[PERF] Qwen3-next MTP speedup (change bool mask indexing to index_select /...

[PERF] Qwen3-next MTP speedup (change bool mask indexing to index_select / index_copy to reduce d2h) (#26437)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>

vllm/model_executor/layers/fla/ops/utils.py

@@ -45,7 +45,7 @@ def tensor_cache(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]
     """
 
     cache_entries: tuple[tuple | None, dict | None, Any] = []
-    cache_size = 4
+    cache_size = 8
 
     @functools.wraps(fn)
     def wrapper(*args: Any, **kwargs: Any) -> Any:

vllm/model_executor/models/qwen3_next.py

@@ -423,7 +423,7 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
             (query, key),
         )
         value = rearrange(value, "l (h d) -> 1 l h d", d=self.head_v_dim)
-        return query, key, value
+        return query.contiguous(), key.contiguous(), value.contiguous()
 
     def forward(
         self,
@@ -455,7 +455,8 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
         spec_query_start_loc = attn_metadata.spec_query_start_loc
         non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
         spec_sequence_masks = attn_metadata.spec_sequence_masks
-        spec_token_masks = attn_metadata.spec_token_masks
+        spec_token_indx = attn_metadata.spec_token_indx
+        non_spec_token_indx = attn_metadata.non_spec_token_indx
         spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor  # noqa: E501
         non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor  # noqa: E501
         self_kv_cache = self.kv_cache[forward_context.virtual_engine]
@@ -463,8 +464,6 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
         ssm_state = self_kv_cache[1]
         num_actual_tokens = attn_metadata.num_actual_tokens
         num_accepted_tokens = attn_metadata.num_accepted_tokens
-        if spec_token_masks is not None:
-            spec_token_masks = spec_token_masks[:num_actual_tokens]
 
         # 1. Set up dimensions for reshapes later
         projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states[:num_actual_tokens])
@@ -487,8 +486,8 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
                 mixed_qkv_spec = mixed_qkv
                 mixed_qkv_non_spec = None
             else:
-                mixed_qkv_spec = mixed_qkv[spec_token_masks]
-                mixed_qkv_non_spec = mixed_qkv[~spec_token_masks]
+                mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
+                mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx)
         else:
             mixed_qkv_spec = None
             mixed_qkv_non_spec = mixed_qkv
@@ -558,10 +557,10 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
                 g_non_spec = None
                 beta_non_spec = None
             else:
-                g_spec = g[:, spec_token_masks]
-                beta_spec = beta[:, spec_token_masks]
-                g_non_spec = g[:, ~spec_token_masks]
-                beta_non_spec = beta[:, ~spec_token_masks]
+                g_spec = g.index_select(1, spec_token_indx)
+                beta_spec = beta.index_select(1, spec_token_indx)
+                g_non_spec = g.index_select(1, non_spec_token_indx)
+                beta_non_spec = beta.index_select(1, non_spec_token_indx)
         else:
             g_spec = None
             beta_spec = None
@@ -638,8 +637,9 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
                 dtype=core_attn_out_non_spec.dtype,
                 device=core_attn_out_non_spec.device,
             )
-            core_attn_out[:, spec_token_masks] = core_attn_out_spec
-            core_attn_out[:, ~spec_token_masks] = core_attn_out_non_spec
+            core_attn_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
+            core_attn_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)
+
         elif spec_sequence_masks is not None:
             core_attn_out = core_attn_out_spec
         else:

vllm/v1/attention/backends/gdn_attn.py

@@ -47,9 +47,9 @@ class GDNAttentionMetadata:
         None  # shape: [batch - num_spec_decodes,]
     )
     spec_sequence_masks: torch.Tensor | None = None  # shape: [batch,]
-    spec_token_masks: torch.Tensor | None = (
-        None  # shape: [num_prefill_tokens + num_decode_tokens,]
-    )
+    spec_token_indx: torch.Tensor | None = None
+    non_spec_token_indx: torch.Tensor | None = None
+
     num_accepted_tokens: torch.Tensor | None = None  # shape: [batch,]
 
     # The following attributes are for triton implementation of causal_conv1d
@@ -105,9 +105,14 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
             dtype=torch.bool,
             device=device,
         )
-        self.spec_token_masks = torch.empty(
+        self.spec_token_indx = torch.empty(
             (self.decode_cudagraph_max_bs * (self.num_spec + 1),),
-            dtype=torch.bool,
+            dtype=torch.int32,
+            device=device,
+        )
+        self.non_spec_token_indx = torch.empty(
+            (self.decode_cudagraph_max_bs * (self.num_spec + 1),),
+            dtype=torch.int32,
             device=device,
         )
         self.spec_query_start_loc = torch.empty(
@@ -166,7 +171,8 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                 split_decodes_and_prefills(m, decode_threshold=1)
             )
             num_spec_decode_tokens = 0
-            spec_token_masks = None
+            spec_token_indx = None
+            non_spec_token_indx = None
             spec_state_indices_tensor = None
             non_spec_state_indices_tensor = m.block_table_tensor[:, 0]
             spec_query_start_loc = None
@@ -180,18 +186,23 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
             num_prefills = non_spec_query_lens.size(0) - num_decodes
             num_decode_tokens = num_decodes
             num_prefill_tokens = non_spec_query_lens.sum().item() - num_decode_tokens
+            num_spec_decode_tokens = (
+                query_lens.sum().item() - num_prefill_tokens - num_decode_tokens
+            )
 
             if num_prefills == 0 and num_decodes == 0:
-                spec_token_masks = torch.ones(
-                    (
-                        min(
+                spec_token_size = min(
                     num_spec_decodes * (self.num_spec + 1),
                     query_start_loc[-1].item(),
                 )
-                    ),
-                    dtype=torch.bool,
+                spec_token_indx = torch.arange(
+                    spec_token_size,
+                    dtype=torch.int32,
                     device=query_start_loc.device,
                 )
+                non_spec_token_indx = torch.empty(
+                    0, dtype=torch.int32, device=query_start_loc.device
+                )
                 spec_state_indices_tensor = m.block_table_tensor[:, : self.num_spec + 1]
                 non_spec_state_indices_tensor = None
                 spec_query_start_loc = query_start_loc
@@ -200,6 +211,11 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                 spec_token_masks = torch.repeat_interleave(
                     spec_sequence_masks, query_lens
                 )
+                index = torch.argsort(spec_token_masks)
+                num_non_spec_tokens = num_prefill_tokens + num_decode_tokens
+                non_spec_token_indx = index[:num_non_spec_tokens]
+                spec_token_indx = index[num_non_spec_tokens:]
+
                 spec_state_indices_tensor = m.block_table_tensor[
                     spec_sequence_masks, : self.num_spec + 1
                 ]
@@ -226,9 +242,6 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                     out=non_spec_query_start_loc[1:],
                 )
 
-            num_spec_decode_tokens = (
-                query_lens.sum().item() - num_prefill_tokens - num_decode_tokens
-            )
             assert num_accepted_tokens is not None
             num_accepted_tokens = num_accepted_tokens[spec_sequence_masks]
 
@@ -274,12 +287,18 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
             spec_sequence_masks = self.spec_sequence_masks[:batch_size]
             spec_sequence_masks[num_spec_decodes:].fill_(False)
 
-            assert spec_token_masks is not None
-            self.spec_token_masks[: spec_token_masks.size(0)].copy_(
-                spec_token_masks, non_blocking=True
+            assert non_spec_token_indx is not None and spec_token_indx is not None
+            self.non_spec_token_indx[: non_spec_token_indx.size(0)].copy_(
+                non_spec_token_indx, non_blocking=True
+            )
+            non_spec_token_indx = self.non_spec_token_indx[
+                : non_spec_token_indx.size(0)
+            ]
+
+            self.spec_token_indx[: spec_token_indx.size(0)].copy_(
+                spec_token_indx, non_blocking=True
             )
-            spec_token_masks = self.spec_token_masks[:num_actual_tokens]
-            spec_token_masks[spec_token_masks.size(0) :].fill_(False)
+            spec_token_indx = self.spec_token_indx[: spec_token_indx.size(0)]
 
             self.spec_query_start_loc[: num_spec_decodes + 1].copy_(
                 spec_query_start_loc, non_blocking=True
@@ -332,7 +351,8 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
             spec_state_indices_tensor=spec_state_indices_tensor,
             non_spec_state_indices_tensor=non_spec_state_indices_tensor,
             spec_sequence_masks=spec_sequence_masks,
-            spec_token_masks=spec_token_masks,
+            spec_token_indx=spec_token_indx,
+            non_spec_token_indx=non_spec_token_indx,
             num_accepted_tokens=num_accepted_tokens,
             nums_dict=nums_dict,
             batch_ptr=batch_ptr,