[Cleanup] Remove obsolete spec decoding compatibility logic (#32003)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-01-09 05:44:18 +00:00
committed by GitHub
parent 7a05d2dc65
commit 29ce48221c
8 changed files with 45 additions and 75 deletions
-10
View File
@@ -55,10 +55,8 @@ class NgramProposer:
# This usually takes less than 1 second.
self.propose(
[[]] * 1024,
[""] * 1024,
np.zeros(1024, dtype=np.int32),
np.zeros((1024, self.max_model_len), dtype=np.int32),
set(),
)
def batch_propose(
@@ -132,10 +130,8 @@ class NgramProposer:
def propose(
self,
sampled_token_ids: list[list[int]],
req_ids: list[str],
num_tokens_no_spec: np.ndarray,
token_ids_cpu: np.ndarray,
spec_decode_unsupported_reqs: set,
) -> list[list[int]]:
# find which requests need ngram proposals
valid_ngram_requests = []
@@ -145,12 +141,6 @@ class NgramProposer:
# Skip speculative decoding.
continue
# Skip requests that require sampling parameters that are not
# supported with speculative decoding.
req_id = req_ids[i]
if req_id in spec_decode_unsupported_reqs:
continue
num_tokens = num_tokens_no_spec[i]
if num_tokens >= self.max_model_len:
# Skip requests that have already reached the max model length.
-6
View File
@@ -46,13 +46,7 @@ class SuffixDecodingProposer:
draft_token_ids.append([])
continue
# Skip requests that require sampling parameters that are not
# supported with speculative decoding.
req_id = input_batch.req_ids[i]
if req_id in input_batch.spec_decode_unsupported_reqs:
draft_token_ids.append([])
continue
num_tokens = input_batch.num_tokens_no_spec[i]
if num_tokens >= self.max_model_len:
# Skip requests that have already reached the max model length.
-14
View File
@@ -1,21 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.sampling_params import SamplingParams
from vllm.triton_utils import tl, triton
_SAMPLING_EPS = 1e-5
def is_spec_decode_unsupported(sampling_params: SamplingParams) -> bool:
"""True if request is incompatible with speculative decoding"""
return (
sampling_params.frequency_penalty != 0.0
or sampling_params.presence_penalty != 0.0
or sampling_params.repetition_penalty != 1.0
or sampling_params.min_p > _SAMPLING_EPS
or sampling_params.logprobs is not None
)
@triton.jit
def eagle_prepare_inputs_padded_kernel(
-7
View File
@@ -22,7 +22,6 @@ from vllm.v1.sample.logits_processor import (
MoveDirectionality,
)
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.spec_decode.utils import is_spec_decode_unsupported
from vllm.v1.utils import copy_slice
from vllm.v1.worker.block_table import MultiGroupBlockTable
@@ -176,9 +175,6 @@ class InputBatch:
self.top_k_cpu = self.top_k_cpu_tensor.numpy()
self.top_k_reqs: set[str] = set()
# IDs of requests which do not support spec decoding
self.spec_decode_unsupported_reqs: set[str] = set()
# Frequency penalty related data structures
self.frequency_penalties = torch.empty(
(max_num_reqs,), dtype=torch.float, device=device
@@ -346,8 +342,6 @@ class InputBatch:
self.block_table.add_row(request.block_ids, req_index)
if sampling_params := request.sampling_params:
if self.is_spec_decode and is_spec_decode_unsupported(sampling_params):
self.spec_decode_unsupported_reqs.add(req_id)
if sampling_params.sampling_type == SamplingType.GREEDY:
# Should avoid division by zero later when apply_temperature.
self.temperature_cpu[req_index] = 0.0
@@ -510,7 +504,6 @@ class InputBatch:
self.random_reqs.discard(req_id)
self.top_p_reqs.discard(req_id)
self.top_k_reqs.discard(req_id)
self.spec_decode_unsupported_reqs.discard(req_id)
self.frequency_penalties_reqs.discard(req_id)
self.presence_penalties_reqs.discard(req_id)
self.repetition_penalties_reqs.discard(req_id)
-2
View File
@@ -3620,10 +3620,8 @@ class GPUModelRunner(
assert isinstance(self.drafter, NgramProposer)
draft_token_ids = self.drafter.propose(
sampled_token_ids,
self.input_batch.req_ids,
self.input_batch.num_tokens_no_spec,
self.input_batch.token_ids_cpu,
self.input_batch.spec_decode_unsupported_reqs,
)
elif spec_config.method == "suffix":
assert isinstance(sampled_token_ids, list)