Skip to main content

​
1. Model introduction

Qwen-Image is Alibaba’s MMDiT text-to-image model. In miles-diffusion it is the reference image recipe: the launcher is aligned flag-for-flag with flow_grpo’s pickscore_qwenimage config, so a run here is directly comparable to the published baseline. Key highlights for RL training:
  • flow_grpo-aligned. Per-prompt reward mean with a batch-wide std (--globalize-reward-std alone) is exactly flow_grpo’s PerPromptStatTracker with global_std=True; gaussian LoRA init and adam-beta2 0.999 match likewise.
  • RoPE caches rebuilt on CUDA. diffusers builds QwenEmbedRope’s frequency tables on CPU while sglang-d builds them on CUDA; the fp32 ULP difference drifts every block. In one frozen-weight comparison, noise_pred mean |Δ| was approximately 2e-2. postprocess_model_after_materialize rebuilds the caches on device after FSDP wrapping.
  • Variable-length text. Conditioning is padded per batch with the attention mask derived from txt_seq_lens — the mask itself is never transmitted from rollout.

​
2. Supported variants

Any checkpoint whose name matches qwen-image resolves to the same config. For a differently-named local directory — the usual case — add --diffusion-model-family qwen_image.

​
3. Family config

Registered in miles/backends/fsdp_utils/configs/qwen_image.py:

​
4. Launch

Canonical recipe: scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py — 4 colocate GPUs + 1 PickScore GPU, 512×512, PickScore reward. Status: 📈 V — Verified
Overrides go through the dataclass CLI, e.g. --num-rollout 50 --extra-args "--diffusion-kl-beta 0.02".

​
5. Recipe notes

  • Of the 10 denoising steps, only steps 3-4 are trained: --diffusion-sde-window-range 3,5 with --diffusion-num-sde-steps 2 leaves the window nowhere to move, so every sample trains the same two mid-schedule steps as SDE while the rest run ODE.
  • Batch arithmetic: 32 prompts × 16 samples = 512 samples per rollout, split into two optimizer steps of 256 samples; the 2 trained SDE steps expand those to 1024 train pairs.
  • --rollout-patch-group sgld applies the diffusers-op-parity patches on the engine. It costs a little rollout throughput and buys a much smaller train/log_prob_mean_abs_diff.

​
6. Pairs well with