✨ Add vLLM 0.11.0 support

.github/workflows/test.yml

@@ -75,7 +75,7 @@ jobs:
         include:
           - vllm_version:
               name: "vLLM:lowest"
-              repo: "git+https://github.com/vllm-project/vllm --tag v0.10.1.1"
+              repo: "git+https://github.com/vllm-project/vllm --tag v0.10.2"
             test_suite:
               name: "backward compat"
               markers: "compat or (cpu and basic)"

pyproject.toml

@@ -13,7 +13,7 @@ license = {text = "Apache 2"}
 dependencies = [
     "fms-model-optimizer[fp8]>=0.6.0",
     "ibm-fms>=1.4.0,<2.0",
-    "vllm>=0.10.1.1,<=0.10.2",
+    "vllm>=0.10.2,<=0.11.0",
     "pytest-mock>=3.15.0",
 ]
 requires-python = ">=3.11"

tests/spyre_util.py

@@ -1,4 +1,3 @@
-import inspect
 import json
 import math
 import os
@@ -367,31 +366,14 @@ def create_random_request(
         assert (len(prompt_token_ids) == num_tokens
                 ), f"need {num_tokens} but got {len(prompt_token_ids)}"
 
-    # temporary backward compat code for 0.10.1.1
-    annotations = inspect.getfullargspec(Request).annotations
-    extra_args = {}  # noqa
-    if ('multi_modal_hashes' in annotations):
-        extra_args.update({
-            'multi_modal_hashes': None,
-        })
-    if ('multi_modal_placeholders' in annotations):
-        extra_args.update({
-            'multi_modal_placeholders': None,
-        })
-    if ('multi_modal_kwargs' in annotations):
-        extra_args.update({
-            'multi_modal_kwargs': None,
-        })
-
     return Request(request_id=str(request_id),
                    prompt_token_ids=prompt_token_ids,
                    sampling_params=sampling_params,
                    eos_token_id=None,
                    arrival_time=0,
                    lora_request=None,
                    pooling_params=None,
-                   cache_salt=None,
-                   **extra_args)
+                   cache_salt=None)
 
 
 def skip_unsupported_tp_size(size: int, backend: str):

tests/utils/test_upstream_compatibility.py

@@ -1,83 +1,35 @@
-import inspect
 import os
 
 import pytest
+from vllm.v1.core.sched.output import NewRequestData
+from vllm.v1.kv_cache_interface import FullAttentionSpec
+
+from vllm_spyre.compat_utils import dataclass_fields
 
 pytestmark = pytest.mark.compat
 
 VLLM_VERSION = os.getenv("TEST_VLLM_VERSION", "default")
 
 
-@pytest.mark.cpu
-def test_init_distributed_environment():
-    """Tests whether vllm's init_distributed_environment
-    has the custom timeout argument"""
-    from vllm.distributed import init_distributed_environment
-
-    annotations = inspect.getfullargspec(
-        init_distributed_environment).annotations
+def test_mm_inputs():
 
     if VLLM_VERSION == "vLLM:lowest":
-        assert 'timeout' \
-                not in annotations, ("we should remove compat code which is now"
-                " part of released vllm version")
-
-
-def test_request():
-
-    from vllm.v1.request import Request
-
-    annotations = inspect.getfullargspec(Request).annotations
-
-    if VLLM_VERSION == "vLLM:main":
-        assert 'multi_modal_kwargs' not in annotations
-        assert 'multi_modal_hashes' not in annotations
-        assert 'multi_modal_placeholders' not in annotations
-    elif VLLM_VERSION == "vLLM:lowest":
-        assert 'multi_modal_hashes' in annotations
-        assert 'multi_modal_placeholders' in annotations
-        assert 'multi_modal_placeholders' in annotations
-        # The compat code introduced in the PR below can now be removed:
-        # https://github.com/vllm-project/vllm-spyre/pull/463
-
-
-def test_model_runner_output():
-
-    from vllm.v1.outputs import ModelRunnerOutput
+        # Can remove "mm_kwargs", "mm_hashes", "mm_positions"
+        # (replaced by mm_features)
+        assert 'mm_kwargs' in dataclass_fields(NewRequestData)
 
-    annotations = inspect.getfullargspec(ModelRunnerOutput).annotations
 
-    if VLLM_VERSION == "vLLM:main":
-        assert 'spec_token_ids' not in annotations
-    elif VLLM_VERSION == "vLLM:lowest":
-        assert 'spec_token_ids' in annotations
-        # The compat code introduced in the PR below can now be removed:
-        # https://github.com/vllm-project/vllm-spyre/pull/463
-
-
-def test_pooling_metadata():
-
-    from vllm.v1.pool.metadata import PoolingMetadata
-
-    has_build_pooling_cursor = getattr(PoolingMetadata, "build_pooling_cursor",
-                                       False)
-
-    if VLLM_VERSION == "vLLM:main":
-        assert has_build_pooling_cursor
-    elif VLLM_VERSION == "vLLM:lowest":
-        assert not has_build_pooling_cursor
-        # The compat code introduced in the PR below can now be removed:
-        # https://github.com/vllm-project/vllm-spyre/pull/463
-
-
-def test_scheduler_output():
+def test_get_sampler():
+    if VLLM_VERSION == "vLLM:lowest":
+        try:
+            from vllm.model_executor.layers.sampler import (  # # noqa
+                get_sampler)
+        except ImportError as e:
+            raise AssertionError(
+                "Remove backwards compatibility for get_sampler") from e
 
-    from vllm.v1.core.sched.output import SchedulerOutput
-    annotations = inspect.getfullargspec(SchedulerOutput).annotations
 
-    if VLLM_VERSION == "vLLM:main":
-        assert 'free_encoder_mm_hashes' in annotations
-    elif VLLM_VERSION == "vLLM:lowest":
-        assert 'free_encoder_mm_hashes' not in annotations
-        # The compat code introduced in the PR below can now be removed:
-        # https://github.com/vllm-project/vllm-spyre/pull/463
+def test_use_mla():
+    if VLLM_VERSION == "vLLM:lowest":
+        # Can remove backwards compatibility for use_mla
+        assert "use_mla" in dataclass_fields(FullAttentionSpec)

vllm_spyre/model_executor/model_loader/spyre.py

@@ -13,10 +13,11 @@
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.model_loader.weight_utils import (
     download_weights_from_hf)
-from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.v1.outputs import SamplerOutput
+from vllm.v1.sample.metadata import SamplingMetadata
+from vllm.v1.sample.sampler import Sampler
 
 import vllm_spyre.envs as envs_spyre
 import vllm_spyre.utils as utils_spyre
@@ -60,7 +61,13 @@ def __init__(
         self.logits_processor = LogitsProcessor(
             vllm_config.model_config.hf_config.vocab_size,
             logits_as_input=True)
-        self.sampler = get_sampler()
+
+        try:
+            ## Temporary backwards compatibility for 0.10.2
+            from vllm.model_executor.layers.sampler import get_sampler
+            self.sampler = get_sampler()
+        except (ImportError, ModuleNotFoundError):
+            self.sampler = Sampler()
 
         # boolean tensor of length batch size with indices:
         # True for unfinished sequences and