@@ -9336,4 +9336,4 @@
     - "Cap --max-num-batched-tokens at 8192 on every TP2 and TP4 point, replacing 16384 on the thirteen points that still carried it; TP2 c64 was already 8192. The prefill chunk is now uniform across the arm, matching the B300 DSv4.1-Flash arm, which uses 8192 at every concurrency, and the B200 TP4 arm, which uses it from c8 up. The earlier 16384 was the upstream recipe default, trimmed only where a point ran short of KV; this cap is for decode tail latency, not for memory."
     - "A prefill chunk scheduled into the same engine step as a decode batch stalls every decoding sequence for the length of that chunk, so the stall scales with the chunk. Comparing the two TP4 ep1 arms of run 37080504692 (MI355X, 16384) and run 37070984585 (B200, 8192): MI355X ITL p50/p90 is 4.43/6.86 ms at c32 and 8.95/19.76 ms at c64, against 3.18/4.52 ms and 5.26/7.00 ms on B200. The p90-minus-p50 spread is 2.43 ms against 1.34 ms at c32 and 10.81 ms against 1.74 ms at c64, while the medians differ by a roughly constant 1.0-1.25 ms that chunk size does not reach."
     - "Expected to raise p90 interactivity from c16 up, where the spread term is 41% (c16), 47% (c32) and 71% (c64) of the decode gap, and to leave c1-c8 flat, where it is 11-12%. Throughput should not move at c1-c32, which the AgentX trace paces rather than the GPU (lane duty cycle 17-22%), and may rise at c64, where MI355X already reaches 49.9%. TTFT may regress slightly because the same prefill is cut into more chunks; TTFT does not enter the interactivity metric. No controlled A/B for this knob exists in the published runs, so the sweep decides."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3735