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.
62 lines
1.7 KiB
Python
62 lines
1.7 KiB
Python
from dataclasses import dataclass
|
|
from typing import List, Optional
|
|
|
|
|
|
@dataclass
|
|
class BeamSearchSequence:
|
|
"""A sequence for beam search.
|
|
It keeps track of the tokens and the log probability of the sequence.
|
|
The text field is optional and will only be filled when the sequence is
|
|
about to be returned to the user.
|
|
"""
|
|
# The tokens includes the prompt.
|
|
tokens: List[int]
|
|
cum_logprob: float = 0.0
|
|
text: Optional[str] = None
|
|
|
|
|
|
@dataclass
|
|
class BeamSearchOutput:
|
|
"""The output of beam search.
|
|
It contains the list of the best beam search sequences.
|
|
The length of the list is equal to the beam width.
|
|
"""
|
|
sequences: List[BeamSearchSequence]
|
|
|
|
|
|
class BeamSearchInstance:
|
|
|
|
def __init__(self, prompt_tokens: List[int]):
|
|
self.beams: List[BeamSearchSequence] = [
|
|
BeamSearchSequence(tokens=prompt_tokens)
|
|
]
|
|
self.completed: List[BeamSearchSequence] = []
|
|
|
|
|
|
def get_beam_search_score(
|
|
tokens: List[int],
|
|
cumulative_logprob: float,
|
|
eos_token_id: int,
|
|
length_penalty: float = 1.0,
|
|
) -> float:
|
|
"""Calculate the beam search score with length penalty.
|
|
|
|
Adapted from
|
|
|
|
https://github.com/huggingface/transformers/blob/ccb92be23def445f2afdea94c31286f84b89eb5b/src/transformers/generation/beam_search.py#L938
|
|
"""
|
|
seq_len = len(tokens)
|
|
if tokens[-1] == eos_token_id:
|
|
seq_len -= 1
|
|
|
|
return cumulative_logprob / (seq_len**length_penalty)
|
|
|
|
|
|
def create_sort_beams_key_function(eos_token_id: int, length_penalty: float):
|
|
|
|
def sort_beams_key(x: BeamSearchSequence) -> float:
|
|
return get_beam_search_score(x.tokens, x.cum_logprob, eos_token_id,
|
|
length_penalty)
|
|
|
|
return sort_beams_key
|