Support flashinfer for Gemma3 prefill

Gemma3 uses bidirectional attention for images. Flashinfer
supports custom masks. Hook up the mask with flashinfer, so that we do
not have to use the slower SDPA implementation for prefills with images.

Author: danieldk

server/text_generation_server/layers/attention/flashinfer.py

@@ -45,6 +45,7 @@ def use_prefill_with_paged_kv_state(
     state: flashinfer.BatchPrefillWithPagedKVCacheWrapper,
     block_tables: torch.Tensor,
     cu_seqlens: torch.Tensor,
+    custom_mask: Optional[torch.Tensor],
     input_lengths: torch.Tensor,
     num_heads: int,
     num_kv_heads: int,
@@ -88,6 +89,7 @@ def use_prefill_with_paged_kv_state(
             paged_kv_indptr=indptr,
             paged_kv_indices=block_tables,
             paged_kv_last_page_len=last_page_len,
+            custom_mask=custom_mask,
             num_qo_heads=num_heads,
             num_kv_heads=num_kv_heads,
             head_dim_qk=head_size,

server/text_generation_server/models/custom_modeling/flash_gemma3_modeling.py

@@ -45,6 +45,7 @@
 from text_generation_server.layers.layernorm import (
     FastRMSNorm,
 )
+from text_generation_server.models.globals import ATTENTION
 from text_generation_server.utils.weights import UnquantizedWeight
 from transformers.activations import ACT2FN
 from text_generation_server.layers.attention import (
@@ -248,7 +249,7 @@ def forward(
 
         # Prefill
         if cu_seqlen_prefill is not None:
-            if attention_mask is None:
+            if attention_mask is None or ATTENTION == "flashinfer":
                 # flash attention
                 attn_output = attention(
                     query=query,
@@ -701,8 +702,16 @@ def __init__(self, prefix, config, weights):
         )
 
     def get_attention_mask(
-        self, input_ids, max_s, cu_seqlen_prefill, dtype, image_token_mask
+        self,
+        input_ids: torch.Tensor,
+        cu_seqlen_prefill: Optional[torch.Tensor],
+        dtype: torch.dtype,
+        bool_mask: bool = False,
     ):
+        image_token_mask = (input_ids == self.config.image_token_index).to(
+            input_ids.device
+        )
+
         device = input_ids.device
         min_dtype = torch.finfo(dtype).min
 
@@ -748,9 +757,10 @@ def get_attention_mask(
         )
         full_attention_mask[:, :, :, :sequence_length] = combined_mask
 
-        final_attention_mask = torch.where(full_attention_mask, 0, min_dtype).to(device)
-
-        return final_attention_mask
+        if bool_mask:
+            return full_attention_mask
+        else:
+            return torch.where(full_attention_mask, 0, min_dtype).to(device)
 
     def forward(
         self,
@@ -793,10 +803,8 @@ def forward(
             )
             attention_mask = self.get_attention_mask(
                 input_ids,
-                max_s,
                 cu_seqlen_prefill,
                 inputs_embeds.dtype,
-                image_token_mask,
             )
         # Use flash attention for text-only input
         # else:

server/text_generation_server/models/flash_causal_lm.py

@@ -2434,6 +2434,7 @@ def _forward_context(
         input_lengths_tensor: torch.Tensor,
         cache_lengths_tensor: torch.Tensor,
         state: Optional[Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
     ) -> ContextManager:
         if ATTENTION != "flashinfer":
             return nullcontext()
@@ -2450,6 +2451,7 @@ def _forward_context(
                 ),
                 block_tables=block_tables,
                 cu_seqlens=cu_seqlen_prefill,
+                custom_mask=attention_mask,
                 input_lengths=input_lengths_tensor + cache_lengths_tensor,
                 num_heads=self.num_heads,
                 num_kv_heads=self.num_kv_heads,

server/text_generation_server/models/vlm_causal_lm.py

@@ -5,6 +5,7 @@
 from opentelemetry import trace
 from typing import Iterable, Optional, Tuple, List, Type, Dict
 
+from torch.nn import attention
 from transformers import PreTrainedTokenizerBase
 from transformers.image_processing_utils import select_best_resolution
 from text_generation_server.pb import generate_pb2
@@ -485,6 +486,14 @@ def forward(
                 )
                 batch.position_ids = position_ids
 
+        if self.model.config.model_type == "gemma3" and cu_seqlen_prefill is not None:
+            # Get the mask, needed for flashinfer.
+            attention_mask = self.model.get_attention_mask(
+                input_ids, cu_seqlen_prefill, self.dtype, bool_mask=True
+            ).reshape(-1)
+        else:
+            attention_mask = None
+
         # Try to find an associated cuda graph
         bs = input_ids.shape[0]
         sorted_padded_bs = sorted([k for k in self.cuda_graphs.keys() if k >= bs])
@@ -508,6 +517,7 @@ def forward(
                 cu_seqlen_prefill=cu_seqlen_prefill,
                 input_lengths_tensor=input_lengths,
                 cache_lengths_tensor=cache_lengths_tensor,
+                attention_mask=attention_mask,
             ):
                 seqlen = Seqlen(
                     input_lengths=input_lengths,