Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ use static compile

Browse Source
This commit is contained in:
Disty0
2025-07-20 16:25:57 +03:00
parent 27ce0dea9a
commit 25a4731a97
2 changed files with 12 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/dequantizer.py
+5 -4
View File
@@ -176,10 +176,11 @@ dequantizer_dict = {
if shared.opts.sdnq_dequantize_compile:
try:
torch._dynamo.config.cache_size_limit = max(8192, torch._dynamo.config.cache_size_limit)
dequantize_asymmetric_compiled = torch.compile(dequantize_asymmetric, fullgraph=True)
dequantize_symmetric_compiled = torch.compile(dequantize_symmetric, fullgraph=True)
dequantize_packed_int_asymmetric_compiled = torch.compile(dequantize_packed_int_asymmetric, fullgraph=True)
dequantize_packed_int_symmetric_compiled = torch.compile(dequantize_packed_int_symmetric, fullgraph=True)
torch._dynamo.config.accumulated_recompile_limit = max(8192, torch._dynamo.config.accumulated_recompile_limit)
dequantize_asymmetric_compiled = torch.compile(dequantize_asymmetric, fullgraph=True, dynamic=False)
dequantize_symmetric_compiled = torch.compile(dequantize_symmetric, fullgraph=True, dynamic=False)
dequantize_packed_int_asymmetric_compiled = torch.compile(dequantize_packed_int_asymmetric, fullgraph=True, dynamic=False)
dequantize_packed_int_symmetric_compiled = torch.compile(dequantize_packed_int_symmetric, fullgraph=True, dynamic=False)
except Exception as e:
shared.log.warning(f"Quantization: type=sdnq Dequantize using torch.compile is not available: {e}")
dequantize_asymmetric_compiled = dequantize_asymmetric
modules/sdnq/forward.py
+7 -6
View File
@@ -399,11 +399,12 @@ def quantized_conv_transpose_3d_forward(self, input: torch.FloatTensor, output_s
if shared.opts.sdnq_dequantize_compile:
try:
torch._dynamo.config.cache_size_limit = max(8192, torch._dynamo.config.cache_size_limit)
int8_matmul = torch.compile(int8_matmul, fullgraph=True)
fp8_matmul = torch.compile(fp8_matmul, fullgraph=True)
fp8_matmul_tensorwise = torch.compile(fp8_matmul_tensorwise, fullgraph=True)
conv_int8_matmul = torch.compile(conv_int8_matmul, fullgraph=True)
conv_fp8_matmul = torch.compile(conv_fp8_matmul, fullgraph=True)
conv_fp8_matmul_tensorwise = torch.compile(conv_fp8_matmul_tensorwise, fullgraph=True)
torch._dynamo.config.accumulated_recompile_limit = max(8192, torch._dynamo.config.accumulated_recompile_limit)
int8_matmul = torch.compile(int8_matmul, fullgraph=True, dynamic=False)
fp8_matmul = torch.compile(fp8_matmul, fullgraph=True, dynamic=False)
fp8_matmul_tensorwise = torch.compile(fp8_matmul_tensorwise, fullgraph=True, dynamic=False)
conv_int8_matmul = torch.compile(conv_int8_matmul, fullgraph=True, dynamic=False)
conv_fp8_matmul = torch.compile(conv_fp8_matmul, fullgraph=True, dynamic=False)
conv_fp8_matmul_tensorwise = torch.compile(conv_fp8_matmul_tensorwise, fullgraph=True, dynamic=False)
except Exception as e:
shared.log.warning(f"Quantization: type=sdnq MatMul using torch.compile is not available: {e}")