fix: restore flash-attn 2.8.3 — hard import in trellis attention module

requirements.txt

@@ -47,12 +47,9 @@ kornia==0.8.2
 timm==0.9.16
 gradio_litmodel3d==0.0.1
 
-# flash_attn: REMOVED (May 2026)
-# Both HF Space and RunPod use ATTN_BACKEND=xformers. Flash-attn is an optional
-# accelerator that xformers picks up if present — it is NOT required. Installing
-# flash-attn 2.8.3 (the only wheel for torch 2.8) caused RunPod's
-# xformers==0.0.32.post2 to reject it (caps at <=2.8.2). Removing it here means
-# both environments use xformers' built-in attention, which is consistent and fast.
+# flash_attn: HARD DEPENDENCY — trellis/modules/attention/full_attn.py does
+# a direct 'import flash_attn' with no fallback. Must match torch 2.8 + CUDA 12.
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
 
 
 # --- Export / API ---