| # Engineering Notes |
|
|
| This file keeps short, still-useful implementation notes that previously lived in separate top-level markdown files. |
|
|
| ## 1. smolagents `reasoning_content` |
| |
| ### Summary |
| |
| Two separate issues affected reasoning capture with Qwen3 through `OpenAIServerModel` / `OpenAIServerModel`-style tool use: |
| |
| 1. `tool_choice="required"` suppressed `reasoning_content` at the server level. |
| 2. smolagents step serialization dropped `reasoning_content` even when it existed on the live OpenAI/Pydantic response object. |
|
|
| ### What mattered in practice |
|
|
| - For this project, `tool_choice="auto"` is required if we want reasoning traces. |
| - Capturing reasoning from the live model response before smolagents serializes steps is the reliable workaround. |
|
|
| ### Root cause |
|
|
| - `tool_choice="required"` constrained generation enough that llama-server stopped returning reasoning blocks. |
| - smolagents stored the live response object, but later step serialization did not preserve Pydantic extra fields such as `reasoning_content`. |
|
|
| ### Project workaround |
|
|
| We handle this in the collector by: |
|
|
| - passing `tool_choice="auto"` to the model |
| - monkey-patching `model.generate()` to extract `reasoning_content` before step serialization |
|
|
| ### Environment where this was observed |
|
|
| - `smolagents 1.24.0` |
| - OpenAI SDK `1.82.0` |
| - llama-server / llama.cpp serving Qwen3 GGUF models |
| - macOS, Python 3.10 |
|
|
| ## 2. llama.cpp + Qwen3 benchmark/config note |
|
|
| These were the practical serving findings that mattered while tuning local collection runs on an M3 Pro. |
|
|
| ### Main benchmark takeaways |
|
|
| For an ~11.6k token prompt: |
|
|
| | Model | Config | Prompt tok/s | Gen tok/s | |
| |-------|--------|--------------|-----------| |
| | 0.6B | baseline | 977 | 32 | |
| | 0.6B | flash + q8_0 KV | 958 | 70 | |
| | 0.6B | flash only | 1642 | 55 | |
| | 1.7B | baseline | 704 | 25 | |
| | 1.7B | flash + q8_0 KV | 694 | 42 | |
| | 1.7B | flash only | 991 | 37 | |
|
|
| ### Practical flag notes |
|
|
| - `--flash-attn` |
| - Biggest prompt-speed improvement. |
| - `-ctk q8_0 -ctv q8_0` |
| - KV cache quantization improved generation speed and reduced KV memory footprint. |
| - `--jinja --chat-template-file` |
| - Needed when using a custom no-think / reasoning-control template. |
|
|
| ### Working heuristics |
|
|
| - Long prompts, short outputs: |
| - `--flash-attn` |
| - Short prompts, long outputs: |
| - `--flash-attn -ctk q8_0 -ctv q8_0` |
| - Memory-constrained runs: |
| - prefer KV quantization |
|
|
| ### Important caveats |
|
|
| - KV cache quantization availability depends on the serving build/config. |
| - Custom chat templates can materially change reasoning behavior, so they should be recorded in run metadata. |
|
|
