gmm tuning

Prayer3th · Prayer3th · commit 8175710c694c · 2025-09-17T17:48:51.000+08:00
diff --git a/python/sgl_jax/srt/layers/moe.py b/python/sgl_jax/srt/layers/moe.py
@@ -223,12 +223,12 @@ def _expert_parallel_forward_with_shard_map(
 
         # 转换为静态整数参数（在动态 m 值上使用最大值作为安全的静态值）
         static_tiling_gate = (
-            min(optimal_tiling_gate[0], 16384),  # 设置合理的最大值
+            optimal_tiling_gate[0],  # 设置合理的最大值
             optimal_tiling_gate[1],
             optimal_tiling_gate[2],
         )
         static_tiling_down = (
-            min(optimal_tiling_down[0], 16384),
+            optimal_tiling_down[0],
             optimal_tiling_down[1],
             optimal_tiling_down[2],
         )

Original file line number	Diff line number	Diff line change
`@@ -223,12 +223,12 @@ def _expert_parallel_forward_with_shard_map(`
`223`	`223`
`224`	`224`	`# 转换为静态整数参数（在动态 m 值上使用最大值作为安全的静态值）`
`225`	`225`	`static_tiling_gate = (`
`226`		`- min(optimal_tiling_gate[0], 16384), # 设置合理的最大值`
	`226`	`+ optimal_tiling_gate[0], # 设置合理的最大值`
`227`	`227`	`optimal_tiling_gate[1],`
`228`	`228`	`optimal_tiling_gate[2],`
`229`	`229`	`)`
`230`	`230`	`static_tiling_down = (`
`231`		`- min(optimal_tiling_down[0], 16384),`
	`231`	`+ optimal_tiling_down[0],`
`232`	`232`	`optimal_tiling_down[1],`
`233`	`233`	`optimal_tiling_down[2],`
`234`	`234`	`)`