v0.3.5 - model autodetection

Changes in this release:

• Added model name autodetection (if --model is omitted)
• Updated tokenizer argument description

v0.3.4 - minor improvements

Release Summary for v0.3.4:

Changes:

• Added --exit-on-first-fail and --no-results-on-fail flags
• Require transformers 5.2.x for more tokenizer choices

v0.3.3 - bugfixes

Fixes

• Fixed coherence test crash when API returns null content field (e.g. some thinking models)
• Print partial results when interrupted

v0.3.2 - coherence test

Added skippable model coherence testing that runs before benchmarks.

v0.3.1 - SGLang fixes

• Fixed an issue with SGLang not providing prompt tokens in usage stats
• Removed per-requests stats from Markdown table when concurrency == 1

v0.3.0 - JSON improvements

• Added an option to include total throughput and per request time series data to JSON output.
• Added JSON schema.
• Added sample JSON file with embedded documentation (in JSONC format).
• Added peak t/s to the output (for total throughput and per request).
• Added a sample Jupyter Notebook with visualization

v0.2.1 - improvements

• Added peak throughput
• Improved output t/s accuracy

v0.2.0 - major update

A major update. Brings the following functionality:

• Added concurrency testing
• Added JSON and CSV output formats
• Added ability to save benchmarks to file

v0.1.2 - Improved compatibility with Nemotron models

Support different reasoning headers for models like Nemotron Nano.

v0.1.1 - cosmetic changes

Fixed version numbering