From 7edcdd5ae618bf0386c95b43a69d741ff825acb9 Mon Sep 17 00:00:00 2001
From: vorapolsiloai <115975949+vorapolsiloai@users.noreply.github.com>
Date: Tue, 8 Sep 2026 12:49:47 +0300
Subject: [PATCH] [AMD] Skip AITER FP8 ASM prefill when GQA is unsupported
 (#38467)

---
 .../srt/layers/attention/aiter_backend.py     | 18 +++++++++++++
 .../attention/test_aiter_fp8_asm_gqa.py       | 27 +++++++++++++++++++
 2 files changed, 45 insertions(+)
 create mode 100644 test/registered/unit/layers/attention/test_aiter_fp8_asm_gqa.py

diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py
index 3fd9f76a7..3d0522c30 100755
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -154,6 +154,18 @@ class ForwardMetadata:
 _AITER_PARTITION_SIZE_ROCM = 256
 
 
+# AITER's gfx950 FP8 FMHA ASM kernels only cover these GQA ratios. Other
+# ratios (e.g. Qwen3.8-27B 24Q/4KV = 6) must not take the pertensor shortcut.
+_AITER_FP8_ASM_GQA_RATIOS = frozenset({1, 2, 4, 8, 16})
+
+
+def _aiter_fp8_asm_supports_gqa(num_q_heads: int, num_kv_heads: int) -> bool:
+    """Whether AITER's FP8 FMHA ASM kernel supports this GQA ratio."""
+    if num_kv_heads <= 0 or num_q_heads % num_kv_heads != 0:
+        return False
+    return (num_q_heads // num_kv_heads) in _AITER_FP8_ASM_GQA_RATIOS
+
+
 def _asm_context_prefill_gather_indices(
     kv_indptr: torch.Tensor,
     kv_indices: torch.Tensor,
@@ -2871,6 +2883,9 @@ class AiterAttnBackend(AttentionBackend):
                 and layer.qk_head_dim == 256
                 and layer.v_head_dim == 256
                 and self.kv_cache_dtype == fp8_dtype
+                and _aiter_fp8_asm_supports_gqa(
+                    layer.tp_q_head_num, layer.tp_k_head_num
+                )
                 and not self.kv_cache_is_vectorized_5d
                 and self.forward_metadata.max_kv_len is not None
             ):
@@ -2936,6 +2951,9 @@ class AiterAttnBackend(AttentionBackend):
                 and layer.qk_head_dim == 256
                 and layer.v_head_dim == 256
                 and self.kv_cache_dtype == fp8_dtype
+                and _aiter_fp8_asm_supports_gqa(
+                    layer.tp_q_head_num, layer.tp_k_head_num
+                )
             ):
                 q_c = q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim)
                 k_c = k.contiguous().view(-1, layer.tp_k_head_num, layer.head_dim)
diff --git a/test/registered/unit/layers/attention/test_aiter_fp8_asm_gqa.py b/test/registered/unit/layers/attention/test_aiter_fp8_asm_gqa.py
new file mode 100644
index 000000000..381313bd2
--- /dev/null
+++ b/test/registered/unit/layers/attention/test_aiter_fp8_asm_gqa.py
@@ -0,0 +1,27 @@
+"""Tests for the AITER FP8 FMHA ASM GQA routing guard."""
+
+import unittest
+
+from sglang.srt.layers.attention.aiter_backend import _aiter_fp8_asm_supports_gqa
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestAiterFp8AsmSupportsGqa(unittest.TestCase):
+    def test_supported_asm_gqa_ratios(self):
+        for ratio in (1, 2, 4, 8, 16):
+            with self.subTest(ratio=ratio):
+                self.assertTrue(_aiter_fp8_asm_supports_gqa(ratio * 4, 4))
+
+    def test_rejects_qwen38_27b_gqa6(self):
+        self.assertFalse(_aiter_fp8_asm_supports_gqa(24, 4))
+
+    def test_rejects_other_unsupported_ratios(self):
+        for num_q_heads, num_kv_heads in ((12, 4), (32, 1), (24, 0), (23, 4), (0, 4)):
+            with self.subTest(num_q_heads=num_q_heads, num_kv_heads=num_kv_heads):
+                self.assertFalse(_aiter_fp8_asm_supports_gqa(num_q_heads, num_kv_heads))
+
+
+if __name__ == "__main__":
+    unittest.main()