ef6abf3dc7
Adds ALL files needed for Dockerfile build:
- qwen3_6_scripts/ (baseline patches + our optimizations)
- vllm/ (full vllm package)
- paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
- Dockerfile + computility-run.yaml
Our optimizations vs baseline:
1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
Triton try/fallback, V2 heuristic, threshold 32K→64K
2. paged_attention_v2_pytorch.py: fills NotImplementedError,
single-bmm Phase 1 (195 launches → 3)
3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
max-num-batched-tokens 8192→16384
This repo can now be submitted to dev.modelhub.org.cn as-is.
57 lines
2.0 KiB
Python
57 lines
2.0 KiB
Python
from abc import ABC, abstractmethod
|
|
from typing import List, Optional, Set, Tuple
|
|
|
|
from vllm.model_executor.layers.sampler import SamplerOutput
|
|
from vllm.sequence import ExecuteModelRequest
|
|
from vllm.spec_decode.interfaces import SpeculativeProposer
|
|
from vllm.worker.worker_base import LoraNotSupportedWorkerBase
|
|
|
|
|
|
class ProposerWorkerBase(LoraNotSupportedWorkerBase, SpeculativeProposer):
|
|
"""Interface for proposer workers"""
|
|
|
|
@abstractmethod
|
|
def sampler_output(
|
|
self,
|
|
execute_model_req: ExecuteModelRequest,
|
|
sample_len: int,
|
|
# A set containing all sequence IDs that were assigned bonus tokens
|
|
# in their last forward pass. This set is used to backfill the KV cache
|
|
# with the key-value pairs of the penultimate token in the sequences.
|
|
# This parameter is only used by the MultiStepWorker, which relies on
|
|
# the KV cache for token generation. It is not used by workers that
|
|
# do not utilize the KV cache.
|
|
seq_ids_with_bonus_token_in_last_step: Set[int]
|
|
) -> Tuple[Optional[List[SamplerOutput]], bool]:
|
|
raise NotImplementedError
|
|
|
|
def set_include_gpu_probs_tensor(self) -> None:
|
|
"""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"""
|
|
|
|
def execute_model(
|
|
self,
|
|
execute_model_req: Optional[ExecuteModelRequest] = None
|
|
) -> List[SamplerOutput]:
|
|
"""get_spec_proposals is used to get the proposals"""
|
|
return []
|
|
|
|
def determine_num_available_blocks(self) -> Tuple[int, int]:
|
|
"""This is never called on the proposer, only the target model"""
|
|
raise NotImplementedError
|
|
|
|
def initialize_cache(self, num_gpu_blocks: int,
|
|
num_cpu_blocks: int) -> None:
|
|
pass
|
|
|
|
def get_cache_block_size_bytes(self) -> int:
|
|
return 0
|