v2.5.1 修复解码速度荒谬值：兼容 vLLM reasoning 字段思维链 + 幻影token防护(usage有token但流无正文时解码速度置空)

llm_providers.py

@@ -63,6 +63,10 @@ def _metrics(start, first_token_at, end, prompt_tokens, output_tokens,
     total_ms = (end - start) * 1000
     prefill = (prompt_tokens / (ttft_ms / 1000)) if prompt_tokens and ttft_ms > 0 else None
     decode = (output_tokens / (decode_ms / 1000)) if output_tokens and decode_ms > 0 else None
+    # 防“幻影 token”：usage 报了输出 token，但整个流没收到任何文本（如推理全用满 max_tokens
+    # 或提供商只回 usage 不正文），此时解码时间无意义，不报出数百万的荒谬速度
+    if output_tokens and decode is not None and not output_chars:
+        decode = None
     return {
         "prompt_tokens": int(prompt_tokens or 0),
         "output_tokens": int(output_tokens or 0),
@@ -132,8 +136,9 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
             event_count += 1
             if obj.get("choices"):
                 delta = obj["choices"][0].get("delta") or {}
-                # 兼容推理型模型：Qwen3/DeepSeek 思维链在 reasoning_content
-                piece = delta.get("content") or delta.get("reasoning_content") or ""
+                # 兼容推理型模型思维链：DeepSeek 用 reasoning_content，部分 vLLM(Qwen3.8-27B-FP8) 用 reasoning，Anthropic/Gemini 分别在各自适配器处理
+                piece = (delta.get("content") or delta.get("reasoning_content")
+                         or delta.get("reasoning") or "")
                 if piece:
                     if first_token_at is None:
                         first_token_at = time.time()