From 7d5a6bc5cb404333d3485617578437c464c5e922 Mon Sep 17 00:00:00 2001
From: Sy03 <1370724210@qq.com>
Date: Fri, 2 Oct 2026 22:19:38 +0800
Subject: [PATCH] [Doc] Clarify CosyVoice3 lazy graph latency and memory
 headroom

Signed-off-by: Sy03 <1370724210@qq.com>
---
 .../deploy/cosyvoice3_packed_streaming_optimized_standard.yaml  | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/vllm_omni/deploy/cosyvoice3_packed_streaming_optimized_standard.yaml b/vllm_omni/deploy/cosyvoice3_packed_streaming_optimized_standard.yaml
index 48bcc562b..0d738a0b1 100644
--- a/vllm_omni/deploy/cosyvoice3_packed_streaming_optimized_standard.yaml
+++ b/vllm_omni/deploy/cosyvoice3_packed_streaming_optimized_standard.yaml
@@ -1,6 +1,8 @@
 # Opt-in single-Hopper streaming throughput profile: AR32 / codec32, Flow10.
 # Use --hf-overrides '{"cosyvoice3_sampling_mode": "standard"}' and stream=true.
 # F0 and HiFT stay FP32; only deterministic HiFT decode is graphed.
+# HiFT graphs capture on the third use of each exact shape, adding latency then.
+# The codec graph pool is not capped by gpu_memory_utilization; leave memory headroom.
 base_config: cosyvoice3_packed_streaming.yaml
 stages:
   - stage_id: 0