
• Qwen2.5 BF16 compute/KV smoke: passed
• TinyLLM/Transformers generation alignment: exact match on both direct prompts
• Offline suite: TinyLLM, Transformers, and vLLM completed all four workloads with no skips; all pairwise token-ID comparisons were exact
• Open-loop suite: five workloads completed 200/200 requests each; lifecycle events were complete, latencies nonnegative, and percentiles monotonic
• Regression preset: passed

Reproducibility

The benchmark candidate is 25e2355921b033abb89091d15f768c57c715c63c with git_dirty=false. Later commits contain only documentation, report JSON, metadata, and replacement of the old public report. The full raw artifact tarball and SHA-256 sidecar will be attached to the signed GitHub Release after this PR is merged and CI is green.