Mar 16

1. 12:04

   feat: spectator mode - all visitors see the same streaming generation

   haiyangliu0
   b1e1a05
2. 11:50

   fix: move k_lens to GPU in SDPA fallback (tested locally)

   haiyangliu0
   7d73321
3. 11:44

   fix: build proper attention mask in SDPA fallback for text cross-attention

   haiyangliu0
   6c6483b
4. 11:37

   perf: increase server frame buffer target to 16 for batch fetching

   haiyangliu0
   7237651
5. 11:34

   revert: remove float16 override, bfloat16 is safer (float16 overflow risk)

   haiyangliu0
   2cefd84
6. 11:20

   perf: use float16 instead of bfloat16 in SDPA for T4 tensor core acceleration

   haiyangliu0
   80c3e53
7. 10:47

   fix: adapt model_manager to HF model API (no schedule_config/cfg_config dicts)

   haiyangliu0
   47551f4
8. 10:39

   fix: cast SDPA output back to original dtype + idempotent patching

   haiyangliu0
   e843211
9. 10:34

   fix: patch flash_attention with SDPA fallback for T4 (no flash-attn)

   haiyangliu0
   bb7e158
10. 09:45

    feat: initial FloodDiffusion streaming demo for HF Space⁠

    haiyangliu0 +1: Claude Opus 4.6 (1M context)
    a4f8eb3