fix: reorder DSA indexer dual-stream ops to avoid CUDA graph stream explosion (#30025)

Co-authored-by: Khoa Pham <264503018+kpham-sgl@users.noreply.github.com>

python/sglang/srt/environ.py

@@ -668,7 +668,7 @@ class Envs:
     SGLANG_USE_FUSED_METADATA_COPY = EnvBool(True)
     SGLANG_DSA_USE_FUSED_METADATA_GENERATION = EnvBool(True)
     SGLANG_DSA_TOPK_BROADCAST = EnvBool(False)
-    SGLANG_DISABLE_DSA_INDEXER_FUSION = EnvBool(True)
+    SGLANG_DISABLE_DSA_INDEXER_FUSION = EnvBool(False)
 
     # sgl-kernel
     SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK = EnvBool(False)

python/sglang/srt/layers/attention/dsa/dsa_indexer.py

@@ -739,6 +739,13 @@ class Indexer(MultiPlatformOp):
 
         current_stream.wait_stream(self.alt_stream)
         self.alt_stream.wait_stream(current_stream)
+        q_fp8, weights = fused_q_indexer_rope_first_quant(
+            q.contiguous(),
+            weights_raw,
+            q_scale_gate,
+            self._indexer_cos_sin_cache,
+            positions,
+        )
         with torch.cuda.stream(self.alt_stream):
             self._fused_k_prepare_and_store(
                 key,
@@ -749,14 +756,6 @@ class Indexer(MultiPlatformOp):
                 out_cache_loc=out_cache_loc,
             )
 
-        q_fp8, weights = fused_q_indexer_rope_first_quant(
-            q.contiguous(),
-            weights_raw,
-            q_scale_gate,
-            self._indexer_cos_sin_cache,
-            positions,
-        )
-
         current_stream.wait_stream(self.alt_stream)
         return q_fp8, weights