[AMD] [GLM5] skip redundant -inf pre-fill of HIP indexer MQA-logits (#28757)

Co-authored-by: Raiden-Makoto <Raiden-Makoto@users.noreply.github.com>

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

@@ -870,8 +870,18 @@ class Indexer(MultiPlatformOp):
                     from aiter.ops.triton.fp8_mqa_logits import fp8_mqa_logits
 
                     kv, scale = kv_fp8
+                    # Match the CUDA deep_gemm path (clean_logits=False): the topk
+                    # transform masks invalid positions via ks/ke/lengths, so the
+                    # -inf pre-fill of the [tokens x seq_len_kv] logits buffer is
+                    # redundant and grows quadratically with context length.
                     logits = fp8_mqa_logits(
-                        q_fp8[:q_offset], kv, scale, weights[:q_offset], ks, ke
+                        q_fp8[:q_offset],
+                        kv,
+                        scale,
+                        weights[:q_offset],
+                        ks,
+                        ke,
+                        clean_logits=False,
                     )
                 else:
                     logits = deep_gemm.fp8_mqa_logits(
@@ -915,6 +925,7 @@ class Indexer(MultiPlatformOp):
                     from aiter.ops.triton.fp8_mqa_logits import fp8_mqa_logits
 
                     kv, scale = kv_fp8
+                    # clean_logits=False: topk transform handles masking (see above)
                     logits_chunk = fp8_mqa_logits(
                         q_fp8[start:end],
                         kv,
@@ -922,6 +933,7 @@ class Indexer(MultiPlatformOp):
                         weights[start:end],
                         ks[start:end],
                         ke[start:end],
+                        clean_logits=False,
                     )
                 else:
                     logits_chunk = deep_gemm.fp8_mqa_logits(