[Feature] Integrate Suffix Spec Decoding (#4045)

### What this PR does / why we need it?
This PR integrate suffix decoding (https://arxiv.org/abs/2411.04975)
from vllm (https://github.com/vllm-project/vllm/pull/25784)

#
Suffix Decoding is a dynamic n-gram matching method that:

1. Uses suffix trees to generate speculative tokens quickly using branch
frequency counts.
2. Can keep a history of prior model responses, which tends to work very
well with repetitive agentic use cases.
3. Can be dynamically updated with newly generated tokens, and FIFO
eviction of older requests.
#
### Does this PR introduce _any_ user-facing change?
This feature should be implemented as opt-in and remain seamless for
users who do not require suffix speculative decoding.

For users who wish to enable it, they must first install
arctic-inference:
`pip install arctic-inference
`

After installation, the suffix speculative decoding feature can be
enabled using the following speculative config:
`--speculative_config '{"method": "suffix", "num_speculative_tokens":
5}'
`

### How was this patch tested?
This PR is currently being tested on vLLM
main:83f478bb19
 with PR https://github.com/vllm-project/vllm/pull/25784

In our previous testing, suffix decoding achieved a 13%-30% throughput
improvement over n-gram on the sonnet dataset, tested on vllm-ascend
v0.9.1 with concurrency ranging from 2 to 40.

- vLLM version: v0.11.2

---------

Signed-off-by: fluctlux <38945811+fluctlux@users.noreply.github.com>

vllm_ascend/spec_decode/__init__.py

@@ -19,6 +19,7 @@
 from vllm_ascend.spec_decode.eagle_proposer import EagleProposer
 from vllm_ascend.spec_decode.mtp_proposer import MtpProposer
 from vllm_ascend.spec_decode.ngram_proposer import NgramProposer
+from vllm_ascend.spec_decode.suffix_proposer import SuffixDecodingProposer
 from vllm_ascend.torchair.torchair_mtp_proposer import TorchairMtpProposer
 
 
@@ -35,6 +36,8 @@ def get_spec_decode_method(method,
         if is_torchair_graph:
             return TorchairMtpProposer(vllm_config, device, runner)
         return MtpProposer(vllm_config, device, runner)
+    elif method == 'suffix':
+        return SuffixDecodingProposer(vllm_config, device, runner)
     else:
         raise ValueError("Unknown speculative decoding method: "
                          f"{method}")

vllm_ascend/spec_decode/interface.py

@@ -14,6 +14,7 @@ class SpecDcodeType(enum.Enum):
     EAGLE = 1
     EAGLE3 = 2
     MTP = 4
+    SUFFIX = 5
 
 
 class Proposer:
@@ -51,4 +52,4 @@ class Proposer:
                            attn_metadata=None,
                            aux_hidden_states: torch.Tensor = None):
         """Called by execute_model in model_runner"""
-        raise NotImplementedError
+        raise NotImplementedError

vllm_ascend/spec_decode/suffix_proposer.py

@@ -0,0 +1,45 @@
+import torch
+from vllm.config import CUDAGraphMode
+from vllm.v1.spec_decode.suffix_decoding import \
+    SuffixDecodingProposer as VllmSuffixDecodingProposer
+
+from vllm_ascend.spec_decode.interface import Proposer, SpecDcodeType
+
+
+class SuffixDecodingProposer(VllmSuffixDecodingProposer, Proposer):
+
+    def __init__(self, vllm_config, device, runner):
+        super().__init__(vllm_config)
+        self.name = SpecDcodeType.SUFFIX
+        self.device = device
+        self.runner = runner
+
+    def load_model(self, *args, **kwargs):
+        # No model to load.
+        pass
+
+    @torch.inference_mode()
+    def dummy_run(self,
+                  num_tokens,
+                  with_prefill=None,
+                  skip_attn=None,
+                  num_reqs=None,
+                  num_tokens_across_dp=None,
+                  aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
+                  batch_descriptor=None,
+                  dummy_compute_logits=lambda hidden_states: None):
+        pass
+
+    def generate_token_ids(self,
+                           valid_sampled_token_ids,
+                           sampling_metadata=None,
+                           scheduler_output=None,
+                           spec_decode_metadata=None,
+                           positions=None,
+                           num_scheduled_tokens=None,
+                           hidden_states=None,
+                           attn_metadata=None,
+                           aux_hidden_states=None) -> list[list[int]]:
+        draft_token_ids = self.propose(self.runner.input_batch,
+                                       valid_sampled_token_ids)
+        return draft_token_ids