less dram claim

Author: root

python/sglang/srt/layers/attention/attention_registry.py

@@ -56,6 +56,7 @@ def create_aiter_backend(runner):
 def create_wave_backend(runner):
     try:
         from sglang.srt.layers.attention.wave_backend import WaveAttnBackend
+
         return WaveAttnBackend(runner)
     except ImportError as e:
         if "wave_lang" in str(e):

test/srt/hicache/test_hicache_mla.py

@@ -13,9 +13,11 @@
 
 _is_hip = is_hip()
 if _is_hip:
-    hicache_args = ["--hicache-size", 200]
+    hicache_args = ["--hicache-size", 40]
+    mem_frac_args = ["--mem-fraction-static", 0.2]
 else:
     hicache_args = ["--hicache-ratio", 2]
+    mem_frac_args = []
 
 
 class TestHierarchicalMLA(CustomTestCase):
@@ -32,6 +34,7 @@ def setUpClass(cls):
                 "--enable-hierarchical-cache",
             ]
             + hicache_args,
+            + mem_frac_args,
         )
 
     @classmethod

test/srt/hicache/test_hicache_storage.py

@@ -26,9 +26,9 @@ def setUpClass(cls):
             other_args=[
                 "--enable-hierarchical-cache",
                 "--mem-fraction-static",
-                0.7,
+                0.7 if not _is_hip else 0.2,
                 "--hicache-size",
-                100 if not _is_hip else 200,
+                100 if not _is_hip else 40,
                 "--page-size",
                 "64",
                 "--hicache-storage-backend",