[perf]: fuse the MiniMax H3 VSA gate combine, accumulate in place at inference

torch.addcmul keeps one full-sequence temporary under grad instead of two,
and when the sparse output is not tracked by autograd the gate branch
accumulates into it directly, allocating nothing. The product is no longer
rounded to bf16 before the add, so gated values move by at most half a bf16
ulp (towards the fp32 value). test_vsa_h3_backward.py passes on the Triton
and FA4 CuTe backends on a GB200.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NjFimeedTQWxgqSzP5xo4j

Author: SolitaryThinker

fastvideo/attention/backends/video_sparse_attn_h3.py

@@ -763,11 +763,18 @@ def forward(  # type: ignore[override]
             out_c = torch.matmul(torch.softmax(scores, dim=-1), v_pooled)  # [B, H, n_tiles, D]
             out_c = out_c.permute(0, 2, 1, 3).to(out.dtype)  # [B, n_tiles, H, D]
             batch, seq_len, heads, dim = out.shape
-            # Out-of-place: on the CuTe backend ``out`` is the tensor FA4's
-            # autograd node saved for its backward, so an in-place add here
-            # bumps its version counter and backward dies with "one of the
-            # variables needed for gradient computation has been modified".
             out_tiled = out.view(batch, n_tiles, tile_elems, heads, dim)
             gate_tiled = logical_gate.view(batch, n_tiles, tile_elems, heads, dim)
-            out = (out_tiled + out_c.unsqueeze(2) * gate_tiled).view(batch, seq_len, heads, dim)
+            out_c = out_c.unsqueeze(2)
+            if out.requires_grad or gate_tiled.dtype != out.dtype:
+                # Out-of-place: ``out`` is the tensor the attention kernel's
+                # autograd node saved for its backward, so an in-place add here
+                # bumps its version counter and backward dies with "one of the
+                # variables needed for gradient computation has been modified".
+                # Fused addcmul keeps one full-sequence temporary instead of two.
+                out = torch.addcmul(out_tiled, out_c, gate_tiled).view(batch, seq_len, heads, dim)
+            else:
+                # Inference: ``out`` is a fresh kernel output nobody else holds,
+                # so accumulate into it and allocate nothing.
+                out_tiled.addcmul_(out_c, gate_tiled)
         return out