bugfix: correct attn output with base 2 or e (#28840)

Signed-off-by: augusto.yjh <augusto.yjh@antgroup.com>

vllm/attention/ops/common.py

@@ -21,6 +21,7 @@ def _correct_attn_cp_out_kernel(
     lse_idx,
     HEAD_DIM: tl.constexpr,
     N_ROUNDED: tl.constexpr,
+    IS_BASE_E: tl.constexpr,
 ):
     """
     Apply the all-gathered lses to correct each local rank's attention
@@ -55,9 +56,14 @@ def _correct_attn_cp_out_kernel(
     lse_max = tl.max(lse, axis=0)
     lse_max = tl.where(lse_max == -float("inf"), 0, lse_max)
     lse -= lse_max
-    lse_exp = tl.exp(lse)
+    if IS_BASE_E:
-    lse_acc = tl.sum(lse_exp, axis=0)
+        lse_exp = tl.exp(lse)
-    lse = tl.log(lse_acc)
+        lse_acc = tl.sum(lse_exp, axis=0)
+        lse = tl.log(lse_acc)
+    else:
+        lse_exp = tl.exp2(lse)
+        lse_acc = tl.sum(lse_exp, axis=0)
+        lse = tl.log2(lse_acc)
     lse += lse_max
 
     lse_offsets = batch_idx * lses_stride_B + head_idx * lses_stride_H
@@ -81,7 +87,7 @@ def _correct_attn_cp_out_kernel(
         -float("inf"),
         lse_finally,
     )
-    factor = tl.exp(lse_finally)
+    factor = tl.exp(lse_finally) if IS_BASE_E else tl.exp2(lse_finally)
     output = tl.load(outputs_ptr + output_offsets)
     output = output * factor
 
@@ -102,7 +108,11 @@ class CPTritonContext:
 
 
 def correct_attn_out(
-    out: torch.Tensor, lses: torch.Tensor, cp_rank: int, ctx: CPTritonContext
+    out: torch.Tensor,
+    lses: torch.Tensor,
+    cp_rank: int,
+    ctx: CPTritonContext,
+    is_lse_base_on_e: bool = True,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """Correct the attention output using the all-gathered lses.
 
@@ -163,8 +173,7 @@ def correct_attn_out(
         l_sH,
         cp_rank,
     )
-    const_args = {"HEAD_DIM": D, "N_ROUNDED": N}
+    const_args = {"HEAD_DIM": D, "N_ROUNDED": N, "IS_BASE_E": is_lse_base_on_e}
-
     ctx.call_kernel(_correct_attn_cp_out_kernel, grid, *regular_args, **const_args)
     return out, lse
 
@@ -174,6 +183,7 @@ def _cp_lse_common(
     cp_attn_lse: torch.Tensor,
     cp_group: GroupCoordinator,
     ctx: CPTritonContext | None = None,
+    is_lse_base_on_e=True,
 ):
     """
     cp_attn_out: [ B, H, D ]
@@ -193,7 +203,13 @@ def _cp_lse_common(
 
     cp_attn_lse = cp_attn_lse.contiguous()
     lses = cp_group.all_gather(cp_attn_lse, dim=0).view_as(lses)
-    out, lse = correct_attn_out(cp_attn_out, lses, cp_group.rank_in_group, ctx)
+    out, lse = correct_attn_out(
+        cp_attn_out,
+        lses,
+        cp_group.rank_in_group,
+        ctx,
+        is_lse_base_on_e=is_lse_base_on_e,
+    )
     return out, lse
 
 
@@ -203,12 +219,15 @@ def cp_lse_ag_out_rs(
     cp_group: GroupCoordinator,
     ctx: CPTritonContext | None = None,
     return_lse: bool = False,
+    is_lse_base_on_e=True,
 ):
     """
     cp_attn_out: [ B, H, D ]
     cp_attn_lse: [ B, H ]
     """
-    out, lse = _cp_lse_common(cp_attn_out, cp_attn_lse, cp_group, ctx=ctx)
+    out, lse = _cp_lse_common(
+        cp_attn_out, cp_attn_lse, cp_group, ctx=ctx, is_lse_base_on_e=is_lse_base_on_e
+    )
     out = cp_group.reduce_scatter(out, dim=1)
 
     if return_lse:
@@ -225,12 +244,15 @@ def cp_lse_ag_out_ar(
     cp_group: GroupCoordinator,
     ctx: CPTritonContext | None = None,
     return_lse: bool = False,
+    is_lse_base_on_e=True,
 ):
     """
     cp_attn_out: [ B, H, D ]
     cp_attn_lse: [ B, H ]
     """
-    out, lse = _cp_lse_common(cp_attn_out, cp_attn_lse, cp_group, ctx=ctx)
+    out, lse = _cp_lse_common(
+        cp_attn_out, cp_attn_lse, cp_group, ctx=ctx, is_lse_base_on_e=is_lse_base_on_e
+    )
     out = cp_group.all_reduce(out)
 
     if return_lse:

vllm/v1/attention/backends/flashinfer.py

@@ -249,7 +249,11 @@ class BatchDCPPrefillWrapper:
             return_lse=True,
         )
         output_context, lse_context = cp_lse_ag_out_rs(
-            output_context_tmp, lse_context_tmp, get_dcp_group(), return_lse=True
+            output_context_tmp,
+            lse_context_tmp,
+            get_dcp_group(),
+            return_lse=True,
+            is_lse_base_on_e=False,
         )
         lse_context = lse_context.transpose(0, 1).contiguous()
 
@@ -1335,7 +1339,10 @@ class FlashInferImpl(AttentionImpl):
                         return_lse=True,
                     )
                     output[:num_decode_tokens] = cp_lse_ag_out_rs(
-                        output_tmp, lse, get_dcp_group()
+                        output_tmp,
+                        lse,
+                        get_dcp_group(),
+                        is_lse_base_on_e=False,
                     )
                 else:
                     decode_wrapper.run(

vllm/v1/attention/backends/mla/common.py

@@ -2057,7 +2057,12 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
 
             # correct dcp attn_out with lse.
             if self.dcp_world_size > 1:
-                attn_out = cp_lse_ag_out_rs(attn_out, lse, get_dcp_group())
+                attn_out = cp_lse_ag_out_rs(
+                    attn_out,
+                    lse,
+                    get_dcp_group(),
+                    is_lse_base_on_e=not self._use_fi_prefill,
+                )
 
             # v_up projection
             self._v_up_proj(attn_out, out=output[:num_decode_tokens])