[mickqian]

Motivation

Two blocking device reads sit on the host's critical path:

1. Cosmos3, every step. The denoising loop checked the CFG window, and the Transfer control-CFG window, with t.item() on the device timestep. So the host waited for each step's kernels to finish before it launched the next step.
2. Qwen-Image-2.1 edit, every request. The edit path built the VAE latent mean/std with new_tensor right after launching the condition-image VAE encode. That pageable upload made the host wait for the encode before it could start the prompt's image preprocessing.

Modifications

• cosmos3.py: the loop takes one host copy of the timestep schedule (timesteps.tolist()), and _cfg_active_at reads the window from those floats. The per-step decisions and kernels are unchanged.
• qwen_image21.py: QwenImage21EncodingStage uploads the latent mean/std once per (dtype, device) and reuses them, so the preprocessing overlaps the encode. The normalization is bit-identical.

Tests:

• test_cosmos3_rollout.py::test_cfg_window_check_reads_host_timesteps: runs the denoising loop with a [400, 1000] CFG window over timesteps whose item() / float() / bool() raise. The window must still choose batched CFG at 999, 700 and 400, and the cond branch alone at 399 and 100. It fails on main.
• test_qwen_image21.py::test_condition_latents_reuse_uploaded_latent_stats: two edit requests must produce the reference-normalized condition latents bitwise, and the second request must reuse the first one's mean/std tensors (checked by identity).

No files are shared with #45942, #45943, #45944 or #45948.

Accuracy Tests

The output pixels are identical across arms for both models.

Speed Tests and Profiling

Server benchmark, 2 interleaved rounds x 5 repeats, e2e. Each commit was measured on its own against the same base.

case	GPUs	base (round 1 / 2)	this PR (round 1 / 2)
Cosmos3-Nano T2I 720p, eager, CFG parallel	2xH200	0.928 / 0.929 s	0.920 / 0.920 s (−8 ms)
Qwen-Image-2.1 edit 1024	1xH200	5.300 / 5.302 s	5.285 / 5.284 s (−17 ms)

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #39900393237
Latest PR Test (Extra): ❌ Run #39900393310
Latest PR Test (AMD ROCm 10): ⏳ Run #39900393090

[mickqian]

/tag-and-rerun-ci