optimize: avoid fla l2-norm recompilation by token count (#31558)

python/sglang/kernels/ops/attention/fla/l2norm.py

@@ -51,13 +51,12 @@ def l2norm_fwd_kernel1(
 #     ],
 #     key=["D", "NB"],
 # )
-@triton.jit
+@triton.jit(do_not_specialize=["T"])
 def l2norm_fwd_kernel(
     x,
     y,
     eps,
-    NB: tl.constexpr,
-    T: tl.constexpr,
+    T,
     D: tl.constexpr,
     BT: tl.constexpr,
     BD: tl.constexpr,
@@ -91,7 +90,6 @@ def l2norm_fwd(
         raise RuntimeError("This layer doesn't support feature dim >= 64KB.")
 
     if D <= 512:
-        NB = triton.cdiv(T, 2048)
 
         def grid(meta):
             return (triton.cdiv(T, meta["BT"]),)
@@ -100,7 +98,6 @@ def l2norm_fwd(
             x,
             y,
             eps,
-            NB=NB,
             T=T,
             D=D,
             BD=BD,