Fix off-by-one in completion_tokens count in generate_stream

[Chessing234]

Summary

In generate_stream(), the loop for i in range(max_new_tokens) generates a token and appends it to output_ids before yielding. However, the usage dict reports completion_tokens: i instead of i + 1. Since i is 0-indexed and the token is already appended by the time the yield executes, this undercounts by exactly 1 on every streaming chunk and on the final response.

Example: After the first token is generated (i=0), the response reports completion_tokens: 0 instead of 1. After N tokens, it reports N-1 instead of N.

Fix: Use i + 1 for completion_tokens and total_tokens in both the streaming yield (line 282-283) and the final yield (line 303-304).

Reference: The vLLM worker (vllm_worker.py) correctly uses len(output.token_ids) for the same field, confirming the intended semantics.

Test plan

• Verify with a simple generation request that completion_tokens matches the actual number of tokens in the response
• Confirm total_tokens = prompt_tokens + completion_tokens holds

🤖 Generated with Claude Code