pullrequests Search Results · bucket language:JavaScript language:JavaScript language:Python language:JavaScript
Filter by
897k results
Summary
- Defensive liveness fix for the bucketed ZMQ weight-transfer protocol: bound each ACK wait and propagate receiver
exceptions to the sender through explicit error ACKs.
- Expose ...
... local histograms, exact finite
boundaries, and explicit rejection of incompatible grids.
The initial infinity-bucket cases failed on the original implementation; later mixed-client cases reproduced ...
... relative path and artifact
fingerprint stay the same.
marin-us-east-02a is rejecting writes while its account is capacity suspended. SkyRL previously placed temporary data in
the shared 08a bucket ...
agent-generated
agentic-lint
weaver
... ). LeakyLenses decays exponentially with e-folding thickness LensDecayLength
(blue). B: Time-mean density (last 50 of a 100 year run) from CFM (Crocus, bucket) with a synthetic forcing at a site
with ...
Summary
Add an opt-in bucket size that bounds the non-tensor-parallel gradient all-reduce s buffer.
Symptom Reproduction
- Symptom: On a 64x H200 MoE RL run with TP8 and sequence parallelism ...
...
default is zero.
- draft_prefill_use_decode_buckets reuses decode batch buckets for draft-prefill CUDA graphs to reduce retained
attention workspaces. Its default is false.
The changes stay in ...
What does this PR do?
This PR adds support for HF Buckets in Trainer. Instead of storing checkpoints locally or on the hub repository, we can
leverage HF Buckets for that (S3 like storage + XET ...