TensorSharp 3.2.1.0

What's Changed

• support MoE offload to cpu by @zhongkaifu in #128
• add grammer check in json mode by @zhongkaifu in #129
• merge two tables into a single table by @zhongkaifu in #131

Full Changelog: v3.2.0.0...v3.2.1.0

What's Changed

• support MoE offload to cpu by @zhongkaifu in #128
• add grammer check in json mode by @zhongkaifu in #129
• merge two tables into a single table by @zhongkaifu in #131

Full Changelog: v3.2.0.0...v3.2.1.0

TensorSharp 3.2.0.0

What's Changed

• optimize cuda backend performance by @zhongkaifu in #94
• update documents to include dotnet installation by @zhongkaifu in #95
• Added tensor parallelism (both local and basic peer to peer clustering) by @simonsoanes in #93
• docs: document multi-GPU / multi-node tensor parallelism across all markdown docs by @zhongkaifu in #96
• Fixed MTP under tensor parallelism on Qwen 3.6 by @simonsoanes in #97
• Feature/fix vision encoder in qwen cuda backend by @zhongkaifu in #98
• fix qwen moe stuck problem in tp by @zhongkaifu in #99
• support tensor parallelism in ggml backend by @zhongkaifu in #100
• fix qwen image encode error in tensor parallelism in ggml cuda backend by @zhongkaifu in #101
• support gpt-oss tensor parallelism in ggml backend by @zhongkaifu in #102
• Feature/qwen35 tp fused prefill ggml backend by @zhongkaifu in #103
• Fix ggml pool freeing recycled blocks with the caller's size, not the… by @zhongkaifu in #104
• docs: align README, docs, and website with GGML-backend tensor parallelism by @zhongkaifu in #105
• Add DeepSeek V4 Flash (deepseek4) support via native whole-model exec… by @zhongkaifu in #106
• support deepseek v4 cpu backend by @zhongkaifu in #107
• DeepSeek V4 Flash: fused custom-op backend + pipelined prefill (faster than llama.cpp) by @zhongkaifu in #109
• Fix Qwen3.6 MTP speculative decoding crashes with quantized KV caches… by @zhongkaifu in #110
• fix gemma4 and qwen template and loading issue by @zhongkaifu in #111
• Fix multi-GPU model load hang: probe NCCL before trusting it, prefaul… by @zhongkaifu in #114
• DSV4: decode index-gather sparse attention + larger prefill chunks by @zhongkaifu in #115
• DSV4: serve through TensorSharp.Server with continuous batching by @zhongkaifu in #116
• update direct cuda backend to support deepseek v4 by @zhongkaifu in #117
• fix deepseek v4 loading issue in direct cuda backend by @zhongkaifu in #118
• update ptx by @zhongkaifu in #119
• refactoring deepseek v4 model by @zhongkaifu in #120
• support dspark by @zhongkaifu in #121
• bug fix 113 by @zhongkaifu in #122
• bug fix by @zhongkaifu in #124
• update documents by @zhongkaifu in #125
• bug fix by @zhongkaifu in #126

Full Changelog: v3.1.2.0...v3.2.0.0

What's Changed

• optimize cuda backend performance by @zhongkaifu in #94
• update documents to include dotnet installation by @zhongkaifu in #95
• Added tensor parallelism (both local and basic peer to peer clustering) by @simonsoanes in #93
• docs: document multi-GPU / multi-node tensor parallelism across all markdown docs by @zhongkaifu in #96
• Fixed MTP under tensor parallelism on Qwen 3.6 by @simonsoanes in #97
• Feature/fix vision encoder in qwen cuda backend by @zhongkaifu in #98
• fix qwen moe stuck problem in tp by @zhongkaifu in #99
• support tensor parallelism in ggml backend by @zhongkaifu in #100
• fix qwen image encode error in tensor parallelism in ggml cuda backend by @zhongkaifu in #101
• support gpt-oss tensor parallelism in ggml backend by @zhongkaifu in #102
• Feature/qwen35 tp fused prefill ggml backend by @zhongkaifu in #103
• Fix ggml pool freeing recycled blocks with the caller's size, not the… by @zhongkaifu in #104
• docs: align README, docs, and website with GGML-backend tensor parallelism by @zhongkaifu in #105
• Add DeepSeek V4 Flash (deepseek4) support via native whole-model exec… by @zhongkaifu in #106
• support deepseek v4 cpu backend by @zhongkaifu in #107
• DeepSeek V4 Flash: fused custom-op backend + pipelined prefill (faster than llama.cpp) by @zhongkaifu in #109
• Fix Qwen3.6 MTP speculative decoding crashes with quantized KV caches… by @zhongkaifu in #110
• fix gemma4 and qwen template and loading issue by @zhongkaifu in #111
• Fix multi-GPU model load hang: probe NCCL before trusting it, prefaul… by @zhongkaifu in #114
• DSV4: decode index-gather sparse attention + larger prefill chunks by @zhongkaifu in #115
• DSV4: serve through TensorSharp.Server with continuous batching by @zhongkaifu in #116
• update direct cuda backend to support deepseek v4 by @zhongkaifu in #117
• fix deepseek v4 loading issue in direct cuda backend by @zhongkaifu in #118
• update ptx by @zhongkaifu in #119
• refactoring deepseek v4 model by @zhongkaifu in #120
• support dspark by @zhongkaifu in #121
• bug fix 113 by @zhongkaifu in #122
• bug fix by @zhongkaifu in #124
• update documents by @zhongkaifu in #125
• bug fix by @zhongkaifu in #126

Full Changelog: v3.1.2.0...v3.2.0.0