From f6777b168d57d13ed43c32d56c1e08ad4d0022f3 Mon Sep 17 00:00:00 2001 From: Dity0 Date: Mon, 24 Aug 2026 19:34:54 +0300 Subject: [PATCH] SDNQ Atten: add quantize_fp32 --- extensions-builtin/sdnq | 2 +- modules/attention.py | 1 + modules/ui_definitions.py | 1 + 3 files changed, 3 insertions(+), 1 deletion(-) diff --git a/extensions-builtin/sdnq b/extensions-builtin/sdnq index 3986a375f..b99d9af15 160000 --- a/extensions-builtin/sdnq +++ b/extensions-builtin/sdnq @@ -1 +1 @@ -Subproject commit 3986a375fc419945538063fd468f2c7069c18575 +Subproject commit b99d9af158a81fdd713dcbe31c9eb36f2332e54c diff --git a/modules/attention.py b/modules/attention.py index 1aad29692..4f331a147 100644 --- a/modules/attention.py +++ b/modules/attention.py @@ -38,6 +38,7 @@ def set_sdnq_attention(): smooth_k=shared.opts.sdnq_attention_smooth_k, use_hadamard=shared.opts.sdnq_attention_use_hadamard, hadamard_group_size=shared.opts.sdnq_attention_hadamard_group_size, + quantize_fp32=shared.opts.sdnq_attention_quantize_fp32, use_fp16_accum=shared.opts.sdnq_attention_use_fp16_accum, ) else: diff --git a/modules/ui_definitions.py b/modules/ui_definitions.py index f77ec04d1..ed0dab9c6 100644 --- a/modules/ui_definitions.py +++ b/modules/ui_definitions.py @@ -258,6 +258,7 @@ def create_settings(cmd_opts): "sdnq_attention_smooth_k": OptionInfo(True, "SDNQ Attention use Smooth K", gr.Checkbox), "sdnq_attention_use_hadamard": OptionInfo(False, "SDNQ Attention use Hadamard", gr.Checkbox), "sdnq_attention_use_fp16_accum": OptionInfo(False, "SDNQ Attention use FP16 Accumulation", gr.Checkbox), + "sdnq_attention_quantize_fp32": OptionInfo(True, "SDNQ Attention Quantize FP32", gr.Checkbox), "sdnq_attention_matmul_type": OptionInfo("enabled", "SDNQ Attention MatMul type", gr.Radio, {"choices": sdnq_matmul_modes}), "sdnq_attention_pv_matmul_type": OptionInfo("disabled", "SDNQ Attention PV MatMul type", gr.Radio, {"choices": sdnq_matmul_modes}), "sdnq_attention_hadamard_group_size": OptionInfo(256, "SDNQ Attention Hadamard Group Size", gr.Slider, {"minimum": 4, "maximum": 1024, "step": 1}),