[Speculative decoding] [Multi-Step] decouple should_modify_greedy_probs_inplace (#6971)

tests/samplers/test_sampler.py

@@ -1,7 +1,7 @@
 import itertools
 import random
 from typing import Dict, List, Optional, Tuple
-from unittest.mock import patch
+from unittest.mock import Mock, patch
 
 import pytest
 import torch
@@ -703,3 +703,28 @@ def test_sampler_repetition_penalty_mixed(device: str):
 
     assert tokens1[0] == tokens2[1]
     assert tokens1[1] == tokens2[0]
+
+
+@pytest.mark.parametrize("device", CUDA_DEVICES)
+def test_sampler_include_gpu_probs_tensor(device: str):
+    set_random_seed(42)
+    torch.set_default_device(device)
+    batch_size = random.randint(1, 256)
+    _, fake_logits, sampler = _prepare_test(batch_size)
+    sampler.include_gpu_probs_tensor = True
+    sampler.should_modify_greedy_probs_inplace = False
+
+    sampling_params = SamplingParams(temperature=0)
+
+    mock_inplace = Mock()
+    with patch(
+            "vllm.model_executor.layers.sampler._modify_greedy_probs_inplace",
+            mock_inplace):
+
+        sampler_output = _do_sample(batch_size, fake_logits, sampler,
+                                    sampling_params, device)
+        mock_inplace.assert_not_called()
+
+    assert sampler_output.sampled_token_probs is not None
+    assert sampler_output.logprobs is not None
+    assert sampler_output.sampled_token_ids is not None

vllm/lora/layers.py

@@ -1067,6 +1067,10 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
     def include_gpu_probs_tensor(self):
         return self.base_layer.include_gpu_probs_tensor
 
+    @property
+    def should_modify_greedy_probs_inplace(self):
+        return self.base_layer.should_modify_greedy_probs_inplace
+
     def create_lora_weights(
         self,
         max_loras: int,

vllm/model_executor/layers/sampler.py

@@ -51,6 +51,7 @@ class Sampler(nn.Module):
         # containing the sampled token ids and probabilities. This is used by
         # speculative decoding.
         self.include_gpu_probs_tensor = False
+        self.should_modify_greedy_probs_inplace = False
 
     def _init_sampling_tensors(
         self,
@@ -177,8 +178,7 @@ class Sampler(nn.Module):
         This is used by speculative decoding, which requires that the sampling
         method be encoded into the probability distribution.
         """
-        # Modify greedy probs if include_gpu_probs_tensor is set.
+        return self.should_modify_greedy_probs_inplace
-        return self.include_gpu_probs_tensor
 
 
 def _get_bin_counts_and_mask(

vllm/spec_decode/medusa_worker.py

@@ -35,6 +35,9 @@ class MedusaWorker(NonLLMProposerWorkerBase, Worker):
     def set_include_gpu_probs_tensor(self):
         pass
 
+    def set_should_modify_greedy_probs_inplace(self):
+        pass
+
     @torch.inference_mode()
     def sampler_output(
         self,

vllm/spec_decode/multi_step_worker.py

@@ -46,6 +46,10 @@ class MultiStepWorker(Worker, ProposerWorkerBase):
         # Need include_gpu_probs_tensor for MultiStepWorker
         self.model_runner.model.sampler.include_gpu_probs_tensor = True
 
+    def set_should_modify_greedy_probs_inplace(self) -> None:
+        self.model_runner.model.sampler.should_modify_greedy_probs_inplace = (
+            True)
+
     @torch.inference_mode()
     def sampler_output(
         self,

vllm/spec_decode/proposer_worker_base.py

@@ -28,6 +28,10 @@ class ProposerWorkerBase(LoraNotSupportedWorkerBase, SpeculativeProposer):
         """Implementation optional"""
         pass
 
+    def set_should_modify_greedy_probs_inplace(self) -> None:
+        """Implementation optional"""
+        pass
+
 
 class NonLLMProposerWorkerBase(ProposerWorkerBase, ABC):
     """Proposer worker which does not use a model with kvcache"""

vllm/spec_decode/smaller_tp_proposer_worker.py

@@ -83,6 +83,12 @@ class SmallerTpProposerWorker(ProposerWorkerBase):
         # Need include_gpu_probs_tensor for multi_step_worker
         self._worker.set_include_gpu_probs_tensor()
 
+    def set_should_modify_greedy_probs_inplace(self) -> None:
+        if self._is_dummy:
+            return
+
+        self._worker.set_should_modify_greedy_probs_inplace()
+
     def load_model(self) -> None:
         if self._is_dummy:
             return

vllm/spec_decode/spec_decode_worker.py

@@ -295,7 +295,10 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
         """
         (self.scorer_worker.model_runner.model.sampler.include_gpu_probs_tensor
          ) = True
+        (self.scorer_worker.model_runner.model.sampler.
+         should_modify_greedy_probs_inplace) = True
         self.proposer_worker.set_include_gpu_probs_tensor()
+        self.proposer_worker.set_should_modify_greedy_probs_inplace()
 
     def determine_num_available_blocks(self) -> Tuple[int, int]:
         """Determine the number of cache blocks to use.