Repository navigation
Highlights
- 🤖 5 new VLA models: PI0-FAST, GR00T-N1.7, Qwen-Drive, SmolVLA, WALL-OSS-0.5
- 🧠 New VLM & decision models: Qwen3.8-27B (NVFP4, with image inference) and Decision 3.0 mini
- 🍎 Apple Metal backend: native MLX runtime with Qwen3-0.6B and MiniCPM5-2B ports
What's Changed
- feat(skills): add model implementation analysis workflow by @zealoct in #55
- feat(cuda): 识别 GB10 (sm_121) 为 Blackwell sm100 家族 by @li-xiu-qi in #60
- fix(core, qwen3vl): 修复因果 Softmax 网格溢出并支持 Qwen3-VL 4B/8B 多尺寸适配 by @li-xiu-qi in #57
- Refactor Apx-cuda by @ChenxiHe231 in #53
- feat: add GR00T N1.7 VLA runtime by @Casten-Wang in #42
- docs: update result of LIBERO-10 Thor by @ChenxiHe231 in #71
- doc: add credits and acknowledgements by @zealoct in #75
- feat(pi0-fast): BF16 inference in a persistent arena and doc the module (70% SR on LIBERO-10, 32.2 ms per token) by @Kimho666 in #62
- perf(attention): FlashAttention-2 prefill for vision and GQA on sm120/121 by @li-xiu-qi in #64
- docs(cuda): add L3 operator catalog by @ChenxiHe231 in #69
- feat: adjust cutlass kernel hierarchy by @ChenxiHe231 in #76
- perf(vision): block-parallel reduction for layer_norm_bf16 by @li-xiu-qi in #66
- Add attention kernel to cuda-new by @ChenxiHe231 in #77
- refactor: separate PI0.5 model, runner and weights by @wong-1994 in #50
- fix(cuda): make reduction scratch reuse explicit and fix attention races by @zealoct in #80
- docs: update pi05 int8 inference latency on RTX 4090 platform by @Kimho666 in #84
- docs(cuda): document native framework contracts by @ChenxiHe231 in #85
- feat(qwen-drive): Qwen-Drive on RTX 4090, Jetson AGX Orin and Jetson AGX Thor by @qhy991 in #78
- fix(cuda): preserve captured KV cache addresses on reset by @zealoct in #89
- refactor(qwen-drive): migrate to the shared model architecture and VLA lifecycle by @wong-1994 in #87
- feat(pi0-fast): add FP8 E4M3 decode with measured activation calibration and a pruned LM head by @Kimho666 in #86
- fix(cuda-new): replay dynamic KV decode graphs by @ChenxiHe231 in #93
- build: promote apxinf-cuda-new to a workspace crate beside apxinf-cuda by @ChenxiHe231 in #98
- Qwen3.8-27B-NVFP4: kernels + operator precision gates (PR-A of 2) by @ChenxiHe231 in #92
- doc(pi0-fast): update pi0-fast BF16 and FP8 latency results for Jetson Thor/Orin and RTX 4090 platfoms by @Kimho666 in #97
- perf(gr00t): optimize Thor BF16/FP8 and Orin BF16/W8A8 by @Casten-Wang in #101
- Qwen3.8-27B-NVFP4: apxinf-model integration behind LlmTrait (PR-B of 2) by @ChenxiHe231 in #99
- fix(gr00t): bind FP8 calibration to processor metadata by @Casten-Wang in #106
- build: drop the quantized-dtypes feature and fix the default workspace build by @zealoct in #104
- perf(qwen-drive): integrate native AOT kernels and padded BF16 direct planning by @wong-1994 in #94
- feat(SmolVLA): add support for SmolVLA BF16/FP16 inference on Jetson Thor/Orin, RTX 4090 and Jetson Xavier NX platforms by @Kimho666 in #110
- build(cuda): parallelize nvcc translation units and cache objects by @ChenxiHe231 in #113
- Pi05 with cuda new by @ChenxiHe231 in #88
- fix(build): gate the CUDA coexistence test on both CUDA crates by @zealoct in #115
- feat(hip): add Device::Hip and an apxinf-hip backend skeleton (re-land of #105) by @zealoct in #118
- docs: define native MLX backend and Mac model migration contracts by @qhy991 in #111
- perf(cuda-new): vectorize NVFP4 quantization and phase graph capture by @ChenxiHe231 in #107
- feat(mlx): add native Metal backend and Qwen3/MiniCPM5 ports by @qhy991 in #112
- feat(cuda-new): add Decision 3 mini inference by @ChenxiHe231 in #131
- doc: Update README with new features and hardware support by @Kimho666 in #132
- feat(cuda-new): stage seven native translations and enable Qwen3.8 by @ChenxiHe231 in #130
- feat(hip): HIP runtime, memory and the operators Backend requires by @zealoct in #119
- feat(walloss): model-runner layout, multi-resolution input, faster BF16 vision by @wong-1994 in #138
- feat(qwen38): add multimodal image inference by @ChenxiHe231 in #129
- feat(hip_pi05): run PI0.5 on HIP (gfx1151, BF16, eager) by @zealoct in #164
- docs: rule for where backend-native model ports live by @zealoct in #165
- build(hip): vendor Composable Kernel's FMHA forward subset by @zealoct in #177
- refactor(mlx): let Metal families own model_variant admission by @qhy991 in #133
New Contributors
- @li-xiu-qi made their first contribution in #60
- @Casten-Wang made their first contribution in #42
- @qhy991 made their first contribution in #78
Full Changelog: v0.1...v0.2