Add Qwen3.6 validation and benchmark coverage

tools/bench/speed.py

@@ -35,6 +35,7 @@ class SpeedRow:
     scenario: str          # e.g. "single/short", "concurrent-4"
     requests: int
     completion_tokens_total: int
+    prompt_tokens_mean: float
     wall_s: float
     ttft_ms_p50: float
     ttft_ms_p95: float
@@ -64,6 +65,7 @@ def _summarize(system: str, scenario: str, results: list[StreamResult], wall_s:
         scenario=scenario,
         requests=len(results),
         completion_tokens_total=total_tokens,
+        prompt_tokens_mean=(statistics.mean(r.prompt_tokens for r in ok) if ok else 0.0),
         wall_s=wall_s,
         ttft_ms_p50=_percentile(ttft_ms, 50),
         ttft_ms_p95=_percentile(ttft_ms, 95),
@@ -82,6 +84,18 @@ async def run_single_stream(
     rows: list[SpeedRow] = []
     raw: list[dict[str, Any]] = []
     for bucket, prompt in SPEED_PROMPTS.items():
+        # Prefill kernels/graphs can be shape-specific. Warm each prompt shape
+        # twice so p95 does not accidentally report one-time graph setup.
+        for _ in range(2):
+            await chat_concurrent(
+                ep.base_url, ep.model_id, [[{"role": "user", "content": prompt}]],
+                max_tokens=cfg.speed_max_tokens,
+                temperature=0.0,
+                api_key=ep.api_key,
+                timeout=cfg.request_timeout_s,
+                concurrency=1,
+                extra_body=ep.extra_body,
+            )
         messages = [[{"role": "user", "content": prompt}]] * cfg.speed_prompts
         results, wall = await chat_concurrent(
             ep.base_url, ep.model_id, messages,
@@ -98,6 +112,7 @@ async def run_single_stream(
                 "system": ep.name, "scenario": f"single/{bucket}", "i": i,
                 "ttft_s": r.ttft_s, "tpot_s": r.tpot_s,
                 "completion_tokens": r.completion_tokens,
+                "prompt_tokens": r.prompt_tokens,
                 "e2e_s": r.e2e_s, "error": r.error,
                 "finish_reason": r.finish_reason,
             })
@@ -131,6 +146,7 @@ async def run_concurrent(
                 "system": ep.name, "scenario": f"concurrent-{c}", "i": i,
                 "ttft_s": r.ttft_s, "tpot_s": r.tpot_s,
                 "completion_tokens": r.completion_tokens,
+                "prompt_tokens": r.prompt_tokens,
                 "e2e_s": r.e2e_s, "error": r.error,
                 "finish_reason": r.finish_reason,
             })