[Model] Add JambaClassfication model

docs/source/models/supported_models.rst

@@ -437,6 +437,9 @@ Classification
     - ✅︎
     - ✅︎
 
+  * - :code:`JambaForSequenceClassification`
+    - Jamba
+
 .. note::
     As an interim measure, these models are supported in both offline and online inference via Embeddings API.
 

vllm/model_executor/models/jamba.py

@@ -16,15 +16,17 @@
                                                RowParallelLinear)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.mamba.mamba_mixer import MambaMixer
+from vllm.model_executor.layers.pooler import Pooler, PoolingType
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.mamba_cache import (MambaCacheManager,
                                                     MambaCacheParams)
+from vllm.model_executor.pooling_metadata import PoolingMetadata
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors
+from vllm.sequence import IntermediateTensors, PoolerOutput
 
 from .interfaces import HasInnerState, SupportsLoRA
 from .utils import maybe_prefix
@@ -548,3 +550,35 @@ def _is_moe_layer(name: str):
             "experts",
             "router",
         ]])
+
+
+class JambaForSequenceClassification(JambaForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        config = vllm_config.model_config.hf_config
+        num_labels: int = config.num_labels
+        score_bias: bool = getattr(config, 'score_bias', False)
+        self.score = nn.Linear(config.hidden_size, num_labels, bias=score_bias)
+
+        pooler_config = vllm_config.model_config.pooler_config
+        self._pooler = Pooler.from_config_with_defaults(
+            pooler_config,
+            pooling_type=PoolingType.LAST,
+            normalize=False,
+            softmax=False)
+
+    def pooler(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> Optional[PoolerOutput]:
+        hidden_states = hidden_states.float()
+        logits = self.score(hidden_states)
+        return self._pooler(logits, pooling_metadata)
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        # TODO: The reward weights themselves have float32 accuracy data, we
+        # would like to load them in fp32 to get that extra precision.
+        super().load_weights(weights)
+        self.score = self.score.float()

vllm/model_executor/models/registry.py

@@ -108,6 +108,7 @@
     "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
     "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
     "DeciLMForCausalLM": ("decilm", "DeciLMForCausalLM"),
+    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
     "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
     "GlmForCausalLM": ("glm", "GlmForCausalLM"),
     "LlamaModel": ("llama", "LlamaForCausalLM"),

vllm/worker/pooling_model_runner.py

@@ -91,6 +91,10 @@ def execute_model(
         ]
 
         multi_modal_kwargs = model_input.multi_modal_kwargs or {}
+        seqlen_agnostic_kwargs = {
+            "finished_requests_ids": model_input.finished_requests_ids,
+            "request_ids_to_seq_ids": model_input.request_ids_to_seq_ids,
+        } if self.has_inner_state else {}
         if (self.observability_config is not None
                 and self.observability_config.collect_model_forward_time):
             model_forward_start = torch.cuda.Event(enable_timing=True)
@@ -110,7 +114,8 @@ def execute_model(
                 intermediate_tensors=intermediate_tensors,
                 **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
                                              device=self.device),
-                **cross_enc_kwargs)
+                **cross_enc_kwargs,
+                **seqlen_agnostic_kwargs)
 
         if (self.observability_config is not None
                 and self.observability_config.collect_model_forward_time):