media Hugging Face Forums · 3d ago · open_models

TraceML reproduces LeRobot regression showing DataLoader fetch time drop from 1375ms to 1.5ms

from English

TraceML, an open-source PyTorch training diagnosis tool, was used to reproduce a known LeRobot training regression and measure the impact of upstream fix #2408. The analysis revealed that while most DataLoader fetches were quick, recurring fetches previously took over five seconds.

Step time decreased from approximately 1500ms to 117ms after the fix.
CPU DataLoader fetch time dropped from ~1375ms to 1.5ms in the first pair and from ~1427ms to 1.3ms in the second pair.
Compute time remained stable around 110–116ms, confirming the workload shifted from INPUT-BOUND to COMPUTE-BOUND.

The tool helps identify where extra time goes when training slows down after a change by instrumenting the training loop and comparing timings between two runs.

Importance 1/3 Trust 1/3 Hugging Face Forums Inference efficiency

Read original

media MarkTechPost · just now Live

DeepSeek releases DeepSeek-V4.1-Flash with 1M context and FP4 KV cache

DeepSeek AI has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model featuring a 1-million-token context window and an FP4 KV cache that reduces global cache footprint to 890 bytes per token. The model utilizes a causal encoder-decoder architecture and Compressed Sparse Attention 2 (CSA2) to significantly lower memory requirements while maintaining performance.

github llama.cpp · 1h ago · 1 view Live

DeepSeek releases DeepSeek-V4.1-Flash with 1M context and FP4 KV cache

llama.cpp b10889 release avoids V cache allocation for indexer

llama.cpp b10888 adds Vulkan command-buffer debug labels for GPU profilers

llama.cpp b10886 adds Q1_0 vector intrinsic support for s390x

llama.cpp b10883 fixes Vulkan matrix multiplication and restores Ampere optimizations