perf: bound throughput by corrected ttft

src/aiconfigurator/sdk/picking.py

@@ -70,10 +70,12 @@ def _build_disagg_summary_dict(
     Returns:
         Dict with keys matching ``common.ColumnsDisagg``.
     """
-    seq_s = min(
-        prefill_summary_dict["seq/s"] * prefill_num_worker * prefill_degradation_factor,
-        decode_summary_dict["seq/s"] * decode_num_worker * decode_degradation_factor,
-    )
+    prefill_capacity = prefill_summary_dict["seq/s"] * prefill_num_worker * prefill_degradation_factor
+    decode_capacity = decode_summary_dict["seq/s"] * decode_num_worker * decode_degradation_factor
+    prefill_ttft_capacity = prefill_summary_dict["bs"] * prefill_num_worker * 1000.0 / prefill_summary_dict["ttft"]
+
+    seq_s = min(prefill_capacity, decode_capacity, prefill_ttft_capacity)
+
     prefill_gpus = prefill_summary_dict["pp"] * prefill_summary_dict["tp"] * prefill_summary_dict["dp"]
     decode_gpus = decode_summary_dict["pp"] * decode_summary_dict["tp"] * decode_summary_dict["dp"]
     num_total_gpus = prefill_gpus * prefill_num_worker + decode_gpus * decode_num_worker

src/aiconfigurator/sdk/sweep.py

@@ -112,10 +112,10 @@ def _rate_match_dict(
     d = decode_summary_dict
     osl = p["osl"]
 
-    seq_s = min(
-        p["seq/s"] * prefill_num_worker * prefill_degradation,
-        d["seq/s"] * decode_num_worker * decode_degradation,
-    )
+    prefill_capacity = p["seq/s"] * prefill_num_worker * prefill_degradation
+    decode_capacity = d["seq/s"] * decode_num_worker * decode_degradation
+    prefill_ttft_capacity = p["bs"] * prefill_num_worker * 1000.0 / p["ttft"]
+    seq_s = min(prefill_capacity, decode_capacity, prefill_ttft_capacity)
     prefill_gpus = p["pp"] * p["tp"] * p["dp"]
     decode_gpus = d["pp"] * d["tp"] * d["dp"]
     num_total_gpus = prefill_gpus * prefill_num_worker + decode_gpus * decode_num_worker

tests/unit/sdk/sweep/test_rate_match_parity.py

@@ -140,6 +140,18 @@ def test_rate_match_with_custom_degradation_factors():
         )
 
 
+def test_rate_match_prefill_ttft_capacity_can_bound_throughput():
+    p = _make_prefill_dict(bs=1, ttft=500.0, **{"seq/s": 100.0})
+    d = _make_decode_dict(**{"seq/s": 100.0})
+
+    new_result = _rate_match_dict(p, 2, d, 2)
+    old_result = _build_disagg_summary_dict(p, 2, d, 2)
+
+    assert new_result["seq/s"] == 4.0
+    for key in new_result:
+        assert new_result[key] == old_result[key]
+
+
 def test_rate_match_zero_osl_does_not_divide_by_zero():
     """request_latency uses max(osl - 1, 0); osl=1 keeps decode_time=0."""
     p = _make_prefill_dict(osl=1)