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.
49 lines
1.6 KiB
Python
49 lines
1.6 KiB
Python
"""Block manager utils."""
|
|
from vllm.sequence import SequenceGroup
|
|
from vllm.utils import (STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE,
|
|
STR_NOT_IMPL_ENC_DEC_SWA)
|
|
|
|
|
|
def _get_block_mgr_sliding_window_attr(block_mgr):
|
|
'''
|
|
BlockManagerV1 and BlockManagerV2 have slightly different
|
|
members related to sliding window attention (SWA). This
|
|
function extracts the appropriate member to use for determining
|
|
whether SWA is enabled.
|
|
|
|
Arguments:
|
|
|
|
* block_mgr: BlockManagerV1 or BlockManagerV2 instance
|
|
'''
|
|
|
|
if hasattr(block_mgr, 'block_sliding_window'):
|
|
return block_mgr.block_sliding_window
|
|
if hasattr(block_mgr, 'max_block_sliding_window'):
|
|
return block_mgr.max_block_sliding_window
|
|
|
|
raise AttributeError("Block manager instance has neither " + \
|
|
"block_sliding_window nor " + \
|
|
"max_block_sliding_window attributes.")
|
|
|
|
|
|
def check_no_caching_or_swa_for_blockmgr_encdec(
|
|
block_mgr, seq_group: SequenceGroup) -> None:
|
|
'''
|
|
Enforce that prefix caching & sliding-window attention (SWA)
|
|
are currently unsupported *specifically* for encoder/decoder models.
|
|
|
|
Raises NotImplementedError if unsupported scenario is detected.
|
|
|
|
Arguments:
|
|
|
|
* block_mgr: BlockSpaceManager instance
|
|
* seq_group: SequenceGroup passed to block_mgr
|
|
'''
|
|
|
|
if seq_group.is_encoder_decoder():
|
|
if _get_block_mgr_sliding_window_attr(block_mgr) is not None:
|
|
raise NotImplementedError(STR_NOT_IMPL_ENC_DEC_SWA)
|
|
|
|
if block_mgr.enable_caching:
|
|
raise NotImplementedError(STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE)
|