Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

IPEX enable int_mm fallback for torch 2.8

Browse Source
This commit is contained in:
Disty0
2025-08-03 18:38:13 +03:00
parent 84bccc07ca
commit 2cdb15b1ce
1 changed files with 8 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/intel/ipex/__init__.py
+8 -6
View File
@@ -126,13 +126,15 @@ def ipex_init(): # pylint: disable=too-many-statements
torch.cuda.List = torch.xpu.List
if torch_version < 2.9:
# torch._int_mm via onednn is supposed to land on pytorch with torch 2.8 or 2.9
# ipex 2.7+ has experimental torch._int_mm support but uses the cpu with torch.compile and also runs as slow as onednn.qlinear
if (not has_ipex or torch_version <= 2.7) and hasattr(torch.ops, "onednn") and hasattr(torch.ops.onednn, "qlinear_pointwise"):
def onednn_mm(x: torch.Tensor, y: torch.Tensor, output_dtype=torch.float32):
# supports int8, fp32, fp16, and bf16 matmul with accumulation using a different float dtype
# torch._int_mm via onednn quantized matmul is supported with torch 2.9
# ipex 2.7+ has the same torch._int_mm support as torch 2.9 but doesn't support torch.compile
# torch._int_mm directly uses onednn quantized matmul
# onednn qlinear is a wrapper around onednn quantized matmul
if hasattr(torch.ops, "onednn") and hasattr(torch.ops.onednn, "qlinear_pointwise"):
def onednn_mm(x: torch.Tensor, y: torch.Tensor):
# supports int8, fp32, fp16, and bf16 matmul with accumulation using a different dtype
# int8 matmul with onednn is slower than 16 bit with dim_size < 4096
return torch.ops.onednn.qlinear_pointwise(x, 1.0, 0, y, torch.ones(1, device=y.device), torch.zeros(1, device=y.device), None, 1.0, 0, output_dtype, "none", [], "none")
return torch.ops.onednn.qlinear_pointwise.default(x, 1.0, 0, y, torch.ones(1, device=y.device), torch.zeros(1, device=y.device), None, 1.0, 0, torch.float32, "none", [], "none")
torch._int_mm = onednn_mm
try:
# torch.compile fix