Skip to content

pullrequests Search Results · bucket language:JavaScript language:Java language:Python language:JavaScript language:JavaScript

Filter by

1M results  (504 ms)

1M results

TPU decode performance probe (R047): per-batch-bucket torch.compile of DeepInfer s decode forward on torch_tpu, and the fixes found while measuring it on a real-weight 4-layer cut of DeepSeek-V4.1 ...

📌 Description Add MoE autotuning buckets at 320, 384 and 448 tokens. Previously, inputs just above 256 used the tactic tuned at 512. For example, 264 tokens now maps to 320. Update both capped and ...

... (removing INLINE STATS) 1.861 s (*) DSL nested Top N with other bucket 2.040 s ESQL nested Top N with other bucket (LEFT OUTER JOIN ...

Three documentation locations now state that the other bucket in the filed split also includes cards whose filer field is absent (no known filer): (1) the _RANGE_FILED_DESCRIPTION constant in ...

... every device tensor at the model input s padded length T, which the torchtitan trainer rounds to a few buckets on TPU (#4213): - log-probs and entropy are scored for all T rows in fixed-size chunks ...

dependencies
python