dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

test docker build

Browse Source
This commit is contained in:
root
2026-08-19 03:09:44 +00:00
parent 7d74ad4e8e
commit c8e8473367
1 changed files with 4 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
qwen3_6_scripts/paged_attn.py
View File

@@ -2290,43 +2290,7 @@ class PagedAttention:
kv_caches: List[torch.Tensor], kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor, src_to_dists: torch.Tensor,
) -> None: ) -> None:
# BI100 CoreX 3.2.3: ixformer has no copy_blocks binding. key_caches = [kv_cache[0] for kv_cache in kv_caches]
# Use xllm_cache.block_copy kernel — single launch for all layers. value_caches = [kv_cache[1] for kv_cache in kv_caches]
if src_to_dists.numel() == 0: ops.copy_blocks(key_caches, value_caches, src_to_dists)
return
import importlib.util as _ilu
import os as _os
# paged_attn.py lives at VLLM_ROOT/attention/ops/paged_attn.py
# xllm_cache.so lives at VLLM_ROOT/xllm_cache.so
_vllm_root = _os.path.dirname(_os.path.dirname(_os.path.dirname(
_os.path.abspath(__file__))))
_so = _os.path.join(_vllm_root, "xllm_cache.so")
_spec = _ilu.spec_from_file_location("xllm_cache", _so)
_mod = _ilu.module_from_spec(_spec)
_spec.loader.exec_module(_mod)
_xllm_block_copy = _mod.block_copy
device = src_to_dists.device
n = src_to_dists.size(0)
# 1:1 mapping → each pair is its own group
src_indices = src_to_dists[:, 0].to(torch.int32).contiguous()
dst_indices = src_to_dists[:, 1].to(torch.int32).contiguous()
cum_sum = torch.arange(1, n + 1, dtype=torch.int32, device=device)
# Build per-layer cache pointer tensors
key_ptrs = torch.tensor(
[kv[0].data_ptr() for kv in kv_caches],
dtype=torch.int64, device=device)
val_ptrs = torch.tensor(
[kv[1].data_ptr() for kv in kv_caches],
dtype=torch.int64, device=device)
numel_per_block = kv_caches[0][0][0].numel()
cache_dtype = kv_caches[0][0].scalar_type()
_xllm_block_copy(key_ptrs, val_ptrs,
src_indices, dst_indices, cum_sum,
numel_per_block, cache_dtype)