v0.0.9

What's Changed

• Add the official AWS launch blog post and reformat the AWS section to awesome-gpt-oss.md by @neoremind in #197
• Metal: simplify Model type check in Python wrapper by @Maratyszcza in #202
• Metal: move max_batch_tokens from Model into Context by @Maratyszcza in #201
• GPT-OSS Reinforcement Finetuning Example by @danielhanchen in #208
• add stateful jupyter notebook option and make browser more reliable by @dkundel-openai in #199
• Removing GPU-CPU sync during MoE MLP in Prefill by @ibahmed-oai in #218
• Eliding copying the KV after RoPE to the KV cache during prefill by @ibahmed-oai in #221
• fix: remove duplicated variable causing Metal build failure by @hukurou0 in #220
• Add user-agent to headers for tracking by @smboy in #235

New Contributors

• @neoremind made their first contribution in #197
• @danielhanchen made their first contribution in #208
• @hukurou0 made their first contribution in #220
• @smboy made their first contribution in #235

Full Changelog: v0.0.8...v0.0.9

v0.0.8

What's Changed

• Adding metal moe kernels for prefill by @ibahmed-oai in #191
• Speeding up prefill with new moe kernels by @ibahmed-oai in #192
• Metal: fix resource leak in prefill by @Maratyszcza in #194
• Metal: support sharded checkpoints in the converter by @Maratyszcza in #195
• Polish the python code and browser experience by @dkundel-openai in #198

Full Changelog: v0.0.7...v0.0.8

v0.0.7

What's Changed

• Evals: correctly pass temperature/max_tokens when using Responses API by @Maratyszcza in #174
• Metal: move sampling to GPU by @Maratyszcza in #175
• Metal: benchmark generation of 100 tokens instead of 1 by @Maratyszcza in #178
• Metal: support generating multiple tokens at once by @Maratyszcza in #179
• Adding prefill benchmarking for metal backend by @ibahmed-oai in #181
• Metal: tune threadgroup sizes by @Maratyszcza in #180
• Metal: Adding optimized dense matmul kernel to optimize prefill perf by @ibahmed-oai in #183
• Metal: fused QKV projection (matmul+RoPE) kernel by @Maratyszcza in #184
• [Bugfix]Capture stderr for python tool with uv as backend by @wuhang2014 in #182

New Contributors

• @ibahmed-oai made their first contribution in #181
• @wuhang2014 made their first contribution in #182

Full Changelog: v0.0.6...v0.0.7

v0.0.6

What's Changed

• Metal: add end-to-end benchmarks by @Maratyszcza in #161
• Metal: simplify and optimize Reponses API adapter by @Maratyszcza in #162
• Metal: fix KV-cache invalidation after reset+append by @Maratyszcza in #163
• Increase max output tokens in Reponses API to 131K by @Maratyszcza in #165
• Remove requirement on maximum Python version by @Maratyszcza in #167
• Move Lemonade to AMD section of awesome-gpt-oss by @danielholanda in #164
• Added VLLM Offline Serve working code. by @hrithiksagar-tih in #150
• Metal: indicate threadgroup is a multiple of simdgroup by @Maratyszcza in #168
• Metal: mlock model weights in memory by @Maratyszcza in #170
• Add You.com as tool for browser by @bojanbabic in #171

New Contributors

• @hrithiksagar-tih made their first contribution in #150
• @bojanbabic made their first contribution in #171

Full Changelog: v0.0.5...v0.0.6

v0.0.5

What's Changed

• Update awesome-gpt-oss.md with llama.cpp by @dkundel-openai in #148
• Update README.md by @dkundel-openai in #154
• Added Tensorfuse (AWS) guide by @samagra14 in #118
• Add Lemonade to awesome-gpt-oss by @danielholanda in #117
• Add uv python backend by @heheda12345 in #156

New Contributors

• @samagra14 made their first contribution in #118
• @danielholanda made their first contribution in #117

Full Changelog: v0.0.4...v0.0.5

v0.0.4

What's Changed

• fix streamlit & ollama demo. Add python tool by @dkundel-openai in #131
• Add some links to awesome-gpt-oss.md by @hiyouga in #28
• fix: fix f-string unmatched '(' bug in streamlit_chat.py by @liuzhiqi71 in #31
• Fix start_q use in upper bound calculation by @peterbell10 in #136
• Process tokens in Context lazily by @Maratyszcza in #138
• Fix for bug where / in model name causes evals to fail by @simonw in #142
• [README] Rename with_browser() to with_browser_tool() in README by @xiaohk in #140
• Update attention kernel to use TensorDescriptor by @peterbell10 in #137
• Parallelize SDPA across multiple simdgroups by @Maratyszcza in #144
• chore: release 0.0.4 by @dkundel-openai in #145

New Contributors

• @hiyouga made their first contribution in #28
• @liuzhiqi71 made their first contribution in #31
• @peterbell10 made their first contribution in #136
• @simonw made their first contribution in #142
• @xiaohk made their first contribution in #140

Full Changelog: v0.0.3...v0.0.4