[Spec Decode] Move `SpecDecodeBaseProposer` out of `eagle.py` (#40732)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

tests/v1/spec_decode/test_eagle.py

@@ -741,9 +741,9 @@ def test_set_inputs_first_pass_parallel_drafting():
 @pytest.mark.parametrize("pp_size", [1, 2])
 @pytest.mark.parametrize("use_distinct_embed_tokens", [True, False])
 @pytest.mark.parametrize("use_distinct_lm_head", [True, False])
-@mock.patch("vllm.v1.spec_decode.eagle.get_pp_group")
-@mock.patch("vllm.v1.spec_decode.eagle.get_layers_from_vllm_config")
-@mock.patch("vllm.v1.spec_decode.eagle.get_model")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_pp_group")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_layers_from_vllm_config")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_model")
 def test_load_model(
     mock_get_model,
     mock_get_layers,

tests/v1/spec_decode/test_mtp.py

@@ -61,9 +61,9 @@ def _create_mtp_proposer(num_speculative_tokens: int) -> EagleProposer:
     return EagleProposer(vllm_config=vllm_config, device=DEVICE_TYPE)
 
 
-@mock.patch("vllm.v1.spec_decode.eagle.get_pp_group")
-@mock.patch("vllm.v1.spec_decode.eagle.get_layers_from_vllm_config")
-@mock.patch("vllm.v1.spec_decode.eagle.get_model")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_pp_group")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_layers_from_vllm_config")
+@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_model")
 def test_mtp_load_model_unified(mock_get_model, mock_get_layers, mock_get_pp_group):
     """Test MTP-specific model loading with unified model approach."""
 

vllm/v1/spec_decode/dflash.py

@@ -11,7 +11,7 @@ from vllm.forward_context import set_forward_context
 from vllm.logger import init_logger
 from vllm.triton_utils import triton
 from vllm.v1.attention.backend import CommonAttentionMetadata
-from vllm.v1.spec_decode.eagle import SpecDecodeBaseProposer
+from vllm.v1.spec_decode.llm_base_proposer import SpecDecodeBaseProposer
 from vllm.v1.spec_decode.utils import copy_and_expand_dflash_inputs_kernel
 
 logger = init_logger(__name__)

vllm/v1/spec_decode/draft_model.py

@@ -9,7 +9,7 @@ from vllm.config import VllmConfig
 from vllm.config.utils import replace
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model
-from vllm.v1.spec_decode.eagle import SpecDecodeBaseProposer
+from vllm.v1.spec_decode.llm_base_proposer import SpecDecodeBaseProposer
 
 logger = init_logger(__name__)
 

vllm/v1/worker/cpu_model_runner.py

@@ -65,16 +65,16 @@ class CPUModelRunner(GPUModelRunner):
 
         # Speculative decoding fallbacks
         import vllm.v1.sample.rejection_sampler
-        import vllm.v1.spec_decode.eagle
+        import vllm.v1.spec_decode.llm_base_proposer
         import vllm.v1.spec_decode.utils
 
-        vllm.v1.spec_decode.eagle.eagle_prepare_inputs_padded_kernel = (
+        vllm.v1.spec_decode.llm_base_proposer.eagle_prepare_inputs_padded_kernel = (
             cpu_tl.eagle_prepare_inputs_padded_kernel
         )
-        vllm.v1.spec_decode.eagle.eagle_prepare_next_token_padded_kernel = (
+        vllm.v1.spec_decode.llm_base_proposer.eagle_prepare_next_token_padded_kernel = (
             cpu_tl.eagle_prepare_next_token_padded_kernel
         )
-        vllm.v1.spec_decode.eagle.copy_and_expand_eagle_inputs_kernel = (
+        vllm.v1.spec_decode.llm_base_proposer.copy_and_expand_eagle_inputs_kernel = (
             cpu_tl.copy_and_expand_eagle_inputs_kernel
         )
         vllm.v1.spec_decode.utils.eagle_step_slot_mapping_metadata_kernel = (