[Bugfix] Fix ModernBert load & Enable sliding window attention for bidirectional attention. (#22637)

Signed-off-by: wang.yuqi <noooop@126.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Max de Bayser <mbayser@br.ibm.com>

tests/models/language/pooling/test_gte.py

@@ -4,10 +4,11 @@ from typing import Any
 
 import pytest
 
-from ...utils import (CLSPoolingEmbedModelInfo, EmbedModelInfo,
+from ...utils import (CLSPoolingEmbedModelInfo, CLSPoolingRerankModelInfo,
-                      LASTPoolingEmbedModelInfo, check_transformers_version)
+                      EmbedModelInfo, LASTPoolingEmbedModelInfo,
+                      RerankModelInfo, check_transformers_version)
 from .embed_utils import correctness_test_embed_models
-from .mteb_utils import mteb_test_embed_models
+from .mteb_utils import mteb_test_embed_models, mteb_test_rerank_models
 
 MODELS = [
     ########## BertModel
@@ -58,6 +59,14 @@ MODELS = [
                               enable_test=False),
 ]
 
+RERANK_MODELS = [
+    # classifier_pooling: mean
+    CLSPoolingRerankModelInfo(
+        "Alibaba-NLP/gte-reranker-modernbert-base",
+        architecture="ModernBertForSequenceClassification",
+        enable_test=True),
+]
+
 
 @pytest.mark.parametrize("model_info", MODELS)
 def test_embed_models_mteb(hf_runner, vllm_runner,
@@ -88,3 +97,9 @@ def test_embed_models_correctness(hf_runner, vllm_runner,
 
     correctness_test_embed_models(hf_runner, vllm_runner, model_info,
                                   example_prompts, vllm_extra_kwargs)
+
+
+@pytest.mark.parametrize("model_info", RERANK_MODELS)
+def test_rerank_models_mteb(hf_runner, vllm_runner,
+                            model_info: RerankModelInfo) -> None:
+    mteb_test_rerank_models(hf_runner, vllm_runner, model_info)

vllm/model_executor/models/modernbert.py

@@ -26,8 +26,7 @@ from vllm.model_executor.pooling_metadata import PoolingMetadata
 from vllm.sequence import IntermediateTensors
 from vllm.tasks import PoolingTask
 
-from .interfaces import (SupportsCrossEncoding, SupportsV0Only,
+from .interfaces import SupportsCrossEncoding, default_pooling_type
-                         default_pooling_type)
 from .utils import WeightsMapper, maybe_prefix
 
 
@@ -93,16 +92,14 @@ class ModernBertAttention(nn.Module):
             bias=config.attention_bias,
         )
 
+        sliding_window = None
         if layer_id % config.global_attn_every_n_layers != 0:
-            self.local_attention = (config.local_attention // 2,
+            sliding_window = config.local_attention // 2
-                                    config.local_attention // 2)
+            rope_theta = config.local_rope_theta if config.local_rope_theta \
+                    is not None else config.global_rope_theta
         else:
-            self.local_attention = (-1, -1)
+            rope_theta = config.global_rope_theta
 
-        rope_theta = config.global_rope_theta
-        if self.local_attention != (
-                -1, -1) and config.local_rope_theta is not None:
-            rope_theta = config.local_rope_theta
         self.rotary_emb = ModernBertRotaryEmbedding(config=config,
                                                     head_size=self.head_dim,
                                                     dim=self.head_dim,
@@ -111,7 +108,8 @@ class ModernBertAttention(nn.Module):
                               self.head_dim,
                               self.scaling,
                               prefix=f"{layer_id}.attn",
-                              attn_type=AttentionType.ENCODER_ONLY)
+                              attn_type=AttentionType.ENCODER_ONLY,
+                              per_layer_sliding_window=sliding_window)
         self.Wo = RowParallelLinear(config.hidden_size,
                                     config.hidden_size,
                                     bias=config.attention_bias)
@@ -278,6 +276,7 @@ class ModernBertPooler(Pooler):
         return self.pooling.get_pooling_updates(task)
 
     def _head(self, pooled_output: torch.Tensor):
+        pooled_output = pooled_output.to(self.dense.weight.dtype)
         return self.norm(self.act(self.dense(pooled_output)))
 
     def forward(
@@ -296,8 +295,7 @@ class ModernBertPooler(Pooler):
 
 
 @default_pooling_type("CLS")
-class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
+class ModernBertForSequenceClassification(nn.Module, SupportsCrossEncoding):
-                                          SupportsCrossEncoding):
 
     is_pooling_model = True
 
@@ -308,6 +306,7 @@ class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
         self.model = ModernBertModel(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "modernbert"))
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+        self.pooling = ModernBertPooler(config)
 
         pooler_config = vllm_config.model_config.pooler_config
         assert pooler_config is not None
@@ -317,14 +316,14 @@ class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
             Pooler.for_encode(pooler_config),
             "classify":
             ClassifierPooler(
-                pooling=ModernBertPooler(config),
+                pooling=self.pooling,
                 classifier=self.classifier,
                 act_fn=ClassifierPooler.act_fn_for_seq_cls(
                     vllm_config.model_config),
             ),
             "score":
             ClassifierPooler(
-                pooling=ModernBertPooler(config),
+                pooling=self.pooling,
                 classifier=self.classifier,
                 act_fn=ClassifierPooler.act_fn_for_cross_encoder(
                     vllm_config.model_config),
@@ -353,7 +352,7 @@ class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
             if name.startswith("head"):
-                param = params_dict["_pooler.pooler." + name[len("head") + 1:]]
+                param = params_dict["pooling." + name[len("head") + 1:]]
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
@@ -368,5 +367,5 @@ class ModernBertForSequenceClassification(nn.Module, SupportsV0Only,
         return self.model(
             input_ids=input_ids,
             inputs_embeds=inputs_embeds,
-            position_ids=positions,
+            positions=positions,
         )

vllm/v1/attention/backends/flash_attn.py

@@ -384,6 +384,8 @@ class FlashAttentionImpl(AttentionImpl):
         self.alibi_slopes = alibi_slopes
         if sliding_window is None:
             self.sliding_window = (-1, -1)
+        elif attn_type == AttentionType.ENCODER_ONLY:
+            self.sliding_window = (sliding_window - 1, sliding_window - 1)
         else:
             self.sliding_window = (sliding_window - 1, 0)
         self.kv_cache_dtype = kv_cache_dtype

vllm/v1/worker/gpu_model_runner.py

@@ -826,7 +826,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         # Prepare encoder attention metadata separately
         # (encoder layers are not in KV cache groups)
         if self.is_encoder_only_model:
-            common_attn_metadata, encoder_attn_metadata = \
+
+            per_layer_metadata = \
                 self._build_encoder_only_attn_metadata(
                 scheduler_output)
 
@@ -835,6 +836,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 self.vllm_config, Attention)
             for layer_name, attn_module in attention_layers.items():
                 if attn_module.attn_type == AttentionType.ENCODER_ONLY:
+                    common_attn_metadata, encoder_attn_metadata =\
+                        per_layer_metadata[layer_name]
                     attn_metadata[layer_name] = encoder_attn_metadata
 
         # Prepare the attention metadata for each KV cache group and make layers
@@ -2683,30 +2686,41 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         # Check if model is encoder-only
         block_size = self.vllm_config.cache_config.block_size
         use_mla = self.vllm_config.model_config.use_mla
-        attn_specs = list[AttentionSpec]()
+        attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
-        for attn_module in attn_layers.values():
+        for layer_name, attn_module in attn_layers.items():
 
             if attn_module.attn_type == AttentionType.ENCODER_ONLY:
-                assert attn_module.sliding_window is None, "Sliding "
+                if attn_module.sliding_window is None:
-                "window attention is not supported for encoder-only models"
+                    attn_spec: AttentionSpec = FullAttentionSpec(
+                        block_size=block_size,
+                        num_kv_heads=attn_module.num_kv_heads,
+                        head_size=attn_module.head_size,
+                        dtype=self.kv_cache_dtype,
+                        use_mla=use_mla)
+                else:
+                    attn_spec = SlidingWindowSpec(
+                        block_size=block_size,
+                        num_kv_heads=attn_module.num_kv_heads,
+                        head_size=attn_module.head_size,
+                        dtype=self.kv_cache_dtype,
+                        sliding_window=attn_module.sliding_window,
+                        use_mla=use_mla)
+                attn_specs[attn_spec].append(layer_name)
 
-                attn_specs.append(
-                    FullAttentionSpec(block_size=block_size,
-                                      num_kv_heads=attn_module.num_kv_heads,
-                                      head_size=attn_module.head_size,
-                                      dtype=self.kv_cache_dtype,
-                                      use_mla=use_mla))
             else:
                 raise ValueError("Expected only encoder-only layers")
 
         if len(attn_specs) > 0:
-            assert len(attn_specs) == len(attn_layers), \
+            total_layers = 0
+            for attn_spec, layer_names in attn_specs.items():
+
+                attn_backends = get_attn_backends_for_layers(layer_names)
+                total_layers += len(layer_names)
+
+                self.attn_groups.append(
+                    create_attn_groups(attn_backends, attn_spec))
+            assert total_layers == len(attn_layers), \
                 "All or none of the layers are expected to be encoder-only"
-
-            attn_backends = get_attn_backends_for_layers(attn_layers.keys())
-
-            self.attn_groups.append(
-                create_attn_groups(attn_backends, attn_specs[0]))
             self.is_encoder_only_model = True
 
     def calculate_reorder_batch_threshold(self) -> None:
@@ -3071,7 +3085,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
 
     def _build_encoder_only_attn_metadata(
             self, scheduler_output: "SchedulerOutput") -> \
-                tuple[CommonAttentionMetadata, Any]:
+                dict[str, tuple[CommonAttentionMetadata, Any]]:
         """Prepare encoder attention metadata for encoder-only models.
 
         Args:
@@ -3088,10 +3102,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
         max_num_scheduled_tokens = max(tokens)
 
-        # Use the first attention metadata builder
-        # to create encoder attention metadata
-        builder = self.attn_groups[0][0].metadata_builder
-
         dummy_block_table = torch.zeros((num_reqs, 1),
                                         dtype=torch.int32,
                                         device=self.device)
@@ -3099,22 +3109,38 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                                          dtype=torch.int32,
                                          device=self.device)
 
-        common_metadata = CommonAttentionMetadata(
+        group_metadata = dict[str, tuple[CommonAttentionMetadata, Any]]()
-            query_start_loc=self.query_start_loc[:num_reqs + 1],
-            query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs + 1],
-            seq_lens=self.seq_lens[:num_reqs],
-            seq_lens_cpu=self.seq_lens_cpu[:num_reqs],
-            num_computed_tokens_cpu=self.input_batch.
-            num_computed_tokens_cpu_tensor[:num_reqs],
-            num_reqs=num_reqs,
-            num_actual_tokens=total_num_scheduled_tokens,
-            max_query_len=max_num_scheduled_tokens,
-            block_table_tensor=dummy_block_table,
-            slot_mapping=dummy_slot_mapping,
-            causal=False,
-        )
 
-        return common_metadata, builder.build(
+        for attn_group_list in self.attn_groups:
-            common_prefix_len=0,  # No cascade for encoder
+
-            common_attn_metadata=common_metadata,
+            assert len(attn_group_list) == 1
-        )
+            attn_group = attn_group_list[0]
+
+            # Use the first attention metadata builder
+            # to create encoder attention metadata
+            builder = attn_group.metadata_builder
+
+            common_metadata = CommonAttentionMetadata(
+                query_start_loc=self.query_start_loc[:num_reqs + 1],
+                query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs + 1],
+                seq_lens=self.seq_lens[:num_reqs],
+                seq_lens_cpu=self.seq_lens_cpu[:num_reqs],
+                num_computed_tokens_cpu=self.input_batch.
+                num_computed_tokens_cpu_tensor[:num_reqs],
+                num_reqs=num_reqs,
+                num_actual_tokens=total_num_scheduled_tokens,
+                max_query_len=max_num_scheduled_tokens,
+                block_table_tensor=dummy_block_table,
+                slot_mapping=dummy_slot_mapping,
+                causal=False,
+            )
+
+            metadata = builder.build(
+                common_prefix_len=0,  # No cascade for encoder
+                common_attn_metadata=common_metadata,
+            )
+
+            for layer_name in attn_group.layer_names:
+                group_metadata[layer_name] = (common_metadata, metadata)
+
+        return group_metadata