068e129cb2
Codex flagged that the legacy chained mtp+ngram path emits --draft-min 48 while --draft-max is suppressed and later reused for MTP (spec_draft_n_max, typically 2/3). That produces an inverted legacy ngram range (--draft-min 48 --draft-max 2/3) on affected binaries, which can break or effectively disable ngram-mod for auto CPU MTP loads and forced mtp+ngram requests. Both --draft-min and --draft-max are generic flags on legacy llama-server builds, so either one would race with MTP's own values. Gate the pair together: when chain_with_mtp=True on the legacy flavor we drop both flags and rely on MTP's emission for the chained range. Standalone ngram still emits both, preserving a valid min<=max window. Updated test_build_ngram_mod_flags_legacy_chained_omits_draft_max (now omits_draft_min_and_max) and added a min<=max guard on the standalone case. Full suite (test_llama_cpp_mtp_detection + test_llama_server_args) passes locally: 229 / 229. |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| _html_to_md.py | ||
| anthropic_compat.py | ||
| audio_codecs.py | ||
| chat_template_helpers.py | ||
| defaults.py | ||
| external_provider.py | ||
| inference.py | ||
| key_exchange.py | ||
| llama_cpp.py | ||
| llama_server_args.py | ||
| mlx_inference.py | ||
| orchestrator.py | ||
| pricing.py | ||
| providers.py | ||
| safetensors_agentic.py | ||
| tool_call_parser.py | ||
| tools.py | ||
| worker.py | ||