Fix correctness issue for triton decoding kernel (#2479)

python/sglang/srt/layers/attention/triton_ops/decode_attention.py

@@ -32,7 +32,7 @@
 logger = logging.getLogger(__name__)
 
 # TODO: Remove this when triton>=3.2.0. This issue will not affect performance and accuracy.
-logger.warn(
+logger.warning(
     "The following error message 'operation scheduled before its operands' can be ignored."
 )
 
@@ -474,6 +474,7 @@ def _decode_grouped_att_m_fwd(
 def _fwd_kernel_stage2(
     Mid_O,
     O,
+    B_Seqlen,
     stride_mid_ob,
     stride_mid_oh,
     stride_mid_os,
@@ -486,6 +487,8 @@ def _fwd_kernel_stage2(
     cur_batch = tl.program_id(0)
     cur_head = tl.program_id(1)
 
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+
     offs_d = tl.arange(0, BLOCK_DV)
     mask_d = offs_d < Lv
 
@@ -497,19 +500,24 @@ def _fwd_kernel_stage2(
     offs_logic = cur_batch * stride_mid_ob + cur_head * stride_mid_oh + Lv
 
     for split_kv_id in range(0, NUM_KV_SPLITS):
-        tv = tl.load(
-            Mid_O + offs_v + split_kv_id * stride_mid_os, mask=mask_d, other=0.0
-        )
-        tlogic = tl.load(Mid_O + offs_logic + split_kv_id * stride_mid_os)
-        n_e_max = tl.maximum(tlogic, e_max)
+        kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+        split_kv_start = kv_len_per_split * split_kv_id
+        split_kv_end = tl.minimum(split_kv_start + kv_len_per_split, cur_batch_seq_len)
 
-        old_scale = tl.exp(e_max - n_e_max)
-        acc *= old_scale
-        exp_logic = tl.exp(tlogic - n_e_max)
-        acc += exp_logic * tv
+        if split_kv_end > split_kv_start:
+            tv = tl.load(
+                Mid_O + offs_v + split_kv_id * stride_mid_os, mask=mask_d, other=0.0
+            )
+            tlogic = tl.load(Mid_O + offs_logic + split_kv_id * stride_mid_os)
+            n_e_max = tl.maximum(tlogic, e_max)
 
-        e_sum = e_sum * old_scale + exp_logic
-        e_max = n_e_max
+            old_scale = tl.exp(e_max - n_e_max)
+            acc *= old_scale
+            exp_logic = tl.exp(tlogic - n_e_max)
+            acc += exp_logic * tv
+
+            e_sum = e_sum * old_scale + exp_logic
+            e_max = n_e_max
 
     tl.store(
         O + cur_batch * stride_obs + cur_head * stride_oh + offs_d,
@@ -523,6 +531,7 @@ def _decode_softmax_reducev_fwd(
     q,
     o,
     v_buffer,
+    b_seq_len,
     num_kv_splits,
 ):
     batch, head_num = q.shape[0], q.shape[1]
@@ -541,6 +550,7 @@ def _decode_softmax_reducev_fwd(
     _fwd_kernel_stage2[grid](
         logits,
         o,
+        b_seq_len,
         logits.stride(0),
         logits.stride(1),
         logits.stride(2),
@@ -580,7 +590,7 @@ def decode_attention_fwd_normal(
         sm_scale,
         logit_cap,
     )
-    _decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, num_kv_splits)
+    _decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, b_seq_len, num_kv_splits)
 
 
 def decode_attention_fwd_grouped(
@@ -608,7 +618,7 @@ def decode_attention_fwd_grouped(
         sm_scale,
         logit_cap,
     )
-    _decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, num_kv_splits)
+    _decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, b_seq_len, num_kv_splits)
 
 
 def decode_attention_fwd(

test/srt/test_triton_attention_kernels.py

@@ -232,9 +232,9 @@ def test_decode_attention(self):
         for B, H_Q, H_KV, D in configs:
             self._test_decode_attention_once(B, H_Q, H_KV, D)
 
-    def _test_grouped_decode_attention_once(self, B, H_Q, H_KV, D, D_V):
+    def _test_grouped_decode_attention_once(self, B, S, H_Q, H_KV, D, D_V):
         dtype = torch.bfloat16
-        seq_len = 128  # This represents the number of tokens already in the sequence
+        seq_len = S  # This represents the number of tokens already in the sequence
         total_tokens = B * seq_len
         sm_scale = 1.0 / (D**0.5)
         num_kv_splits = 8
@@ -300,6 +300,7 @@ def _test_grouped_decode_attention_once(self, B, H_Q, H_KV, D, D_V):
         self.assertTrue(torch.allclose(o, o_grouped, atol=3e-2))
 
     def test_grouped_decode_attention(self):
+        seq_lens = [5, 100, 128, 500]
         configs = [
             (2, 16, 16, 64, 64),
             (2, 16, 1, 64, 64),
@@ -309,8 +310,9 @@ def test_grouped_decode_attention(self):
             (2, 128, 1, 576, 512),
         ]
 
-        for B, H_Q, H_KV, D, D_V in configs:
-            self._test_grouped_decode_attention_once(B, H_Q, H_KV, D, D_V)
+        for S in seq_lens:
+            for B, H_Q, H_KV, D, D_V in configs:
+                self._test_grouped_decode_attention_once(B, S, H_Q, H_KV, D, D_V)
 
 
 if __name__ == "__main__":