v0.5.0

What's Changed

• refactor: remove cross-component e2e test by @hlin99 in #83
• docs: add strict bot policies (non-loop) + update trigger workflow by @hlin99 in #85
• docs: unified structure — LICENSE, README, CONTRIBUTING by @hlin99 in #86

Full Changelog: v0.4.0...v0.5.0

v0.4.0

What's New

OpenAI API Spec Compliance

• Parameter range validation (temperature, top_p, frequency_penalty, presence_penalty, n, best_of)
• response_format support: json_object and json_schema for chat completions
• Embedding encoding_format=base64 support (little-endian)
• Stop sequences correctly skipped for JSON-formatted responses

vLLM Backend Compatibility

• Accept all vLLM-specific sampling params (top_k, min_p, use_beam_search, etc.) without error
• Response fields: stop_reason, service_tier, kv_transfer_params
• best_of and echo on chat completions

Scheduling

• response_format works in both scheduled and non-scheduled paths

Tests

• 19 new API compliance tests (TC14-TC17)
• 285 total tests passing

v0.3.0

Changes since v0.2.0

• Cap output tokens by max_model_len (vLLM behavior)
• E2e proxy integration tests
• E2e dependency changed to xpyd-proxy>=1.2.0
• 6 regression tests for max_model_len cap
• 7 e2e tests for proxy integration

v0.2.0

Changes

• Flat layout: moved src/xpyd_sim/ → xpyd_sim/ (matching xPyD-proxy convention)
• Includes all bug fixes from v0.1.1

v0.1.1

Bug fixes for issues #48-#65:

• render_dummy_text cycles tokens for any max_tokens value
• Streaming requests logged to JSONL
• max_model_len enforced in all code paths
• prefill/app.py: /ping and /metrics endpoints
• Warmup penalty works with scheduling
• /v1/completions routes through scheduler
• Scheduler sync-safe (no event loop dependency)
• CLI arg priority fixed (_TrackAction subparser bug)
• Batch metrics always present in /metrics
• /debug/batch returns zero-state when scheduling disabled
• Prefill batch FIFO ordering fixed
• PyPI publish workflow added