Skip to content

qwen4exp: Qwen3.8-Flash-Next long-context decode, +66% at 131k #9

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
firelzrd wants to merge 1 commit into
base: release/v0.7.5-staging
Choose a base branch
from
Open

qwen4exp: Qwen3.8-Flash-Next long-context decode, +66% at 131k #9

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions src/models/qwen4exp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -713,8 +713,12 @@ ggml_tensor * llama_model_qwen4exp::graph::build_qsa_top_k(
members->nb[2], members->nb[3], i*members->nb[1]);
fresh = fresh ? ggml_add(ctx0, fresh, slice) : slice;
}
fresh = ggml_scale(ctx0, fresh, 1.0f/(float) r);
cb(fresh, "indexer_k_pooled", il);
// no ggml_scale by 1/r here: the only consumer is the RMS norm below, and RMS norm is
// scale invariant. rms(x*s) = x*s / sqrt(mean(x^2)*s^2 + eps) = x / sqrt(mean(x^2) + eps/s^2),
// so dropping the divide only moves the epsilon from eps to r^2*eps -- 1e-6 to 1.6e-5 against
// a mean square of order 1. the scale was a full read and write of [idx_dim, n_blocks] f32 per
// layer per ubatch: 2.25 ms of a 97 ms decode step at 131k context.
cb(fresh, "indexer_k_sum", il);

// count blocks along ne1: rms_norm launches gridDim.y = ne2, capped at 65535, and 262144/4 = 65536
fresh = ggml_reshape_3d(ctx0, fresh, idx_dim, n_recomp*n_stream, 1);
Expand Down