--- a/src/device/generate.py
+++ b/src/device/generate.py
@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 import os
+import re
 import sys
 import subprocess
 from dataclasses import dataclass
@@ -428,6 +429,19 @@
       cond = "(defined(__gfx90a__) || defined(__gfx942__) || defined(__gfx950__) || defined(__gfx1250__)) && defined(ENABLE_LL128)"
   elif fn.acc == "1":
       cond = "defined(__gfx942__) || defined(__gfx950__) || defined(__gfx1250__)"
+
+  if fn.ty in ("f8e4m3", "f8e5m2"):
+      # The struct-based fp8 fallback trips an assertion in vanilla llvm-24's
+      # AMDGPU backend (LegalizeDAG); only build fp8 kernels on targets with
+      # hardware fp8 (see src/include/rccl_float8.h). The guard must stay a
+      # flat OR list: dl_evaluate_guard() cannot express AND.
+      fp8 = ("gfx942", "gfx950", "gfx1200", "gfx1201", "gfx1250")
+      ll128 = " && defined(ENABLE_LL128)" if (cond and "ENABLE_LL128" in cond) else ""
+      archs = [a for a in re.findall(r"__gfx[0-9a-z]+__", cond or "") if a.strip("_") in fp8]
+      if not archs:
+          cond = "defined(__gfx0000__)"  # never built for any target
+      else:
+          cond = " || ".join("defined(%s)" % a for a in archs) + ll128
   return cond
 
 # Build the mangled function symbol suffix. The user-buffer registration mode is