Skip to content

ggml-hrx: fence global F32 output before Qwen Q8 packing #174

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

Merged
AaronStGeorge merged 4 commits into from
Nov 3, 2026
Merged

ggml-hrx: fence global F32 output before Qwen Q8 packing #174

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0f796f7
ggml-hrx: preserve Qwen MoE fused Q8 row in workgroup storage
AaronStGeorge Nov 3, 2026
2902974
tests: check fused Qwen3-Coder Q8 publication in model trace
AaronStGeorge Nov 3, 2026
c5b0153
tests: validate Q8 trace discovery and fixture size
AaronStGeorge Nov 3, 2026
fe12a25
ggml-hrx: use global fence for Qwen Q8 publication
AaronStGeorge Nov 3, 2026
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
3 changes: 2 additions & 1 deletion ...-hrx/kernel-corpus/kernels/loom-libs/motifs/routing/weighted_reduce_next_rmsnorm_f32.loom
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,7 +111,8 @@ template.def<@ggml.routing.weighted_reduce_next_rmsnorm_f32> device requires [#t
// Q8 consumers. The reload keeps the legacy reduction and RMSNorm motif exact.
template.def<@ggml.routing.weighted_reduce_next_rmsnorm_f32_publish_q8> device requires [#target.subgroup.size<32>] @ggml_routing_weighted_reduce_next_rmsnorm_f32_publish_q8(%token_count: index, %route_weights: buffer, %routed_output: buffer, %hidden_state: buffer, %next_norm_weight: buffer, %next_projection_input: buffer, %q8_output: buffer) {
template.apply<@ggml.routing.weighted_reduce_next_rmsnorm_f32>(%token_count, %route_weights, %routed_output, %hidden_state, %next_norm_weight, %next_projection_input) : (index, buffer, buffer, buffer, buffer, buffer)
kernel.barrier<workgroup> scope(workgroup) ordering(acq_rel)
// The packer reads global F32 output written by other workitems.
kernel.barrier<global> scope(workgroup) ordering(acq_rel)

%workitem = kernel.workitem.id<x> : index
%token0 = kernel.workgroup.id<x> : index
Expand Down
Loading