[diffusion] multi-platform: support diffusion on amd and fix encoder loading on MI325 (#13760)

Co-authored-by: Sabre Shao <sabre.shao@amd.com>
Co-authored-by: Yusheng (Ethan) Su <yushengsu.thu@gmail.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com>
Co-authored-by: xsun <sunxiao04@gmail.com>

python/sglang/multimodal_gen/configs/models/dits/base.py

@@ -28,6 +28,7 @@ class DiTArchConfig(ArchConfig):
             AttentionBackendEnum.SLIDING_TILE_ATTN,
             AttentionBackendEnum.SAGE_ATTN,
             AttentionBackendEnum.FA,
+            AttentionBackendEnum.AITER,
             AttentionBackendEnum.TORCH_SDPA,
             AttentionBackendEnum.VIDEO_SPARSE_ATTN,
             AttentionBackendEnum.VMOBA_ATTN,

python/sglang/multimodal_gen/configs/sample/sampling_params.py

@@ -6,6 +6,7 @@ import dataclasses
 import hashlib
 import json
 import math
+import os
 import os.path
 import re
 import time
@@ -201,6 +202,11 @@ class SamplingParams:
         if self.height is None:
             self.height_not_provided = True
 
+        # Allow env var to override num_inference_steps (for faster CI testing on AMD)
+        env_steps = os.environ.get("SGLANG_TEST_NUM_INFERENCE_STEPS")
+        if env_steps is not None and self.num_inference_steps is not None:
+            self.num_inference_steps = int(env_steps)
+
     def check_sampling_param(self):
         if self.prompt_path and not self.prompt_path.endswith(".txt"):
             raise ValueError("prompt_path must be a txt file")