14 Pull requests merged by 9 people

• Flash Attention now explicitly checks whether it is in decoding mode.

  #1026 merged Feb 28, 2025

• External KV input for _update_layer_kwargs

  #1025 merged Feb 27, 2025

• Minor changes to Checkpointer

  #1024 merged Feb 27, 2025

• Add an option to break ties for top_k_logits when k = 1

  #1022 merged Feb 27, 2025

• Add priority_class as a launch flag

  #1020 merged Feb 26, 2025

• Fix TypeError in calcualte_goodput.py

  #1023 merged Feb 26, 2025

• Emulate flash attentnion unittests on CPU.

  #1021 merged Feb 26, 2025

• Add additional_node_networks config to support multi-nic for v6e

  #1019 merged Feb 25, 2025

• Fix crash in log-mel frontend when waveform samples are integers.

  #1017 merged Feb 25, 2025

• Add LoadBalancer to GKE replicatedJob

  #1015 merged Feb 25, 2025

• Expose jax.lax.scan's unroll option to Repeat layer

  #1016 merged Feb 25, 2025

• Handle None bias in BiasAndResidual

  #1018 merged Feb 25, 2025

• Allows a required value in a config_for_{function,class} to be specified via **kwargs in instantiate().

  #1013 merged Feb 24, 2025

• fix dtype in frontend pre emphasis

  #1014 merged Feb 23, 2025

2 Pull requests opened by 2 people

• [CI] Enable 4- and 8-device Flash Attention layers tests.

  #1027 opened Feb 28, 2025

• Refactor Mask in Attention

  #1028 opened Feb 28, 2025

3 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Basic FP8 Training Support

  #992 commented on Feb 27, 2025 • 14 new comments

• Improvements and fixes to gradient accumulation

  #993 commented on Feb 28, 2025 • 6 new comments

• Add GKE A3 Ultra support

  #940 commented on Feb 28, 2025 • 1 new comment