$ the-wire · showcase
python-genai 3.5.0 ships Live Interactions, Mosaic GPU hits K=96 FP4
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
python-genai cut v3.5.0 with self-hosted BYOS environments and experimental WebSocket/WebRTC Live Interactions, while JAX reworked scatter codegen and Pallas semaphore aliasing on the GPU side.
[Mosaic GPU] Support K=96 block-scaled FP4 tcgen05.mma on SM103. google/jax
On SM103 (sm_103a, GB300) dense 4-bit block-scaled tcgen05.mma ops now take K=96 steps in addition to K=64, cutting MMA instructions per K tile by up to 1.5x and lifting 2-CTA throughput from about 9.73 to up to 14.62 PFLOP/s on 256x256x768 tiles. That figure holds for mxf4 and mxf4nvf4 with SMEM operands on this one part; K=64 remains the path everywhere else.
v3.5.0 googleapis/python-genai
by release-please[bot]
The 3.5.0 release introduces self-hosted BYOS mode with system_instruction, custom_config, and environment_key across the SDKs and CLI, and generates Live Interactions types plus an auth-token client. Both land as public API surface, so pinning to v3.4.0 is the safe move until you have exercised the new fields.
feat(interactions): add experimental Live Interactions (WebSocket and WebRTC) SDK support googleapis/python-genai
Live Interactions support arrives over WebSocket and WebRTC, flagged experimental in the PR title. Experimental means the transport surface can move under you; treat it as a preview integration rather than a stable dependency.
[Pallas/Mosaic GPU] Add `Read` and `WriteEffect`s to necessary semaphore primitives. google/jax
Pallas/Mosaic GPU semaphore primitives now carry Read and WriteEffects. Without the write effect, a semaphore threaded into a kernel was not correctly aliased, so kernels that pass semaphores in should re-check after this lands.
Add mesh.jax_partitioning(True) as the API to enter into partitioning locally. Users can still do `jax.config.update('jax_partitioning', True)` but maybe we can lift that into `jax.set_mesh` later. google/jax
mesh.jax_partitioning(True) is now the local entry point into partitioning mode; jax.config.update('jax_partitioning', True) still works, with the PR noting it may later be lifted into jax.set_mesh. The long tail was routine: JAX gained a 0-collective scatter path along sharded inserted_window_dims where each shard shifts scatter_indices by -axis_index * shard_size, python-genai now retries goo...