From c7ab779ac102063b32f9f9cf7a19e252d2a80a18 Mon Sep 17 00:00:00 2001 From: Peter Hicks Date: Fri, 23 Oct 2026 18:41:22 +0000 Subject: [PATCH] Reduce the learning-rate note to the recommendation The note cited 1.0e-5 (4B) and 2.0e-6 (0.8B), the only mention of either draft in the doc, so a reader met two rates for two models the page never introduces. The Configuration table above already states the rate this section used. All the note needs to add is the recommendation to sweep. Co-Authored-By: Claude Opus 5 Signed-off-by: Peter Hicks --- examples/speculative_decoding/doc/external_draft.md | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/examples/speculative_decoding/doc/external_draft.md b/examples/speculative_decoding/doc/external_draft.md index 477f706bd07..27752acb9d4 100644 --- a/examples/speculative_decoding/doc/external_draft.md +++ b/examples/speculative_decoding/doc/external_draft.md @@ -240,10 +240,8 @@ torchrun --nproc_per_node 8 main.py \ | Epochs | 1 | | Serving policy | top-k 20, top-p 0.95, temperature 1.0 | -> **Note on the learning rate:** the vocabulary-matched drafts trained against the same dump -> used 1.0e-5 (4B) and 2.0e-6 (0.8B). Each rate was selected by sweeping against held-out -> acceptance; the ordering does not transfer to another target or dump, so re-sweep for your -> own pair. +> **Note on the learning rate:** we recommend sweeping against held-out acceptance to find +> the best rate for your own draft and target. ## Open Items