Make draft attention backends extensible (#35932)

Co-authored-by: Yichao Fu <yichaofu@meta.com>

python/sglang/srt/server_args.py

@@ -208,6 +208,17 @@ ATTENTION_BACKEND_CHOICES = [
     "intel_xpu",
 ]
 
+# trtllm_mha is valid for decode-only dense-MQA drafts. DFLASH rejects it
+# earlier when its per-layer attention requirements are not met.
+DRAFT_ATTENTION_BACKEND_CHOICES = [
+    "flashinfer",
+    "fa3",
+    "fa4",
+    "triton",
+    "ascend",
+    "trtllm_mha",
+]
+
 # Attention backends whose kernels read the chunked prefix-cache layout.
 # Out-of-tree platforms may extend this list (via
 # add_chunked_prefix_cache_attention_backend) before ServerArgs construction;
@@ -406,6 +417,10 @@ def add_attention_backend_choices(choices):
     ATTENTION_BACKEND_CHOICES.extend(choices)
 
 
+def add_draft_attention_backend_choices(choices):
+    DRAFT_ATTENTION_BACKEND_CHOICES.extend(choices)
+
+
 def add_chunked_prefix_cache_attention_backend(backend_name):
     CHUNKED_PREFIX_CACHE_SUPPORTED_ATTENTION_BACKENDS.append(backend_name)