[CPU] Upgrade CPU backend to torch-2.6 (#13381)
Signed-off-by: jiang1.li <jiang1.li@intel.com> Co-authored-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
@@ -17,7 +17,7 @@ class _PagedAttention:
|
||||
|
||||
@staticmethod
|
||||
def get_supported_head_sizes() -> List[int]:
|
||||
return [32, 64, 80, 96, 112, 128, 256]
|
||||
return [32, 64, 80, 96, 112, 128, 192, 256]
|
||||
|
||||
@staticmethod
|
||||
def get_kv_cache_shape(
|
||||
|
||||
@@ -254,10 +254,11 @@ def _run_worker_process(
|
||||
# online (in situ) tuning is enabled.
|
||||
# Offline tuning API (record_untuned_is_enabled()) only
|
||||
# available in PyTorch 2.6 or later.
|
||||
import torch.cuda.tunable as tunable
|
||||
if (tunable.is_enabled() and tunable.tuning_is_enabled()
|
||||
and not tunable.record_untuned_is_enabled()):
|
||||
tunable.write_file()
|
||||
if torch.cuda.is_available():
|
||||
import torch.cuda.tunable as tunable
|
||||
if (tunable.is_enabled() and tunable.tuning_is_enabled()
|
||||
and not tunable.record_untuned_is_enabled()):
|
||||
tunable.write_file()
|
||||
|
||||
logger.info("Worker exiting")
|
||||
|
||||
|
||||
@@ -193,10 +193,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
|
||||
global_num_experts: int = -1,
|
||||
expert_map: Optional[torch.Tensor] = None,
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
scoring_func: str = "softmax",
|
||||
e_score_correction_bias: Optional[torch.Tensor] = None,
|
||||
activation: str = "silu",
|
||||
**kwargs,
|
||||
):
|
||||
assert custom_routing_function is None
|
||||
assert activation == "silu", f"{activation} is not supported."
|
||||
return layer.ipex_fusion(
|
||||
x,
|
||||
@@ -206,6 +207,9 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
|
||||
renormalize,
|
||||
topk_group,
|
||||
num_expert_group,
|
||||
custom_routing_function,
|
||||
scoring_func,
|
||||
e_score_correction_bias,
|
||||
)
|
||||
|
||||
def forward_tpu(
|
||||
|
||||
@@ -121,6 +121,9 @@ class CpuPlatform(Platform):
|
||||
# Disable torch async compiling which won't work with daemonic processes
|
||||
os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"
|
||||
|
||||
# MLA attention is not supported
|
||||
os.environ["VLLM_MLA_DISABLE"] = "1"
|
||||
|
||||
# Intel OpenMP setting
|
||||
ld_prealod_str = os.getenv("LD_PRELOAD", "")
|
||||
if "libiomp5.so" in ld_prealod_str:
|
||||
|
||||
Reference in New Issue
Block a user