v0.2.6

What's Changed

• [ascend] fix paged_prefill and use max_batchsize as max capture size by @tangzhiyi11 in #294
• [ascend] update multinode doc by @tangzhiyi11 in #296
• Fix refact by @jinminxi104 in #297
• Patch rotary_embedding builder and DlinferSoftmaxTopKImpl by @tangzhiyi11 in #301
• patch profile npu by @jinminxi104 in #300
• [fix]fix camb run qwen vl by @huaxiaofen in #298
• fix neg seqlen by @jinminxi104 in #302
• bump to 0.2.6 by @jinminxi104 in #303

New Contributors

• @huaxiaofen made their first contribution in #298

Full Changelog: v0.2.5...v0.2.6

bump version to 0.2.5

What's Changed

• [Maca] Fix qwenvl and internvl using mcoplib by @wanfengcxz in #279
• refactor code by @yao-fengchen in #282
• [ascend] patch dptp moe by @tangzhiyi11 in #285
• fix qwen vl mask temporarily by @yao-fengchen in #286
• [ascend] support w8a8 using torch_npu ops by @wanfengcxz in #284
• [ascend] disable compile dicp and use torch_npu replay by @tangzhiyi11 in #287
• fix qwen vl vision part flash-attention by @yao-fengchen in #288
• [ascend] add multinodes doc by @tangzhiyi11 in #290
• Copilot/add patch for ascend in lmdeploy by @jinminxi104 in #291
• Copilot/update capture batch size logic by @jinminxi104 in #292
• Copilot/bump version to 025 by @jinminxi104 in #289

New Contributors

• @wanfengcxz made their first contribution in #279

Full Changelog: v0.2.4...v0.2.5

bump version to 0.2.4

v0.2.4

bump version to 0.2.4 (#283)

dlinfer release v0.2.3.post2

• 支持华为A3
• 支持华为aclgraph

dlinfer release v0.2.2

• 华为
  • 支持A3上的Qwen系列
  • 修复图模式下的内存泄漏问题
• 沐曦
  • Refine代码。适配最新软件栈。

dlinfer release v0.2.1.post2

华为

• 多卡下默认使用ray，修复稳定性问题
• 其他bug fix

v0.1.8

华为

• MoE优化

What's Changed

• revert ci by @jinminxi104 in #198
• [ascend]remove attention patch in lmdeploy by @yao-fengchen in #194
• [feat]support ascend w8a8 graph_mode by @yao-fengchen in #191
• Refactor rope by @yao-fengchen in #199
• change ci setting by @jinminxi104 in #202
• fix ascend awq by @yao-fengchen in #204
• [ascend]Optimize moe by @yao-fengchen in #203
• Bump018 by @jinminxi104 in #209

Full Changelog: v0.1.7...v0.1.8

dlinfer release v0.1.7

华为&沐曦

• 支持多节点
• 支持mla，支持dsv2

What's Changed

• [maca] support deepseekv2 for maca backend. by @Reinerzhou in #133
• [camb]fix fused_moe param order by @JackWeiw in #173
• [camb]add w8a8 support by @JackWeiw in #176
• [ascend] add new_empty op and opt reshape ops by @tangzhiyi11 in #178
• fix glm-4v by @jinminxi104 in #184
• revert ci by @jinminxi104 in #187
• default enable npu launch by @jinminxi104 in #188
• [ascend] support multi ndoes by @tangzhiyi11 in #190
• [maca] support multinode for maca. by @Reinerzhou in #192
• [ascend] remove allreduce group in eager mode by @tangzhiyi11 in #193
• [ascend]support mla by @yao-fengchen in #182
• bump version to 0.1.7 by @jinminxi104 in #195

Full Changelog: v0.1.6...v0.1.7

dlinfer release v0.1.6

• performance up on Ascend platform!
  • reaches almost the same performance as MindIE on most dense models with both single and multiple cards.

dlinfer release v0.1.5

• 华为，沐曦，寒武纪上均支持InternLM3-8B
• 沐曦
  • 支持cudagraph
• 寒武纪
  • 完成代码合并，支持InternLM/InternVL/LLaMA/Qwen(VL)系列/Mixtral8x7B