fix for https://github.com/vllm-project/vllm/issues/17746

git.diff

diff --git a/tests/v1/engine/test_processor.py b/tests/v1/engine/test_processor.py
new file mode 100644
index 000000000..c793290d4
--- /dev/null
+++ b/tests/v1/engine/test_processor.py
@@ -0,0 +1,196 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+import math
+import time
+from typing import Optional
+
+import pytest
+
+from vllm.config import (CacheConfig, DeviceConfig, LoRAConfig,
+                         ModelConfig, VllmConfig)
+from vllm.lora.request import LoRARequest
+from vllm.sampling_params import SamplingParams
+from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
+from vllm.v1.engine.processor import Processor
+
+
+@pytest.fixture
+def opt_125m_huggingface_id():
+    return "facebook/opt-125m"
+
+
+@pytest.fixture
+def fake_lora_adapter_files(tmp_path, opt_125m_huggingface_id):
+    adapter_config = {
+        "base_model_name_or_path": opt_125m_huggingface_id,
+        "bias": "none",
+        "fan_in_fan_out": False,
+        "inference_mode": True,
+        "init_lora_weights": True,
+        "lora_alpha": 16,
+        "lora_dropout": 0.05,
+        "peft_type": "LORA",
+        "r": 8,
+        "target_modules": [
+            "q_proj",
+            "v_proj"
+        ],
+        "task_type": "CAUSAL_LM"
+    }
+    with open(tmp_path / "adapter_config.json", "w") as f:
+        json.dump(adapter_config, f)
+
+    tokenizer_config = {
+        "add_bos_token": True,
+        "add_eos_token": False,
+        "added_tokens_decoder": {
+            "50300": {
+                "content": "_A_",
+                "lstrip": False,
+                "normalized": True,
+                "rstrip": False,
+                "single_word": False,
+                "special": False
+            },
+            "50301": {
+                "content": "_B_",
+                "lstrip": False,
+                "normalized": True,
+                "rstrip": False,
+                "single_word": False,
+                "special": False
+            },
+        },
+        "additional_special_tokens": [],
+        "bos_token": "<s>",
+        "chat_template": "{{ bos_token }}{% for message in messages %}{% endfor %}",
+        "clean_up_tokenization_spaces": False,
+        "eos_token": "</s>",
+        "legacy": True,
+        "model_max_length": 1000000000000000019884624838656,
+        "pad_token": "</s>",
+        "sp_model_kwargs": {},
+        "spaces_between_special_tokens": False,
+        "tokenizer_class": "LlamaTokenizer",
+        "unk_token": "<unk>",
+        "use_default_system_prompt": False
+    }
+    with open(tmp_path / "tokenizer_config.json", "w") as f:
+        json.dump(tokenizer_config, f)
+
+    base_vocab = {f"tok_{i}": i for i in range (3,50300)}
+    added_vocab = {
+        "<unk>": 0,
+        "<s>": 1,
+        "</s>": 2,
+        "_A_": 50300,
+        "_B_": 50301,
+    }
+    vocab = {**base_vocab, **added_vocab}
+    tokenizer = {
+        "version": "1.0",
+        "added_tokens": [
+            {
+                "id": 0,
+                "content": "<unk>",
+                "single_word": False,
+                "lstrip": False,
+                "rstrip": False,
+                "normalized": False,
+                "special": True
+            },
+            {
+                "id": 1,
+                "content": "<s>",
+                "single_word": False,
+                "lstrip": False,
+                "rstrip": False,
+                "normalized": False,
+                "special": True
+            },
+            {
+                "id": 2,
+                "content": "</s>",
+                "single_word": False,
+                "lstrip": False,
+                "rstrip": False,
+                "normalized": False,
+                "special": True
+            },
+            {
+                "id": 50300,
+                "content": "_A_",
+                "single_word": False,
+                "lstrip": False,
+                "rstrip": False,
+                "normalized": True,
+                "special": False
+            },
+            {
+                "id": 50301,
+                "content": "_B_",
+                "single_word": False,
+                "lstrip": False,
+                "rstrip": False,
+                "normalized": True,
+                "special": False
+            },
+        ],
+        "model": {
+            "type": "BPE",
+            "unk_token": "<unk>",
+            "fuse_unk": True,
+            "byte_fallback": True,
+            "ignore_merges": False,
+            "vocab": vocab,
+            "merges": [],
+        },
+    }
+    with open(tmp_path / "tokenizer.json", "w") as f:
+        json.dump(tokenizer, f)
+
+    with open(tmp_path / "adapter_model.bin", "wb") as f:
+        f.write("".encode("utf-8"))
+
+    return tmp_path
+
+
+def test_allowed_token_ids_with_lora_vocab(opt_125m_huggingface_id, fake_lora_adapter_files):
+    model = opt_125m_huggingface_id
+    model_config = ModelConfig(
+        model=model,
+        task="auto",
+        tokenizer=model,
+        tokenizer_mode="auto",
+        trust_remote_code=True,
+        dtype="float16",
+        seed=42,
+    )
+    cache_config = CacheConfig(
+        block_size=16,
+        gpu_memory_utilization=0.9,
+        swap_space=0,
+        cache_dtype="auto",
+    )
+    device_config = DeviceConfig()
+
+    lora_config = LoRAConfig()
+    vllm_config = VllmConfig(
+        model_config=model_config,
+        cache_config=cache_config,
+        device_config=device_config,
+        lora_config=lora_config,
+    )
+
+    tokenizer = init_tokenizer_from_configs(
+        model_config=vllm_config.model_config,
+        scheduler_config=vllm_config.scheduler_config,
+        lora_config=vllm_config.lora_config)
+    processor = Processor(vllm_config, tokenizer)
+
+    # We define tokens 50300, 50301 in our fake lora adapter
+    lora_token_ids=[50300, 50301]
+    sampling_params = SamplingParams(allowed_token_ids=lora_token_ids)
+    lora_request = LoRARequest("1", 1, str(fake_lora_adapter_files))
+    processor._validate_sampling_params(sampling_params, lora_request)
diff --git a/vllm/v1/engine/processor.py b/vllm/v1/engine/processor.py
index 27d70a781..2aa19f8bb 100644
--- a/vllm/v1/engine/processor.py
+++ b/vllm/v1/engine/processor.py
@@ -74,6 +74,7 @@ class Processor:
     def _validate_sampling_params(
         self,
         params: SamplingParams,
+        lora_request: Optional[LoRARequest],
     ) -> None:
         self._validate_structured_output(params)
         self._validate_logit_bias(params)
@@ -82,7 +83,8 @@ class Processor:
             return
         if not params.allowed_token_ids:
             raise ValueError("allowed_token_ids is not None and empty!")
-        vocab_size = self.model_config.get_vocab_size()
+        tokenizer = self.tokenizer.get_lora_tokenizer(lora_request)
+        vocab_size = len(tokenizer)
         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!")
@@ -122,6 +124,7 @@ class Processor:
     def _validate_params(
         self,
         params: Union[SamplingParams, PoolingParams],
+        lora_request: Optional[LoRARequest],
     ):
         """
         Validate supported SamplingParam.
@@ -132,7 +135,7 @@ class Processor:
             raise ValueError("V1 does not yet support Pooling models.")
 
         self._validate_logprobs(params)
-        self._validate_sampling_params(params)
+        self._validate_sampling_params(params, lora_request)
         self._validate_supported_sampling_params(params)
 
     def _validate_lora(self, lora_request: Optional[LoRARequest]) -> None:
@@ -207,7 +210,7 @@ class Processor:
         # TODO(woosuk): Support pooling models.
         # TODO(woosuk): Support encoder-decoder models.
         self._validate_lora(lora_request)
-        self._validate_params(params)
+        self._validate_params(params, lora_request)
         if priority != 0:
             raise ValueError("V1 does not support priority yet.")
         if trace_headers is not None: