Add qwen235b prefill-only tuning support

src/aituner/llm.py

@@ -182,6 +182,11 @@ def build_prompt(
         objective_notes.append(
             "This study is decode-only. The engine uses a KV decode benchmark connector, so TTFT is informational only unless an explicit TTFT rule is configured."
         )
+    if study.trace.completion_tokens_override is not None:
+        objective_notes.append(
+            "The trace replay overrides completion length for every request. "
+            f"min_tokens=max_tokens={study.trace.completion_tokens_override}."
+        )
     if study.slo.ttft_rule is None:
         objective_notes.append("There is no TTFT SLO for this study.")
     if study.slo.tpot_rule is None:
@@ -243,6 +248,7 @@ def build_prompt(
                 "trace": {
                     "window_id": study.trace.window_id,
                     "request_mode": study.trace.request_mode,
+                    "completion_tokens_override": study.trace.completion_tokens_override,
                     "input_length_filter": (
                         {
                             "min_input_tokens": study.trace.input_length_filter.min_input_tokens,

src/aituner/spec.py

@@ -317,6 +317,7 @@ class TraceSpec:
     window_id: str
     trace_file_override: str | None
     request_mode: str
+    completion_tokens_override: int | None
     u_field: str
     timestamp_field: str
     max_concurrency: int
@@ -331,9 +332,17 @@ class TraceSpec:
     def from_dict(cls, data: Mapping[str, Any]) -> "TraceSpec":
         max_requests = data.get("max_requests_per_probe")
         synthetic_prompt_cap = data.get("synthetic_prompt_cap_tokens")
+        completion_tokens_override = data.get("completion_tokens_override")
         request_mode = str(data.get("request_mode") or "chat").strip().lower()
         if request_mode not in {"chat", "decode_only"}:
             raise SpecError("trace.request_mode must be one of: chat, decode_only.")
+        if completion_tokens_override is not None:
+            completion_tokens_override = _require_int(
+                completion_tokens_override,
+                context="trace.completion_tokens_override",
+            )
+            if completion_tokens_override < 0:
+                raise SpecError("trace.completion_tokens_override must be >= 0.")
         return cls(
             windows_path=_require_str(data.get("windows_path"), context="trace.windows_path"),
             window_id=_require_str(data.get("window_id"), context="trace.window_id"),
@@ -341,6 +350,7 @@ class TraceSpec:
             if data.get("trace_file_override")
             else None,
             request_mode=request_mode,
+            completion_tokens_override=completion_tokens_override,
             u_field=str(data.get("u_field") or "sampling_u").strip(),
             timestamp_field=str(data.get("timestamp_field") or "timestamp").strip(),
             max_concurrency=_require_int(

src/aituner/trace.py

@@ -204,7 +204,11 @@ def load_trace_requests(study: StudySpec, *, study_spec_path: Path) -> tuple[Win
                 "stream": True,
                 "stream_options": {"include_usage": True},
             }
-            completion_tokens = _coerce_completion_tokens(row)
+            completion_tokens = (
+                study.trace.completion_tokens_override
+                if study.trace.completion_tokens_override is not None
+                else _coerce_completion_tokens(row)
+            )
             if completion_tokens is not None:
                 body["min_tokens"] = completion_tokens
                 body["max_tokens"] = completion_tokens