修改dyanmic_dual_level_mx_quant的demo & 修改grouped_dyanmic_mx

已合并

修改dyanmic_dual_level_mx_quant的demo & 修改grouped_dyanmic_mx_quant的非连续说明 #8100

季骏创建于 7月30日从ji-jun1/ops-nn： dynamic_dual_level_mx_quant 合入到 cann/ops-nn： master

修改dyanmic_dual_level_mx_quant的demo & 修改grouped_dyanmic_mx_quant的非连续说明 #8100

已合并

季骏创建于 7月30日

从ji-jun1/ops-nn dynamic_dual_level_mx_quant合入到cann/ops-nn master

共 5 个文件变更+28-16

Mquant/dynamic_dual_level_mx_quant/docs/aclnnDynamicDualLevelMxQuant.md+2-2

	@@ -405,7 +405,7 @@ aclnnStatus aclnnDynamicDualLevelMxQuant(
405		405
406	// 2. 构造输入与输出，需要根据API的接口自定义构造	406	// 2. 构造输入与输出，需要根据API的接口自定义构造
407	std::vector<int64_t> xShape = {1, 512};	407	std::vector<int64_t> xShape = {1, 512};
408	- std::vector<int64_t> smoothScaleOptionalShape = {1};	408	+ std::vector<int64_t> smoothScaleOptionalShape = {512};
409	std::vector<int64_t> yOutShape = {1, 512};	409	std::vector<int64_t> yOutShape = {1, 512};
410	std::vector<int64_t> level0ScaleOutShape = {1, 1};	410	std::vector<int64_t> level0ScaleOutShape = {1, 1};
411	std::vector<int64_t> level1ScaleOutShape = {1, 8, 2};	411	std::vector<int64_t> level1ScaleOutShape = {1, 8, 2};
	@@ -422,7 +422,7 @@ aclnnStatus aclnnDynamicDualLevelMxQuant(
422		422
423	// 对应 BF16 的值 (0->0, 16640->8, 17024->64, 17408->512)	423	// 对应 BF16 的值 (0->0, 16640->8, 17024->64, 17408->512)
424	std::vector<uint16_t> xHostData(512, 16640);	424	std::vector<uint16_t> xHostData(512, 16640);
425	- std::vector<uint16_t> smoothScaleOptionalHostData = {0};	425	+ std::vector<uint16_t> smoothScaleOptionalHostData(512, 0);
426	// 对应 float4_e2m1 的值 (0->0, 72->4, 96->32, 120->256)	426	// 对应 float4_e2m1 的值 (0->0, 72->4, 96->32, 120->256)
427	std::vector<uint8_t> yOutHostData(512, 0);	427	std::vector<uint8_t> yOutHostData(512, 0);
428	// 对应 float32 的值 (0->0)	428	// 对应 float32 的值 (0->0)

Mquant/dynamic_dual_level_mx_quant/examples/arch35/test_aclnn_dynamic_dual_level_mx_quant.cpp+15-11

Mquant/grouped_dynamic_mx_quant/docs/aclnnGroupedDynamicMxQuant.md+1-1

	@@ -167,7 +167,7 @@ aclnnStatus aclnnGroupedDynamicMxQuant(
167	<td>FLOAT8_E8M0</td>	167	<td>FLOAT8_E8M0</td>
168	<td>ND</td>	168	<td>ND</td>
169	<td>3</td>	169	<td>3</td>
170	- <td>√</td>	170	+ <td>×</td>
171	</tr>	171	</tr>
172	<tr>	172	<tr>
173	<td>workspaceSize</td>	173	<td>workspaceSize</td>

Mquant/grouped_dynamic_mx_quant/docs/aclnnGroupedDynamicMxQuantV2.md+2-2

	@@ -243,7 +243,7 @@ aclnnStatus aclnnGroupedDynamicMxQuantV2(
243	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>	243	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>
244	<td>ND</td>	244	<td>ND</td>
245	<td>2</td>	245	<td>2</td>
246	- <td>√</td>	246	+ <td>FLOAT8支持非连续，FLOAT4不支持非连续</td>
247	</tr>	247	</tr>
248	<tr>	248	<tr>
249	<td>mxscale (aclTensor*)</td>	249	<td>mxscale (aclTensor*)</td>
	@@ -253,7 +253,7 @@ aclnnStatus aclnnGroupedDynamicMxQuantV2(
253	<td>FLOAT8_E8M0</td>	253	<td>FLOAT8_E8M0</td>
254	<td>ND</td>	254	<td>ND</td>
255	<td>3</td>	255	<td>3</td>
256	- <td>√</td>	256	+ <td>×</td>
257	</tr>	257	</tr>
258	<tr>	258	<tr>
259	<td>workspaceSize</td>	259	<td>workspaceSize</td>

Mquant/grouped_dynamic_mx_quant/op_api/aclnn_grouped_dynamic_mx_quant_v2.cpp+8-0

	@@ -176,6 +176,13 @@ aclnnStatus aclnnGroupedDynamicMxQuantV2GetWorkspaceSize(const aclTensor* x, con
176	return ACLNN_SUCCESS;	176	return ACLNN_SUCCESS;
177	}	177	}
178		178
		179	+ // 当输出y是FP4时，不支持非连续
		180	+ if ((y->GetDataType() == op::DataType::DT_FLOAT4_E2M1 \|\| y->GetDataType() == op::DataType::DT_FLOAT4_E1M2) &&
		181	+ !IsContiguous(y)) {
		182	+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "When the data type of y is float4, it must be contiguous");
		183	+ return ACLNN_ERR_PARAM_INVALID;
		184	+ }
		185	+
179	// x如果非连续，需要转连续	186	// x如果非连续，需要转连续
180	auto selfContiguous = l0op::Contiguous(x, uniqueExecutor.get());	187	auto selfContiguous = l0op::Contiguous(x, uniqueExecutor.get());
181	CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);	188	CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
	@@ -190,6 +197,7 @@ aclnnStatus aclnnGroupedDynamicMxQuantV2GetWorkspaceSize(const aclTensor* x, con
190	// 如果出参y是非连续Tensor，需要把计算完的连续Tensor转非连续	197	// 如果出参y是非连续Tensor，需要把计算完的连续Tensor转非连续
191	auto viewCopyResult0 = l0op::ViewCopy(yOut, y, uniqueExecutor.get());	198	auto viewCopyResult0 = l0op::ViewCopy(yOut, y, uniqueExecutor.get());
192	CHECK_RET(viewCopyResult0 != nullptr, ACLNN_ERR_INNER_NULLPTR);	199	CHECK_RET(viewCopyResult0 != nullptr, ACLNN_ERR_INNER_NULLPTR);
		200	+
193	if (!IsContiguous(mxscale)) {	201	if (!IsContiguous(mxscale)) {
194	OP_LOGE(ACLNN_ERR_PARAM_INVALID, "mxscale must be contiguous.");	202	OP_LOGE(ACLNN_ERR_PARAM_INVALID, "mxscale must be contiguous.");
195	return ACLNN_ERR_PARAM_INVALID;	203	return ACLNN_ERR_PARAM_INVALID;

	@@ -1,12 +1,16 @@
1	/**	1	/**
2	- * This program is free software, you can redistribute it and/or modify.
3	* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	* Copyright (c) 2026 Huawei Technologies Co., Ltd.
4	- * This file is a part of the CANN Open Software.	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
5	- * Licensed under CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
6	* Please refer to the License for details. You may not use this file except in compliance with the License.	5	* Please refer to the License for details. You may not use this file except in compliance with the License.
7	- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
8	- * BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. See LICENSE in the root of	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
9	- * the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
		9	+ */
		10	+
		11	+/*!
		12	+ * \file test_aclnn_dynamic_dual_level_mx_quant.cpp
		13	+ * \brief
10	*/	14	*/
11		15
12	#include <iostream>	16	#include <iostream>
	@@ -95,7 +99,7 @@ int aclnnDynamicDualLevelMxQuantTest(int32_t deviceId, aclrtStream& stream)
95		99
96	// 2. 构造输入与输出，需要根据API的接口自定义构造	100	// 2. 构造输入与输出，需要根据API的接口自定义构造
97	std::vector<int64_t> xShape = {1, 512};	101	std::vector<int64_t> xShape = {1, 512};
98	- std::vector<int64_t> smoothScaleOptionalShape = {1};	102	+ std::vector<int64_t> smoothScaleOptionalShape = {512};
99	std::vector<int64_t> yOutShape = {1, 512};	103	std::vector<int64_t> yOutShape = {1, 512};
100	std::vector<int64_t> level0ScaleOutShape = {1, 1};	104	std::vector<int64_t> level0ScaleOutShape = {1, 1};
101	std::vector<int64_t> level1ScaleOutShape = {1, 8, 2};	105	std::vector<int64_t> level1ScaleOutShape = {1, 8, 2};
	@@ -112,12 +116,12 @@ int aclnnDynamicDualLevelMxQuantTest(int32_t deviceId, aclrtStream& stream)
112		116
113	// 对应 BF16 的值 (0->0, 16640->8, 17024->64, 17408->512)	117	// 对应 BF16 的值 (0->0, 16640->8, 17024->64, 17408->512)
114	std::vector<uint16_t> xHostData(512, 16640);	118	std::vector<uint16_t> xHostData(512, 16640);
115	- std::vector<uint16_t> smoothScaleOptionalHostData = {0};	119	+ std::vector<uint16_t> smoothScaleOptionalHostData(512, 0);
116	// 对应 float4_e2m1 的值 (0->0, 72->4, 96->32, 120->256)	120	// 对应 float4_e2m1 的值 (0->0, 72->4, 96->32, 120->256)
117	std::vector<uint8_t> yOutHostData(512, 0);	121	std::vector<uint8_t> yOutHostData(512, 0);
118	// 对应 float32 的值 (0->0)	122	// 对应 float32 的值 (0->0)
119	std::vector<float> level0ScaleOutHostData = {{0}};	123	std::vector<float> level0ScaleOutHostData = {{0}};
120	- //对应float8_e8m0的值(128->2)	124	+ // 对应float8_e8m0的值(128->2)
121	std::vector<std::vector<std::vector<uint8_t>>> level1ScaleOutHostData(	125	std::vector<std::vector<std::vector<uint8_t>>> level1ScaleOutHostData(
122	1, std::vector<std::vector<uint8_t>>(8, std::vector<uint8_t>(2, 0)));	126	1, std::vector<std::vector<uint8_t>>(8, std::vector<uint8_t>(2, 0)));
123	const char* roundModeOptional = "rint";	127	const char* roundModeOptional = "rint";
	@@ -179,7 +183,7 @@ int aclnnDynamicDualLevelMxQuantTest(int32_t deviceId, aclrtStream& stream)
179	ret = aclnnDynamicDualLevelMxQuant(workspaceAddr, workspaceSize, executor, stream);	183	ret = aclnnDynamicDualLevelMxQuant(workspaceAddr, workspaceSize, executor, stream);
180	CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDynamicDualLevelMxQuant failed. ERROR: %d\n", ret); return ret);	184	CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDynamicDualLevelMxQuant failed. ERROR: %d\n", ret); return ret);
181		185
182	- //（固定写法）同步等待任务执行结束	186	+ // （固定写法）同步等待任务执行结束
183	ret = aclrtSynchronizeStream(stream);	187	ret = aclrtSynchronizeStream(stream);
184	CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);	188	CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
185		189
	@@ -226,4 +230,4 @@ int main()
226		230
227	Finalize(deviceId, stream);	231	Finalize(deviceId, stream);
228	return 0;	232	return 0;
229	-}	233	+}