Block a user
perf(neuron): speculative decoding with a small same-family drafter
Design (2026-06-13) — picking this up after #23
Feasibility gate ✅ and a drafter correction
Speculative decoding requires the drafter and target to share a tokenizer/vocabulary…
grenade
deleted branch feat/18-single-gpu-vision-chunked from helexa/helexa
2026-06-13 05:23:10 +00:00
Chunk the single-GPU vision prefill (parity with TP)
feat(neuron): chunk the single-GPU vision prefill (parity with TP) (#18)
feat(neuron): chunk the single-GPU vision prefill (parity with TP) (#18)
grenade
created branch feat/18-single-gpu-vision-chunked in helexa/helexa
2026-06-13 05:17:14 +00:00
docs(learnings): source-control P1 + P2 sprint learnings
docs(learnings): source-control P1 + P2 sprint learnings
Reduce TP=2 Q6K cold-load time for Qwen3.6-27B (~5 min today)
Reduce TP=2 Q6K cold-load time for Qwen3.6-27B (~5 min today)
Closing numbers. Landed via #40 (parallel in-situ quantization + the per-phase timing this issue asked for first).
Before/after — Qwen3.6-27B Q6K TP=2 cold load on beast (service start → `TP…
ci: Build neuron-ampere intermittently fails — ptxas segfault in candle-flash-attn sm_86 kernels
538cc87