[DSA] Support trtllm sparse mla kernel for prefill batches (#21783)

python/sglang/srt/layers/attention/nsa_backend.py

@@ -1297,6 +1297,7 @@ class NativeSparseAttnBackend(
                 cos_sin_cache,
                 is_neox,
                 llama_4_scaling,
+                is_prefill=True,
             )
 
         if k is not None:
@@ -1929,6 +1930,7 @@ class NativeSparseAttnBackend(
         cos_sin_cache: Optional[torch.Tensor] = None,
         is_neox: Optional[bool] = False,
         llama_4_scaling: Optional[torch.Tensor] = None,
+        is_prefill: bool = False,
     ) -> torch.Tensor:
         """Forward using TRT-LLM sparse MLA kernel."""
         import flashinfer.decode
@@ -1990,6 +1992,13 @@ class NativeSparseAttnBackend(
 
         if envs.SGLANG_NSA_FUSE_TOPK.get():
             page_table_1 = topk_indices
+        elif is_prefill:
+            page_table_1 = transform_index_page_table_prefill(
+                page_table=metadata.page_table_1,
+                topk_indices=topk_indices,
+                extend_lens_cpu=metadata.nsa_extend_seq_lens_list,
+                page_size=1,
+            )
         else:
             page_table_1 = transform_index_page_table_decode(
                 page_table=metadata.page_table_1,

python/sglang/srt/server_args.py

@@ -1453,9 +1453,6 @@ class ServerArgs:
             if self.dp_size == 1 and major >= 10:
                 self.nsa_prefill_backend = "trtllm"
                 self.nsa_decode_backend = "trtllm"
-                logger.warning(
-                    "Flashmla is not supported on Blackwell device without DP attention. Set NSA prefill/decode backends to trtllm, which runs fast but loses a little accuracy."
-                )
             else:
                 # flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics
                 if not user_set_prefill:
@@ -1524,14 +1521,6 @@ class ServerArgs:
                         logger.warning(
                             f"Set dense attention kv len threshold to model index_topk={envs.SGLANG_NSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD.get()} for DeepSeek with DSA."
                         )
-                    if self.nsa_prefill_backend == "trtllm":
-                        # We temporarily set the threshold to 128k to avoid IMA error. Should be removed after supporting flashmla prefill impl with trtllm decode impl.
-                        envs.SGLANG_NSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD.set(
-                            128 * 1024
-                        )
-                        logger.warning(
-                            "TRTLLM sparse MLA kernel requires MHA as prefill impl, the threshold for dense attention is overridden. This will be fixed in the future."
-                        )
                 if self.is_attention_backend_not_set():
                     self.attention_backend = "nsa"
                     logger.info("Use nsa attention backend for DeepSeek with DSA.")

python/sglang/test/run_eval.py

@@ -20,10 +20,10 @@ from sglang.test.simple_eval_common import (
 def get_thinking_kwargs(args):
     thinking_mode = getattr(args, "thinking_mode", None)
     if thinking_mode in THINKING_MODE_CHOICES:
-        if thinking_mode == "deepseek-v3":
+        if thinking_mode in ["deepseek-v3", "kimi-k2"]:
             thinking_param = "thinking"
         else:
-            # Qwen3
+            # All models other than dpsk v3/kimi_k2
             thinking_param = "enable_thinking"
         return {thinking_param: True}
     return {}
@@ -267,7 +267,7 @@ def run_eval(args):
     return metrics
 
 
-THINKING_MODE_CHOICES = ["deepseek-v3", "qwen3"]
+THINKING_MODE_CHOICES = ["deepseek-v3", "qwen-3", "glm-45", "kimi-k2"]
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()