[Bugfix][V1] Fix allowed_token_ids for v1 Sampler (#14169)

Signed-off-by: Lu Fang <lufang@fb.com>

vllm/v1/engine/processor.py

@@ -92,10 +92,12 @@ class Processor:
             return
         if params.allowed_token_ids is None:
             return
-        if not all(0 <= tid < self.model_config.vocab_size
-                   for tid in params.allowed_token_ids):
+        if not params.allowed_token_ids:
+            raise ValueError("allowed_token_ids is not None and empty!")
+        vocab_size = self.model_config.get_vocab_size()
+        if not all(0 <= tid < vocab_size for tid in params.allowed_token_ids):
             raise ValueError(
-                "allowed_token_ids contains out-of-vocab token id")
+                "allowed_token_ids contains out-of-vocab token id!")
 
     def process_inputs(
         self,

vllm/v1/worker/gpu_input_batch.py

@@ -199,6 +199,8 @@ class InputBatch:
         self.logit_bias: list[Optional[dict[int,
                                             float]]] = [None] * max_num_reqs
         self.has_allowed_token_ids: set[str] = set()
+        # NOTE(lufang): In the mask tensor, if the corresponding token allowed,
+        # the value is False. Since we use masked_fill_ to set -inf.
         self.allowed_token_ids_mask: Optional[torch.Tensor] = None
         self.allowed_token_ids_mask_cpu_tensor: Optional[torch.Tensor] = None
 
@@ -300,6 +302,7 @@ class InputBatch:
             self.has_allowed_token_ids.add(req_id)
             if self.allowed_token_ids_mask_cpu_tensor is None:
                 # Lazy allocation for this tensor, which can be large.
+                # False means we don't fill with -inf.
                 self.allowed_token_ids_mask = torch.zeros(self.max_num_reqs,
                                                           self.vocab_size,
                                                           dtype=torch.bool,
@@ -309,8 +312,10 @@ class InputBatch:
                     self.vocab_size,
                     dtype=torch.bool,
                     device="cpu")
+            self.allowed_token_ids_mask_cpu_tensor[req_index] = True
+            # False means we don't fill with -inf.
             self.allowed_token_ids_mask_cpu_tensor[req_index][
-                sampling_params.allowed_token_ids] = True
+                sampling_params.allowed_token_ids] = False
 
         # Add request lora ID
         if request.lora_request:
@@ -359,6 +364,7 @@ class InputBatch:
         self.logit_bias[req_index] = None
         self.has_allowed_token_ids.discard(req_id)
         if self.allowed_token_ids_mask_cpu_tensor is not None:
+            # False means we don't fill with -inf.
             self.allowed_token_ids_mask_cpu_tensor[req_index].fill_(False)
         return req_index