From 176da1bbddbed865759d97942cf8038fdac16e82 Mon Sep 17 00:00:00 2001
From: strgrb <zhangkaihong.zkh@antgroup.com>
Date: Sat, 24 Jan 2026 13:35:14 +0800
Subject: [PATCH] Fix: mistake sigmoid in kda (#17508)

---
 python/sglang/srt/models/kimi_linear.py       |   3 +-
 test/registered/attention/test_kda_kernels.py | 123 ++++++++++++++++++
 2 files changed, 125 insertions(+), 1 deletion(-)
 create mode 100644 test/registered/attention/test_kda_kernels.py

diff --git a/python/sglang/srt/models/kimi_linear.py b/python/sglang/srt/models/kimi_linear.py
index 7efb8c2e9..9f6513654 100644
--- a/python/sglang/srt/models/kimi_linear.py
+++ b/python/sglang/srt/models/kimi_linear.py
@@ -314,14 +314,15 @@ class KimiDeltaAttention(nn.Module):
             self.v_conv1d.weight.size(0), self.v_conv1d.weight.size(2)
         )
 
-        beta = self.b_proj(hidden_states)[0].float().sigmoid()
         forget_gate = self.f_b_proj(self.f_a_proj(hidden_states)[0])[0]
 
+        beta = self.b_proj(hidden_states)[0].float()
         # fused_kda_gate is fused to KimiLinearAttentionBackend with decode
         if not forward_batch.forward_mode.is_decode():
             forget_gate = fused_kda_gate(
                 forget_gate, self.A_log, self.head_dim, g_bias=self.dt_bias
             )
+            beta = beta.sigmoid()
         beta = beta.unsqueeze(0)
         forget_gate = forget_gate.unsqueeze(0)
 
diff --git a/test/registered/attention/test_kda_kernels.py b/test/registered/attention/test_kda_kernels.py
new file mode 100644
index 000000000..e9c712db7
--- /dev/null
+++ b/test/registered/attention/test_kda_kernels.py
@@ -0,0 +1,123 @@
+import unittest
+
+import torch
+
+from sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent import (
+    fused_sigmoid_gating_delta_rule_update,
+)
+from sglang.srt.layers.attention.fla.kda import fused_kda_gate, fused_recurrent_kda
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=30, suite="stage-b-test-large-1-gpu")
+
+
+@unittest.skipIf(not torch.cuda.is_available(), "Test requires CUDA")
+class TestKDAFusedSigmoidGatingRecurrent(unittest.TestCase):
+    def setUp(self):
+        self.token_num = 4
+        self.query_start_loc = torch.tensor([0, 1, 2, 3, 4], device="cuda")
+        self.cache_indices = torch.tensor([0, 2, 5, 8], device="cuda")
+        self.local_num_heads = 8
+        self.head_dim = 128
+        self.cache_len = 64
+
+        self.A_log = torch.randn(
+            1, 1, self.local_num_heads, 1, dtype=torch.float32, device="cuda"
+        )
+        self.a = torch.randn(
+            1,
+            self.token_num,
+            self.local_num_heads * self.head_dim,
+            dtype=torch.bfloat16,
+            device="cuda",
+        )
+        self.dt_bias = torch.randn(
+            self.local_num_heads * self.head_dim, dtype=torch.bfloat16, device="cuda"
+        )
+        self.softplus_beta = 1.0
+        self.softplus_threshold = 20.0
+        self.q = torch.randn(
+            1,
+            self.token_num,
+            self.local_num_heads,
+            self.head_dim,
+            dtype=torch.bfloat16,
+            device="cuda",
+        )
+        self.k = torch.randn(
+            1,
+            self.token_num,
+            self.local_num_heads,
+            self.head_dim,
+            dtype=torch.bfloat16,
+            device="cuda",
+        )
+        self.v = torch.randn(
+            1,
+            self.token_num,
+            self.local_num_heads,
+            self.head_dim,
+            dtype=torch.bfloat16,
+            device="cuda",
+        )
+        self.beta = torch.randn(
+            1, self.token_num, self.local_num_heads, dtype=torch.bfloat16, device="cuda"
+        )
+
+        self.ssm_states = torch.zeros(
+            self.cache_len,
+            self.local_num_heads,
+            self.head_dim,
+            self.head_dim,
+            dtype=torch.float32,
+            device="cuda",
+        )
+
+    def run_fused(self):
+        ssm_states = self.ssm_states.clone()
+        core_attn_out = fused_sigmoid_gating_delta_rule_update(
+            A_log=self.A_log,
+            dt_bias=self.dt_bias,
+            q=self.q,
+            k=self.k,
+            v=self.v,
+            a=self.a,
+            b=self.beta,
+            initial_state_source=ssm_states,
+            initial_state_indices=self.cache_indices,
+            cu_seqlens=self.query_start_loc,
+            use_qk_l2norm_in_kernel=True,
+            softplus_beta=self.softplus_beta,
+            softplus_threshold=self.softplus_threshold,
+            is_kda=True,
+        )
+        return core_attn_out, ssm_states[self.cache_indices]
+
+    def run_kda(self):
+        b = self.beta.float().sigmoid()
+        g = fused_kda_gate(self.a, self.A_log, self.head_dim, g_bias=self.dt_bias)
+        initial_state = self.ssm_states[self.cache_indices].clone()
+        core_attn_out, last_state = fused_recurrent_kda(
+            q=self.q,
+            k=self.k,
+            v=self.v,
+            g=g,
+            beta=b,
+            initial_state=initial_state,
+            use_qk_l2norm_in_kernel=True,
+            cu_seqlens=self.query_start_loc,
+        )
+        return core_attn_out, last_state
+
+    def test_kda_fused_sigmoid_gating_recurrent(self):
+        core_attn_out, last_state = self.run_fused()
+        core_attn_out_ref, last_state_ref = self.run_kda()
+        abs_diff_out = (core_attn_out - core_attn_out_ref).abs().max()
+        abs_diff_state = (last_state - last_state_ref).abs().max()
+        print(f"{abs_diff_out=}, {abs_diff_state=}")
+        self.assertTrue(torch.allclose(core_attn_out, core_attn_out_ref))
+        self.assertTrue(torch.allclose(last_state, last_state_ref))
+
+
+if __name__ == "__main__":
+    unittest.main()