feat: 3-tier ixformer flash prefill dispatch + OpenCompass max_tokens clamp + n>1 fanout + index sanitizer

paged_attn.py (+218 lines):
  - Tier 0: ixformer flash_attn_varlen_func (cu_seqlens packed)
  - Tier 0.5: ixformer flash_attn_func (non-varlen, batch layout)
  - Tier 1: CoreXFA2 3-mode dispatch (packed/paged/chunked)
  - Tier 2 fallback: existing Python Q-tiling (unchanged)
  - Import chain: flash_attn_func + CoreXFA2Class + varlen

serving_chat.py (+23 lines):
  - max_tokens clamp: fixes OpenCompass 0 score (5 benchmarks all 400)
  - n>1 fanout: remove temperature==0 restriction for t2_n_2 FAIL

api_server.py (+33 lines):
  - HTTP middleware: strip index from messages before pydantic validation
  - Fixes ValidatorIterator 0.index Extra inputs are not permitted x6

qwen3_6_scripts/api_server.py

@@ -903,6 +903,39 @@ def build_app(args: Namespace) -> FastAPI:
         allow_headers=args.allowed_headers,
     )
 
+    @app.middleware("http")
+    async def sanitize_chat_body(request: Request, call_next):
+        """Strip fields from chat messages that vLLM's pydantic models reject.
+
+        Some replay datasets include ``index`` on messages (used by OpenAI
+        streaming deltas but forbidden by the non-streaming request schema).
+        Stripping it here avoids a ValidatorIterator 400 before our handler
+        even runs.
+        """
+        if (request.method == "POST"
+                and request.url.path.endswith("/v1/chat/completions")):
+            content_type = request.headers.get("content-type", "")
+            if "json" in content_type or not content_type:
+                try:
+                    body = await request.json()
+                    changed = False
+                    for msg in body.get("messages", []) if isinstance(body, dict) else []:
+                        if isinstance(msg, dict) and "index" in msg:
+                            del msg["index"]
+                            changed = True
+                    if changed:
+                        import json as _json
+                        raw = _json.dumps(body).encode("utf-8")
+
+                        async def patched_body():
+                            return raw
+
+                        request._body = raw
+                        request._receive = patched_body  # noqa
+                except Exception:
+                    pass
+        return await call_next(request)
+
     @app.exception_handler(RequestValidationError)
     async def validation_exception_handler(raw_request, exc):
         _bi100_log_request_validation_4xx(raw_request, exc)