Searcharxiv⌕ Search

arXiv subjects

Shing Wai Pun

Publications and source records attributed to Shing Wai Pun.

2 recordsLinked to original sources

Optimize the Cob, Then the Kernel: Optimizing Beam Search for Generative Recommendation Inference

Generative recommendation (GR) inference differs substantially from conventional LLM serving, featuring long input sequences, wide beam search, and only a few fixed output tokens. Existing LLM serving engines such as vLLM process beams as batched independent sequences, causing redundant KV cache accesses over the shared prompt prefix. Furthermore, vLLM executes each beam search generation step as a separate engine invocation, incurring host-side beam processing overhead between decoding steps and substantial runtime KV block management in PagedAttention. Using Tree Attention, we concatenate beams along the sequence dimension to eliminate redundant KV accesses, and further optimize execution through pre-generated partial-attention masks, static KV cache allocation, loop-unrolled generation dispatch, TorchAir graph compilation, automatic operator fusion, and vocabulary-constrained decoding. Without the use of custom kernels, our approach reduces end-to-end latency by up to 25x over vLLM on Ascend NPUs, and achieves generation latency comparable to a SOTA solution using custom kernels, showing that the generation pipeline should be optimized holistically before resorting to custom kernel development.

cs.LG↗

Taming Wild Branches: Overcoming Hard-to-Predict Branches using the Bullseye Predictor

Branch prediction is key to the performance of out-of-order processors. While the CBP-2016 winner TAGE-SC-L combines geometric-history tables, a statistical corrector, and a loop predictor, over half of its remaining mispredictions stem from a small set of hard-to-predict (H2P) branches. These branches occur under diverse global histories, causing repeated thrashing in TAGE and eviction before usefulness counters can mature. Prior work shows that simply enlarging the tables offers only marginal improvement. We augment a 159 KB TAGE-SC-L predictor with a 28 KB H2P-targeted subsystem called the Bullseye predictor. It identifies problematic PCs using a set-associative H2P Identification Table (HIT) and steers them to one of two branch-specific perceptrons, one indexed by hashed local history and the other by folded global history. A short trial phase tracks head-to-head accuracy in an H2P cache. A branch becomes perceptron-resident only if the perceptron's sustained accuracy and output magnitude exceed dynamic thresholds, after which TAGE updates for that PC are suppressed to reduce pollution. The HIT, cache, and perceptron operate fully in parallel with TAGE-SC-L, providing higher fidelity on the H2P tail. This achieves an average MPKI of 3.4045 and CycWpPKI of 145.09.

cs.AR↗