Starstruk Logo MarkStarstruk

v0.5.0

dancasas/awesome-virtual-try-onv0.5.0Mar 24, 2025by EricLBuehler

AI Summary

v0.5.0 adds support for several new models like Gemma 3 and Qwen 2.5 VL, while introducing native tool calling and Tensor Parallelism capabilities. The release focuses heavily on performance, featuring a 30x speedup in ISQ times on Metal and integrating FlashAttention V3 into the PagedAttention system.

Key Highlights

  • Support for Gemma 3, Qwen 2.5 VL, and Phi 4 Multimodal models.
  • Native tool calling support for major models like Llama 3.1 and Mistral Nemo.
  • Introduction of Tensor Parallelism (NCCL) for distributed inference.
  • Integration of FlashAttention V3 into PagedAttention for improved performance.
  • Significant performance boost with a 30x reduction in ISQ times on Metal.

New Features

  • Improved streaming UX
  • Faster fp8 blockwise dequantization
  • PagedAttention prompt chunking support
  • Speculative decoding revamp
  • UQFF text/vision model API

Full Release Notes

## Highlights

Blog post: https://huggingface.co/blog/EricB/mistralrs-v0-5-0

Thank you to all contributors for this release! This release includes the following highlights but also countless improvements, fixes, and optimizations.

- Support for many **more models**:
  - Gemma 3
  - Qwen 2.5 VL
  - Mistral Small 3.1
  - Phi 4 Multimodal (image only)
- **Native tool calling support** for:
  - Llama 3.1/3.2/3.3
  - Mistral Small 3
  - Mistral Nemo
  - Hermes 2 Pro
  - Hermes 3
- **Tensor Parallelism** support (NCCL)!
- **FlashAttention V3** support and integration in PagedAttention
- **30x** reduction in ISQ times on Metal!
- Revamped **prefix cacher system**

## What's Changed
* Allow using library in CurrentThread runtime by @sgrebnov in https://github.com/EricLBuehler/mistral.rs/pull/1082
* Improve accuracy of uqff auto device map by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1084
* DeepSeekV3 sigmoid support by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1092
* GPU-accelerated sampling (+5% decode perf) by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1094
* Fix missing perceiver_config in qwen2vl by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1096
* More topk methods for deepseek 2/3 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1097
* More accurate layer size computation for deepseek 2/3 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1098
* Improve streaming UX by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1102
* Faster fp8 blockwise dequant by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1100
* DS2/3 paged attn by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1103
* Faster bincount by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1104
* PagedAttention prompt chunking support by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1105
* Refactor server SSE by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1107
* PagedAttention + FlashAttention (and FlashAttention V3) by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1109
* Take KEEP_ALIVE_INTERVAL into account by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1111
* Refactor enable of flash attn by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1110
* Fix imatrix isq quantize_onto by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1112
* Tensor parallelism and pipeline parallelism by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1113
* Bump openssl from 0.10.69 to 0.10.70 by @dependabot in https://github.com/EricLBuehler/mistral.rs/pull/1121
* Allow chat streaming to use tools by @Jeadie in https://github.com/EricLBuehler/mistral.rs/pull/1088
* New file format for imatrix: `.cimatrix` by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1004
* Fix isq with bias for column parallel by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1128
* Multi-node support for tensor parallelism by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1125
* Add an NCCL feature flag by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1129
* Fix mistral 2501 gguf by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1131
* Add jinja strftime_now function by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1132
* Multiple models multi node by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1136
* Remove unexpected cp behavior by @jncraton in https://github.com/EricLBuehler/mistral.rs/pull/1141
* Revamp speculative decoding! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1027
* Fuse MLP mul-and-act  by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1142
* Short-circuit dry sampling: +6% T/s by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1143
* Integrate fused MLP mul-act for more models! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1144
* Use cudarc 0.13.5 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1145
* Handle HF_HUB_CACHE env var by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1146
* FlashAttention V2/V3 metadata with support for device location by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1148
* FP8 blockwise dequant cuda kernel by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1149
* Blockwise FP8 CUDA for cc < 800 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1150
* Fix chat sampling response by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1154
* Multiple processes for TP by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1152
* Ensure we do not bind the port for daemon processes by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1158
* Handle CUDA_NVCC_FLAGS in flash attn v3 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1160
* build fix for arm. by @jamesvren in https://github.com/EricLBuehler/mistral.rs/pull/1164
* Working PrefixCacherV2! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1168
* Implement Phi-4 Multimodal! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1163
* No extra split/cat pair in rope by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1169
* Remove gpu<>cpu sync for faster long-context by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1170
* Refactor NCCL device mappers by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1172
* Bump ring from 0.17.11 to 0.17.13 by @dependabot in https://github.com/EricLBuehler/mistral.rs/pull/1179
* DSV3/R1 fixes by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1173
* Fix diffusion device mapping by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1187
* Internal abstraction for distributed op by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1188
* Make Sequence::set_toks more safe by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1190
* Fix CI tests out of storage? by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1191
* Internal abstraction for distributed op by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1189
* Fix build_cuda_all.yaml CI by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1193
* Support tensor parallelism for vision models! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1194
* Always pass _USE_MATH_DEFINES for CUDA by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1195
* Remove matmul via f16 framework by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1196
* Remove API for matmul_via_f16 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1197
* Add UQFF text/vision model API by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1198
* Complete qwen2_5_vl, and some fixes by @brrr in https://github.com/EricLBuehler/mistral.rs/pull/1184
* Implement Gemma 3 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1201
* Add Gemma 3 vision support! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1202
* Manually fixup sentencepiece detok by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1204
* More vision models with TP by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1200
* Fix topology link in the docs by @etiennebalit in https://github.com/EricLBuehler/mistral.rs/pull/1205
* Gemma3 1b support and optimized rotating cache by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1206
* Improve rotating kv cache, prefix cacher system by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1207
* Better handling for kvcache set_len by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1208
* Update deps and use rand 0.9 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1210
* Update hf hub dep, add initial blockwise fp8 GEMM tests by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1212
* Growable RotatingKvCache and fixes for Phi-4 mini by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1215
* Gemma 3 cuda fixes by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1217
* Add pydantic schema examples! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1219
* Sliding window attention fixes by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1220
* adapt to rig crate as client by @benliao in https://github.com/EricLBuehler/mistral.rs/pull/1214
* Implement Mistral 3! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1221
* Metal SDPA with masking by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1225
* Send [DONE] SSE chunk per openai spec by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1226
* Fix handling of device when compiled for but disabled nccl by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1227
* Fix nccl blocking case by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1228
* Native Llama, Mistral Small 3.1, Mistral Nemo, Hermes 2 Pro, Hermes 3 tool calling! by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1229
* OpenAI API compatability fixes by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1230
* [Breaking] Automatic server logging by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1231
* Use default stream for flash attn by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1232
* Bump version to 0.5.0 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/1233

## New Contributors
* @sgrebnov made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1082
* @jncraton made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1141
* @jamesvren made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1164
* @brrr made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1184
* @etiennebalit made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1205
* @benliao made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/1214

**Full Changelog**: https://github.com/EricLBuehler/mistral.rs/compare/v0.4.0...v0.5.0