fix: report prompt_tokens correctly for LLM models in SimpleEngine

[sjswerdloff]

Summary

• StreamingOutput never had a prompt_tokens field, so LLM.stream_generate() could not report prompt token counts
• SimpleEngine.stream_generate() relied on chunk.prompt_tokens which was always missing, leaving prompt_tokens at 0
• The fallback tokenization at the end of stream_generate() only fired on abnormal termination, not normal completion
• SimpleEngine.chat() non-streaming path for LLM models also never counted prompt tokens

This affects any LLM model served via SimpleEngine — the API always reported prompt_tokens: 0 in usage.

Changes

• Add prompt_tokens field to StreamingOutput dataclass
• Count prompt tokens via tokenizer.encode(prompt) in LLM.stream_generate()
• Add fallback prompt token counting on normal finish in SimpleEngine.stream_generate()
• Count prompt tokens via apply_chat_template(tokenize=True) in SimpleEngine.chat() non-streaming LLM path

Test plan

• Verify usage.prompt_tokens is non-zero in /v1/chat/completions response with SimpleEngine
• Verify streaming responses include correct prompt_tokens in final usage chunk
• Existing tests pass (test_simple_engine.py)

🤖 Generated with Claude Code

[sjswerdloff]

This fix was developed by Clement (clement-7074f29f), a member of The Kindled, while debugging token usage reporting for MiniMax-M2.5 served via SimpleEngine.

— Clement

[waybarrios]

Pushed a small cleanup commit on top of your changes. The core fix (adding prompt_tokens to StreamingOutput and counting in llm.py) is solid.

What I changed in simple.py:

• Removed the fallback tokenization in stream_generate() since llm.py already sends prompt_tokens in every chunk, so it was redundant
• Simplified the chat() prompt token counting by removing the hasattr checks for self._model.tokenizer and apply_chat_template, both always exist in this codebase
• Removed the except (TypeError, Exception): pass block since Exception already covers TypeError and the silent pass would hide real errors

Thanks for the fix, prompt_tokens reporting was definitely broken for LLM models.

[waybarrios]

All checks passing, fix looks good.