FLUX INT8 ConvRot 验证报告

RTX 3090 (Ampere, sm_86) | ComfyUI :8188 | 2026-08-08

✅ 验证结果：INT8 ConvRot 完全可用

INT8 Cold Start

45s

INT8 Warm Run

30s

FP8 Baseline

165s

Warm 加速比

5.5×

Warm Cache 下 INT8 vs FP8

5.5× 加速

30s (INT8 warm) vs 165s (FP8 cold with fp8 T5)

⚠️ FP8 cold start 含 fp8 T5 冷加载。INT8 warm 用 fp16 T5 (已缓存)。纯推理步数对比 = INT8 native vs FP8 emulated。

INT8 ConvRot - Cold 45s

INT8 ConvRot cold

flux1-dev.safetensors (23GB BF16) + film-storyboard LoRA + INT8 ConvRot | seed=42

FP8 Baseline 165s

FP8 baseline

flux1-dev-fp8.safetensors (12GB FP8) + film-storyboard LoRA | seed=42

INT8 ConvRot - Warm 30s

INT8 warm

Same INT8 pipeline, model cached | seed=99, different prompt

验证清单

✅ INT8ModelAdapter 节点注册 (7 个 INT8 节点)
✅ quantization_mode: int8_convrot
✅ BF16 模型加载 (flux1-dev.safetensors 23GB)
✅ Ampere 兼容 (无 fp8e4nv 崩溃)
✅ INT8 量化成功 (~10s)
✅ 20 步采样完成
✅ VAEDecode 输出正常
✅ 图像质量正常 (1024×576, 1.3MB)
✅ Warm run 30s (加速明显)
✅ KAP 集成已有 (commit 94597ad1)

关键参数 (验证通过)

参数	值	说明
Node	`INT8ModelAdapter`	非 "Enable INT8 on MODEL"
enable_quantization	`always`	非 boolean True
quantization_mode	`int8_convrot`	Hadamard rotation + INT8
model_type	`auto`	自动检测 FLUX 架构
small_batch_fallback	`only_small_layers`	非 boolean False
runtime_backend	`torch_int_mm`	非 "auto"
Model file	`flux1-dev.safetensors`	BF16 23GB, 非 FP8
weight_dtype	`default`	加载原生 BF16
LoRA 位置	INT8 之前	bake_loaded_loras=true 合并

工作流拓扑

UNETLoader(flux1-dev.safetensors, default)
→ LoraLoader(film-storyboard, 0.8)
→ INT8ModelAdapter(int8_convrot, bake_loaded_loras=true)
→ ModelSamplingFlux(shift=1.73)
→ BasicScheduler(normal, 20 steps)
→ CFGGuider(cfg=3.5) + KSamplerSelect(euler)
→ SamplerCustomAdvanced
→ VAEDecode → SaveImage