chore: add support for vllm v0.19.0

.github/workflows/test.yml

@@ -74,7 +74,7 @@ jobs:
           # Lower bound support
           - vllm_version:
               name: "vLLM:lowest"
-              repo: "git+https://github.com/vllm-project/vllm --tag v0.18.0"
+              repo: "git+https://github.com/vllm-project/vllm --tag v0.19.0"
             test_suite:
               name: "backward compat"
               markers: "compat or (cpu and basic and not quantized)"

pyproject.toml

@@ -15,7 +15,8 @@ dependencies = [
     "ibm-fms>=1.9.0,<2",
     # NB: use strict < with the next patch version to not exclude versions with
     # build metadata suffixes
-    "vllm>=0.18.0,<0.18.2",
+    "vllm>=0.19.0,<0.19.1",
+
     # Specific torch version overrides handled by uv
     "torch",
     "torchvision",
@@ -75,8 +76,8 @@ override-dependencies = [
     # requests is used by many dependencies, make sure it's patched for CVEs
     "requests>=2.32.4",
 
-    # temporarily bypass vllm 0.18.0 compressed-tensors==0.13.0 requirement
-    # TODO: remove once vllm pins correct version bounds
+    # vllm 0.18.0 pins compressed-tensors==0.13.0; override to 0.14.0.1
+    # TODO: remove once minimum vllm is bumped past 0.18.0
     "compressed-tensors==0.14.0.1",
 ]
 # This adds constraints to all dependent build environments, which will ensure everything is built
@@ -86,7 +87,7 @@ build-constraint-dependencies = ["torch==2.10.0"]
 extra-build-variables = { vllm = { VLLM_TARGET_DEVICE = "empty" } }
 
 [tool.uv.sources]
-vllm = { git = "https://github.com/vllm-project/vllm", rev = "v0.18.1" }
+vllm = { git = "https://github.com/vllm-project/vllm", rev = "v0.19.0" }
 torch = [
   { index = "pytorch-cpu" },
 ]

tests/utils/test_platform_validation.py

@@ -12,7 +12,8 @@
 
 
 from vllm import SamplingParams
-from vllm.inputs.data import token_inputs
+
+from vllm.inputs import tokens_input
 from vllm.pooling_params import PoolingParams
 from vllm.sampling_params import StructuredOutputsParams
 from vllm_spyre.platform import SpyrePlatform
@@ -43,7 +44,7 @@ def test_strips_structured_outputs(self):
 
         assert params.structured_outputs is not None
 
-        SpyrePlatform.validate_request(token_inputs(prompt_token_ids=[0]), params)
+        SpyrePlatform.validate_request(tokens_input(prompt_token_ids=[0]), params)
 
         assert params.structured_outputs is None
 
@@ -53,7 +54,7 @@ def test_logs_warning_when_stripping(self, caplog_vllm_spyre):
             max_tokens=20, structured_outputs=StructuredOutputsParams(json_object=True)
         )
 
-        SpyrePlatform.validate_request(token_inputs(prompt_token_ids=[0]), params)
+        SpyrePlatform.validate_request(tokens_input(prompt_token_ids=[0]), params)
 
         assert len(caplog_vllm_spyre.records) > 0
         warning_record = caplog_vllm_spyre.records[0]
@@ -74,7 +75,7 @@ def test_strips_different_structured_output_types(self, structured_output):
 
         assert params.structured_outputs is not None
 
-        SpyrePlatform.validate_request(token_inputs(prompt_token_ids=[0]), params)
+        SpyrePlatform.validate_request(tokens_input(prompt_token_ids=[0]), params)
 
         assert params.structured_outputs is None
 
@@ -96,7 +97,7 @@ def test_preserves_other_sampling_params(self):
             "top_k": params.top_k,
         }
 
-        SpyrePlatform.validate_request(token_inputs(prompt_token_ids=[0]), params)
+        SpyrePlatform.validate_request(tokens_input(prompt_token_ids=[0]), params)
 
         # Verify other params are unchanged
         assert params.max_tokens == original_values["max_tokens"]
@@ -111,7 +112,7 @@ def test_does_not_affect_pooling_params(self):
         pooling_params = PoolingParams()
 
         # Should not raise any errors and should return early
-        SpyrePlatform.validate_request(token_inputs(prompt_token_ids=[0]), pooling_params)
+        SpyrePlatform.validate_request(tokens_input(prompt_token_ids=[0]), pooling_params)
 
         # PoolingParams don't have structured_outputs, so just verify no exception
         assert True  # If we got here, the early return worked

tests/v1/core/test_scheduler_structured_outputs.py

@@ -10,10 +10,12 @@
 from unittest.mock import Mock, patch
 from vllm import SamplingParams
 from vllm.sampling_params import StructuredOutputsParams
-from vllm.v1.request import Request, RequestStatus
 from vllm.v1.core.sched.request_queue import FCFSRequestQueue
+from vllm.v1.request import Request, RequestStatus
 from vllm_spyre.v1.core.scheduler import ChunkedPrefillSpyreScheduler
 
+_WAITING_FOR_GRAMMAR = RequestStatus.WAITING_FOR_STRUCTURED_OUTPUT_GRAMMAR
+
 
 pytestmark = pytest.mark.skip_global_cleanup
 
@@ -80,7 +82,7 @@ def test_scheduler_strips_structured_output_request(self, mocked_scheduler, capl
 
         # Verify structured_output_request is set
         assert request.structured_output_request is not None
-        assert request.status == RequestStatus.WAITING_FOR_FSM
+        assert request.status == _WAITING_FOR_GRAMMAR
 
         # Add request to waiting queue
         mocked_scheduler.waiting.append(request)
@@ -154,7 +156,7 @@ def test_scheduler_handles_multiple_requests_with_structured_outputs(
         # Verify all have structured_output_request set
         for request in requests:
             assert request.structured_output_request is not None
-            assert request.status == RequestStatus.WAITING_FOR_FSM
+            assert request.status == _WAITING_FOR_GRAMMAR
 
         # Call the actual schedule method
         mocked_scheduler.schedule()