dev

ascend-robot[Bugfix][quant]Switch MXFP4QuantFA float path to npu_fusion_attention

7491360d创建于 7月15日历史提交

文件	最后提交记录	最后更新时间
__init__.py	[Feature][quant]Cover MXFP4 quantization updates Co-authored-by: lijinxi<lijinxi2@huawei.com> # message auto-generated for no-merge-commit merge: !365 merge dev into dev [Feature][quant]Cover MXFP4 quantization updates Created-by: weixin_44144262 Commit-by: lijinxi Merged-by: ascend-robot Description: # Purpose 新增部署端mxfp4 fa逻辑 # Test Plan 测试Wan2.2 e2e端到端精度，性能 # Test Report  See merge request: Ascend/MindIE-SD!365	1 个月前
config.py	[Quant]Optimize online quantization config and docs Co-authored-by: lijinxi<lijinxi2@huawei.com> # message auto-generated for no-merge-commit merge: !401 merge dev into dev [Quant]Optimize online quantization config and docs Created-by: weixin_44144262 Commit-by: lijinxi Merged-by: ascend-robot Description: # Purpose 更新在线量化文档和功能优化 新增fa和mm算法配置 新增回退手段 # Test Plan 测试Wan2.2 e2e # Test Report  See merge request: Ascend/MindIE-SD!401	1 个月前
layer.py	[Bugfix][quant]Switch MXFP4QuantFA float path to npu_fusion_attention Co-authored-by: shijia10<shijia40@huawei.com> # message auto-generated for no-merge-commit merge: !454 merge quant into dev [Bugfix][quant]Switch MXFP4QuantFA float path to npu_fusion_attention Created-by: shijia10 Commit-by: shijia10 Merged-by: ascend-robot Description: <!-- PR描述模板更新日期：20251224 --> # Which issue(s) this PR fixes or accomplishes # Purpose 将 MXFP4QuantFA._forward_float 中的注意力算子从 fused_infer_attention_score_v2 切换为标准的 torch_npu.npu_fusion_attention： - 参数名对齐新算子：num_query_heads -> head_num、softmax_scale -> scale、pre_tokens/next_tokens -> pre_tockens/next_tockens - 移除 out_dtype 显式指定 - 移除 _crop_fa_output 后处理，直接返回算子输出 - 影响文件：mindiesd/quantization/layer.py # Test Plan > 建议在 NPU 上对 MXFP4QuantFA 的 _forward_float 路径进行验证： > 1. 不同 input_layout（BNSD / BSND / BSH 等）下输出 shape 是否与改动前一致 > 2. 与改动前基线做数值精度对比，确认 npu_fusion_attention 默认输出 dtype 满足 MXFP4 精度要求 > 3. 确认 pre_tockens / next_tockens 的 mask 行为与原 pre_tokens / next_tokens 一致 # Test Report > 精度测试结果  See merge request: Ascend/MindIE-SD!454	1 个月前
mode.py	[feature]FA MXFP8 特性提交，新注册标识字符串“MXFP8_DYNAMIC”，并同步提交量化代码类 Co-authored-by: AzGewhin961224<wangzihe9@h-partners.com> # message auto-generated for no-merge-commit merge: !375 merge dev into dev [feature]FA MXFP8 特性提交，新注册标识字符串“MXFP8_DYNAMIC”，并同步提交量化代码类 Created-by: LagavulinNeat Commit-by: AzGewhin961224 Merged-by: ascend-robot Description: # Purpose 接入FA MXFP8_DYNAMIC量化算法： 读取模型权重描述文件，识别对应attn层使能MXFP8_DYNAMIC量化算法，保存对应attn层的旋转矩阵 为对应的attn添加FA量化算法'fa_quant'：在传入FA算子前，先将Q、K进行@旋转&MXFP8量化 # Test Plan > 设计了哪些测试内容，指导他人如何对你的PR进行测试\ > Apply information to show others your test design and how to test your Pull Request 模型侧通过对应层是否包含'fa_quant"成员判断是否使能FA量化，并传入对应layout的Q/K/V，观察量化前后的加速比和精度损失情况。 参考test_quantize.py：test_add_fa_quant_with_valid_layer # Test Report Wan2.2-T2V-A14B测试情况：精度肉眼观测无问题，单算子性能接近FP8 tiling256 UT运行结果： `` python test_quantize.py .................../usr/local/lib/python3.11/site-packages/mindiesd/quantization/layer.py:185: UserWarning: The use of x.T on tensors of dimension other than 2 to reverse their shape is deprecated and it will throw an error in a future release. Consider x.mT to transpose batches of matrices or x.permute(*torch.arange(x.ndim - 1, -1, -1)) to reverse the dimensions of a tensor. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:4416.) weight_scale = get_quant_weight(weights, f'{prefix}.weight_scale').T.to(self.dtype) .....s............... ---------------------------------------------------------------------- Ran 40 tests in 0.179s OK (skipped=1) `` See merge request: Ascend/MindIE-SD!375	1 个月前
quantize.py	[Quant]Optimize online quantization config and docs Co-authored-by: lijinxi<lijinxi2@huawei.com> # message auto-generated for no-merge-commit merge: !401 merge dev into dev [Quant]Optimize online quantization config and docs Created-by: weixin_44144262 Commit-by: lijinxi Merged-by: ascend-robot Description: # Purpose 更新在线量化文档和功能优化 新增fa和mm算法配置 新增回退手段 # Test Plan 测试Wan2.2 e2e # Test Report  See merge request: Ascend/MindIE-SD!401	1 个月前
utils.py	[Quant]Optimize online quantization config and docs Co-authored-by: lijinxi<lijinxi2@huawei.com> # message auto-generated for no-merge-commit merge: !401 merge dev into dev [Quant]Optimize online quantization config and docs Created-by: weixin_44144262 Commit-by: lijinxi Merged-by: ascend-robot Description: # Purpose 更新在线量化文档和功能优化 新增fa和mm算法配置 新增回退手段 # Test Plan 测试Wan2.2 e2e # Test Report  See merge request: Ascend/MindIE-SD!401	1 个月前