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.
64 lines
2.1 KiB
Python
64 lines
2.1 KiB
Python
from enum import IntEnum
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from vllm.model_executor.pooling_metadata import (PoolingMetadata,
|
|
PoolingTensors)
|
|
from vllm.sequence import EmbeddingSequenceGroupOutput, PoolerOutput
|
|
|
|
|
|
class PoolingType(IntEnum):
|
|
"""Enumeration for different types of pooling methods."""
|
|
LAST = 0
|
|
ALL = 1
|
|
|
|
|
|
class Pooler(nn.Module):
|
|
"""A layer that pools specific information from hidden states.
|
|
|
|
This layer does the following:
|
|
1. Extracts specific tokens or aggregates data based on pooling method.
|
|
2. Normalizes output if specified.
|
|
3. Returns structured results as `PoolerOutput`.
|
|
|
|
Attributes:
|
|
pooling_type: The type of pooling to use (LAST, AVERAGE, MAX).
|
|
normalize: Whether to normalize the pooled data.
|
|
"""
|
|
|
|
def __init__(self, pooling_type: PoolingType, normalize: bool):
|
|
super().__init__()
|
|
self.pooling_type = pooling_type
|
|
self.normalize = normalize
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
pooling_metadata: PoolingMetadata,
|
|
) -> PoolerOutput:
|
|
"""Pools specific information from hidden states based on metadata."""
|
|
prompt_lens = PoolingTensors.from_pooling_metadata(
|
|
pooling_metadata, hidden_states.device).prompt_lens
|
|
|
|
if self.pooling_type == PoolingType.LAST:
|
|
last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
|
|
pooled_data = hidden_states[last_token_flat_indices]
|
|
elif self.pooling_type == PoolingType.ALL:
|
|
offset = 0
|
|
pooled_data = []
|
|
for prompt_len in prompt_lens:
|
|
pooled_data.append(hidden_states[offset:offset + prompt_len])
|
|
offset += prompt_len
|
|
else:
|
|
raise ValueError(f"Invalid pooling type: {self.pooling_type}")
|
|
|
|
if self.normalize:
|
|
pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
|
|
|
|
pooled_outputs = [
|
|
EmbeddingSequenceGroupOutput(data.tolist()) for data in pooled_data
|
|
]
|
|
|
|
return PoolerOutput(outputs=pooled_outputs)
|