Add DeepSeek V3/R1 shared experts fusion (#4918)

2025-04-04 16:59:29 +08:00
parent 6ff9c6a5e7
commit 924ca7c92c
14 changed files with 536 additions and 36 deletions
--- a/sgl-kernel/tests/test_moe_align.py
+++ b/sgl-kernel/tests/test_moe_align.py
@@ -144,7 +144,7 @@ def moe_align_block_size_triton(
            [32, 64, 128, 256],  # block_size
            [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096],  # num_tokens
            [1, 2, 4, 8, 16, 32, 64],  # topk
-            [64, 160, 256],  #  num_experts
+            [64, 160, 256, 257, 260, 264],  #  num_experts
        )
    ),
 )