fix: redo timings with kv caching enabled + add end-to-end plots

README.md

@@ -70,6 +70,17 @@ model, so the pipeline, text encoder, and VAE stay unchanged.
 > - The first distillation removing need for prompt upsampling into the LoRA, thus significantly reducing end-to-end generation.
 > The v0.1 adapters remain available.
 
+<table>
+  <tr>
+    <td width="50%">
+      <img src="assets/P-Q-Image-2.1-t2i-1024-end-to-end.png" alt="Estimated end-to-end text-to-image latency at 1024 × 1024" width="100%">
+    </td>
+    <td width="50%">
+      <img src="assets/P-Q-Image-2.1-i2i-1024-end-to-end.png" alt="Estimated end-to-end editing latency at 1024 × 1024 with one input image" width="100%">
+    </td>
+  </tr>
+</table>
+
 ## 6-step variants
 
 Pick **6-steps** for faithful base-model behavior, or **6-steps-upsampled** for
@@ -220,11 +231,6 @@ pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
 
 
 <table>
-  <tr>
-    <td>
-      <img src="assets/P-Q-Image-2.1-t2i-1024-all.png" alt="Text-to-image generation latency at 1024 pixels across base, 8-step, 6-step, 6-step-upsampled, and 5-step adapters" width="100%">
-    </td>
-  </tr>
   <tr>
     <td>
       <img src="assets/P-Q-Image-2.1-v02-generation.png" alt="Six-step generation latency at 1024 and 2048 pixels" width="100%">
@@ -237,10 +243,17 @@ pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
   </tr>
 </table>
 
-One NVIDIA H100 80GB, BF16, batch size 1. Median of 3 after one warmup, seed 42.
+One NVIDIA H100 80GB, BF16, batch size 1. v0.2 generation timings
+report the median of six measured requests pooled across both adapters (three
+per adapter), excluding one warmup per adapter and case.
 Generation includes prompt encoding, reference-image processing, denoising, and
 VAE decode. It excludes loading, warmup, fixture prep, saving, and network.
-No CFG, compilation, CPU offload, or diffusion KV caching.
+No CFG, compilation, or CPU offload. **Diffusion KV caching is enabled**, except
+for the case marked with an asterisk below.
+
+\* **2048 × 2048 editing with 3 input images:** KV caching was disabled due to
+memory constraints on the H100 80GB. The 2048 multi-image cases use expandable
+CUDA memory segments (`PYTORCH_ALLOC_CONF=expandable_segments:True`).
 
 Prompt upsampling is timed separately with
 [`Qwen-Image-2.1-PE-T2I`](https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I)
@@ -249,7 +262,7 @@ It covers request prep through parsed text, not image generation. vLLM runs
 eager, without prefix or image-processor caching, and with a fresh image ID so
 vision encoding is always counted. Autoregressive KV cache stays on.
 
-Text-to-image uses the official Qwen model-card prompt. Editing uses a fixed
+Text-to-image uses a fixed neon-sign prompt. Editing uses a fixed
 prompt and 1–3 reference images, square-cropped and resized. These cases are
 not matched to the v0.1 charts, and adding the two medians is not an
 end-to-end time.