[fix] baseline4 ops_sh fix
This commit is contained in:
@@ -1,33 +0,0 @@
|
|||||||
from patch_utils import package_root, replace_once
|
|
||||||
|
|
||||||
|
|
||||||
CUSTOM_OPS = package_root("vllm") / "_custom_ops.py"
|
|
||||||
|
|
||||||
CLEAN_BLOCK = """\
|
|
||||||
def copy_blocks(key_caches: List[torch.Tensor],
|
|
||||||
value_caches: List[torch.Tensor],
|
|
||||||
block_mapping: torch.Tensor) -> None:
|
|
||||||
ixf_F.copy_blocks(key_caches, value_caches, block_mapping)
|
|
||||||
"""
|
|
||||||
|
|
||||||
COMPATIBLE_BLOCK = """\
|
|
||||||
def copy_blocks(key_caches: List[torch.Tensor],
|
|
||||||
value_caches: List[torch.Tensor],
|
|
||||||
block_mapping: torch.Tensor) -> None:
|
|
||||||
# BI100 CoreX 3.2.3 exposes vllm_copy_blocks, not copy_blocks.
|
|
||||||
_fn = getattr(ixf_F, "copy_blocks", None) or getattr(
|
|
||||||
ixf_F, "vllm_copy_blocks", None)
|
|
||||||
if _fn is None:
|
|
||||||
raise RuntimeError(
|
|
||||||
"ixformer exposes neither copy_blocks nor vllm_copy_blocks")
|
|
||||||
_fn(key_caches, value_caches, block_mapping)
|
|
||||||
"""
|
|
||||||
|
|
||||||
|
|
||||||
replace_once(
|
|
||||||
CUSTOM_OPS,
|
|
||||||
CLEAN_BLOCK,
|
|
||||||
COMPATIBLE_BLOCK,
|
|
||||||
required=True,
|
|
||||||
already_contains="BI100 CoreX 3.2.3 exposes vllm_copy_blocks",
|
|
||||||
)
|
|
||||||
@@ -63,3 +63,32 @@ replace_once(
|
|||||||
required=True,
|
required=True,
|
||||||
already_contains="BI100 CoreX 3.2.3 exposes vllm_swap_blocks",
|
already_contains="BI100 CoreX 3.2.3 exposes vllm_swap_blocks",
|
||||||
)
|
)
|
||||||
|
|
||||||
|
# ---------- copy_blocks: same naming mismatch ----------
|
||||||
|
COPY_CLEAN = """\
|
||||||
|
def copy_blocks(key_caches: List[torch.Tensor],
|
||||||
|
value_caches: List[torch.Tensor],
|
||||||
|
block_mapping: torch.Tensor) -> None:
|
||||||
|
ixf_F.copy_blocks(key_caches, value_caches, block_mapping)
|
||||||
|
"""
|
||||||
|
|
||||||
|
COPY_COMPAT = """\
|
||||||
|
def copy_blocks(key_caches: List[torch.Tensor],
|
||||||
|
value_caches: List[torch.Tensor],
|
||||||
|
block_mapping: torch.Tensor) -> None:
|
||||||
|
# BI100 CoreX 3.2.3 exposes vllm_copy_blocks, not copy_blocks.
|
||||||
|
_fn = getattr(ixf_F, "copy_blocks", None) or getattr(
|
||||||
|
ixf_F, "vllm_copy_blocks", None)
|
||||||
|
if _fn is None:
|
||||||
|
raise RuntimeError(
|
||||||
|
"ixformer exposes neither copy_blocks nor vllm_copy_blocks")
|
||||||
|
_fn(key_caches, value_caches, block_mapping)
|
||||||
|
"""
|
||||||
|
|
||||||
|
replace_once(
|
||||||
|
CUSTOM_OPS,
|
||||||
|
COPY_CLEAN,
|
||||||
|
COPY_COMPAT,
|
||||||
|
required=True,
|
||||||
|
already_contains="BI100 CoreX 3.2.3 exposes vllm_copy_blocks",
|
||||||
|
)
|
||||||
@@ -149,9 +149,8 @@ cp ./bi100_profile.py "${VLLM_ROOT}/bi100_profile.py"
|
|||||||
cp ./block_major_kv_cache.py "${VLLM_ROOT}/block_major_kv_cache.py"
|
cp ./block_major_kv_cache.py "${VLLM_ROOT}/block_major_kv_cache.py"
|
||||||
cp ./gdn_prefix.py "${VLLM_ROOT}/gdn_prefix.py"
|
cp ./gdn_prefix.py "${VLLM_ROOT}/gdn_prefix.py"
|
||||||
|
|
||||||
build_stage "installing CoreX paged-KV swap/copy compatibility"
|
build_stage "installing CoreX paged-KV swap compatibility"
|
||||||
python3 ./patch_corex_swap_blocks.py
|
python3 ./patch_corex_swap_blocks.py
|
||||||
python3 ./patch_corex_copy_blocks.py
|
|
||||||
python3 ./patch_block_major_cache_engine.py
|
python3 ./patch_block_major_cache_engine.py
|
||||||
python3 ./patch_worker_cache_transfer_order.py
|
python3 ./patch_worker_cache_transfer_order.py
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user